From 4ece2a50a3513a28339d3bda67ef1149d0dfb827 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E9=83=91=E6=99=A8=E5=8D=89?= Date: Wed, 29 May 2024 10:55:16 +0800 Subject: [PATCH] Revert last two commits and Sync patch from openeuler/gcc --- 0142-crc-loop-optimization-initial.patch | 2332 ++++++++ ...y-if-conversion-of-simple-arithmetic.patch | 109 + ...low-matching-uaddsub-overflow-for-wi.patch | 236 + 0145-Match-double-sized-mul-pattern.patch | 488 ++ ...rc32-Optimization-in-Gzip-For-crc32-.patch | 2354 ++++++++ ...nd-correct-costs-for-cmlt-generation.patch | 194 + 0148-Introduce-RTL-ifcvt-enhancements.patch | 502 ++ ...e-check-for-pointer-aliasing-during-.patch | 239 + ...ropagation-of-permutations-in-fwprop.patch | 1050 ++++ ...Fix-bugs-and-add-tests-for-RTL-ifcvt.patch | 381 ++ 0152-Add-LLC-Allocation-Pass.patch | 4905 +++++++++++++++++ 0153-LLC-add-extending-outer-loop.patch | 1285 +++++ gcc.spec | 46 +- 13 files changed, 14108 insertions(+), 13 deletions(-) create mode 100644 0142-crc-loop-optimization-initial.patch create mode 100644 0143-Perform-early-if-conversion-of-simple-arithmetic.patch create mode 100644 0144-Add-option-to-allow-matching-uaddsub-overflow-for-wi.patch create mode 100644 0145-Match-double-sized-mul-pattern.patch create mode 100644 0146-LOOP-CRC32-Add-Crc32-Optimization-in-Gzip-For-crc32-.patch create mode 100644 0147-add-insn-defs-and-correct-costs-for-cmlt-generation.patch create mode 100644 0148-Introduce-RTL-ifcvt-enhancements.patch create mode 100644 0149-Add-more-flexible-check-for-pointer-aliasing-during-.patch create mode 100644 0150-Implement-propagation-of-permutations-in-fwprop.patch create mode 100644 0151-Fix-bugs-and-add-tests-for-RTL-ifcvt.patch create mode 100644 0152-Add-LLC-Allocation-Pass.patch create mode 100644 0153-LLC-add-extending-outer-loop.patch diff --git a/0142-crc-loop-optimization-initial.patch b/0142-crc-loop-optimization-initial.patch new file mode 100644 index 0000000..61d5ae5 --- /dev/null +++ b/0142-crc-loop-optimization-initial.patch @@ -0,0 +1,2332 @@ +From 2716abb1a4de2a4edf06d2f1877d9b76a88e5807 Mon Sep 17 00:00:00 2001 +From: bule +Date: Thu, 15 Dec 2022 14:34:16 +0800 +Subject: [PATCH 05/33] crc loop optimization initial + +--- + gcc/Makefile.in | 1 + + gcc/common.opt | 4 + + gcc/doc/invoke.texi | 6 +- + gcc/match.pd | 169 +++++ + gcc/passes.def | 1 + + gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c | 85 +++ + .../tree-ssa/loop-crc-1.c.042t.loop_crc | 90 +++ + gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c | 88 +++ + gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c | 85 +++ + gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c | 89 +++ + .../tree-ssa/loop-crc-4.c.042t.loop_crc | 0 + .../loop-crc-calculation-check-fail.c | 156 +++++ + ...crc-calculation-check-fail.c.042t.loop_crc | 64 ++ + .../loop-crc-calculation-check-fail.s | 329 +++++++++ + .../gcc.dg/tree-ssa/loop-crc-loop-form-fail.c | 111 +++ + .../gcc.dg/tree-ssa/loop-crc-sucess.c | 84 +++ + .../tree-ssa/loop-crc-table-check-fail.c | 113 +++ + gcc/timevar.def | 1 + + gcc/tree-pass.h | 1 + + gcc/tree-ssa-loop-crc.c | 644 ++++++++++++++++++ + 20 files changed, 2120 insertions(+), 1 deletion(-) + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c.042t.loop_crc + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c + create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c + create mode 100644 gcc/tree-ssa-loop-crc.c + +diff --git a/gcc/Makefile.in b/gcc/Makefile.in +index 3f06b8907..2a59acfbe 100644 +--- a/gcc/Makefile.in ++++ b/gcc/Makefile.in +@@ -1592,6 +1592,7 @@ OBJS = \ + tree-ssa-loop-manip.o \ + tree-ssa-loop-niter.o \ + tree-ssa-loop-array-widen-compare.o \ ++ tree-ssa-loop-crc.o \ + tree-ssa-loop-prefetch.o \ + tree-ssa-loop-split.o \ + tree-ssa-loop-unswitch.o \ +diff --git a/gcc/common.opt b/gcc/common.opt +index 6f0ed7cea..a286a2628 100644 +--- a/gcc/common.opt ++++ b/gcc/common.opt +@@ -1065,6 +1065,10 @@ Common Report Var(flag_array_widen_compare) Optimization + Extends types for pointers to arrays to improve array comparsion performance. + In some extreme situations this may result in unsafe behavior. + ++floop-crc ++Common Report Var(flag_loop_crc) Optimization ++do the loop crc conversion. ++ + fauto-inc-dec + Common Report Var(flag_auto_inc_dec) Init(1) Optimization + Generate auto-inc/dec instructions. +diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi +index 7498758b0..52018617a 100644 +--- a/gcc/doc/invoke.texi ++++ b/gcc/doc/invoke.texi +@@ -460,7 +460,7 @@ Objective-C and Objective-C++ Dialects}. + -fno-allocation-dce -fallow-store-data-races @gol + -fassociative-math -fauto-profile -fauto-profile[=@var{path}] @gol + -farray-widen-compare -fauto-inc-dec -fbranch-probabilities @gol +--fcaller-saves @gol ++-fcaller-saves -floop-crc @gol + -fcombine-stack-adjustments -fconserve-stack @gol + -fcompare-elim -fcprop-registers -fcrossjumping @gol + -fcse-follow-jumps -fcse-skip-blocks -fcx-fortran-rules @gol +@@ -9722,6 +9722,10 @@ extreme situations this may result in unsafe behavior. + This option may generate better or worse code; results are highly dependent + on the structure of loops within the source code. + ++@item -floop-crc ++@opindex floop-crc ++Do the loop crc conversion ++ + @item -fdce + @opindex fdce + Perform dead code elimination (DCE) on RTL@. +diff --git a/gcc/match.pd b/gcc/match.pd +index 01f81b063..87b316953 100644 +--- a/gcc/match.pd ++++ b/gcc/match.pd +@@ -3487,6 +3487,175 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) + ) + #endif + ++ ++#if GIMPLE ++(if (canonicalize_math_p ()) ++/* These patterns are mostly used by PHIOPT to move some operations outside of ++ the if statements. They should be done late because it gives jump threading ++ and few other passes to reduce what is going on. */ ++/* a ? x op POW2 : x -> x op (a ? POW2 : 0). */ ++ (for op (plus minus bit_ior bit_xor lshift rshift lrotate rrotate) ++ (simplify ++ (cond @0 (op:s @1 INTEGER_CST@2) @1) ++ /* powerof2cst */ ++ (if (INTEGRAL_TYPE_P (type) && integer_pow2p (@2)) ++ (with { ++ tree shift = build_int_cst (integer_type_node, tree_log2 (@2)); ++ } ++ (op @1 (lshift (convert (convert:boolean_type_node @0)) { shift; }))) ++ ) ++ ) ++ ) ++) ++#endif ++ ++#if GIMPLE ++/* These patterns are mostly used by FORWPROP to move some operations outside of ++ the if statements. They should be done late because it gives jump threading ++ and few other passes to reduce what is going on. */ ++/* Mul64 is defined as a multiplication algorithm which compute two 64-bit integers to one 128-bit integer ++ (i64 ResLo, i64 ResHi) = Mul64(i64 In0, i64 In1) { ++ In0Lo = In0(D) & 4294967295; ++ In0Hi = In0(D) >> 32; ++ In1Lo = In1(D) & 4294967295; ++ In1Hi = In1(D) >> 32; ++ Mull_01 = In0Hi * In1Lo; ++ Addc = In0Lo * In1Hi + Mull_01; ++ addc32 = Addc << 32; ++ ResLo = In0Lo * In1Lo + addc32; ++ ResHi = ((long unsigned int) (addc32 > ResLo)) + ++ (((long unsigned int) (Mull_01 > Addc)) << 32) + (Addc >> 32) + In0Hi * In1Hi; ++ } */ ++ (simplify ++ (plus ++ (plus ++ (convert ++ (gt @10 ++ (plus ++ (mult @4 @6) ++ (lshift@10 @9 @3)))) ++ (lshift ++ (convert ++ (gt @8 @9)) @3)) ++ (plus@11 ++ (rshift ++ (plus@9 ++ (mult (bit_and@4 SSA_NAME@0 @2) @7) ++ (mult@8 @5 (bit_and@6 SSA_NAME@1 INTEGER_CST@2))) @3) ++ (mult (rshift@5 SSA_NAME@0 @3) ++ (rshift@7 SSA_NAME@1 INTEGER_CST@3)))) ++ (if (INTEGRAL_TYPE_P (type) && INTEGRAL_TYPE_P (TREE_TYPE (@0)) && types_match (@0, @1) && ++ TYPE_PRECISION (type) == 64) ++ (with { ++ tree i128_type = build_nonstandard_integer_type (128, TYPE_UNSIGNED (type)); ++ tree shift = build_int_cst (integer_type_node, 64); ++ //direct_internal_fn_supported_p (UMULH, type, OPTIMIZE_FOR_BOTH) ++ } ++ (convert:type (rshift ++ (mult (convert:i128_type @0) (convert:i128_type @1)) { shift; }))) ++ ) ++ ) ++ ++ /* (i64 ResLo, i64 ResHi) = Mul64(i64 In0, i64 In1) { ++ In0Lo = In0(D) & 4294967295; ++ In0Hi = In0(D) >> 32; ++ In1Lo = In1(D) & 4294967295; ++ In1Hi = In1(D) >> 32; ++ Mull_01 = In0Hi * In1Lo; ++ Addc = In0Lo * In1Hi + Mull_01; ++ addc32 = Addc << 32; ++ ResLo = In0(D) * In1(D); ++ ResHi = ((long unsigned int) (addc32 > ResLo)) + ++ (((long unsigned int) (Mull_01 > Addc)) << 32) + (Addc >> 32) + In0Hi * In1Hi; ++ } */ ++ (simplify ++ (plus ++ (plus ++ (convert ++ (gt (lshift@10 @9 @3) ++ (mult @0 @1))) ++ (lshift ++ (convert ++ (gt @8 @9)) @3)) ++ (plus@11 ++ (rshift ++ (plus@9 ++ (mult (bit_and@4 SSA_NAME@0 @2) @7) ++ (mult@8 @5 (bit_and@6 SSA_NAME@1 INTEGER_CST@2))) @3) ++ (mult (rshift@5 SSA_NAME@0 @3) ++ (rshift@7 SSA_NAME@1 INTEGER_CST@3)))) ++ (if (INTEGRAL_TYPE_P (type) && INTEGRAL_TYPE_P (TREE_TYPE (@0)) && types_match (@0, @1) && ++ TYPE_PRECISION (type) == 64) ++ (with { ++ tree i128_type = build_nonstandard_integer_type (128, TYPE_UNSIGNED (type)); ++ tree shift = build_int_cst (integer_type_node, 64); ++ //direct_internal_fn_supported_p (UMULH, type, OPTIMIZE_FOR_BOTH) ++ } ++ (convert:type (rshift ++ (mult (convert:i128_type @0) (convert:i128_type @1)) { shift; }))) ++ ) ++ ) ++#endif ++ ++#if GIMPLE ++/* These patterns are mostly used by FORWPROP to move some operations outside of ++ the if statements. They should be done late because it gives jump threading ++ and few other passes to reduce what is going on. */ ++ /* ++ In0Lo = In0(D) & 4294967295; ++ In0Hi = In0(D) >> 32; ++ In1Lo = In1(D) & 4294967295; ++ In1Hi = In1(D) >> 32; ++ Addc = In0Lo * In1Hi + In0Hi * In1Lo; ++ addc32 = Addc << 32; ++ ResLo = In0Lo * In1Lo + addc32 ++ */ ++ (simplify ++ (plus (mult @4 @5) ++ (lshift ++ (plus ++ (mult (bit_and@4 SSA_NAME@0 @2) (rshift SSA_NAME@1 @3)) ++ (mult (rshift SSA_NAME@0 @3) (bit_and@5 SSA_NAME@1 INTEGER_CST@2))) INTEGER_CST@3)) ++ (if (INTEGRAL_TYPE_P (type) && INTEGRAL_TYPE_P (TREE_TYPE (@0)) && types_match (@0, @1) && ++ TYPE_PRECISION (type) == 64) ++ (with { ++ tree i128_type = build_nonstandard_integer_type (128, TYPE_UNSIGNED (type)); ++ tree shift = build_int_cst (integer_type_node, 64); ++ //direct_internal_fn_supported_p (UMULH, type, OPTIMIZE_FOR_BOTH) ++ } ++ (mult (convert:type @0) (convert:type @1))) ++ ) ++ ) ++#endif ++ ++ ++#if GIMPLE ++/* Try to match */ ++ /* ++_4 = (int) _3; //NOP_EXPR (SSA_NAME @2) ++_5 = _4 ^ c_10; //BIT_XOR_EXPR (SSA_NAME@1, SSA_NAME) ++_6 = _5 & 255; //BIT_XOR_EXPR (SSA_NAME, INTEGER_CST@3) ++ */ ++(match (crc_match_index @1 @2 @3) ++ (bit_and (bit_xor (nop SSA_NAME@2) SSA_NAME@1) INTEGER_CST@3) ++ (if (INTEGRAL_TYPE_P (type) && tree_to_uhwi(@3) == 255)) ++) ++ ++#endif ++ ++#if GIMPLE ++/* Try to match */ ++ /* ++_8 = c_12 >> 8; // RSHIFT_EXPR (SSA_NAME @1, INTEGER_CST @2) ++c_19 = _7 ^ _8; // BIT_XOR_EXPR (SSA_NAME@3, SSA_NAME) ++ */ ++(match (crc_match_res @1 @2 @3) ++ (bit_xor SSA_NAME@3 (rshift SSA_NAME@1 INTEGER_CST@2)) ++ (if (INTEGRAL_TYPE_P (type) && tree_to_uhwi(@2) == 8)) ++) ++ ++#endif ++ + /* Simplification moved from fold_cond_expr_with_comparison. It may also + be extended. */ + /* This pattern implements two kinds simplification: +diff --git a/gcc/passes.def b/gcc/passes.def +index ea50db086..7abd946ce 100644 +--- a/gcc/passes.def ++++ b/gcc/passes.def +@@ -92,6 +92,7 @@ along with GCC; see the file COPYING3. If not see + NEXT_PASS (pass_cd_dce); + NEXT_PASS (pass_phiopt, true /* early_p */); + NEXT_PASS (pass_array_widen_compare); ++ NEXT_PASS (pass_loop_crc); + NEXT_PASS (pass_tail_recursion); + NEXT_PASS (pass_convert_switch); + NEXT_PASS (pass_cleanup_eh); +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c +new file mode 100644 +index 000000000..07f9e01ec +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c +@@ -0,0 +1,85 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++ ++ulg updcrc(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ } while (--n); ++ } ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++/* { dg-final { scan-tree-dump-times "Processing loop" 1 "loop_crc"} } */ ++/* { dg-final { scan-tree-dump-times "the loop can be optimized" 1 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc +new file mode 100644 +index 000000000..c726059f3 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc +@@ -0,0 +1,90 @@ ++ ++;; Function updcrc (updcrc, funcdef_no=0, decl_uid=3687, cgraph_uid=1, symbol_order=1) ++ ++;; 2 loops found ++;; ++;; Loop 0 ++;; header 0, latch 1 ++;; depth 0, outer -1 ++;; nodes: 0 1 2 3 6 4 7 5 ++;; ++;; Loop 1 ++;; header 4, latch 7 ++;; depth 1, outer 0 ++;; nodes: 4 7 ++;; 2 succs { 5 3 } ++;; 3 succs { 6 5 } ++;; 6 succs { 4 } ++;; 4 succs { 7 5 } ++;; 7 succs { 4 } ++;; 5 succs { 1 } ++ ++Starting the loop_crc pass ++====================================== ++Processing loop 1: ++====================================== ++;; ++;; Loop 1 ++;; header 4, latch 7 ++;; depth 1, outer 0 ++;; nodes: 4 7 ++ ++ ++The 1th loop form is success matched,and the loop can be optimized. ++updcrc (uch * s, unsigned int n) ++{ ++ static ulg crc = 4294967295; ++ register ulg c; ++ unsigned char _2; ++ long unsigned int _3; ++ long unsigned int _4; ++ long unsigned int _5; ++ long unsigned int _6; ++ long unsigned int _7; ++ ulg _21; ++ ++ : ++ if (s_12(D) == 0B) ++ goto ; [INV] ++ else ++ goto ; [INV] ++ ++ : ++ c_14 = crc; ++ if (n_15(D) != 0) ++ goto ; [INV] ++ else ++ goto ; [INV] ++ ++ : ++ ++ : ++ # s_8 = PHI ++ # n_9 = PHI ++ # c_10 = PHI ++ s_16 = s_8 + 1; ++ _2 = *s_8; ++ _3 = (long unsigned int) _2; ++ _4 = _3 ^ c_10; ++ _5 = _4 & 255; ++ _6 = crc_32_tab[_5]; ++ _7 = c_10 >> 8; ++ c_17 = _6 ^ _7; ++ n_18 = n_9 + 4294967295; ++ if (n_18 != 0) ++ goto ; [INV] ++ else ++ goto ; [INV] ++ ++ : ++ goto ; [100.00%] ++ ++ : ++ # c_11 = PHI <4294967295(2), c_14(3), c_17(4)> ++ crc = c_11; ++ _21 = c_11 ^ 4294967295; ++ return _21; ++ ++} ++ ++ +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c +new file mode 100644 +index 000000000..f73c4d550 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c +@@ -0,0 +1,88 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++ ++ulg updcrc(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ for (int i = 0; i < 5; i++) { ++ c++; ++ } ++ ++ } while (--n); ++ } ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++/* { dg-final { scan-tree-dump-times "Wrong loop form for crc matching." 1 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c +new file mode 100644 +index 000000000..70eb1b814 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c +@@ -0,0 +1,85 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++ ++ulg updcrc(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ } while (--n || c != 0) ; ++ } ++ crc = c; ++exit1: ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++/* { dg-final { scan-tree-dump-times "Wrong loop form for crc matching." 1 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c +new file mode 100644 +index 000000000..1d7e0a319 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c +@@ -0,0 +1,89 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++int test[5] = {0}; ++ ++ulg updcrc(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) * test[c%5]; ++ } while (--n) ; ++ } ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ test[c%5] = c; ++ } while (--n) ; ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++/* { dg-final { scan-tree-dump-times "Table check fail. not only single array is read." 2 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c.042t.loop_crc b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c.042t.loop_crc +new file mode 100644 +index 000000000..e69de29bb +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c +new file mode 100644 +index 000000000..71b25f537 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c +@@ -0,0 +1,156 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++ ++int test[5] = {0}; ++ ++ulg updcrc(s, n) ++ uch *s; ++ unsigned n; ++{ ++ register ulg c; ++ ++ static ulg crc = (ulg)0xffffffffL; ++ int a = 0; ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ a++; ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) ; ++ } while (--n) ; ++ } ++ crc = c; ++ return c ^ 0xffffffffL*a; ++} ++ ++ulg updcrc1(s, n) ++ uch *s; ++ unsigned n; ++{ ++ register ulg c; ++ ++ static ulg crc = (ulg)0xffffffffL; ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ unsigned n_back = n; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) ; ++ n = n - 2; ++ } while (n != 0) ; ++ } ++ ++ crc = c; ++ return c ^ 0xffffffffL; ++} ++ ++ulg updcrc2(s, n) ++ uch *s; ++ unsigned n; ++{ ++ register ulg c; ++ ++ static ulg crc = (ulg)0xffffffffL; ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ unsigned n_back = n; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) + 1; ++ } while (--n) ; ++ } ++ ++ crc = c; ++ return c ^ 0xffffffffL; ++} ++/* ++ulg updcrc3(s, n) ++ uch *s; ++ int n; ++{ ++ register ulg c; ++ ++ static ulg crc = (ulg)0xffffffffL; ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ --n; ++ } while (n ) ; ++ } ++ ++ crc = c; ++ return c ^ 0xffffffffL; ++}*/ ++/* { dg-final { scan-tree-dump-times "num of phi noeds check failed." 1 "loop_crc"} } */ ++/* { dg-final { scan-tree-dump-times "evolution pattern check failed." 1 "loop_crc"} } */ ++/* { dg-final { scan-tree-dump-times "calculation pattern check failed." 1 "loop_crc"} } */ ++ +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc +new file mode 100644 +index 000000000..6d52a8684 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc +@@ -0,0 +1,64 @@ ++ ++;; Function updcrc3 (updcrc3, funcdef_no=0, decl_uid=3687, cgraph_uid=1, symbol_order=1) ++ ++;; 2 loops found ++;; ++;; Loop 0 ++;; header 0, latch 1 ++;; depth 0, outer -1 ++;; nodes: 0 1 2 3 4 5 ++;; ++;; Loop 1 ++;; header 4, latch 4 ++;; depth 1, outer 0 ++;; nodes: 4 ++;; 2 succs { 5 3 } ++;; 3 succs { 4 5 } ++;; 4 succs { 4 } ++;; 5 succs { 1 } ++ ++Starting the loop_crc pass ++====================================== ++Processing loop 1: ++====================================== ++;; ++;; Loop 1 ++;; header 4, latch 4 ++;; depth 1, outer 0 ++;; nodes: 4 ++ ++ ++ ++Wrong loop form for crc matching. ++updcrc3 (uch * s, unsigned int n) ++{ ++ unsigned int n_back; ++ static ulg crc = 4294967295; ++ register ulg c; ++ ulg _22; ++ ++ : ++ if (s_12(D) == 0B) ++ goto ; [INV] ++ else ++ goto ; [INV] ++ ++ : ++ c_14 = crc; ++ if (n_15(D) != 0) ++ goto ; [INV] ++ else ++ goto ; [INV] ++ ++ : ++ goto ; [100.00%] ++ ++ : ++ # c_11 = PHI <4294967295(2), c_14(3)> ++ crc = c_11; ++ _22 = c_11 ^ 4294967295; ++ return _22; ++ ++} ++ ++ +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s +new file mode 100644 +index 000000000..cae934bfe +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s +@@ -0,0 +1,329 @@ ++ .arch armv8-a ++ .file "loop-crc-calculation-check-fail.c" ++ .text ++ .section .rodata ++ .align 3 ++ .type crc_32_tab, %object ++ .size crc_32_tab, 2048 ++crc_32_tab: ++ .xword 0 ++ .xword 1996959894 ++ .xword 3993919788 ++ .xword 2567524794 ++ .xword 124634137 ++ .xword 1886057615 ++ .xword 3915621685 ++ .xword 2657392035 ++ .xword 249268274 ++ .xword 2044508324 ++ .xword 3772115230 ++ .xword 2547177864 ++ .xword 162941995 ++ .xword 2125561021 ++ .xword 3887607047 ++ .xword 2428444049 ++ .xword 498536548 ++ .xword 1789927666 ++ .xword 4089016648 ++ .xword 2227061214 ++ .xword 450548861 ++ .xword 1843258603 ++ .xword 4107580753 ++ .xword 2211677639 ++ .xword 325883990 ++ .xword 1684777152 ++ .xword 4251122042 ++ .xword 2321926636 ++ .xword 335633487 ++ .xword 1661365465 ++ .xword 4195302755 ++ .xword 2366115317 ++ .xword 997073096 ++ .xword 1281953886 ++ .xword 3579855332 ++ .xword 2724688242 ++ .xword 1006888145 ++ .xword 1258607687 ++ .xword 3524101629 ++ .xword 2768942443 ++ .xword 901097722 ++ .xword 1119000684 ++ .xword 3686517206 ++ .xword 2898065728 ++ .xword 853044451 ++ .xword 1172266101 ++ .xword 3705015759 ++ .xword 2882616665 ++ .xword 651767980 ++ .xword 1373503546 ++ .xword 3369554304 ++ .xword 3218104598 ++ .xword 565507253 ++ .xword 1454621731 ++ .xword 3485111705 ++ .xword 3099436303 ++ .xword 671266974 ++ .xword 1594198024 ++ .xword 3322730930 ++ .xword 2970347812 ++ .xword 795835527 ++ .xword 1483230225 ++ .xword 3244367275 ++ .xword 3060149565 ++ .xword 1994146192 ++ .xword 31158534 ++ .xword 2563907772 ++ .xword 4023717930 ++ .xword 1907459465 ++ .xword 112637215 ++ .xword 2680153253 ++ .xword 3904427059 ++ .xword 2013776290 ++ .xword 251722036 ++ .xword 2517215374 ++ .xword 3775830040 ++ .xword 2137656763 ++ .xword 141376813 ++ .xword 2439277719 ++ .xword 3865271297 ++ .xword 1802195444 ++ .xword 476864866 ++ .xword 2238001368 ++ .xword 4066508878 ++ .xword 1812370925 ++ .xword 453092731 ++ .xword 2181625025 ++ .xword 4111451223 ++ .xword 1706088902 ++ .xword 314042704 ++ .xword 2344532202 ++ .xword 4240017532 ++ .xword 1658658271 ++ .xword 366619977 ++ .xword 2362670323 ++ .xword 4224994405 ++ .xword 1303535960 ++ .xword 984961486 ++ .xword 2747007092 ++ .xword 3569037538 ++ .xword 1256170817 ++ .xword 1037604311 ++ .xword 2765210733 ++ .xword 3554079995 ++ .xword 1131014506 ++ .xword 879679996 ++ .xword 2909243462 ++ .xword 3663771856 ++ .xword 1141124467 ++ .xword 855842277 ++ .xword 2852801631 ++ .xword 3708648649 ++ .xword 1342533948 ++ .xword 654459306 ++ .xword 3188396048 ++ .xword 3373015174 ++ .xword 1466479909 ++ .xword 544179635 ++ .xword 3110523913 ++ .xword 3462522015 ++ .xword 1591671054 ++ .xword 702138776 ++ .xword 2966460450 ++ .xword 3352799412 ++ .xword 1504918807 ++ .xword 783551873 ++ .xword 3082640443 ++ .xword 3233442989 ++ .xword 3988292384 ++ .xword 2596254646 ++ .xword 62317068 ++ .xword 1957810842 ++ .xword 3939845945 ++ .xword 2647816111 ++ .xword 81470997 ++ .xword 1943803523 ++ .xword 3814918930 ++ .xword 2489596804 ++ .xword 225274430 ++ .xword 2053790376 ++ .xword 3826175755 ++ .xword 2466906013 ++ .xword 167816743 ++ .xword 2097651377 ++ .xword 4027552580 ++ .xword 2265490386 ++ .xword 503444072 ++ .xword 1762050814 ++ .xword 4150417245 ++ .xword 2154129355 ++ .xword 426522225 ++ .xword 1852507879 ++ .xword 4275313526 ++ .xword 2312317920 ++ .xword 282753626 ++ .xword 1742555852 ++ .xword 4189708143 ++ .xword 2394877945 ++ .xword 397917763 ++ .xword 1622183637 ++ .xword 3604390888 ++ .xword 2714866558 ++ .xword 953729732 ++ .xword 1340076626 ++ .xword 3518719985 ++ .xword 2797360999 ++ .xword 1068828381 ++ .xword 1219638859 ++ .xword 3624741850 ++ .xword 2936675148 ++ .xword 906185462 ++ .xword 1090812512 ++ .xword 3747672003 ++ .xword 2825379669 ++ .xword 829329135 ++ .xword 1181335161 ++ .xword 3412177804 ++ .xword 3160834842 ++ .xword 628085408 ++ .xword 1382605366 ++ .xword 3423369109 ++ .xword 3138078467 ++ .xword 570562233 ++ .xword 1426400815 ++ .xword 3317316542 ++ .xword 2998733608 ++ .xword 733239954 ++ .xword 1555261956 ++ .xword 3268935591 ++ .xword 3050360625 ++ .xword 752459403 ++ .xword 1541320221 ++ .xword 2607071920 ++ .xword 3965973030 ++ .xword 1969922972 ++ .xword 40735498 ++ .xword 2617837225 ++ .xword 3943577151 ++ .xword 1913087877 ++ .xword 83908371 ++ .xword 2512341634 ++ .xword 3803740692 ++ .xword 2075208622 ++ .xword 213261112 ++ .xword 2463272603 ++ .xword 3855990285 ++ .xword 2094854071 ++ .xword 198958881 ++ .xword 2262029012 ++ .xword 4057260610 ++ .xword 1759359992 ++ .xword 534414190 ++ .xword 2176718541 ++ .xword 4139329115 ++ .xword 1873836001 ++ .xword 414664567 ++ .xword 2282248934 ++ .xword 4279200368 ++ .xword 1711684554 ++ .xword 285281116 ++ .xword 2405801727 ++ .xword 4167216745 ++ .xword 1634467795 ++ .xword 376229701 ++ .xword 2685067896 ++ .xword 3608007406 ++ .xword 1308918612 ++ .xword 956543938 ++ .xword 2808555105 ++ .xword 3495958263 ++ .xword 1231636301 ++ .xword 1047427035 ++ .xword 2932959818 ++ .xword 3654703836 ++ .xword 1088359270 ++ .xword 936918000 ++ .xword 2847714899 ++ .xword 3736837829 ++ .xword 1202900863 ++ .xword 817233897 ++ .xword 3183342108 ++ .xword 3401237130 ++ .xword 1404277552 ++ .xword 615818150 ++ .xword 3134207493 ++ .xword 3453421203 ++ .xword 1423857449 ++ .xword 601450431 ++ .xword 3009837614 ++ .xword 3294710456 ++ .xword 1567103746 ++ .xword 711928724 ++ .xword 3020668471 ++ .xword 3272380065 ++ .xword 1510334235 ++ .xword 755167117 ++ .text ++ .align 2 ++ .global updcrc3 ++ .type updcrc3, %function ++updcrc3: ++.LFB0: ++ .cfi_startproc ++ str x19, [sp, -48]! ++ .cfi_def_cfa_offset 48 ++ .cfi_offset 19, -48 ++ str x0, [sp, 24] ++ str w1, [sp, 20] ++ ldr x0, [sp, 24] ++ cmp x0, 0 ++ bne .L2 ++ mov x19, 4294967295 ++ b .L3 ++.L2: ++ adrp x0, crc.0 ++ add x0, x0, :lo12:crc.0 ++ ldr x19, [x0] ++ ldr w0, [sp, 20] ++ str w0, [sp, 44] ++ ldr w0, [sp, 20] ++ cmp w0, 0 ++ beq .L3 ++.L4: ++ ldr x0, [sp, 24] ++ add x1, x0, 1 ++ str x1, [sp, 24] ++ ldrb w0, [x0] ++ and x0, x0, 255 ++ eor x0, x19, x0 ++ and x1, x0, 255 ++ adrp x0, crc_32_tab ++ add x0, x0, :lo12:crc_32_tab ++ ldr x1, [x0, x1, lsl 3] ++ lsr x0, x19, 8 ++ eor x19, x1, x0 ++ ldr w0, [sp, 20] ++ sub w0, w0, #1 ++ str w0, [sp, 20] ++ ldr w0, [sp, 20] ++ cmp w0, 999 ++ bls .L4 ++.L3: ++ adrp x0, crc.0 ++ add x0, x0, :lo12:crc.0 ++ str x19, [x0] ++ eor x0, x19, 4294967295 ++ ldr x19, [sp], 48 ++ .cfi_restore 19 ++ .cfi_def_cfa_offset 0 ++ ret ++ .cfi_endproc ++.LFE0: ++ .size updcrc3, .-updcrc3 ++ .data ++ .align 3 ++ .type crc.0, %object ++ .size crc.0, 8 ++crc.0: ++ .xword 4294967295 ++ .ident "GCC: (Kunpeng gcc 10.3.1-2.3.0.b006) 10.3.1" ++ .section .note.GNU-stack,"",@progbits +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c +new file mode 100644 +index 000000000..b59704e31 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c +@@ -0,0 +1,111 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++ ++/* check when the loop have a innor loop, should fail. */ ++ulg updcrc(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ for (int i = 0; i < 5; i++) { ++ c++; ++ } ++ ++ } while (--n); ++ } ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++ ++/* check when the loop have a second backedge, should fail. */ ++ulg updcrc1(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ } while (--n || c != 0) ; ++ } ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++/* { dg-final { scan-tree-dump-times "Wrong loop form for crc matching." 2 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c +new file mode 100644 +index 000000000..e1e16eaf2 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c +@@ -0,0 +1,84 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++ ++ulg updcrc(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ } while (--n); ++ } ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++/* { dg-final { scan-tree-dump-times "the loop can be optimized" 1 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c +new file mode 100644 +index 000000000..f03a4fa82 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c +@@ -0,0 +1,113 @@ ++/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ ++/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ ++ ++#include ++#include ++typedef unsigned long ulg; ++typedef unsigned char uch; ++ ++static const ulg crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf1L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++int test[5] = {0}; ++ ++/* check when the loop is doing more then 1 array read or writing an array, both should fail. */ ++ulg updcrc(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) * test[c%5]; ++ } while (--n) ; ++ } ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ test[c%5] = c; ++ } while (--n) ; ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++ ++/* check when the loop is not working on a correct crc_table. should fail. */ ++ulg updcrc1(s, n) ++ uch *s; /* pointer to bytes to pump through */ ++ unsigned n; /* number of bytes in s[] */ ++{ ++ register ulg c; /* temporary variable */ ++ ++ static ulg crc = (ulg)0xffffffffL; /* shift register contents */ ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ } while (--n) ; ++ } ++ crc = c; ++ return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ ++} ++/* { dg-final { scan-tree-dump-times "Table check fail. not only single array is read." 2 "loop_crc"} } */ ++/* { dg-final { scan-tree-dump-times "Table check fail. Table not matching." 1 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/timevar.def b/gcc/timevar.def +index 2814b14f2..ba86a1b7b 100644 +--- a/gcc/timevar.def ++++ b/gcc/timevar.def +@@ -215,6 +215,7 @@ DEFTIMEVAR (TV_TREE_COPY_RENAME , "tree rename SSA copies") + DEFTIMEVAR (TV_TREE_SSA_VERIFY , "tree SSA verifier") + DEFTIMEVAR (TV_TREE_STMT_VERIFY , "tree STMT verifier") + DEFTIMEVAR (TV_TREE_ARRAY_WIDEN_COMPARE, "tree array widen compare") ++DEFTIMEVAR (TV_TREE_LOOP_CRC, "tree loop crc") + DEFTIMEVAR (TV_TREE_SWITCH_CONVERSION, "tree switch conversion") + DEFTIMEVAR (TV_TREE_SWITCH_LOWERING, "tree switch lowering") + DEFTIMEVAR (TV_TREE_RECIP , "gimple CSE reciprocals") +diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h +index 3cdc12466..027f8992d 100644 +--- a/gcc/tree-pass.h ++++ b/gcc/tree-pass.h +@@ -437,6 +437,7 @@ extern gimple_opt_pass *make_pass_phiopt (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_forwprop (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_phiprop (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_array_widen_compare (gcc::context *ctxt); ++extern gimple_opt_pass *make_pass_loop_crc (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_tree_ifcombine (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_dse (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_nrv (gcc::context *ctxt); +diff --git a/gcc/tree-ssa-loop-crc.c b/gcc/tree-ssa-loop-crc.c +new file mode 100644 +index 000000000..4982384c6 +--- /dev/null ++++ b/gcc/tree-ssa-loop-crc.c +@@ -0,0 +1,644 @@ ++/* Array widen compare. ++ Copyright (C) 2022-2022 Free Software Foundation, Inc. ++ ++This file is part of GCC. ++ ++GCC is free software; you can redistribute it and/or modify it ++under the terms of the GNU General Public License as published by the ++Free Software Foundation; either version 3, or (at your option) any ++later version. ++ ++GCC is distributed in the hope that it will be useful, but WITHOUT ++ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or ++FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License ++for more details. ++ ++You should have received a copy of the GNU General Public License ++along with GCC; see the file COPYING3. If not see ++. */ ++ ++#include "config.h" ++#include "system.h" ++#include "coretypes.h" ++#include "backend.h" ++#include "target.h" ++#include "tree.h" ++#include "gimple.h" ++#include "tree-pass.h" ++#include "gimple-ssa.h" ++#include "tree-pretty-print.h" ++#include "fold-const.h" ++#include "gimplify.h" ++#include "gimple-iterator.h" ++#include "tree-ssa-loop-manip.h" ++#include "tree-ssa-loop.h" ++#include "ssa.h" ++#include "tree-into-ssa.h" ++#include "cfganal.h" ++#include "cfgloop.h" ++#include "gimple-pretty-print.h" ++#include "tree-cfg.h" ++#include "cgraph.h" ++#include "print-tree.h" ++#include "cfghooks.h" ++#include "gimple-fold.h" ++ ++/* Match.pd function to match the ctz expression. */ ++extern bool gimple_crc_match_index (tree, tree *, tree (*)(tree)); ++extern bool gimple_crc_match_res (tree, tree *, tree (*)(tree)); ++ ++static gimple *crc_table_read_stmt = NULL; ++ ++ ++/* The loop form check will check the entire loop control flow ++ It should be a loop that: ++ 1. a do-while loop with header and latch only with no other control flow inside the loop ++ 2. have only one exiting edge ++ 3. have only one back edge and one entry edge ++*/ ++static bool ++crc_loop_form_check (class loop *loop) ++{ ++ if (loop->num_nodes > 2 || loop->inner) ++ return false; ++ // should only have 1 exit edge ++ vec edges; ++ edges = get_loop_exit_edges (loop); ++ if (edges.length() != 1) ++ return false; ++ ++ // The header should have only 2 incoming edges ++ // One of them is the preheader edge and the other is the backedge from the latch ++ if (EDGE_COUNT (loop->header->preds) != 2) ++ return false; ++ edge e1 = EDGE_PRED (loop->header, 0); ++ edge e2 = EDGE_PRED (loop->header, 1); ++ ++ if ((e1->src == loop->latch && e2->src->loop_father != loop) ++ || (e2->src == loop->latch && e1->src->loop_father != loop)) ++ return true; ++ ++ return false; ++} ++ ++/* Check there is only one array is read in the loop. ++ Return the only array as crc_table. */ ++static bool ++only_one_array_read (class loop *loop, tree &crc_table) ++{ ++ gimple_stmt_iterator gsi; ++ gimple *stmt; ++ bool res = false; ++ for (gsi = gsi_start_bb (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) ++ { ++ stmt = gsi_stmt (gsi); ++ if (stmt == NULL) ++ return false; ++ ++ if (gimple_code (stmt) == GIMPLE_ASSIGN && ++ TREE_CODE(gimple_assign_lhs (stmt)) == ARRAY_REF ) ++ return false; ++ ++ if (gimple_code (stmt) == GIMPLE_ASSIGN && ++ TREE_CODE(gimple_assign_rhs1 (stmt)) == ARRAY_REF) ++ { ++ if (crc_table == NULL) ++ { ++ crc_table = gimple_assign_rhs1 (stmt); ++ crc_table_read_stmt = stmt; ++ res = true; ++ } ++ else ++ return false; ++ } ++ } ++ return res; ++} ++ ++static const unsigned HOST_WIDE_INT crc_32_tab[] = { ++ 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, ++ 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, ++ 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, ++ 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, ++ 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, ++ 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, ++ 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, ++ 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, ++ 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, ++ 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, ++ 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, ++ 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, ++ 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, ++ 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, ++ 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, ++ 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, ++ 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, ++ 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, ++ 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, ++ 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, ++ 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, ++ 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, ++ 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, ++ 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, ++ 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, ++ 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, ++ 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, ++ 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, ++ 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, ++ 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, ++ 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, ++ 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, ++ 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, ++ 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, ++ 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, ++ 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, ++ 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, ++ 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, ++ 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, ++ 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, ++ 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, ++ 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, ++ 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, ++ 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, ++ 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, ++ 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, ++ 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, ++ 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, ++ 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, ++ 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, ++ 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, ++ 0x2d02ef8dL ++}; ++ ++/* Check the content of the array. */ ++static bool ++match_crc_table (tree crc_table) ++{ ++ unsigned HOST_WIDE_INT lb = tree_to_uhwi (array_ref_low_bound (crc_table)); ++ unsigned HOST_WIDE_INT ub = tree_to_uhwi (array_ref_up_bound (crc_table)); ++ unsigned HOST_WIDE_INT es = tree_to_uhwi (array_ref_element_size (crc_table)); ++ if (lb != 0 || ub != 255 || es != 8) ++ return false; ++ ++ tree decl = TREE_OPERAND (crc_table, 0); ++ tree ctor = ctor_for_folding(decl); ++ for (int i = 0; i < 255; i++) { ++ unsigned HOST_WIDE_INT val = tree_to_uhwi (CONSTRUCTOR_ELT (ctor,i)->value); ++ if (crc_32_tab[i] != val) ++ return false; ++ } ++ return true; ++} ++ ++ ++/* Check the crc table. The loop should have only one data reference. ++ And match the data reference with the predefined array. */ ++static bool ++crc_table_check (class loop *loop) ++{ ++ tree crc_table = NULL; ++ if (!only_one_array_read (loop, crc_table)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nTable check fail. not only single array is read.\n"); ++ return false; ++ } ++ if (!match_crc_table (crc_table)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nTable check fail. Table not matching.\n"); ++ return false; ++ } ++ return true; ++} ++ ++/* check whether the evolution pattern of phi is phi = SSA_NAME + target*/ ++static bool ++evolution_pattern_plus_with_p (class loop* loop, gphi *phi, unsigned HOST_WIDE_INT target) ++{ ++ edge backedge = find_edge (loop->latch, loop->header); ++ if (backedge == NULL) ++ return false; ++ tree evolution_node = PHI_ARG_DEF_FROM_EDGE (phi, backedge); ++ gimple *evolution_expr = SSA_NAME_DEF_STMT (evolution_node); ++ ++ if (evolution_expr && (gimple_assign_rhs_code (evolution_expr) == PLUS_EXPR || ++ gimple_assign_rhs_code (evolution_expr) == POINTER_PLUS_EXPR)) ++ { ++ tree rhs1 = gimple_assign_rhs1 (evolution_expr); ++ tree rhs2 = gimple_assign_rhs2 (evolution_expr); ++ if (TREE_CODE (rhs1) == SSA_NAME && TREE_CODE (rhs2) == INTEGER_CST ++ && tree_to_uhwi (rhs2) == target) ++ return true; ++ } ++ return false; ++} ++ ++/* Check whether there are only 3 phi nodes in the header block. ++ Return 3 phi nodes in the capture. */ ++static bool ++check_num_of_phi (basic_block header, gphi *capture[]) ++{ ++ gphi *phi; ++ gphi_iterator gsi; ++ int num_of_phi = 0; ++ ++ for (gsi = gsi_start_phis (header); !gsi_end_p (gsi); gsi_next (&gsi)) ++ { ++ phi = gsi.phi(); ++ if (phi) num_of_phi++; ++ if (num_of_phi > 3) ++ return false; ++ capture[num_of_phi - 1] = phi; ++ } ++ /* phi node should be exactly 3. */ ++ return num_of_phi == 3; ++} ++ ++/* Check the evolution pattern of three phi nodes. ++ Should be one of the node +1 every time (s), one of the node -1 ++ every time (n), and a 3rd one neither (c). Return 3 phi nodes in ++ the capture with the order of s,n,c.*/ ++static bool ++check_evolution_pattern (class loop* loop, gphi *capture[]) ++{ ++ gphi *s=NULL; ++ gphi *n=NULL; ++ gphi *c=NULL; ++ ++ for (int i = 0; i < 3; i++) ++ { ++ if (evolution_pattern_plus_with_p(loop, capture[i], 1)) ++ { ++ if (s != NULL) ++ return false; ++ s = capture[i]; ++ } ++ else if (evolution_pattern_plus_with_p(loop, capture[i], 4294967295)) ++ { ++ if (n != NULL) ++ return false; ++ n = capture[i]; ++ } ++ else ++ { ++ if (c != NULL) ++ return false; ++ c = capture[i]; ++ } ++ } ++ ++ // some envolution pattern cannot find ++ if (!n || !s || !c) ++ return false; ++ ++ capture[0] = s; ++ capture[1] = n; ++ capture[2] = c; ++ return true; ++} ++/* check the calculation pattern before and after the crc_table array read stmt. ++ _7 = crc_32_tab[_6]; ++ The caculation of index _6 should be the result of a sequency of calculation by the s and c ++ The result of the array read _7 should be used to calculate the new c. */ ++static bool ++check_calculation_pattern (class loop* loop, gphi *capture[]) ++{ ++ gphi *s=capture[0]; ++ gphi *c=capture[2]; ++ tree res_ops[3]; ++ tree index = TREE_OPERAND (gimple_assign_rhs1 (crc_table_read_stmt), 1); ++ ++ /* Try to match ++ _4 = (int) _3; //NOP_EXPR (SSA_NAME @2) ++ _5 = _4 ^ c_10; //BIT_XOR_EXPR (SSA_NAME, PHI @1) ++ _6 = _5 & 255; //BIT_XOR_EXPR (SSA_NAME, INTEGER_CST@3) ++ */ ++ ++ if (!gimple_crc_match_index(index, res_ops, NULL)) ++ return false; ++ gimple *s_res_stmt = SSA_NAME_DEF_STMT(res_ops[1]); ++ tree s_res = TREE_OPERAND(gimple_assign_rhs1(s_res_stmt),0); ++ if (res_ops[0] != gimple_phi_result (c) || ++ s_res != gimple_phi_result (s)) ++ return false; ++ ++ /* Try to match ++ _8 = c_12 >> 8; // RSHIFT_EXPR (SSA_NAME @1, INTEGER_CST @2) ++ c_19 = _7 ^ _8; // BIT_XOR_EXPR (SSA_NAME@3, SSA_NAME) ++ */ ++ edge backedge = find_edge(loop->latch, loop->header); ++ tree updated_c = PHI_ARG_DEF_FROM_EDGE (c, backedge); ++ if (!gimple_crc_match_res(updated_c, res_ops, NULL)) ++ return false; ++ if (res_ops[0] != gimple_phi_result (c) ++ || res_ops[2] != gimple_assign_lhs(crc_table_read_stmt)) ++ return false; ++ ++ return true; ++} ++ ++/* check the exit condition is n != 0. */ ++static bool ++check_exit_condition (class loop* loop, gphi *n) ++{ ++ edge backedge = find_edge(loop->latch, loop->header); ++ gimple *cond_stmt = gsi_stmt (gsi_last_bb (loop->header)); ++ if (!cond_stmt || gimple_code (cond_stmt) != GIMPLE_COND || gimple_cond_code (cond_stmt) != NE_EXPR ++ || gimple_cond_lhs (cond_stmt) != PHI_ARG_DEF_FROM_EDGE (n, backedge) ++ || tree_to_uhwi(gimple_cond_rhs (cond_stmt)) != 0) ++ return false; ++ ++ return true; ++} ++ ++/* Check the loop body. The loop body we are trying to match is ++ ++# s_10 = PHI ++# n_11 = PHI ++# c_12 = PHI ++_1 = (int) c_12; ++s_18 = s_10 + 1; ++_3 = *s_10; ++_4 = (int) _3; ++_5 = _1 ^ _4; ++_6 = _5 & 255; ++_7 = crc_32_tab[_6]; ++_8 = c_12 >> 8; ++c_19 = _7 ^ _8; ++n_20 = n_11 + 4294967295; ++if (n_20 != 0) ++ goto ; [INV] ++else ++ goto ; [INV] ++ ++which is doing a very simple calculation ++do { ++ c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++} while (--n); ++ ++In this case ,we don't want this loop to have any other operation inside. ++so the matching condition is ++1. There are only 3 loop variant during each itoration, namely s,c,n, ++ which is limited by the condition that the loop have exactly 3 phi nodes. ++2. The 3 loop variants should have evolution pattern as 1 of the 3 nodes is ++ increased by 1 every itoration, 1 of the 3 nodes is decreased by 1 every itor ++ and the 3rd one is neither. These three tree node SSA value will be captured for ++ the later arithmatic pattern matching ++3. Pattern matching for the index of crc_table ++4. pattern matching for the result of c calcuation after read from crc_table ++5. The exit condition matching. ++ */ ++static bool ++crc_loop_body_check (class loop *loop) ++{ ++ basic_block header = loop->header; ++ gphi *capture[3]; ++ if (!check_num_of_phi(header, capture)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\n num of phi noeds check failed.\n"); ++ return false; ++ } ++ if (!check_evolution_pattern(loop, capture)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\n evolution pattern check failed.\n"); ++ return false; ++ } ++ if (!check_calculation_pattern(loop, capture)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\n calculation pattern check failed.\n"); ++ return false; ++ } ++ if (!check_exit_condition(loop, capture[1] /* n*/)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\n exit condition check failed.\n"); ++ return false; ++ } ++ return true; ++/* gphi *phi; ++ gphi_iterator gsi; ++ int num_of_phi = 0; ++ //s, n, c; ++ //only 3 phi nodes are there, every one of the phi nodes comming from 2 edge only, one from preheader, one from latch ++ // s increase by 1 every itoration ++ // n decrease by 1 every itoration ++ // The final one is c, which is the result, should be used for the start of the later pattern matching ++ for (gsi = gsi_start_phis(loop->header); !gsi_end_p(gsi); gsi_next(&gsi)) ++ { ++ phi = gsi.phi(); ++ ++ if (phi) num_of_phi++; ++ if (num_of_phi > 3) return false; // more then 3 phi node ++ if (gimple_phi_num_args(phi) > 2) // more than 2 edges other then one backedge and one preheader edge ++ return false; ++ //capture[num_of_phi - 1] = gimple_phi_result(phi); ++ capture[num_of_phi - 1] = phi; ++ } ++ if (num_of_phi != 3) return false; // phi node should be 3 */ ++ // Find the envolution pattern for s and n, try to match the identity of these variable ++/* gphi *s=NULL; ++ gphi *n=NULL; ++ gphi *c=NULL; ++ ++ for (int i = 0; i < 3; i++) ++ { ++ if (evolution_pattern_plus_with_p(loop, capture[i], 1)) ++ { ++ if(s != NULL) ++ return false; ++ s = capture[i]; ++ } ++ else if (evolution_pattern_plus_with_p(loop, capture[i], 4294967295)) ++ { ++ if(n != NULL) ++ return false; ++ n = capture[i]; ++ } ++ else ++ { ++ if(c != NULL) ++ return false; ++ c = capture[i]; ++ } ++ } ++ ++ // some envolution pattern cannot find ++ if (!n || !s || !c) ++ return false; ++ gphi *s=capture[0]; ++ gphi *n=capture[1]; ++ gphi *c=capture[2]; ++ tree res_ops[3]; ++ tree index = TREE_OPERAND (gimple_assign_rhs1 (crc_table_read_stmt), 1); ++ ++ /* Try to match ++ _1 = (int) c_12; //NOP_EXPR (SSA_NAME @1) ++ _4 = (int) _3; //NOP_EXPR (SSA_NAME @2) ++ _5 = _1 ^ _4; //BIT_XOR_EXPR (SSA_NAME, SSA_NAME) ++ _6 = _5 & 255; //BIT_XOR_EXPR (SSA_NAME, INTEGER_CST@3) ++ ++ ++ if (!gimple_crc_match_index(index, res_ops, NULL)) ++ return false; ++ gimple *s_res_stmt = SSA_NAME_DEF_STMT(res_ops[1]); ++ tree s_res = TREE_OPERAND(gimple_assign_rhs1(s_res_stmt),0); ++ if (res_ops[0] != gimple_phi_result (c) || ++ s_res != gimple_phi_result (s)) ++ return false; ++ ++ /* ++_8 = c_12 >> 8; // RSHIFT_EXPR (SSA_NAME @1, INTEGER_CST @2) ++c_19 = _7 ^ _8; // BIT_XOR_EXPR (SSA_NAME@3, SSA_NAME) ++ ++ edge backedge = find_edge(loop->latch, loop->header); ++ tree updated_c = PHI_ARG_DEF_FROM_EDGE (c, backedge); ++ if (!gimple_crc_match_res(updated_c, res_ops, NULL)) ++ return false; ++ if (res_ops[0] != gimple_phi_result (c) ++ || res_ops[2] != gimple_assign_lhs(crc_table_read_stmt)) ++ return false; ++ ++ // try match n as the induction variable ++ // The proceed condition for back edge is n != 0 ++ gimple *cond_stmt = gsi_stmt (gsi_last_bb (loop->header)); ++ if (!cond_stmt || gimple_code (cond_stmt) != GIMPLE_COND || gimple_cond_code (cond_stmt) != NE_EXPR ++ || gimple_cond_lhs (cond_stmt) != PHI_ARG_DEF_FROM_EDGE (n, backedge) ++ || tree_to_uhwi(gimple_cond_rhs (cond_stmt)) != 0) ++ return false; ++ ++ return true; ++ */ ++} ++ ++ ++static bool ++match_crc_loop (class loop *loop) ++{ ++ if (!crc_loop_form_check(loop)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nWrong loop form for crc matching.\n"); ++ return false; ++ } ++ if (!crc_table_check(loop)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nWrong crc table for crc matching.\n"); ++ return false; ++ } ++ if (!crc_loop_body_check(loop)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nWrong loop body for crc matching.\n"); ++ return false; ++ } ++ return true; ++} ++ ++/* The main entry of loop crc optimizes. */ ++static unsigned int ++tree_ssa_loop_crc () ++{ ++ unsigned int todo = 0; ++ class loop *loop; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ flow_loops_dump (dump_file, NULL, 1); ++ fprintf (dump_file, "\nStarting the loop_crc pass\n"); ++ } ++ ++ FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "======================================\n"); ++ fprintf (dump_file, "Processing loop %d:\n", loop->num); ++ fprintf (dump_file, "======================================\n"); ++ flow_loop_dump (loop, dump_file, NULL, 1); ++ fprintf (dump_file, "\n\n"); ++ } ++ ++ if (match_crc_loop (loop)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "The %dth loop form is success matched," ++ "and the loop can be optimized.\n", ++ loop->num); ++ } ++ ++ convert_to_new_loop (loop); ++ } ++ } ++ ++ todo |= (TODO_update_ssa); ++ return todo; ++} ++ ++/* Loop crc. */ ++ ++namespace { ++ ++const pass_data pass_data_tree_loop_crc = ++{ ++ GIMPLE_PASS, ++ "loop_crc", ++ OPTGROUP_LOOP, ++ TV_TREE_LOOP_CRC, ++ (PROP_cfg | PROP_ssa), ++ 0, ++ 0, ++ 0, ++ (TODO_update_ssa | TODO_verify_all) ++}; ++ ++class pass_loop_crc : public gimple_opt_pass ++{ ++public: ++ pass_loop_crc (gcc::context *ctxt) ++ : gimple_opt_pass (pass_data_tree_loop_crc, ctxt) ++ {} ++ ++ /* opt_pass methods: */ ++ virtual bool gate (function *); ++ virtual unsigned int execute (function *); ++ ++}; // class pass_loop_crc ++ ++bool ++pass_loop_crc::gate (function *) ++{ ++ return (flag_loop_crc > 0 && optimize >= 3); ++} ++ ++unsigned int ++pass_loop_crc::execute (function *fun) ++{ ++ if (number_of_loops (fun) <= 1) ++ return 0; ++ ++ /* Only supports LP64 data mode. */ ++ if (TYPE_PRECISION (long_integer_type_node) != 64 ++ || POINTER_SIZE != 64 || TYPE_PRECISION (integer_type_node) != 32) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "The current data mode is not supported," ++ "only the LP64 date mode is supported.\n"); ++ return 0; ++ } ++ ++ return tree_ssa_loop_crc (); ++} ++ ++} // anon namespace ++ ++gimple_opt_pass * ++make_pass_loop_crc (gcc::context *ctxt) ++{ ++ return new pass_loop_crc (ctxt); ++} +\ No newline at end of file +-- +2.33.0 + diff --git a/0143-Perform-early-if-conversion-of-simple-arithmetic.patch b/0143-Perform-early-if-conversion-of-simple-arithmetic.patch new file mode 100644 index 0000000..6965a9b --- /dev/null +++ b/0143-Perform-early-if-conversion-of-simple-arithmetic.patch @@ -0,0 +1,109 @@ +From 7acb88ae27eb3e1af0da866d433968143c7754bd Mon Sep 17 00:00:00 2001 +From: Pronin Alexander 00812787 +Date: Thu, 12 Jan 2023 14:52:49 +0300 +Subject: [PATCH 20/33] Perform early if-conversion of simple arithmetic + +--- + gcc/common.opt | 4 ++++ + gcc/match.pd | 25 +++++++++++++++++++ + gcc/testsuite/gcc.dg/ifcvt-gimple.c | 37 +++++++++++++++++++++++++++++ + 3 files changed, 66 insertions(+) + create mode 100644 gcc/testsuite/gcc.dg/ifcvt-gimple.c + +diff --git a/gcc/common.opt b/gcc/common.opt +index 6f0ed7cea..6950756fd 100644 +--- a/gcc/common.opt ++++ b/gcc/common.opt +@@ -1721,6 +1721,10 @@ fif-conversion2 + Common Report Var(flag_if_conversion2) Optimization + Perform conversion of conditional jumps to conditional execution. + ++fif-conversion-gimple ++Common Report Var(flag_if_conversion_gimple) Optimization ++Perform conversion of conditional jumps to branchless equivalents during gimple transformations. ++ + fstack-reuse= + Common Joined RejectNegative Enum(stack_reuse_level) Var(flag_stack_reuse) Init(SR_ALL) Optimization + -fstack-reuse=[all|named_vars|none] Set stack reuse level for local variables. +diff --git a/gcc/match.pd b/gcc/match.pd +index 01f81b063..e98cd02e0 100644 +--- a/gcc/match.pd ++++ b/gcc/match.pd +@@ -3402,6 +3402,31 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) + ) + ) + ) ++ ++(if (flag_if_conversion_gimple) ++ (for simple_op (plus minus bit_and bit_ior bit_xor) ++ (simplify ++ (cond @0 (simple_op @1 INTEGER_CST@2) @1) ++ (switch ++ /* a = cond ? a + 1 : a -> a = a + ((int) cond) */ ++ (if (integer_onep (@2)) ++ (simple_op @1 (convert (convert:boolean_type_node @0)))) ++ /* a = cond ? a + powerof2cst : a -> ++ a = a + ((int) cond) << log2 (powerof2cst) */ ++ (if (INTEGRAL_TYPE_P (type) && integer_pow2p (@2)) ++ (with ++ { ++ tree shift = build_int_cst (integer_type_node, tree_log2 (@2)); ++ } ++ (simple_op @1 (lshift (convert (convert:boolean_type_node @0)) ++ { shift; }) ++ ) ++ ) ++ ) ++ ) ++ ) ++ ) ++) + #endif + + #if GIMPLE +diff --git a/gcc/testsuite/gcc.dg/ifcvt-gimple.c b/gcc/testsuite/gcc.dg/ifcvt-gimple.c +new file mode 100644 +index 000000000..0f7c87e5c +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/ifcvt-gimple.c +@@ -0,0 +1,37 @@ ++/* { dg-do compile } */ ++/* { dg-options "-O2 -fif-conversion-gimple -fdump-tree-optimized" } */ ++ ++int test_int (int optimizable_int) { ++ if (optimizable_int > 5) ++ ++optimizable_int; ++ return optimizable_int; ++} ++ ++int test_int_pow2 (int optimizable_int_pow2) { ++ if (optimizable_int_pow2 <= 4) ++ optimizable_int_pow2 += 1024; ++ return optimizable_int_pow2; ++} ++ ++int test_int_non_pow2 (int not_optimizable_int_non_pow2) { ++ if (not_optimizable_int_non_pow2 == 1) ++ not_optimizable_int_non_pow2 += 513; ++ return not_optimizable_int_non_pow2; ++} ++ ++float test_float (float not_optimizable_float) { ++ if (not_optimizable_float > 5) ++ not_optimizable_float += 1; ++ return not_optimizable_float; ++} ++ ++/* Expecting if-else block in test_float and test_int_non_pow2 only. */ ++/* { dg-final { scan-tree-dump-not "if \\(optimizable" "optimized" } } */ ++/* { dg-final { scan-tree-dump "if \\(not_optimizable_int_non_pow2" "optimized" } } */ ++/* { dg-final { scan-tree-dump "if \\(not_optimizable_float" "optimized" } } */ ++/* { dg-final { scan-tree-dump-times "if " 2 "optimized" } } */ ++/* { dg-final { scan-tree-dump-times "else" 2 "optimized" } } */ ++ ++/* Expecting shifted result only for optimizable_int_pow2. */ ++/* { dg-final { scan-tree-dump-times " << " 1 "optimized" } } */ ++/* { dg-final { scan-tree-dump " << 10;" "optimized" } } */ +-- +2.33.0 + diff --git a/0144-Add-option-to-allow-matching-uaddsub-overflow-for-wi.patch b/0144-Add-option-to-allow-matching-uaddsub-overflow-for-wi.patch new file mode 100644 index 0000000..e5ea737 --- /dev/null +++ b/0144-Add-option-to-allow-matching-uaddsub-overflow-for-wi.patch @@ -0,0 +1,236 @@ +From f788555b23b0b676729bb695af96954fe083e354 Mon Sep 17 00:00:00 2001 +From: Pronin Alexander 00812787 +Date: Tue, 24 Jan 2023 16:43:40 +0300 +Subject: [PATCH 21/33] Add option to allow matching uaddsub overflow for widen + ops too. + +--- + gcc/common.opt | 5 ++ + gcc/testsuite/gcc.dg/uaddsub.c | 143 +++++++++++++++++++++++++++++++++ + gcc/tree-ssa-math-opts.c | 35 +++++++- + 3 files changed, 179 insertions(+), 4 deletions(-) + create mode 100644 gcc/testsuite/gcc.dg/uaddsub.c + +diff --git a/gcc/common.opt b/gcc/common.opt +index 6950756fd..c2f01bbc0 100644 +--- a/gcc/common.opt ++++ b/gcc/common.opt +@@ -2989,6 +2989,11 @@ freciprocal-math + Common Report Var(flag_reciprocal_math) SetByCombined Optimization + Same as -fassociative-math for expressions which include division. + ++fuaddsub-overflow-match-all ++Common Report Var(flag_uaddsub_overflow_match_all) ++Match unsigned add/sub overflow even if the target does not support ++the corresponding instruction. ++ + ; Nonzero means that unsafe floating-point math optimizations are allowed + ; for the sake of speed. IEEE compliance is not guaranteed, and operations + ; are allowed to assume that their arguments and results are "normal" +diff --git a/gcc/testsuite/gcc.dg/uaddsub.c b/gcc/testsuite/gcc.dg/uaddsub.c +new file mode 100644 +index 000000000..96c26d308 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/uaddsub.c +@@ -0,0 +1,143 @@ ++/* { dg-do compile } */ ++/* { dg-options "-O2 -fuaddsub-overflow-match-all -fdump-tree-optimized" } */ ++#include ++ ++typedef unsigned __int128 uint128_t; ++typedef struct uint256_t ++{ ++ uint128_t lo; ++ uint128_t hi; ++} uint256_t; ++ ++uint16_t add16 (uint8_t a, uint8_t b) ++{ ++ uint8_t tmp = a + b; ++ uint8_t overflow = 0; ++ if (tmp < a) ++ overflow = 1; ++ ++ uint16_t res = overflow; ++ res <<= 8; ++ res += tmp; ++ return res; ++} ++ ++uint32_t add32 (uint16_t a, uint16_t b) ++{ ++ uint16_t tmp = a + b; ++ uint16_t overflow = 0; ++ if (tmp < a) ++ overflow = 1; ++ ++ uint32_t res = overflow; ++ res <<= 16; ++ res += tmp; ++ return res; ++} ++ ++uint64_t add64 (uint32_t a, uint32_t b) ++{ ++ uint32_t tmp = a + b; ++ uint32_t overflow = 0; ++ if (tmp < a) ++ overflow = 1; ++ ++ uint64_t res = overflow; ++ res <<= 32; ++ res += tmp; ++ return res; ++} ++ ++uint128_t add128 (uint64_t a, uint64_t b) ++{ ++ uint64_t tmp = a + b; ++ uint64_t overflow = 0; ++ if (tmp < a) ++ overflow = 1; ++ ++ uint128_t res = overflow; ++ res <<= 64; ++ res += tmp; ++ return res; ++} ++ ++uint256_t add256 (uint128_t a, uint128_t b) ++{ ++ uint128_t tmp = a + b; ++ uint128_t overflow = 0; ++ if (tmp < a) ++ overflow = 1; ++ ++ uint256_t res; ++ res.hi = overflow; ++ res.lo = tmp; ++ return res; ++} ++ ++uint16_t sub16 (uint8_t a, uint8_t b) ++{ ++ uint8_t tmp = a - b; ++ uint8_t overflow = 0; ++ if (tmp > a) ++ overflow = -1; ++ ++ uint16_t res = overflow; ++ res <<= 8; ++ res += tmp; ++ return res; ++} ++ ++uint32_t sub32 (uint16_t a, uint16_t b) ++{ ++ uint16_t tmp = a - b; ++ uint16_t overflow = 0; ++ if (tmp > a) ++ overflow = -1; ++ ++ uint32_t res = overflow; ++ res <<= 16; ++ res += tmp; ++ return res; ++} ++ ++uint64_t sub64 (uint32_t a, uint32_t b) ++{ ++ uint32_t tmp = a - b; ++ uint32_t overflow = 0; ++ if (tmp > a) ++ overflow = -1; ++ ++ uint64_t res = overflow; ++ res <<= 32; ++ res += tmp; ++ return res; ++} ++ ++uint128_t sub128 (uint64_t a, uint64_t b) ++{ ++ uint64_t tmp = a - b; ++ uint64_t overflow = 0; ++ if (tmp > a) ++ overflow = -1; ++ ++ uint128_t res = overflow; ++ res <<= 64; ++ res += tmp; ++ return res; ++} ++ ++uint256_t sub256 (uint128_t a, uint128_t b) ++{ ++ uint128_t tmp = a - b; ++ uint128_t overflow = 0; ++ if (tmp > a) ++ overflow = -1; ++ ++ uint256_t res; ++ res.hi = overflow; ++ res.lo = tmp; ++ return res; ++} ++ ++/* { dg-final { scan-tree-dump-times "= .ADD_OVERFLOW \\(a_\[0-9\]+\\(D\\), b_\[0-9\]+\\(D\\)\\)" 5 "optimized" } } */ ++/* { dg-final { scan-tree-dump-times "= .SUB_OVERFLOW \\(a_\[0-9\]+\\(D\\), b_\[0-9\]+\\(D\\)\\)" 5 "optimized" } } */ +diff --git a/gcc/tree-ssa-math-opts.c b/gcc/tree-ssa-math-opts.c +index 4c89fddcf..716bf9e35 100644 +--- a/gcc/tree-ssa-math-opts.c ++++ b/gcc/tree-ssa-math-opts.c +@@ -3290,6 +3290,27 @@ convert_mult_to_fma (gimple *mul_stmt, tree op1, tree op2, + } + } + ++/* Check if the corresponding operation has wider equivalent on the target. */ ++ ++static bool ++wider_optab_check_p (optab op, machine_mode mode, int unsignedp) ++{ ++ machine_mode wider_mode; ++ FOR_EACH_WIDER_MODE (wider_mode, mode) ++ { ++ machine_mode next_mode; ++ if (optab_handler (op, wider_mode) != CODE_FOR_nothing ++ || (op == smul_optab ++ && GET_MODE_WIDER_MODE (wider_mode).exists (&next_mode) ++ && (find_widening_optab_handler ((unsignedp ++ ? umul_widen_optab ++ : smul_widen_optab), ++ next_mode, mode)))) ++ return true; ++ } ++ ++ return false; ++} + + /* Helper function of match_uaddsub_overflow. Return 1 + if USE_STMT is unsigned overflow check ovf != 0 for +@@ -3390,12 +3411,18 @@ match_uaddsub_overflow (gimple_stmt_iterator *gsi, gimple *stmt, + gimple *use_stmt; + + gcc_checking_assert (code == PLUS_EXPR || code == MINUS_EXPR); ++ optab op = code == PLUS_EXPR ? uaddv4_optab : usubv4_optab; ++ machine_mode mode = TYPE_MODE (type); ++ int unsignedp = TYPE_UNSIGNED (type); + if (!INTEGRAL_TYPE_P (type) +- || !TYPE_UNSIGNED (type) ++ || !unsignedp + || has_zero_uses (lhs) +- || has_single_use (lhs) +- || optab_handler (code == PLUS_EXPR ? uaddv4_optab : usubv4_optab, +- TYPE_MODE (type)) == CODE_FOR_nothing) ++ || has_single_use (lhs)) ++ return false; ++ ++ if (optab_handler (op, mode) == CODE_FOR_nothing ++ && (!flag_uaddsub_overflow_match_all ++ || !wider_optab_check_p (op, mode, unsignedp))) + return false; + + FOR_EACH_IMM_USE_FAST (use_p, iter, lhs) +-- +2.33.0 + diff --git a/0145-Match-double-sized-mul-pattern.patch b/0145-Match-double-sized-mul-pattern.patch new file mode 100644 index 0000000..f04d6ce --- /dev/null +++ b/0145-Match-double-sized-mul-pattern.patch @@ -0,0 +1,488 @@ +From 3be7a26a08772d014f54f7b1a0555ccca91115d6 Mon Sep 17 00:00:00 2001 +From: Pronin Alexander 00812787 +Date: Wed, 25 Jan 2023 15:04:07 +0300 +Subject: [PATCH 22/33] Match double sized mul pattern + +--- + gcc/match.pd | 136 +++++++++++++++++++++ + gcc/testsuite/gcc.dg/double_sized_mul-1.c | 141 ++++++++++++++++++++++ + gcc/testsuite/gcc.dg/double_sized_mul-2.c | 62 ++++++++++ + gcc/tree-ssa-math-opts.c | 80 ++++++++++++ + 4 files changed, 419 insertions(+) + create mode 100644 gcc/testsuite/gcc.dg/double_sized_mul-1.c + create mode 100644 gcc/testsuite/gcc.dg/double_sized_mul-2.c + +diff --git a/gcc/match.pd b/gcc/match.pd +index e98cd02e0..74f8ab999 100644 +--- a/gcc/match.pd ++++ b/gcc/match.pd +@@ -6390,3 +6390,139 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) + to the number of trailing zeroes. */ + (match (ctz_table_index @1 @2 @3) + (rshift (mult (bit_and:c (negate @1) @1) INTEGER_CST@2) INTEGER_CST@3)) ++ ++/* Match multiplication with double sized result. ++ ++ Consider the following calculations: ++ arg0 * arg1 = (2^(bit_size/2) * arg0_hi + arg0_lo) ++ * (2^(bit_size/2) * arg1_hi + arg1_lo) ++ arg0 * arg1 = 2^bit_size * arg0_hi * arg1_hi ++ + 2^(bit_size/2) * (arg0_hi * arg1_lo + arg0_lo * arg1_hi) ++ + arg0_lo * arg1_lo ++ ++ The products of high and low parts fits in bit_size values, thus they are ++ placed in high and low parts of result respectively. ++ ++ The sum of the mixed products may overflow, so we need a detection for that. ++ Also it has a bit_size/2 offset, thus it intersects with both high and low ++ parts of result. Overflow detection constant is bit_size/2 due to this. ++ ++ With this info: ++ arg0 * arg1 = 2^bit_size * arg0_hi * arg1_hi ++ + 2^(bit_size/2) * middle ++ + 2^bit_size * possible_middle_overflow ++ + arg0_lo * arg1_lo ++ arg0 * arg1 = 2^bit_size * (arg0_hi * arg1_hi + possible_middle_overflow) ++ + 2^(bit_size/2) * (2^(bit_size/2) * middle_hi + middle_lo) ++ + arg0_lo * arg1_lo ++ arg0 * arg1 = 2^bit_size * (arg0_hi * arg1_hi + middle_hi ++ + possible_middle_overflow) ++ + 2^(bit_size/2) * middle_lo ++ + arg0_lo * arg1_lo ++ ++ The last sum can produce overflow for the high result part. With this: ++ arg0 * arg1 = 2^bit_size * (arg0_hi * arg1_hi + possible_middle_overflow ++ + possible_res_lo_overflow + middle_hi) ++ + res_lo ++ = res_hi + res_lo ++ ++ This formula is quite big to fit into one match pattern with all of the ++ combinations of terms inside it. There are many helpers for better code ++ readability. ++ ++ The simplification basis is res_hi: assuming that res_lo only is not ++ real practical case for such calculations. ++ ++ Overflow handling is done via matching complex calculations: ++ the realpart and imagpart are quite handy here. */ ++/* Match low and high parts of the argument. */ ++(match (double_size_mul_arg_lo @0 @1) ++ (bit_and @0 INTEGER_CST@1) ++ (if (wi::to_wide (@1) ++ == wi::mask (TYPE_PRECISION (type) / 2, false, TYPE_PRECISION (type))))) ++(match (double_size_mul_arg_hi @0 @1) ++ (rshift @0 INTEGER_CST@1) ++ (if (wi::to_wide (@1) == TYPE_PRECISION (type) / 2))) ++ ++/* Match various argument parts products. */ ++(match (double_size_mul_lolo @0 @1) ++ (mult@4 (double_size_mul_arg_lo @0 @2) (double_size_mul_arg_lo @1 @3)) ++ (if (single_use (@4)))) ++(match (double_size_mul_hihi @0 @1) ++ (mult@4 (double_size_mul_arg_hi @0 @2) (double_size_mul_arg_hi @1 @3)) ++ (if (single_use (@4)))) ++(match (double_size_mul_lohi @0 @1) ++ (mult:c@4 (double_size_mul_arg_lo @0 @2) (double_size_mul_arg_hi @1 @3)) ++ (if (single_use (@4)))) ++ ++/* Match complex middle sum. */ ++(match (double_size_mul_middle_complex @0 @1) ++ (IFN_ADD_OVERFLOW@2 (double_size_mul_lohi @0 @1) (double_size_mul_lohi @1 @0)) ++ (if (num_imm_uses (@2) == 2))) ++ ++/* Match real middle results. */ ++(match (double_size_mul_middle @0 @1) ++ (realpart@2 (double_size_mul_middle_complex @0 @1)) ++ (if (num_imm_uses (@2) == 2))) ++(match (double_size_mul_middleres_lo @0 @1) ++ (lshift@3 (double_size_mul_middle @0 @1) INTEGER_CST@2) ++ (if (wi::to_wide (@2) == TYPE_PRECISION (type) / 2 ++ && single_use (@3)))) ++(match (double_size_mul_middleres_hi @0 @1) ++ (rshift@3 (double_size_mul_middle @0 @1) INTEGER_CST@2) ++ (if (wi::to_wide (@2) == TYPE_PRECISION (type) / 2 ++ && single_use (@3)))) ++ ++/* Match low result part. */ ++/* Number of uses may be < 2 in case when we are interested in ++ high part only. */ ++(match (double_size_mul_res_lo_complex @0 @1) ++ (IFN_ADD_OVERFLOW:c@2 ++ (double_size_mul_lolo:c @0 @1) (double_size_mul_middleres_lo @0 @1)) ++ (if (num_imm_uses (@2) <= 2))) ++(match (double_size_mul_res_lo @0 @1) ++ (realpart (double_size_mul_res_lo_complex @0 @1))) ++ ++/* Match overflow terms. */ ++(match (double_size_mul_overflow_check_lo @0 @1 @5) ++ (convert@4 (ne@3 ++ (imagpart@2 (double_size_mul_res_lo_complex@5 @0 @1)) integer_zerop)) ++ (if (single_use (@2) && single_use (@3) && single_use (@4)))) ++(match (double_size_mul_overflow_check_hi @0 @1) ++ (lshift@6 (convert@5 (ne@4 ++ (imagpart@3 (double_size_mul_middle_complex @0 @1)) integer_zerop)) ++ INTEGER_CST@2) ++ (if (wi::to_wide (@2) == TYPE_PRECISION (type) / 2 ++ && single_use (@3) && single_use (@4) && single_use (@5) ++ && single_use (@6)))) ++ ++/* Match all possible permutations for high result part calculations. */ ++(for op1 (double_size_mul_hihi ++ double_size_mul_overflow_check_hi ++ double_size_mul_middleres_hi) ++ op2 (double_size_mul_overflow_check_hi ++ double_size_mul_middleres_hi ++ double_size_mul_hihi) ++ op3 (double_size_mul_middleres_hi ++ double_size_mul_hihi ++ double_size_mul_overflow_check_hi) ++ (match (double_size_mul_candidate @0 @1 @2 @3) ++ (plus:c@2 ++ (plus:c@4 (double_size_mul_overflow_check_lo @0 @1 @3) (op1:c @0 @1)) ++ (plus:c@5 (op2:c @0 @1) (op3:c @0 @1))) ++ (if (single_use (@4) && single_use (@5)))) ++ (match (double_size_mul_candidate @0 @1 @2 @3) ++ (plus:c@2 (double_size_mul_overflow_check_lo @0 @1 @3) ++ (plus:c@4 (op1:c @0 @1) ++ (plus:c@5 (op2:c @0 @1) (op3:c @0 @1)))) ++ (if (single_use (@4) && single_use (@5)))) ++ (match (double_size_mul_candidate @0 @1 @2 @3) ++ (plus:c@2 (op1:c @0 @1) ++ (plus:c@4 (double_size_mul_overflow_check_lo @0 @1 @3) ++ (plus:c@5 (op2:c @0 @1) (op3:c @0 @1)))) ++ (if (single_use (@4) && single_use (@5)))) ++ (match (double_size_mul_candidate @0 @1 @2 @3) ++ (plus:c@2 (op1:c @0 @1) ++ (plus:c@4 (op2:c @0 @1) ++ (plus:c@5 (double_size_mul_overflow_check_lo @0 @1 @3) (op3:c @0 @1)))) ++ (if (single_use (@4) && single_use (@5))))) +diff --git a/gcc/testsuite/gcc.dg/double_sized_mul-1.c b/gcc/testsuite/gcc.dg/double_sized_mul-1.c +new file mode 100644 +index 000000000..4d475cc8a +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/double_sized_mul-1.c +@@ -0,0 +1,141 @@ ++/* { dg-do compile } */ ++/* fif-conversion-gimple and fuaddsub-overflow-match-all are required for ++ proper overflow detection in some cases. */ ++/* { dg-options "-O2 -fif-conversion-gimple -fuaddsub-overflow-match-all -fdump-tree-widening_mul-stats" } */ ++#include ++ ++typedef unsigned __int128 uint128_t; ++ ++uint16_t mul16 (uint8_t a, uint8_t b) ++{ ++ uint8_t a_lo = a & 0xF; ++ uint8_t b_lo = b & 0xF; ++ uint8_t a_hi = a >> 4; ++ uint8_t b_hi = b >> 4; ++ uint8_t lolo = a_lo * b_lo; ++ uint8_t lohi = a_lo * b_hi; ++ uint8_t hilo = a_hi * b_lo; ++ uint8_t hihi = a_hi * b_hi; ++ uint8_t middle = hilo + lohi; ++ uint8_t middle_hi = middle >> 4; ++ uint8_t middle_lo = middle << 4; ++ uint8_t res_lo = lolo + middle_lo; ++ uint8_t res_hi = hihi + middle_hi; ++ res_hi += (res_lo < middle_lo ? 1 : 0); ++ res_hi += (middle < hilo ? 0x10 : 0); ++ uint16_t res = ((uint16_t) res_hi) << 8; ++ res += res_lo; ++ return res; ++} ++ ++uint32_t mul32 (uint16_t a, uint16_t b) ++{ ++ uint16_t a_lo = a & 0xFF; ++ uint16_t b_lo = b & 0xFF; ++ uint16_t a_hi = a >> 8; ++ uint16_t b_hi = b >> 8; ++ uint16_t lolo = a_lo * b_lo; ++ uint16_t lohi = a_lo * b_hi; ++ uint16_t hilo = a_hi * b_lo; ++ uint16_t hihi = a_hi * b_hi; ++ uint16_t middle = hilo + lohi; ++ uint16_t middle_hi = middle >> 8; ++ uint16_t middle_lo = middle << 8; ++ uint16_t res_lo = lolo + middle_lo; ++ uint16_t res_hi = hihi + middle_hi; ++ res_hi += (res_lo < middle_lo ? 1 : 0); ++ res_hi += (middle < hilo ? 0x100 : 0); ++ uint32_t res = ((uint32_t) res_hi) << 16; ++ res += res_lo; ++ return res; ++} ++ ++uint64_t mul64 (uint32_t a, uint32_t b) ++{ ++ uint32_t a_lo = a & 0xFFFF; ++ uint32_t b_lo = b & 0xFFFF; ++ uint32_t a_hi = a >> 16; ++ uint32_t b_hi = b >> 16; ++ uint32_t lolo = a_lo * b_lo; ++ uint32_t lohi = a_lo * b_hi; ++ uint32_t hilo = a_hi * b_lo; ++ uint32_t hihi = a_hi * b_hi; ++ uint32_t middle = hilo + lohi; ++ uint32_t middle_hi = middle >> 16; ++ uint32_t middle_lo = middle << 16; ++ uint32_t res_lo = lolo + middle_lo; ++ uint32_t res_hi = hihi + middle_hi; ++ res_hi += (res_lo < middle_lo ? 1 : 0); ++ res_hi += (middle < hilo ? 0x10000 : 0); ++ uint64_t res = ((uint64_t) res_hi) << 32; ++ res += res_lo; ++ return res; ++} ++ ++uint128_t mul128 (uint64_t a, uint64_t b) ++{ ++ uint64_t a_lo = a & 0xFFFFFFFF; ++ uint64_t b_lo = b & 0xFFFFFFFF; ++ uint64_t a_hi = a >> 32; ++ uint64_t b_hi = b >> 32; ++ uint64_t lolo = a_lo * b_lo; ++ uint64_t lohi = a_lo * b_hi; ++ uint64_t hilo = a_hi * b_lo; ++ uint64_t hihi = a_hi * b_hi; ++ uint64_t middle = hilo + lohi; ++ uint64_t middle_hi = middle >> 32; ++ uint64_t middle_lo = middle << 32; ++ uint64_t res_lo = lolo + middle_lo; ++ uint64_t res_hi = hihi + middle_hi; ++ res_hi += (res_lo < middle_lo ? 1 : 0); ++ res_hi += (middle < hilo ? 0x100000000 : 0); ++ uint128_t res = ((uint128_t) res_hi) << 64; ++ res += res_lo; ++ return res; ++} ++ ++uint64_t mul64_perm (uint32_t a, uint32_t b) ++{ ++ uint32_t a_lo = a & 0xFFFF; ++ uint32_t b_lo = b & 0xFFFF; ++ uint32_t a_hi = a >> 16; ++ uint32_t b_hi = b >> 16; ++ uint32_t lolo = a_lo * b_lo; ++ uint32_t lohi = a_lo * b_hi; ++ uint32_t hilo = a_hi * b_lo; ++ uint32_t hihi = a_hi * b_hi; ++ uint32_t middle = hilo + lohi; ++ uint32_t middle_hi = middle >> 16; ++ uint32_t middle_lo = middle << 16; ++ uint32_t res_lo = lolo + middle_lo; ++ uint32_t res_hi = hihi + middle_hi; ++ res_hi = res_lo < middle_lo ? res_hi + 1 : res_hi; ++ res_hi = middle < hilo ? res_hi + 0x10000 : res_hi; ++ uint64_t res = ((uint64_t) res_hi) << 32; ++ res += res_lo; ++ return res; ++} ++ ++uint128_t mul128_perm (uint64_t a, uint64_t b) ++{ ++ uint64_t a_lo = a & 0xFFFFFFFF; ++ uint64_t b_lo = b & 0xFFFFFFFF; ++ uint64_t a_hi = a >> 32; ++ uint64_t b_hi = b >> 32; ++ uint64_t lolo = a_lo * b_lo; ++ uint64_t lohi = a_lo * b_hi; ++ uint64_t hilo = a_hi * b_lo; ++ uint64_t hihi = a_hi * b_hi; ++ uint64_t middle = hilo + lohi; ++ uint64_t middle_hi = middle >> 32; ++ uint64_t middle_lo = middle << 32; ++ uint64_t res_lo = lolo + middle_lo; ++ uint64_t res_hi = hihi + middle_hi; ++ res_hi = res_lo < middle_lo ? res_hi + 1 : res_hi; ++ res_hi = middle < hilo ? res_hi + 0x100000000 : res_hi; ++ uint128_t res = ((uint128_t) res_hi) << 64; ++ res += res_lo; ++ return res; ++} ++ ++/* { dg-final { scan-tree-dump-times "double sized mul optimized: 1" 6 "widening_mul" } } */ +diff --git a/gcc/testsuite/gcc.dg/double_sized_mul-2.c b/gcc/testsuite/gcc.dg/double_sized_mul-2.c +new file mode 100644 +index 000000000..cc6e5af25 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/double_sized_mul-2.c +@@ -0,0 +1,62 @@ ++/* { dg-do compile } */ ++/* fif-conversion-gimple is required for proper overflow detection ++ in some cases. */ ++/* { dg-options "-O2 -fif-conversion-gimple -fuaddsub-overflow-match-all -fdump-tree-widening_mul-stats" } */ ++#include ++ ++typedef unsigned __int128 uint128_t; ++typedef struct uint256_t ++{ ++ uint128_t lo; ++ uint128_t hi; ++} uint256_t; ++ ++uint64_t mul64_double_use (uint32_t a, uint32_t b) ++{ ++ uint32_t a_lo = a & 0xFFFF; ++ uint32_t b_lo = b & 0xFFFF; ++ uint32_t a_hi = a >> 16; ++ uint32_t b_hi = b >> 16; ++ uint32_t lolo = a_lo * b_lo; ++ uint32_t lohi = a_lo * b_hi; ++ uint32_t hilo = a_hi * b_lo; ++ uint32_t hihi = a_hi * b_hi; ++ uint32_t middle = hilo + lohi; ++ uint32_t middle_hi = middle >> 16; ++ uint32_t middle_lo = middle << 16; ++ uint32_t res_lo = lolo + middle_lo; ++ uint32_t res_hi = hihi + middle_hi; ++ res_hi += (res_lo < middle_lo ? 1 : 0); ++ res_hi += (middle < hilo ? 0x10000 : 0); ++ uint64_t res = ((uint64_t) res_hi) << 32; ++ res += res_lo; ++ return res + lolo; ++} ++ ++uint256_t mul256 (uint128_t a, uint128_t b) ++{ ++ uint128_t a_lo = a & 0xFFFFFFFFFFFFFFFF; ++ uint128_t b_lo = b & 0xFFFFFFFFFFFFFFFF; ++ uint128_t a_hi = a >> 64; ++ uint128_t b_hi = b >> 64; ++ uint128_t lolo = a_lo * b_lo; ++ uint128_t lohi = a_lo * b_hi; ++ uint128_t hilo = a_hi * b_lo; ++ uint128_t hihi = a_hi * b_hi; ++ uint128_t middle = hilo + lohi; ++ uint128_t middle_hi = middle >> 64; ++ uint128_t middle_lo = middle << 64; ++ uint128_t res_lo = lolo + middle_lo; ++ uint128_t res_hi = hihi + middle_hi; ++ res_hi += (res_lo < middle_lo ? 1 : 0); ++ /* Constant is to big warning WA */ ++ uint128_t overflow_tmp = (middle < hilo ? 1 : 0); ++ overflow_tmp <<= 64; ++ res_hi += overflow_tmp; ++ uint256_t res; ++ res.lo = res_lo; ++ res.hi = res_hi; ++ return res; ++} ++ ++/* { dg-final { scan-tree-dump-not "double sized mul optimized" "widening_mul" } } */ +diff --git a/gcc/tree-ssa-math-opts.c b/gcc/tree-ssa-math-opts.c +index 716bf9e35..a81d7501c 100644 +--- a/gcc/tree-ssa-math-opts.c ++++ b/gcc/tree-ssa-math-opts.c +@@ -182,6 +182,9 @@ static struct + + /* Number of divmod calls inserted. */ + int divmod_calls_inserted; ++ ++ /* Number of optimized double sized multiplications. */ ++ int double_sized_mul_optimized; + } widen_mul_stats; + + /* The instance of "struct occurrence" representing the highest +@@ -3708,6 +3711,78 @@ convert_to_divmod (gassign *stmt) + return true; + } + ++/* Pattern matcher for double sized multiplication defined in match.pd. */ ++extern bool gimple_double_size_mul_candidate (tree, tree*, tree (*)(tree)); ++ ++static bool ++convert_double_size_mul (gimple_stmt_iterator *gsi, gimple *stmt) ++{ ++ gimple *use_stmt, *complex_res_lo; ++ gimple_stmt_iterator insert_before; ++ imm_use_iterator use_iter; ++ tree match[4]; // arg0, arg1, res_hi, complex_res_lo ++ tree arg0, arg1, widen_mult, new_type, tmp; ++ tree lhs = gimple_assign_lhs (stmt); ++ location_t loc = UNKNOWN_LOCATION; ++ machine_mode mode; ++ ++ if (!gimple_double_size_mul_candidate (lhs, match, NULL)) ++ return false; ++ ++ new_type = build_nonstandard_integer_type ( ++ TYPE_PRECISION (TREE_TYPE (match[0])) * 2, 1); ++ mode = TYPE_MODE (new_type); ++ ++ /* Early return if the target multiplication doesn't exist on target. */ ++ if (optab_handler (smul_optab, mode) == CODE_FOR_nothing ++ && !wider_optab_check_p (smul_optab, mode, 1)) ++ return false; ++ ++ /* Determine the point where the wide multiplication ++ should be inserted. Complex low res is OK since it is required ++ by both high and low part getters, thus it dominates both of them. */ ++ complex_res_lo = SSA_NAME_DEF_STMT (match[3]); ++ insert_before = gsi_for_stmt (complex_res_lo); ++ gsi_next (&insert_before); ++ ++ /* Create the widen multiplication. */ ++ arg0 = build_and_insert_cast (&insert_before, loc, new_type, match[0]); ++ arg1 = build_and_insert_cast (&insert_before, loc, new_type, match[1]); ++ widen_mult = build_and_insert_binop (&insert_before, loc, "widen_mult", ++ MULT_EXPR, arg0, arg1); ++ ++ /* Find the mult low part getter. */ ++ FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, match[3]) ++ if (gimple_assign_rhs_code (use_stmt) == REALPART_EXPR) ++ BREAK_FROM_IMM_USE_STMT (use_iter); ++ ++ /* Create high and low (if needed) parts extractors. */ ++ /* Low part. */ ++ if (use_stmt) ++ { ++ loc = gimple_location (use_stmt); ++ tmp = build_and_insert_cast (&insert_before, loc, ++ TREE_TYPE (gimple_get_lhs (use_stmt)), ++ widen_mult); ++ gassign *new_stmt = gimple_build_assign (gimple_get_lhs (use_stmt), ++ NOP_EXPR, tmp); ++ gsi_replace (&insert_before, new_stmt, true); ++ } ++ ++ /* High part. */ ++ loc = gimple_location (stmt); ++ tmp = build_and_insert_binop (gsi, loc, "widen_mult_hi", ++ RSHIFT_EXPR, widen_mult, ++ build_int_cst (new_type, ++ TYPE_PRECISION (new_type) / 2)); ++ tmp = build_and_insert_cast (gsi, loc, TREE_TYPE (lhs), tmp); ++ gassign *new_stmt = gimple_build_assign (lhs, NOP_EXPR, tmp); ++ gsi_replace (gsi, new_stmt, true); ++ ++ widen_mul_stats.double_sized_mul_optimized++; ++ return true; ++} ++ + /* Find integer multiplications where the operands are extended from + smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR + where appropriate. */ +@@ -3801,6 +3876,9 @@ math_opts_dom_walker::after_dom_children (basic_block bb) + break; + + case PLUS_EXPR: ++ if (convert_double_size_mul (&gsi, stmt)) ++ break; ++ __attribute__ ((fallthrough)); + case MINUS_EXPR: + if (!convert_plusminus_to_widen (&gsi, stmt, code)) + match_uaddsub_overflow (&gsi, stmt, code); +@@ -3892,6 +3970,8 @@ pass_optimize_widening_mul::execute (function *fun) + widen_mul_stats.fmas_inserted); + statistics_counter_event (fun, "divmod calls inserted", + widen_mul_stats.divmod_calls_inserted); ++ statistics_counter_event (fun, "double sized mul optimized", ++ widen_mul_stats.double_sized_mul_optimized); + + return cfg_changed ? TODO_cleanup_cfg : 0; + } +-- +2.33.0 + diff --git a/0146-LOOP-CRC32-Add-Crc32-Optimization-in-Gzip-For-crc32-.patch b/0146-LOOP-CRC32-Add-Crc32-Optimization-in-Gzip-For-crc32-.patch new file mode 100644 index 0000000..a9a8e94 --- /dev/null +++ b/0146-LOOP-CRC32-Add-Crc32-Optimization-in-Gzip-For-crc32-.patch @@ -0,0 +1,2354 @@ +From 179412c66d0cdd6a48ef1c29acae90908102a1c9 Mon Sep 17 00:00:00 2001 +From: xingyushuai +Date: Mon, 24 Apr 2023 09:34:35 +0800 +Subject: [PATCH 08/13] [LOOP CRC32]Add Crc32 Optimization in Gzip For crc32 + algorithm in APBC int_gzip. Match crc32 lookup table algorithm. An example + for crc32 lookup table alg: ```c do { c = crc_32_tab[((int)c ^ (*s++)) & + 0xff] ^ (c >> 8); } while (--n); + +Usage: `gcc -O3 -march=armv8.1-a -floop-crc yourfile.c` +Node: The cpu you use needs to support the crc32 instructions +--- + gcc/config/aarch64/aarch64-builtins.c | 29 + + gcc/config/aarch64/aarch64-protos.h | 1 + + gcc/config/aarch64/aarch64.c | 12 + + gcc/doc/tm.texi | 9 + + gcc/doc/tm.texi.in | 2 + + gcc/match.pd | 146 +-- + gcc/passes.def | 2 +- + gcc/target.def | 14 + + gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c | 85 -- + .../tree-ssa/loop-crc-1.c.042t.loop_crc | 90 -- + gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c | 88 -- + .../tree-ssa/loop-crc-4.c.042t.loop_crc | 0 + .../loop-crc-calculation-check-fail.c | 156 --- + ...crc-calculation-check-fail.c.042t.loop_crc | 64 -- + .../loop-crc-calculation-check-fail.s | 329 ------- + ...crc-3.c => loop-crc-loop-condition-fail.c} | 6 +- + ...op-crc-4.c => loop-crc-loop-form-fail-2.c} | 7 +- + .../gcc.dg/tree-ssa/loop-crc-loop-form-fail.c | 3 +- + .../gcc.dg/tree-ssa/loop-crc-sucess.c | 7 +- + .../tree-ssa/loop-crc-table-check-fail.c | 3 +- + gcc/tree-ssa-loop-crc.c | 903 +++++++++++++++--- + 21 files changed, 873 insertions(+), 1083 deletions(-) + delete mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c + delete mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc + delete mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c + delete mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c.042t.loop_crc + delete mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c + delete mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc + delete mode 100644 gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s + rename gcc/testsuite/gcc.dg/tree-ssa/{loop-crc-3.c => loop-crc-loop-condition-fail.c} (97%) + rename gcc/testsuite/gcc.dg/tree-ssa/{loop-crc-4.c => loop-crc-loop-form-fail-2.c} (95%) + +diff --git a/gcc/config/aarch64/aarch64-builtins.c b/gcc/config/aarch64/aarch64-builtins.c +index d92157dff..1e8b046da 100644 +--- a/gcc/config/aarch64/aarch64-builtins.c ++++ b/gcc/config/aarch64/aarch64-builtins.c +@@ -441,6 +441,12 @@ typedef struct + #define VAR1(T, N, MAP, A) \ + AARCH64_SIMD_BUILTIN_##T##_##N##A, + ++enum aarch64_crc_builtins{ ++ AARCH64_BUILTIN_CRC32B, ++ AARCH64_BUILTIN_CRC32H, ++ AARCH64_BUILTIN_CRC32W, ++}; ++ + enum aarch64_builtins + { + AARCH64_BUILTIN_MIN, +@@ -1321,6 +1327,29 @@ aarch64_general_builtin_decl (unsigned code, bool) + + return aarch64_builtin_decls[code]; + } ++/* Implement TARGET_GET_CRC_BUILTIN_CODE */ ++unsigned ++get_crc_builtin_code(unsigned code, bool) ++{ ++ if (code > AARCH64_BUILTIN_CRC32W) ++ return AARCH64_BUILTIN_MIN; ++ ++ unsigned res = AARCH64_BUILTIN_MIN; ++ switch (code) { ++ case AARCH64_BUILTIN_CRC32B: ++ res = AARCH64_BUILTIN_crc32b; ++ break; ++ case AARCH64_BUILTIN_CRC32H: ++ res = AARCH64_BUILTIN_crc32h; ++ break; ++ case AARCH64_BUILTIN_CRC32W: ++ res = AARCH64_BUILTIN_crc32w; ++ break; ++ default: ++ break; ++ } ++ return res; ++} + + typedef enum + { +diff --git a/gcc/config/aarch64/aarch64-protos.h b/gcc/config/aarch64/aarch64-protos.h +index 9b6d309a7..a0ca662bc 100644 +--- a/gcc/config/aarch64/aarch64-protos.h ++++ b/gcc/config/aarch64/aarch64-protos.h +@@ -723,6 +723,7 @@ tree aarch64_general_fold_builtin (unsigned int, tree, unsigned int, tree *); + gimple *aarch64_general_gimple_fold_builtin (unsigned int, gcall *); + rtx aarch64_general_expand_builtin (unsigned int, tree, rtx, int); + tree aarch64_general_builtin_decl (unsigned, bool); ++unsigned get_crc_builtin_code(unsigned , bool); + tree aarch64_general_builtin_rsqrt (unsigned int); + tree aarch64_builtin_vectorized_function (unsigned int, tree, tree); + +diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c +index cbdde11b0..b8407c612 100644 +--- a/gcc/config/aarch64/aarch64.c ++++ b/gcc/config/aarch64/aarch64.c +@@ -13735,6 +13735,15 @@ aarch64_builtin_decl (unsigned int code, bool initialize_p) + gcc_unreachable (); + } + ++/* Implement TARGET_GET_CRC_BUILTIN_CODE. */ ++static unsigned ++aarch64_get_crc_builtin_code(unsigned code, bool initialize_p) ++{ ++ unsigned subcode = get_crc_builtin_code(code,initialize_p); ++ unsigned res = subcode << AARCH64_BUILTIN_SHIFT; ++ return res; ++} ++ + /* Return true if it is safe and beneficial to use the approximate rsqrt optabs + to optimize 1.0/sqrt. */ + +@@ -23911,6 +23920,9 @@ aarch64_run_selftests (void) + #undef TARGET_BUILTIN_DECL + #define TARGET_BUILTIN_DECL aarch64_builtin_decl + ++#undef TARGET_GET_CRC_BUILTIN_CODE ++#define TARGET_GET_CRC_BUILTIN_CODE aarch64_get_crc_builtin_code ++ + #undef TARGET_BUILTIN_RECIPROCAL + #define TARGET_BUILTIN_RECIPROCAL aarch64_builtin_reciprocal + +diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi +index 0508fce57..b46418d0b 100644 +--- a/gcc/doc/tm.texi ++++ b/gcc/doc/tm.texi +@@ -11610,6 +11610,15 @@ If @var{code} is out of range the function should return + @code{error_mark_node}. + @end deftypefn + ++@deftypefn {Target Hook} unsigned TARGET_GET_CRC_BUILTIN_CODE (unsigned @var{code}, bool @var{initialize_p}) ++Define this hook to get crc32 builtin code. It should be a function that ++returns the crc32 builtin function code @var{code}. ++If there is no such builtin and it cannot be initialized at this time ++if @var{initialize_p} is true the function should return @code{NULL_TREE}. ++If @var{code} is out of range the function should return ++@code{error_mark_node}. ++@end deftypefn ++ + @deftypefn {Target Hook} rtx TARGET_EXPAND_BUILTIN (tree @var{exp}, rtx @var{target}, rtx @var{subtarget}, machine_mode @var{mode}, int @var{ignore}) + + Expand a call to a machine specific built-in function that was set up by +diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in +index 3b70ea484..2663547c7 100644 +--- a/gcc/doc/tm.texi.in ++++ b/gcc/doc/tm.texi.in +@@ -7941,6 +7941,8 @@ to by @var{ce_info}. + + @hook TARGET_BUILTIN_DECL + ++@hook TARGET_GET_CRC_BUILTIN_CODE ++ + @hook TARGET_EXPAND_BUILTIN + + @hook TARGET_RESOLVE_OVERLOADED_BUILTIN +diff --git a/gcc/match.pd b/gcc/match.pd +index 87b316953..0f92003f7 100644 +--- a/gcc/match.pd ++++ b/gcc/match.pd +@@ -3487,160 +3487,17 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) + ) + #endif + +- +-#if GIMPLE +-(if (canonicalize_math_p ()) +-/* These patterns are mostly used by PHIOPT to move some operations outside of +- the if statements. They should be done late because it gives jump threading +- and few other passes to reduce what is going on. */ +-/* a ? x op POW2 : x -> x op (a ? POW2 : 0). */ +- (for op (plus minus bit_ior bit_xor lshift rshift lrotate rrotate) +- (simplify +- (cond @0 (op:s @1 INTEGER_CST@2) @1) +- /* powerof2cst */ +- (if (INTEGRAL_TYPE_P (type) && integer_pow2p (@2)) +- (with { +- tree shift = build_int_cst (integer_type_node, tree_log2 (@2)); +- } +- (op @1 (lshift (convert (convert:boolean_type_node @0)) { shift; }))) +- ) +- ) +- ) +-) +-#endif +- +-#if GIMPLE +-/* These patterns are mostly used by FORWPROP to move some operations outside of +- the if statements. They should be done late because it gives jump threading +- and few other passes to reduce what is going on. */ +-/* Mul64 is defined as a multiplication algorithm which compute two 64-bit integers to one 128-bit integer +- (i64 ResLo, i64 ResHi) = Mul64(i64 In0, i64 In1) { +- In0Lo = In0(D) & 4294967295; +- In0Hi = In0(D) >> 32; +- In1Lo = In1(D) & 4294967295; +- In1Hi = In1(D) >> 32; +- Mull_01 = In0Hi * In1Lo; +- Addc = In0Lo * In1Hi + Mull_01; +- addc32 = Addc << 32; +- ResLo = In0Lo * In1Lo + addc32; +- ResHi = ((long unsigned int) (addc32 > ResLo)) + +- (((long unsigned int) (Mull_01 > Addc)) << 32) + (Addc >> 32) + In0Hi * In1Hi; +- } */ +- (simplify +- (plus +- (plus +- (convert +- (gt @10 +- (plus +- (mult @4 @6) +- (lshift@10 @9 @3)))) +- (lshift +- (convert +- (gt @8 @9)) @3)) +- (plus@11 +- (rshift +- (plus@9 +- (mult (bit_and@4 SSA_NAME@0 @2) @7) +- (mult@8 @5 (bit_and@6 SSA_NAME@1 INTEGER_CST@2))) @3) +- (mult (rshift@5 SSA_NAME@0 @3) +- (rshift@7 SSA_NAME@1 INTEGER_CST@3)))) +- (if (INTEGRAL_TYPE_P (type) && INTEGRAL_TYPE_P (TREE_TYPE (@0)) && types_match (@0, @1) && +- TYPE_PRECISION (type) == 64) +- (with { +- tree i128_type = build_nonstandard_integer_type (128, TYPE_UNSIGNED (type)); +- tree shift = build_int_cst (integer_type_node, 64); +- //direct_internal_fn_supported_p (UMULH, type, OPTIMIZE_FOR_BOTH) +- } +- (convert:type (rshift +- (mult (convert:i128_type @0) (convert:i128_type @1)) { shift; }))) +- ) +- ) +- +- /* (i64 ResLo, i64 ResHi) = Mul64(i64 In0, i64 In1) { +- In0Lo = In0(D) & 4294967295; +- In0Hi = In0(D) >> 32; +- In1Lo = In1(D) & 4294967295; +- In1Hi = In1(D) >> 32; +- Mull_01 = In0Hi * In1Lo; +- Addc = In0Lo * In1Hi + Mull_01; +- addc32 = Addc << 32; +- ResLo = In0(D) * In1(D); +- ResHi = ((long unsigned int) (addc32 > ResLo)) + +- (((long unsigned int) (Mull_01 > Addc)) << 32) + (Addc >> 32) + In0Hi * In1Hi; +- } */ +- (simplify +- (plus +- (plus +- (convert +- (gt (lshift@10 @9 @3) +- (mult @0 @1))) +- (lshift +- (convert +- (gt @8 @9)) @3)) +- (plus@11 +- (rshift +- (plus@9 +- (mult (bit_and@4 SSA_NAME@0 @2) @7) +- (mult@8 @5 (bit_and@6 SSA_NAME@1 INTEGER_CST@2))) @3) +- (mult (rshift@5 SSA_NAME@0 @3) +- (rshift@7 SSA_NAME@1 INTEGER_CST@3)))) +- (if (INTEGRAL_TYPE_P (type) && INTEGRAL_TYPE_P (TREE_TYPE (@0)) && types_match (@0, @1) && +- TYPE_PRECISION (type) == 64) +- (with { +- tree i128_type = build_nonstandard_integer_type (128, TYPE_UNSIGNED (type)); +- tree shift = build_int_cst (integer_type_node, 64); +- //direct_internal_fn_supported_p (UMULH, type, OPTIMIZE_FOR_BOTH) +- } +- (convert:type (rshift +- (mult (convert:i128_type @0) (convert:i128_type @1)) { shift; }))) +- ) +- ) +-#endif +- +-#if GIMPLE +-/* These patterns are mostly used by FORWPROP to move some operations outside of +- the if statements. They should be done late because it gives jump threading +- and few other passes to reduce what is going on. */ +- /* +- In0Lo = In0(D) & 4294967295; +- In0Hi = In0(D) >> 32; +- In1Lo = In1(D) & 4294967295; +- In1Hi = In1(D) >> 32; +- Addc = In0Lo * In1Hi + In0Hi * In1Lo; +- addc32 = Addc << 32; +- ResLo = In0Lo * In1Lo + addc32 +- */ +- (simplify +- (plus (mult @4 @5) +- (lshift +- (plus +- (mult (bit_and@4 SSA_NAME@0 @2) (rshift SSA_NAME@1 @3)) +- (mult (rshift SSA_NAME@0 @3) (bit_and@5 SSA_NAME@1 INTEGER_CST@2))) INTEGER_CST@3)) +- (if (INTEGRAL_TYPE_P (type) && INTEGRAL_TYPE_P (TREE_TYPE (@0)) && types_match (@0, @1) && +- TYPE_PRECISION (type) == 64) +- (with { +- tree i128_type = build_nonstandard_integer_type (128, TYPE_UNSIGNED (type)); +- tree shift = build_int_cst (integer_type_node, 64); +- //direct_internal_fn_supported_p (UMULH, type, OPTIMIZE_FOR_BOTH) +- } +- (mult (convert:type @0) (convert:type @1))) +- ) +- ) +-#endif +- +- + #if GIMPLE + /* Try to match */ + /* + _4 = (int) _3; //NOP_EXPR (SSA_NAME @2) + _5 = _4 ^ c_10; //BIT_XOR_EXPR (SSA_NAME@1, SSA_NAME) +-_6 = _5 & 255; //BIT_XOR_EXPR (SSA_NAME, INTEGER_CST@3) ++_6 = _5 & 255; //BIT_AND_EXPR (SSA_NAME, INTEGER_CST@3) + */ + (match (crc_match_index @1 @2 @3) + (bit_and (bit_xor (nop SSA_NAME@2) SSA_NAME@1) INTEGER_CST@3) + (if (INTEGRAL_TYPE_P (type) && tree_to_uhwi(@3) == 255)) + ) +- + #endif + + #if GIMPLE +@@ -3653,7 +3510,6 @@ c_19 = _7 ^ _8; // BIT_XOR_EXPR (SSA_NAME@3, SSA_NAME) + (bit_xor SSA_NAME@3 (rshift SSA_NAME@1 INTEGER_CST@2)) + (if (INTEGRAL_TYPE_P (type) && tree_to_uhwi(@2) == 8)) + ) +- + #endif + + /* Simplification moved from fold_cond_expr_with_comparison. It may also +diff --git a/gcc/passes.def b/gcc/passes.def +index 7abd946ce..df7d65733 100644 +--- a/gcc/passes.def ++++ b/gcc/passes.def +@@ -92,7 +92,7 @@ along with GCC; see the file COPYING3. If not see + NEXT_PASS (pass_cd_dce); + NEXT_PASS (pass_phiopt, true /* early_p */); + NEXT_PASS (pass_array_widen_compare); +- NEXT_PASS (pass_loop_crc); ++ NEXT_PASS (pass_loop_crc); + NEXT_PASS (pass_tail_recursion); + NEXT_PASS (pass_convert_switch); + NEXT_PASS (pass_cleanup_eh); +diff --git a/gcc/target.def b/gcc/target.def +index 202056411..34d3561bd 100644 +--- a/gcc/target.def ++++ b/gcc/target.def +@@ -2421,6 +2421,20 @@ If @var{code} is out of range the function should return\n\ + @code{error_mark_node}.", + tree, (unsigned code, bool initialize_p), NULL) + ++/* Initialize (if INITIALIZE_P is true) and return the real code of ++ target-specific built-in function . ++ Return NULL if that is not possible. Return error_mark_node if CODE ++ is outside of the range of valid crc32 codes. */ ++DEFHOOK ++(get_crc_builtin_code, ++ "Define this hook to get crc32 builtin code. It should be a function that\n\ ++returns the crc32 builtin function code @var{code}.\n\ ++If there is no such builtin and it cannot be initialized at this time\n\ ++if @var{initialize_p} is true the function should return @code{NULL_TREE}.\n\ ++If @var{code} is out of range the function should return\n\ ++@code{error_mark_node}.", ++ unsigned , (unsigned code, bool initialize_p), NULL) ++ + /* Expand a target-specific builtin. */ + DEFHOOK + (expand_builtin, +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c +deleted file mode 100644 +index 07f9e01ec..000000000 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c ++++ /dev/null +@@ -1,85 +0,0 @@ +-/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ +-/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ +- +-#include +-#include +-typedef unsigned long ulg; +-typedef unsigned char uch; +- +-static const ulg crc_32_tab[] = { +- 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, +- 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, +- 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, +- 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, +- 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, +- 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, +- 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, +- 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, +- 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, +- 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, +- 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, +- 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, +- 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, +- 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, +- 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, +- 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, +- 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, +- 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, +- 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, +- 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, +- 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, +- 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, +- 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, +- 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, +- 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, +- 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, +- 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, +- 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, +- 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, +- 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, +- 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, +- 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, +- 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, +- 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, +- 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, +- 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, +- 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, +- 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, +- 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, +- 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, +- 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, +- 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, +- 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, +- 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, +- 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, +- 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, +- 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, +- 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, +- 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, +- 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, +- 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, +- 0x2d02ef8dL +-}; +- +-ulg updcrc(s, n) +- uch *s; /* pointer to bytes to pump through */ +- unsigned n; /* number of bytes in s[] */ +-{ +- register ulg c; /* temporary variable */ +- +- static ulg crc = (ulg)0xffffffffL; /* shift register contents */ +- +- if (s == NULL) { +- c = 0xffffffffL; +- } else { +- c = crc; +- if (n) +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); +- } while (--n); +- } +- crc = c; +- return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ +-} +-/* { dg-final { scan-tree-dump-times "Processing loop" 1 "loop_crc"} } */ +-/* { dg-final { scan-tree-dump-times "the loop can be optimized" 1 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc +deleted file mode 100644 +index c726059f3..000000000 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-1.c.042t.loop_crc ++++ /dev/null +@@ -1,90 +0,0 @@ +- +-;; Function updcrc (updcrc, funcdef_no=0, decl_uid=3687, cgraph_uid=1, symbol_order=1) +- +-;; 2 loops found +-;; +-;; Loop 0 +-;; header 0, latch 1 +-;; depth 0, outer -1 +-;; nodes: 0 1 2 3 6 4 7 5 +-;; +-;; Loop 1 +-;; header 4, latch 7 +-;; depth 1, outer 0 +-;; nodes: 4 7 +-;; 2 succs { 5 3 } +-;; 3 succs { 6 5 } +-;; 6 succs { 4 } +-;; 4 succs { 7 5 } +-;; 7 succs { 4 } +-;; 5 succs { 1 } +- +-Starting the loop_crc pass +-====================================== +-Processing loop 1: +-====================================== +-;; +-;; Loop 1 +-;; header 4, latch 7 +-;; depth 1, outer 0 +-;; nodes: 4 7 +- +- +-The 1th loop form is success matched,and the loop can be optimized. +-updcrc (uch * s, unsigned int n) +-{ +- static ulg crc = 4294967295; +- register ulg c; +- unsigned char _2; +- long unsigned int _3; +- long unsigned int _4; +- long unsigned int _5; +- long unsigned int _6; +- long unsigned int _7; +- ulg _21; +- +- : +- if (s_12(D) == 0B) +- goto ; [INV] +- else +- goto ; [INV] +- +- : +- c_14 = crc; +- if (n_15(D) != 0) +- goto ; [INV] +- else +- goto ; [INV] +- +- : +- +- : +- # s_8 = PHI +- # n_9 = PHI +- # c_10 = PHI +- s_16 = s_8 + 1; +- _2 = *s_8; +- _3 = (long unsigned int) _2; +- _4 = _3 ^ c_10; +- _5 = _4 & 255; +- _6 = crc_32_tab[_5]; +- _7 = c_10 >> 8; +- c_17 = _6 ^ _7; +- n_18 = n_9 + 4294967295; +- if (n_18 != 0) +- goto ; [INV] +- else +- goto ; [INV] +- +- : +- goto ; [100.00%] +- +- : +- # c_11 = PHI <4294967295(2), c_14(3), c_17(4)> +- crc = c_11; +- _21 = c_11 ^ 4294967295; +- return _21; +- +-} +- +- +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c +deleted file mode 100644 +index f73c4d550..000000000 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-2.c ++++ /dev/null +@@ -1,88 +0,0 @@ +-/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ +-/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ +- +-#include +-#include +-typedef unsigned long ulg; +-typedef unsigned char uch; +- +-static const ulg crc_32_tab[] = { +- 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, +- 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, +- 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, +- 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, +- 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, +- 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, +- 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, +- 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, +- 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, +- 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, +- 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, +- 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, +- 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, +- 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, +- 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, +- 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, +- 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, +- 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, +- 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, +- 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, +- 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, +- 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, +- 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, +- 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, +- 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, +- 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, +- 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, +- 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, +- 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, +- 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, +- 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, +- 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, +- 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, +- 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, +- 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, +- 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, +- 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, +- 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, +- 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, +- 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, +- 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, +- 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, +- 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, +- 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, +- 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, +- 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, +- 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, +- 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, +- 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, +- 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, +- 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, +- 0x2d02ef8dL +-}; +- +-ulg updcrc(s, n) +- uch *s; /* pointer to bytes to pump through */ +- unsigned n; /* number of bytes in s[] */ +-{ +- register ulg c; /* temporary variable */ +- +- static ulg crc = (ulg)0xffffffffL; /* shift register contents */ +- +- if (s == NULL) { +- c = 0xffffffffL; +- } else { +- c = crc; +- if (n) +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); +- for (int i = 0; i < 5; i++) { +- c++; +- } +- +- } while (--n); +- } +- crc = c; +- return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ +-} +-/* { dg-final { scan-tree-dump-times "Wrong loop form for crc matching." 1 "loop_crc"} } */ +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c.042t.loop_crc b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c.042t.loop_crc +deleted file mode 100644 +index e69de29bb..000000000 +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c +deleted file mode 100644 +index 71b25f537..000000000 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c ++++ /dev/null +@@ -1,156 +0,0 @@ +-/* { dg-do compile { target {{ aarch64*-*-linux* } && lp64 } } } */ +-/* { dg-options "-O3 -mabi=lp64 -floop-crc -fdump-tree-loop_crc-details" } */ +- +-#include +-#include +-typedef unsigned long ulg; +-typedef unsigned char uch; +- +-static const ulg crc_32_tab[] = { +- 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L, +- 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L, +- 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L, +- 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL, +- 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L, +- 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L, +- 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L, +- 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL, +- 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L, +- 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL, +- 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L, +- 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L, +- 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L, +- 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL, +- 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL, +- 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L, +- 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL, +- 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L, +- 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L, +- 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L, +- 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL, +- 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L, +- 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L, +- 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL, +- 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L, +- 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L, +- 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L, +- 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L, +- 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L, +- 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL, +- 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL, +- 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L, +- 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L, +- 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL, +- 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL, +- 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L, +- 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL, +- 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L, +- 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL, +- 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L, +- 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL, +- 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L, +- 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L, +- 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL, +- 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L, +- 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L, +- 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L, +- 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L, +- 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L, +- 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L, +- 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL, +- 0x2d02ef8dL +-}; +- +-int test[5] = {0}; +- +-ulg updcrc(s, n) +- uch *s; +- unsigned n; +-{ +- register ulg c; +- +- static ulg crc = (ulg)0xffffffffL; +- int a = 0; +- if (s == NULL) { +- c = 0xffffffffL; +- } else { +- c = crc; +- if (n) +- do { +- a++; +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) ; +- } while (--n) ; +- } +- crc = c; +- return c ^ 0xffffffffL*a; +-} +- +-ulg updcrc1(s, n) +- uch *s; +- unsigned n; +-{ +- register ulg c; +- +- static ulg crc = (ulg)0xffffffffL; +- if (s == NULL) { +- c = 0xffffffffL; +- } else { +- c = crc; +- unsigned n_back = n; +- if (n) +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) ; +- n = n - 2; +- } while (n != 0) ; +- } +- +- crc = c; +- return c ^ 0xffffffffL; +-} +- +-ulg updcrc2(s, n) +- uch *s; +- unsigned n; +-{ +- register ulg c; +- +- static ulg crc = (ulg)0xffffffffL; +- if (s == NULL) { +- c = 0xffffffffL; +- } else { +- c = crc; +- unsigned n_back = n; +- if (n) +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) + 1; +- } while (--n) ; +- } +- +- crc = c; +- return c ^ 0xffffffffL; +-} +-/* +-ulg updcrc3(s, n) +- uch *s; +- int n; +-{ +- register ulg c; +- +- static ulg crc = (ulg)0xffffffffL; +- if (s == NULL) { +- c = 0xffffffffL; +- } else { +- c = crc; +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); +- --n; +- } while (n ) ; +- } +- +- crc = c; +- return c ^ 0xffffffffL; +-}*/ +-/* { dg-final { scan-tree-dump-times "num of phi noeds check failed." 1 "loop_crc"} } */ +-/* { dg-final { scan-tree-dump-times "evolution pattern check failed." 1 "loop_crc"} } */ +-/* { dg-final { scan-tree-dump-times "calculation pattern check failed." 1 "loop_crc"} } */ +- +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc +deleted file mode 100644 +index 6d52a8684..000000000 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.c.042t.loop_crc ++++ /dev/null +@@ -1,64 +0,0 @@ +- +-;; Function updcrc3 (updcrc3, funcdef_no=0, decl_uid=3687, cgraph_uid=1, symbol_order=1) +- +-;; 2 loops found +-;; +-;; Loop 0 +-;; header 0, latch 1 +-;; depth 0, outer -1 +-;; nodes: 0 1 2 3 4 5 +-;; +-;; Loop 1 +-;; header 4, latch 4 +-;; depth 1, outer 0 +-;; nodes: 4 +-;; 2 succs { 5 3 } +-;; 3 succs { 4 5 } +-;; 4 succs { 4 } +-;; 5 succs { 1 } +- +-Starting the loop_crc pass +-====================================== +-Processing loop 1: +-====================================== +-;; +-;; Loop 1 +-;; header 4, latch 4 +-;; depth 1, outer 0 +-;; nodes: 4 +- +- +- +-Wrong loop form for crc matching. +-updcrc3 (uch * s, unsigned int n) +-{ +- unsigned int n_back; +- static ulg crc = 4294967295; +- register ulg c; +- ulg _22; +- +- : +- if (s_12(D) == 0B) +- goto ; [INV] +- else +- goto ; [INV] +- +- : +- c_14 = crc; +- if (n_15(D) != 0) +- goto ; [INV] +- else +- goto ; [INV] +- +- : +- goto ; [100.00%] +- +- : +- # c_11 = PHI <4294967295(2), c_14(3)> +- crc = c_11; +- _22 = c_11 ^ 4294967295; +- return _22; +- +-} +- +- +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s +deleted file mode 100644 +index cae934bfe..000000000 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-calculation-check-fail.s ++++ /dev/null +@@ -1,329 +0,0 @@ +- .arch armv8-a +- .file "loop-crc-calculation-check-fail.c" +- .text +- .section .rodata +- .align 3 +- .type crc_32_tab, %object +- .size crc_32_tab, 2048 +-crc_32_tab: +- .xword 0 +- .xword 1996959894 +- .xword 3993919788 +- .xword 2567524794 +- .xword 124634137 +- .xword 1886057615 +- .xword 3915621685 +- .xword 2657392035 +- .xword 249268274 +- .xword 2044508324 +- .xword 3772115230 +- .xword 2547177864 +- .xword 162941995 +- .xword 2125561021 +- .xword 3887607047 +- .xword 2428444049 +- .xword 498536548 +- .xword 1789927666 +- .xword 4089016648 +- .xword 2227061214 +- .xword 450548861 +- .xword 1843258603 +- .xword 4107580753 +- .xword 2211677639 +- .xword 325883990 +- .xword 1684777152 +- .xword 4251122042 +- .xword 2321926636 +- .xword 335633487 +- .xword 1661365465 +- .xword 4195302755 +- .xword 2366115317 +- .xword 997073096 +- .xword 1281953886 +- .xword 3579855332 +- .xword 2724688242 +- .xword 1006888145 +- .xword 1258607687 +- .xword 3524101629 +- .xword 2768942443 +- .xword 901097722 +- .xword 1119000684 +- .xword 3686517206 +- .xword 2898065728 +- .xword 853044451 +- .xword 1172266101 +- .xword 3705015759 +- .xword 2882616665 +- .xword 651767980 +- .xword 1373503546 +- .xword 3369554304 +- .xword 3218104598 +- .xword 565507253 +- .xword 1454621731 +- .xword 3485111705 +- .xword 3099436303 +- .xword 671266974 +- .xword 1594198024 +- .xword 3322730930 +- .xword 2970347812 +- .xword 795835527 +- .xword 1483230225 +- .xword 3244367275 +- .xword 3060149565 +- .xword 1994146192 +- .xword 31158534 +- .xword 2563907772 +- .xword 4023717930 +- .xword 1907459465 +- .xword 112637215 +- .xword 2680153253 +- .xword 3904427059 +- .xword 2013776290 +- .xword 251722036 +- .xword 2517215374 +- .xword 3775830040 +- .xword 2137656763 +- .xword 141376813 +- .xword 2439277719 +- .xword 3865271297 +- .xword 1802195444 +- .xword 476864866 +- .xword 2238001368 +- .xword 4066508878 +- .xword 1812370925 +- .xword 453092731 +- .xword 2181625025 +- .xword 4111451223 +- .xword 1706088902 +- .xword 314042704 +- .xword 2344532202 +- .xword 4240017532 +- .xword 1658658271 +- .xword 366619977 +- .xword 2362670323 +- .xword 4224994405 +- .xword 1303535960 +- .xword 984961486 +- .xword 2747007092 +- .xword 3569037538 +- .xword 1256170817 +- .xword 1037604311 +- .xword 2765210733 +- .xword 3554079995 +- .xword 1131014506 +- .xword 879679996 +- .xword 2909243462 +- .xword 3663771856 +- .xword 1141124467 +- .xword 855842277 +- .xword 2852801631 +- .xword 3708648649 +- .xword 1342533948 +- .xword 654459306 +- .xword 3188396048 +- .xword 3373015174 +- .xword 1466479909 +- .xword 544179635 +- .xword 3110523913 +- .xword 3462522015 +- .xword 1591671054 +- .xword 702138776 +- .xword 2966460450 +- .xword 3352799412 +- .xword 1504918807 +- .xword 783551873 +- .xword 3082640443 +- .xword 3233442989 +- .xword 3988292384 +- .xword 2596254646 +- .xword 62317068 +- .xword 1957810842 +- .xword 3939845945 +- .xword 2647816111 +- .xword 81470997 +- .xword 1943803523 +- .xword 3814918930 +- .xword 2489596804 +- .xword 225274430 +- .xword 2053790376 +- .xword 3826175755 +- .xword 2466906013 +- .xword 167816743 +- .xword 2097651377 +- .xword 4027552580 +- .xword 2265490386 +- .xword 503444072 +- .xword 1762050814 +- .xword 4150417245 +- .xword 2154129355 +- .xword 426522225 +- .xword 1852507879 +- .xword 4275313526 +- .xword 2312317920 +- .xword 282753626 +- .xword 1742555852 +- .xword 4189708143 +- .xword 2394877945 +- .xword 397917763 +- .xword 1622183637 +- .xword 3604390888 +- .xword 2714866558 +- .xword 953729732 +- .xword 1340076626 +- .xword 3518719985 +- .xword 2797360999 +- .xword 1068828381 +- .xword 1219638859 +- .xword 3624741850 +- .xword 2936675148 +- .xword 906185462 +- .xword 1090812512 +- .xword 3747672003 +- .xword 2825379669 +- .xword 829329135 +- .xword 1181335161 +- .xword 3412177804 +- .xword 3160834842 +- .xword 628085408 +- .xword 1382605366 +- .xword 3423369109 +- .xword 3138078467 +- .xword 570562233 +- .xword 1426400815 +- .xword 3317316542 +- .xword 2998733608 +- .xword 733239954 +- .xword 1555261956 +- .xword 3268935591 +- .xword 3050360625 +- .xword 752459403 +- .xword 1541320221 +- .xword 2607071920 +- .xword 3965973030 +- .xword 1969922972 +- .xword 40735498 +- .xword 2617837225 +- .xword 3943577151 +- .xword 1913087877 +- .xword 83908371 +- .xword 2512341634 +- .xword 3803740692 +- .xword 2075208622 +- .xword 213261112 +- .xword 2463272603 +- .xword 3855990285 +- .xword 2094854071 +- .xword 198958881 +- .xword 2262029012 +- .xword 4057260610 +- .xword 1759359992 +- .xword 534414190 +- .xword 2176718541 +- .xword 4139329115 +- .xword 1873836001 +- .xword 414664567 +- .xword 2282248934 +- .xword 4279200368 +- .xword 1711684554 +- .xword 285281116 +- .xword 2405801727 +- .xword 4167216745 +- .xword 1634467795 +- .xword 376229701 +- .xword 2685067896 +- .xword 3608007406 +- .xword 1308918612 +- .xword 956543938 +- .xword 2808555105 +- .xword 3495958263 +- .xword 1231636301 +- .xword 1047427035 +- .xword 2932959818 +- .xword 3654703836 +- .xword 1088359270 +- .xword 936918000 +- .xword 2847714899 +- .xword 3736837829 +- .xword 1202900863 +- .xword 817233897 +- .xword 3183342108 +- .xword 3401237130 +- .xword 1404277552 +- .xword 615818150 +- .xword 3134207493 +- .xword 3453421203 +- .xword 1423857449 +- .xword 601450431 +- .xword 3009837614 +- .xword 3294710456 +- .xword 1567103746 +- .xword 711928724 +- .xword 3020668471 +- .xword 3272380065 +- .xword 1510334235 +- .xword 755167117 +- .text +- .align 2 +- .global updcrc3 +- .type updcrc3, %function +-updcrc3: +-.LFB0: +- .cfi_startproc +- str x19, [sp, -48]! +- .cfi_def_cfa_offset 48 +- .cfi_offset 19, -48 +- str x0, [sp, 24] +- str w1, [sp, 20] +- ldr x0, [sp, 24] +- cmp x0, 0 +- bne .L2 +- mov x19, 4294967295 +- b .L3 +-.L2: +- adrp x0, crc.0 +- add x0, x0, :lo12:crc.0 +- ldr x19, [x0] +- ldr w0, [sp, 20] +- str w0, [sp, 44] +- ldr w0, [sp, 20] +- cmp w0, 0 +- beq .L3 +-.L4: +- ldr x0, [sp, 24] +- add x1, x0, 1 +- str x1, [sp, 24] +- ldrb w0, [x0] +- and x0, x0, 255 +- eor x0, x19, x0 +- and x1, x0, 255 +- adrp x0, crc_32_tab +- add x0, x0, :lo12:crc_32_tab +- ldr x1, [x0, x1, lsl 3] +- lsr x0, x19, 8 +- eor x19, x1, x0 +- ldr w0, [sp, 20] +- sub w0, w0, #1 +- str w0, [sp, 20] +- ldr w0, [sp, 20] +- cmp w0, 999 +- bls .L4 +-.L3: +- adrp x0, crc.0 +- add x0, x0, :lo12:crc.0 +- str x19, [x0] +- eor x0, x19, 4294967295 +- ldr x19, [sp], 48 +- .cfi_restore 19 +- .cfi_def_cfa_offset 0 +- ret +- .cfi_endproc +-.LFE0: +- .size updcrc3, .-updcrc3 +- .data +- .align 3 +- .type crc.0, %object +- .size crc.0, 8 +-crc.0: +- .xword 4294967295 +- .ident "GCC: (Kunpeng gcc 10.3.1-2.3.0.b006) 10.3.1" +- .section .note.GNU-stack,"",@progbits +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-condition-fail.c +similarity index 97% +rename from gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c +rename to gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-condition-fail.c +index 70eb1b814..fefa949f9 100644 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-3.c ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-condition-fail.c +@@ -74,12 +74,12 @@ ulg updcrc(s, n) + } else { + c = crc; + if (n) +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ if (n) do { ++ c = crc_32_tab[((int)c ^ (*s++)) & 0xff] ^ (c >> 8); + } while (--n || c != 0) ; + } + crc = c; + exit1: + return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ + } +-/* { dg-final { scan-tree-dump-times "Wrong loop form for crc matching." 1 "loop_crc"} } */ +\ No newline at end of file ++/* { dg-final { scan-tree-dump-times "Wrong loop form for crc matching." 1 "loop_crc"} } */ +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail-2.c +similarity index 95% +rename from gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c +rename to gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail-2.c +index 1d7e0a319..b37446ec5 100644 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-4.c ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail-2.c +@@ -75,8 +75,8 @@ ulg updcrc(s, n) + } else { + c = crc; + if (n) +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8) * test[c%5]; ++ if (n) do { ++ c = crc_32_tab[((int)c ^ (*s++)) & 0xff] ^ (c >> 8); + } while (--n) ; + } + do { +@@ -86,4 +86,5 @@ ulg updcrc(s, n) + crc = c; + return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ + } +-/* { dg-final { scan-tree-dump-times "Table check fail. not only single array is read." 2 "loop_crc"} } */ +\ No newline at end of file ++/* { dg-final { scan-tree-dump-times "Table check fail. not only single array is read." 2 "loop_crc"} } */ ++/* { dg-final { scan-tree-dump-times "Wrong crc table for crc matching." 1 "loop_crc"} } */ +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c +index b59704e31..3dc500a46 100644 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-loop-form-fail.c +@@ -108,4 +108,5 @@ ulg updcrc1(s, n) + crc = c; + return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ + } +-/* { dg-final { scan-tree-dump-times "Wrong loop form for crc matching." 2 "loop_crc"} } */ +\ No newline at end of file ++/* { dg-final { scan-tree-dump-times "Table check fail. not only single array is read." 1 "loop_crc"} } */ ++/* { dg-final { scan-tree-dump-times "Wrong crc table for crc matching." 2 "loop_crc"} } */ +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c +index e1e16eaf2..8b556efc8 100644 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-sucess.c +@@ -73,12 +73,11 @@ ulg updcrc(s, n) + c = 0xffffffffL; + } else { + c = crc; +- if (n) +- do { +- c = crc_32_tab[(c ^ (*s++)) & 0xff] ^ (c >> 8); ++ if (n) do { ++ c = crc_32_tab[((int)c ^ (*s++)) & 0xff] ^ (c >> 8); + } while (--n); + } + crc = c; + return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ + } +-/* { dg-final { scan-tree-dump-times "the loop can be optimized" 1 "loop_crc"} } */ +\ No newline at end of file ++/* { dg-final { scan-tree-dump-times "The 1th loop form is successmatched,and the loop can be optimized." 1 "loop_crc"} } */ +diff --git a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c +index f03a4fa82..de21f4553 100644 +--- a/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c ++++ b/gcc/testsuite/gcc.dg/tree-ssa/loop-crc-table-check-fail.c +@@ -110,4 +110,5 @@ ulg updcrc1(s, n) + return c ^ 0xffffffffL; /* (instead of ~c for 64-bit machines) */ + } + /* { dg-final { scan-tree-dump-times "Table check fail. not only single array is read." 2 "loop_crc"} } */ +-/* { dg-final { scan-tree-dump-times "Table check fail. Table not matching." 1 "loop_crc"} } */ +\ No newline at end of file ++/* { dg-final { scan-tree-dump-times "Wrong crc table for crc matching." 3 "loop_crc"} } */ ++/* { dg-final { scan-tree-dump-times "Table check fail. Table not matching." 1 "loop_crc"} } */ +diff --git a/gcc/tree-ssa-loop-crc.c b/gcc/tree-ssa-loop-crc.c +index 4982384c6..8225c2fa5 100644 +--- a/gcc/tree-ssa-loop-crc.c ++++ b/gcc/tree-ssa-loop-crc.c +@@ -1,5 +1,5 @@ +-/* Array widen compare. +- Copyright (C) 2022-2022 Free Software Foundation, Inc. ++/* loop crc. ++ Copyright (C) 2023-2023 Free Software Foundation, Inc. + + This file is part of GCC. + +@@ -42,13 +42,235 @@ along with GCC; see the file COPYING3. If not see + #include "print-tree.h" + #include "cfghooks.h" + #include "gimple-fold.h" ++#include "diagnostic-core.h" ++ ++/* This pass handles scenarios similar to the following: ++ulg updcrc(s, n) ++ uch *s; ++ unsigned n; ++{ ++ register ulg c; ++ ++ static ulg crc = (ulg)0xffffffffL; ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) do { ++ c = crc_32_tab[((int)c ^ (*s++)) & 0xff] ^ (c >> 8); ++ } while (--n); ++ } ++ crc = c; ++ return c ^ 0xffffffffL; ++} ++ ++If the hardware supports the crc instruction, then the pass completes the ++conversion of the above scenario into: ++ ++#define SIZE_U32 sizeof(uint32_t) ++unsigned long updcrc(s, n) ++ unsigned char *s; ++ unsigned n; ++{ ++ register unsigned long c; ++ ++ static unsigned long crc = (unsigned long)0xffffffffL; ++ ++ if (s == NULL) { ++ c = 0xffffffffL; ++ } else { ++ c = crc; ++ if (n) ++ { ++ uint32_t nn = n/SIZE_U32; ++ do{ ++ c = __crc32w(c,*((uint32_t *)s)); ++ s += SIZE_U32; ++ }while(--nn); ++ if (n & sizeof(uint16_t)) { ++ c = __crc32h(c, *((uint16_t *)s)); ++ s += sizeof(uint16_t); ++ } ++ if (n & sizeof(uint8_t)) ++ c = __crc32b(c, *s); ++ } ++ } ++ crc = c; ++ return c ^ 0xffffffffL; ++} ++ ++This pass is to complete the conversion of such scenarios from the internal ++perspective of the compiler: ++1)match_crc_loop:The function completes the screening of such scenarios; ++2)convert_to_new_loop:The function completes the conversion of ++ origin_loop to new loops, and removes origin_loop; ++3)origin_loop_info: The structure is used to record important information ++ of origin_loop: such as loop exit, initial value of induction ++ variable, etc; ++4) create_new_loops: The function is used as the key content of the pass ++ to complete the creation of new loops. */ + +-/* Match.pd function to match the ctz expression. */ + extern bool gimple_crc_match_index (tree, tree *, tree (*)(tree)); + extern bool gimple_crc_match_res (tree, tree *, tree (*)(tree)); + + static gimple *crc_table_read_stmt = NULL; + ++static gphi* phi_s = NULL; ++static gphi* phi_c = NULL; ++static tree nn_tree = NULL; ++ ++enum aarch64_crc_builtins ++{ ++ AARCH64_BUILTIN_CRC32B, ++ AARCH64_BUILTIN_CRC32H, ++ AARCH64_BUILTIN_CRC32W, ++}; ++ ++/* The useful information of origin loop. */ ++struct origin_loop_info ++{ ++ tree limit; /* The limit index of the array in the old loop. */ ++ tree base_n; /* The initial value of the old loop. */ ++ tree base_s; /* The initial value of the old loop. */ ++ tree base_c; /* The initial value of the old loop. */ ++ edge entry_edge; /* The edge into the old loop. */ ++ edge exit_edge; /* The edge outto the old loop. */ ++ basic_block exit_bb; ++}; ++ ++typedef struct origin_loop_info origin_loop_info; ++ ++static origin_loop_info origin_loop; ++hash_map n_map; ++hash_map nn_map; ++hash_map s_map; ++hash_map c_map; ++hash_map crc_map; ++ ++/* Initialize the origin_loop structure. */ ++static void ++init_origin_loop_structure () ++{ ++ origin_loop.entry_edge = NULL; ++ origin_loop.exit_edge = NULL; ++ origin_loop.exit_bb = NULL; ++ origin_loop.limit = NULL; ++ origin_loop.base_n = NULL; ++ origin_loop.base_s = NULL; ++ origin_loop.base_c = NULL; ++} ++ ++/* Get the edge that first entered the loop. */ ++static edge ++get_loop_preheader_edge (class loop *loop) ++{ ++ edge e; ++ edge_iterator ei; ++ ++ FOR_EACH_EDGE (e, ei, loop->header->preds) ++ if (e->src != loop->latch) ++ break; ++ ++ return e; ++} ++ ++/* Returns true if t is SSA_NAME and user variable exists. */ ++ ++static bool ++ssa_name_var_p (tree t) ++{ ++ if (!t || TREE_CODE (t) != SSA_NAME) ++ return false; ++ if (SSA_NAME_VAR (t)) ++ return true; ++ return false; ++} ++ ++/* Returns true if t1 and t2 are SSA_NAME and belong to the same variable. */ ++ ++static bool ++same_ssa_name_var_p (tree t1, tree t2) ++{ ++ if (!ssa_name_var_p (t1) || !ssa_name_var_p (t2)) ++ return false; ++ if (SSA_NAME_VAR (t1) == SSA_NAME_VAR (t2)) ++ return true; ++ return false; ++} ++ ++/* Get origin loop induction variable upper bound. */ ++ ++static bool ++get_iv_upper_bound (gimple *stmt) ++{ ++ if (origin_loop.limit != NULL || origin_loop.base_n != NULL) ++ return false; ++ ++ tree lhs = gimple_cond_lhs (stmt); ++ tree rhs = gimple_cond_rhs (stmt); ++ ++ if (TREE_CODE (TREE_TYPE (lhs)) != INTEGER_TYPE ++ || TREE_CODE (TREE_TYPE (rhs)) != INTEGER_TYPE) ++ return false; ++ ++ /* TODO: Currently, the input restrictions on lhs and rhs are implemented ++ through PARM_DECL. We may consider relax the restrictions later, and ++ we need to consider the overall adaptation scenario and adding test ++ cases. */ ++ if (ssa_name_var_p (lhs) && TREE_CODE (SSA_NAME_VAR (lhs)) == PARM_DECL) ++ { ++ origin_loop.limit = rhs; ++ origin_loop.base_n = lhs; ++ } ++ else ++ return false; ++ ++ if (origin_loop.limit != NULL && origin_loop.base_n != NULL) ++ return true; ++ ++ return false; ++} ++ ++/* Get origin loop info. */ ++static bool ++get_origin_loop_info(class loop *loop) ++{ ++ vec edges; ++ edges = get_loop_exit_edges (loop); ++ origin_loop.exit_edge = edges[0]; ++ origin_loop.exit_bb = origin_loop.exit_edge->dest; ++ origin_loop.entry_edge = get_loop_preheader_edge(loop); ++ origin_loop.base_s = PHI_ARG_DEF_FROM_EDGE(phi_s,origin_loop.entry_edge); ++ origin_loop.base_c = PHI_ARG_DEF_FROM_EDGE(phi_c,origin_loop.entry_edge); ++ ++ basic_block preheader_bb; ++ preheader_bb = origin_loop.entry_edge->src; ++ ++ if(preheader_bb->preds->length() != 1) ++ return false; ++ ++ edge entry_pre_bb_edge; ++ entry_pre_bb_edge = EDGE_PRED (preheader_bb, 0); ++ ++ basic_block pre_preheader_bb; ++ pre_preheader_bb = entry_pre_bb_edge->src; ++ ++ gimple_stmt_iterator gsi; ++ gimple *stmt; ++ bool get_upper_bound = false; ++ for (gsi = gsi_start_bb (pre_preheader_bb); !gsi_end_p (gsi); gsi_next (&gsi)) ++ { ++ stmt = gsi_stmt (gsi); ++ if (stmt && gimple_code (stmt) == GIMPLE_COND ++ && get_iv_upper_bound (stmt)) { ++ get_upper_bound = true; ++ break; ++ } ++ } ++ ++ return get_upper_bound; ++} + + /* The loop form check will check the entire loop control flow + It should be a loop that: +@@ -102,7 +324,8 @@ only_one_array_read (class loop *loop, tree &crc_table) + if (gimple_code (stmt) == GIMPLE_ASSIGN && + TREE_CODE(gimple_assign_rhs1 (stmt)) == ARRAY_REF) + { +- if (crc_table == NULL) ++ if (crc_table == NULL && ++ gimple_assign_rhs1 (stmt)->base.readonly_flag) + { + crc_table = gimple_assign_rhs1 (stmt); + crc_table_read_stmt = stmt; +@@ -174,15 +397,18 @@ static const unsigned HOST_WIDE_INT crc_32_tab[] = { + static bool + match_crc_table (tree crc_table) + { ++ const unsigned LOW_BOUND = 0; ++ const unsigned UP_BOUND = 255; ++ const unsigned ELEMENT_SIZE = 8; + unsigned HOST_WIDE_INT lb = tree_to_uhwi (array_ref_low_bound (crc_table)); + unsigned HOST_WIDE_INT ub = tree_to_uhwi (array_ref_up_bound (crc_table)); + unsigned HOST_WIDE_INT es = tree_to_uhwi (array_ref_element_size (crc_table)); +- if (lb != 0 || ub != 255 || es != 8) ++ if (lb != LOW_BOUND || ub != UP_BOUND || es != ELEMENT_SIZE) + return false; + + tree decl = TREE_OPERAND (crc_table, 0); + tree ctor = ctor_for_folding(decl); +- for (int i = 0; i < 255; i++) { ++ for (int i = lb; i <= ub; i++) { + unsigned HOST_WIDE_INT val = tree_to_uhwi (CONSTRUCTOR_ELT (ctor,i)->value); + if (crc_32_tab[i] != val) + return false; +@@ -273,6 +499,7 @@ check_evolution_pattern (class loop* loop, gphi *capture[]) + if (s != NULL) + return false; + s = capture[i]; ++ phi_s = s; + } + else if (evolution_pattern_plus_with_p(loop, capture[i], 4294967295)) + { +@@ -285,6 +512,7 @@ check_evolution_pattern (class loop* loop, gphi *capture[]) + if (c != NULL) + return false; + c = capture[i]; ++ phi_c = c; + } + } + +@@ -314,14 +542,19 @@ check_calculation_pattern (class loop* loop, gphi *capture[]) + _5 = _4 ^ c_10; //BIT_XOR_EXPR (SSA_NAME, PHI @1) + _6 = _5 & 255; //BIT_XOR_EXPR (SSA_NAME, INTEGER_CST@3) + */ +- + if (!gimple_crc_match_index(index, res_ops, NULL)) + return false; +- gimple *s_res_stmt = SSA_NAME_DEF_STMT(res_ops[1]); +- tree s_res = TREE_OPERAND(gimple_assign_rhs1(s_res_stmt),0); +- if (res_ops[0] != gimple_phi_result (c) || +- s_res != gimple_phi_result (s)) ++ gimple *s_res_stmt = SSA_NAME_DEF_STMT (res_ops[0]); ++ if (!s_res_stmt) ++ return false; ++ gimple *s_def_stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (s_res_stmt)); ++ if (!s_def_stmt) + return false; ++ tree s_res = TREE_OPERAND (gimple_assign_rhs1 (s_def_stmt), 0); ++ if (res_ops[1] != gimple_phi_result (c) || s_res != gimple_phi_result (s)) ++ { ++ return false; ++ } + + /* Try to match + _8 = c_12 >> 8; // RSHIFT_EXPR (SSA_NAME @1, INTEGER_CST @2) +@@ -333,7 +566,11 @@ check_calculation_pattern (class loop* loop, gphi *capture[]) + return false; + if (res_ops[0] != gimple_phi_result (c) + || res_ops[2] != gimple_assign_lhs(crc_table_read_stmt)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\n gimple_crc_match_res pattern check failed.\n"); + return false; ++ } + + return true; + } +@@ -419,101 +656,91 @@ crc_loop_body_check (class loop *loop) + return false; + } + return true; +-/* gphi *phi; +- gphi_iterator gsi; +- int num_of_phi = 0; +- //s, n, c; +- //only 3 phi nodes are there, every one of the phi nodes comming from 2 edge only, one from preheader, one from latch +- // s increase by 1 every itoration +- // n decrease by 1 every itoration +- // The final one is c, which is the result, should be used for the start of the later pattern matching +- for (gsi = gsi_start_phis(loop->header); !gsi_end_p(gsi); gsi_next(&gsi)) +- { +- phi = gsi.phi(); ++} + +- if (phi) num_of_phi++; +- if (num_of_phi > 3) return false; // more then 3 phi node +- if (gimple_phi_num_args(phi) > 2) // more than 2 edges other then one backedge and one preheader edge +- return false; +- //capture[num_of_phi - 1] = gimple_phi_result(phi); +- capture[num_of_phi - 1] = phi; +- } +- if (num_of_phi != 3) return false; // phi node should be 3 */ +- // Find the envolution pattern for s and n, try to match the identity of these variable +-/* gphi *s=NULL; +- gphi *n=NULL; +- gphi *c=NULL; ++/* Check the prev_bb of prev_bb of loop header. The prev_bb we are trying to match is + +- for (int i = 0; i < 3; i++) +- { +- if (evolution_pattern_plus_with_p(loop, capture[i], 1)) +- { +- if(s != NULL) +- return false; +- s = capture[i]; +- } +- else if (evolution_pattern_plus_with_p(loop, capture[i], 4294967295)) +- { +- if(n != NULL) +- return false; +- n = capture[i]; +- } +- else +- { +- if(c != NULL) +- return false; +- c = capture[i]; +- } +- } ++c_15 = crc; ++if (n_16(D) != 0) ++ goto ; [INV] ++else ++ goto ; [INV] + +- // some envolution pattern cannot find +- if (!n || !s || !c) +- return false; +- gphi *s=capture[0]; +- gphi *n=capture[1]; +- gphi *c=capture[2]; +- tree res_ops[3]; +- tree index = TREE_OPERAND (gimple_assign_rhs1 (crc_table_read_stmt), 1); ++ In this case , we must be sure that the n is not zero. ++ so the match condition is ++ 1、the n is not zero. + +- /* Try to match +- _1 = (int) c_12; //NOP_EXPR (SSA_NAME @1) +- _4 = (int) _3; //NOP_EXPR (SSA_NAME @2) +- _5 = _1 ^ _4; //BIT_XOR_EXPR (SSA_NAME, SSA_NAME) +- _6 = _5 & 255; //BIT_XOR_EXPR (SSA_NAME, INTEGER_CST@3) ++ : ++if (s_13(D) == 0B) ++ goto ; [INV] ++else ++ goto ; [INV] + +- +- if (!gimple_crc_match_index(index, res_ops, NULL)) ++ In this case, we must be sure the s is not NULL. ++ so the match condition is ++ 1、the s is not NULL. ++*/ ++static bool ++crc_prev_bb_of_loop_header_check(class loop *loop) ++{ ++ basic_block header = loop->header; ++ basic_block prev_header_bb = header->prev_bb; ++ if(NULL == prev_header_bb) ++ { + return false; +- gimple *s_res_stmt = SSA_NAME_DEF_STMT(res_ops[1]); +- tree s_res = TREE_OPERAND(gimple_assign_rhs1(s_res_stmt),0); +- if (res_ops[0] != gimple_phi_result (c) || +- s_res != gimple_phi_result (s)) ++ } ++ ++ basic_block prev_prev_header_bb = prev_header_bb->prev_bb; ++ if(NULL == prev_prev_header_bb) ++ { + return false; ++ } ++ ++ gimple_stmt_iterator gsi; ++ gimple *stmt; ++ bool res = false; ++ for (gsi = gsi_start_bb (prev_prev_header_bb); !gsi_end_p (gsi); gsi_next (&gsi)) ++ { ++ stmt = gsi_stmt (gsi); ++ if (stmt == NULL) ++ return false; + +- /* +-_8 = c_12 >> 8; // RSHIFT_EXPR (SSA_NAME @1, INTEGER_CST @2) +-c_19 = _7 ^ _8; // BIT_XOR_EXPR (SSA_NAME@3, SSA_NAME) ++ if (gimple_code (stmt) == GIMPLE_COND && ++ gimple_cond_code(stmt) == NE_EXPR && ++ TREE_CODE(gimple_cond_rhs (stmt)) == INTEGER_CST && ++ tree_int_cst_sgn(gimple_cond_rhs (stmt)) == 0 ) ++ { ++ res = true; ++ break; ++ } ++ } + +- edge backedge = find_edge(loop->latch, loop->header); +- tree updated_c = PHI_ARG_DEF_FROM_EDGE (c, backedge); +- if (!gimple_crc_match_res(updated_c, res_ops, NULL)) +- return false; +- if (res_ops[0] != gimple_phi_result (c) +- || res_ops[2] != gimple_assign_lhs(crc_table_read_stmt)) ++ if(!res) ++ { + return false; ++ } + +- // try match n as the induction variable +- // The proceed condition for back edge is n != 0 +- gimple *cond_stmt = gsi_stmt (gsi_last_bb (loop->header)); +- if (!cond_stmt || gimple_code (cond_stmt) != GIMPLE_COND || gimple_cond_code (cond_stmt) != NE_EXPR +- || gimple_cond_lhs (cond_stmt) != PHI_ARG_DEF_FROM_EDGE (n, backedge) +- || tree_to_uhwi(gimple_cond_rhs (cond_stmt)) != 0) ++ basic_block first_bb = prev_prev_header_bb->prev_bb; ++ if(NULL == first_bb) + return false; +- +- return true; +- */ +-} + ++ for (gsi = gsi_start_bb (first_bb); !gsi_end_p (gsi); gsi_next (&gsi)) ++ { ++ stmt = gsi_stmt (gsi); ++ if (stmt == NULL) ++ return false; ++ ++ if (gimple_code (stmt) == GIMPLE_COND && ++ gimple_cond_code(stmt) == EQ_EXPR && ++ TREE_CODE(gimple_cond_rhs (stmt)) == INTEGER_CST && ++ tree_int_cst_sgn(gimple_cond_rhs (stmt)) == 0 ) ++ { ++ return true; ++ } ++ } ++ ++ return false; ++} + + static bool + match_crc_loop (class loop *loop) +@@ -536,13 +763,463 @@ match_crc_loop (class loop *loop) + fprintf (dump_file, "\nWrong loop body for crc matching.\n"); + return false; + } ++ if(!crc_prev_bb_of_loop_header_check(loop)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nWrong prev basic_blocks of loop header for crc matching.\n"); ++ return false; ++ } ++ ++ init_origin_loop_structure(); ++ if(!get_origin_loop_info(loop)) ++ return false; ++ + return true; + } + ++static void ++create_new_bb (basic_block &new_bb, basic_block after_bb, ++ basic_block dominator_bb, class loop *outer) ++{ ++ new_bb = create_empty_bb (after_bb); ++ add_bb_to_loop (new_bb, outer); ++ set_immediate_dominator (CDI_DOMINATORS, new_bb, dominator_bb); ++} ++ ++static void ++change_preheader_bb(edge entry_edge) ++{ ++ gimple_seq stmts = NULL; ++ gimple_stmt_iterator gsi; ++ gimple* g; ++ tree lhs1; ++ ++ lhs1 = create_tmp_var(TREE_TYPE(origin_loop.base_n),"nn"); ++ lhs1 = make_ssa_name(lhs1); ++ gsi = gsi_last_bb (entry_edge->src); ++ g = gimple_build_assign(lhs1,RSHIFT_EXPR,origin_loop.base_n, ++ build_int_cst (TREE_TYPE (origin_loop.base_n), 2)); ++ gimple_seq_add_stmt(&stmts,g); ++ gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); ++ nn_tree = lhs1; ++ set_current_def(nn_tree, lhs1); ++ nn_map.put (entry_edge->src, lhs1); ++} ++ ++static gphi* ++create_phi_node_for_bb(tree old_name, basic_block bb) ++{ ++ gphi *phi = create_phi_node(NULL_TREE, bb); ++ create_new_def_for(old_name, phi, gimple_phi_result_ptr(phi)); ++ return phi; ++} ++ ++static gimple* ++call_builtin_fun(int code,tree& lhs, tree arg1, tree arg2) ++{ ++ unsigned int builtin_code = targetm.get_crc_builtin_code(code, true);// 根据code获取到正确的builtin_fun_code ++ tree fn = targetm.builtin_decl(builtin_code,true); // get the decl of __builtin_aarch64_crc32w ++ if (!fn || fn == error_mark_node) ++ fatal_error (input_location, ++ "target specific builtin not available"); ++ gimple* call_builtin = gimple_build_call(fn, 2, arg1, arg2); // _40 = __builtin_aarch64_crc32* (_1, _2); ++ lhs = make_ssa_name (unsigned_type_node); ++ gimple_call_set_lhs(call_builtin,lhs); ++ ++ return call_builtin; ++} ++ ++/* Create loop_header and loop_latch for new loop ++ : ++ # s_14 = PHI ++ # c_16 = PHI ++ # nn_19 = PHI ++ _1 = (unsigned int) c_16; ++ _2 = MEM[(uint32_t *)s_14]; ++ _40 = __builtin_aarch64_crc32w (_1, _2); ++ c_29 = (long unsigned int) _40; ++ s_30 = s_14 + 4; ++ nn_31 = nn_19 + 4294967295; ++ if (nn_31 != 0) ++ The IR of bb is as above. */ ++static void ++create_loop_bb(basic_block& loop_bb, basic_block after_bb, ++ basic_block dominator_bb, class loop *outer, edge entry_edge) ++{ ++ gimple_seq stmts = NULL; ++ gimple_stmt_iterator gsi; ++ gimple* g; ++ gphi* phi_s_loop; ++ gphi* phi_c_loop; ++ gphi* phi_nn_loop; ++ ++ create_new_bb(loop_bb, after_bb, dominator_bb, outer); ++ redirect_edge_and_branch(entry_edge, loop_bb); ++ gsi = gsi_last_bb(loop_bb); ++ tree entry_nn = get_current_def(nn_tree); ++ phi_s_loop = create_phi_node_for_bb(origin_loop.base_s, loop_bb); ++ phi_c_loop = create_phi_node_for_bb(origin_loop.base_c, loop_bb); ++ phi_nn_loop = create_phi_node_for_bb(entry_nn, loop_bb); ++ ++ tree res_s = gimple_phi_result(phi_s_loop); ++ tree res_nn = gimple_phi_result(phi_nn_loop); ++ tree lhs1 = gimple_build(&stmts, NOP_EXPR, unsigned_type_node, ++ gimple_phi_result(phi_c_loop)); ++ g = gimple_build_assign(make_ssa_name(unsigned_type_node), ++ fold_build2(MEM_REF,unsigned_type_node,res_s, ++ build_int_cst (build_pointer_type (unsigned_type_node), 0))); ++ gimple_seq_add_stmt(&stmts, g); ++ tree lhs2 = gimple_assign_lhs(g); // _2 = MEM[(uint32_t *)s_14]; ++ unsigned int code = AARCH64_BUILTIN_CRC32W; ++ tree lhs3; ++ gimple* build_crc32w = call_builtin_fun(code,lhs3, lhs1, lhs2); ++ crc_map.put(loop_bb, lhs3); ++ gimple_seq_add_stmt(&stmts,build_crc32w); ++ ++ tree lhs4 = copy_ssa_name(origin_loop.base_c); ++ g = gimple_build_assign(lhs4, NOP_EXPR, lhs3); ++ gimple_seq_add_stmt(&stmts, g); ++ c_map.put(loop_bb, lhs4); ++ ++ tree lhs5 = copy_ssa_name(origin_loop.base_s); ++ g = gimple_build_assign(lhs5, POINTER_PLUS_EXPR, res_s, ++ build_int_cst (sizetype, 4)); ++ gimple_seq_add_stmt(&stmts, g); ++ s_map.put(loop_bb, lhs5); ++ ++ tree lhs6 = copy_ssa_name(nn_tree); ++ g = gimple_build_assign(lhs6, PLUS_EXPR, res_nn, ++ build_int_cst (TREE_TYPE (res_nn), 4294967295)); ++ gimple_seq_add_stmt(&stmts,g); ++ nn_map.put(loop_bb, lhs6); ++ ++ gcond* cond_stmt = gimple_build_cond (NE_EXPR, lhs6, origin_loop.limit, ++ NULL_TREE, NULL_TREE); ++ gimple_seq_add_stmt (&stmts, cond_stmt); ++ gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); ++} ++ ++/* : ++ # c_6 = PHI ++ # s_46 = PHI ++ _44 = n_26(D) & 2; ++ if (_44 != 0) ++ The IR of bb is as above. */ ++static void ++create_cond_bb(basic_block& cond_bb, basic_block after_bb, ++ basic_block dominator_bb, class loop *outer){ ++ gimple_seq stmts = NULL; ++ gimple_stmt_iterator gsi; ++ gphi* phi_s_loop; ++ gphi* phi_c_loop; ++ ++ create_new_bb(cond_bb, after_bb, dominator_bb, outer); ++ gsi = gsi_last_bb(cond_bb); ++ tree entry_nn = get_current_def(nn_tree); ++ phi_s_loop = create_phi_node_for_bb(origin_loop.base_s, cond_bb); ++ phi_c_loop = create_phi_node_for_bb(origin_loop.base_c, cond_bb); ++ tree res_s = gimple_phi_result(phi_s_loop); ++ set_current_def(origin_loop.base_s, res_s); ++ s_map.put(cond_bb, res_s); ++ tree res_c = gimple_phi_result(phi_c_loop); ++ set_current_def(origin_loop.base_c, res_c); ++ c_map.put(cond_bb, res_c); ++ ++ tree lhs1 = gimple_build(&stmts, BIT_AND_EXPR, TREE_TYPE(origin_loop.base_n), ++ origin_loop.base_n, build_int_cst (TREE_TYPE (origin_loop.base_n), 2)); ++ gcond* cond_stmt = gimple_build_cond (NE_EXPR, lhs1, origin_loop.limit, ++ NULL_TREE, NULL_TREE); ++ gimple_seq_add_stmt (&stmts, cond_stmt); ++ gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); ++} ++ ++/* : ++ _7 = MEM[(uint16_t *)s_46]; ++ _41 = __builtin_aarch64_crc32h (_8, _7); ++ c_33 = (long unsigned int) _41; ++ s_34 = s_30 + 2; ++ The IR of bb is as above.*/ ++static void ++create_cond_true_bb(basic_block& cond_true_bb, basic_block after_bb, ++ basic_block dominator_bb, class loop *outer){ ++ gimple_seq stmts = NULL; ++ gimple* g; ++ gimple_stmt_iterator gsi; ++ ++ create_new_bb(cond_true_bb, after_bb, dominator_bb, outer); ++ gsi = gsi_last_bb(cond_true_bb); ++ tree s_46 = *(s_map.get(after_bb)); ++ g = gimple_build_assign(make_ssa_name(short_unsigned_type_node), ++ fold_build2(MEM_REF,short_unsigned_type_node,s_46, ++ build_int_cst (build_pointer_type (short_unsigned_type_node), 0))); ++ gimple_seq_add_stmt(&stmts,g); ++ tree lhs1 = gimple_assign_lhs(g); // _7 = MEM[(uint16_t *)s_46]; ++ unsigned int code = AARCH64_BUILTIN_CRC32H; ++ tree lhs2; ++ gimple* call_builtin = call_builtin_fun(code, lhs2,*(crc_map.get(cond_true_bb->prev_bb->prev_bb)),lhs1); ++ crc_map.put(cond_true_bb,lhs2); ++ gimple_seq_add_stmt(&stmts, call_builtin); ++ ++ tree lhs3 = copy_ssa_name(origin_loop.base_c); ++ g = gimple_build_assign(lhs3, NOP_EXPR, lhs2); ++ gimple_seq_add_stmt(&stmts, g); ++ c_map.put(cond_true_bb, lhs3); ++ ++ tree lhs5 = copy_ssa_name(s_46); ++ g = gimple_build_assign(lhs5, POINTER_PLUS_EXPR, s_46, ++ build_int_cst (sizetype, 2)); // s_30 + 2; ++ gimple_seq_add_stmt(&stmts, g); ++ s_map.put(cond_true_bb, lhs5); ++ ++ gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); ++ s_map.put(cond_true_bb, lhs5); ++} ++ ++/* : ++ # s_15 = PHI ++ # c_17 = PHI ++ _3 = n_26(D) & 1; ++ if (_3 != 0) ++ The IR of bb is as above.*/ ++static void ++create_cond_false_bb(basic_block& cond_false_bb, basic_block after_bb, ++ basic_block dominator_bb, class loop *outer) ++{ ++ gimple_seq stmts = NULL; ++ gimple_stmt_iterator gsi; ++ gphi* phi_s_cond_true_bb; ++ gphi* phi_c_cond_true_bb; ++ ++ create_new_bb(cond_false_bb, after_bb, dominator_bb, outer); ++ make_single_succ_edge(after_bb, cond_false_bb, EDGE_FALLTHRU); ++ ++ tree entry_s = get_current_def(origin_loop.base_s); ++ phi_s_cond_true_bb = create_phi_node_for_bb(entry_s, cond_false_bb); ++ tree entry_c = get_current_def(origin_loop.base_c); ++ phi_c_cond_true_bb = create_phi_node_for_bb(entry_c, cond_false_bb); ++ tree res_s = gimple_phi_result(phi_s_cond_true_bb); ++ set_current_def(origin_loop.base_s, res_s); ++ s_map.put(cond_false_bb, res_s); ++ tree res_c = gimple_phi_result(phi_c_cond_true_bb); ++ set_current_def(origin_loop.base_c, res_c); ++ c_map.put(cond_false_bb, res_c); ++ ++ gsi = gsi_last_bb(cond_false_bb); ++ tree lhs1 = gimple_build(&stmts, BIT_AND_EXPR, TREE_TYPE(origin_loop.base_n), ++ origin_loop.base_n, build_int_cst (TREE_TYPE (origin_loop.base_n), 1)); ++ gcond* cond_stmt = gimple_build_cond (NE_EXPR, lhs1, origin_loop.limit, ++ NULL_TREE, NULL_TREE); ++ gimple_seq_add_stmt (&stmts, cond_stmt); ++ gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); ++} ++ ++/* : ++ _11 = (unsigned int) c_17; ++ _12 = *s_15; ++ _42 = __builtin_aarch64_crc32b (_11, _12); ++ c_36 = (long unsigned int) _42; ++ The IR of bb is as above. */ ++static void ++create_lastcond_true_bb(basic_block& new_bb, basic_block after_bb, ++ basic_block dominator_bb, class loop *outer){ ++ gimple_seq stmts = NULL; ++ gimple_stmt_iterator gsi; ++ gimple* g; ++ ++ create_new_bb(new_bb, after_bb, dominator_bb, outer); ++ gsi = gsi_last_bb(new_bb); ++ ++ tree lhs1 = gimple_build(&stmts, NOP_EXPR, unsigned_type_node, ++ get_current_def(origin_loop.base_c)); ++ tree lhs2; ++ tree s_15 = get_current_def(origin_loop.base_s); ++ g = gimple_build_assign (make_ssa_name (unsigned_char_type_node), ++ fold_build2 (MEM_REF, unsigned_char_type_node, s_15, ++ build_int_cst (TREE_TYPE(s_15), 0))); ++ gimple_seq_add_stmt (&stmts, g); ++ lhs2 = gimple_assign_lhs (g); ++ ++ unsigned int code = AARCH64_BUILTIN_CRC32B; ++ tree lhs3; ++ gimple* call_builtin = call_builtin_fun(code, lhs3, lhs1, lhs2); ++ crc_map.put(new_bb,lhs3); ++ gimple_seq_add_stmt(&stmts,call_builtin); ++ ++ tree lhs4 = copy_ssa_name(origin_loop.base_c); ++ g = gimple_build_assign(lhs4, NOP_EXPR, lhs3); ++ gimple_seq_add_stmt(&stmts, g); ++ c_map.put(new_bb, lhs4); ++ ++ gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); ++} ++ ++static bool ++optional_add_phi_arg(gphi * phi, tree phi_res, tree phi_arg, edge e) ++{ ++ location_t loc; ++ if (same_ssa_name_var_p (phi_arg, phi_res)) ++ { ++ if (virtual_operand_p (phi_arg)) ++ loc = UNKNOWN_LOCATION; ++ else ++ loc = gimple_location (SSA_NAME_DEF_STMT (phi_arg)); ++ add_phi_arg (phi, phi_arg, e, loc); ++ ++ return true; ++ } ++ ++ return false; ++} ++ ++/* Add phi_arg for bb with phi node. */ ++static void ++update_phi_nodes (basic_block bb) ++{ ++ edge e; ++ edge_iterator ei; ++ gphi *phi; ++ gphi_iterator gsi; ++ tree res; ++ ++ for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) ++ { ++ phi = gsi.phi (); ++ res = gimple_phi_result (phi); ++ ++ FOR_EACH_EDGE (e, ei, bb->preds) ++ { ++ if (PHI_ARG_DEF_FROM_EDGE (phi, e)) ++ continue; ++ tree var_c; ++ tree* ptr_var_c = c_map.get (e->src); ++ if(ptr_var_c == NULL) ++ { ++ var_c = origin_loop.base_c; ++ } else { ++ var_c = *ptr_var_c; ++ } ++ if(optional_add_phi_arg(phi, res, var_c, e)) ++ continue; ++ ++ tree var_nn; ++ tree* ptr_var_nn = nn_map.get (e->src); ++ if(ptr_var_nn == NULL) ++ { ++ var_nn = nn_tree; ++ } else { ++ var_nn = *ptr_var_nn; ++ } ++ if(optional_add_phi_arg(phi, res, var_nn, e)) ++ continue; ++ ++ tree var_s; ++ tree* ptr_var_s = s_map.get (e->src); ++ if(ptr_var_s == NULL) ++ { ++ var_s = origin_loop.base_s; ++ } else { ++ var_s = *ptr_var_s; ++ } ++ if(optional_add_phi_arg(phi, res, var_s, e)) ++ continue; ++ } ++ } ++} ++ ++static void ++create_new_loops(edge entry_edge) ++{ ++ class loop* new_loop = NULL; ++ basic_block loop_bb, cond_bb, cond_true_bb, cond_false_bb, lastcond_true_bb; ++ class loop *outer = entry_edge->src->loop_father; ++ change_preheader_bb(entry_edge); ++ ++ create_loop_bb(loop_bb, entry_edge->src, entry_edge->src, outer, entry_edge); ++ create_cond_bb(cond_bb, loop_bb, loop_bb, outer); ++ make_edge(loop_bb, loop_bb, EDGE_TRUE_VALUE); ++ make_edge(loop_bb, cond_bb, EDGE_FALSE_VALUE); ++ update_phi_nodes(loop_bb); ++ ++ new_loop = alloc_loop (); ++ new_loop->header = loop_bb; ++ new_loop->latch = loop_bb; ++ add_loop (new_loop, outer); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\nPrint byte new loop %d:\n", new_loop->num); ++ flow_loop_dump (new_loop, dump_file, NULL, 1); ++ fprintf (dump_file, "\n\n"); ++ } ++ ++ create_cond_true_bb(cond_true_bb, cond_bb, cond_bb, outer); ++ make_edge(cond_bb, cond_true_bb, EDGE_TRUE_VALUE); ++ create_cond_false_bb(cond_false_bb, cond_true_bb, cond_bb, outer); ++ make_edge(cond_bb, cond_false_bb, EDGE_FALSE_VALUE); ++ update_phi_nodes(cond_bb); ++ update_phi_nodes(cond_false_bb); ++ create_lastcond_true_bb(lastcond_true_bb, cond_false_bb, cond_false_bb, outer); ++ make_edge(cond_false_bb, lastcond_true_bb, EDGE_TRUE_VALUE); ++ make_edge(cond_false_bb, origin_loop.exit_bb, EDGE_FALSE_VALUE); ++ make_single_succ_edge(lastcond_true_bb, origin_loop.exit_bb, EDGE_FALLTHRU); ++ ++ update_phi_nodes(origin_loop.exit_bb); ++ remove_edge(origin_loop.exit_edge); ++} ++ ++/* Clear information about the original loop. */ ++static void ++remove_origin_loop(class loop* loop) ++{ ++ basic_block* body = get_loop_body_in_dom_order(loop); ++ unsigned n = loop->num_nodes; ++ for(int i = 0; i < n; ++i) ++ { ++ delete_basic_block(body[i]); ++ } ++ free(body); ++ delete_loop(loop); ++} ++ ++/* Make sure that the dominance relationship of the newly inserted cfg ++ is not missing. */ ++static void ++update_loop_dominator(cdi_direction dir) ++{ ++ gcc_assert (dom_info_available_p (dir)); ++ ++ basic_block bb; ++ FOR_EACH_BB_FN (bb, cfun) ++ { ++ basic_block imm_bb = get_immediate_dominator (dir, bb); ++ if (!imm_bb || bb == origin_loop.exit_bb) ++ { ++ set_immediate_dominator (CDI_DOMINATORS, bb, ++ recompute_dominator (CDI_DOMINATORS, bb)); ++ continue; ++ } ++ } ++} ++ ++/* Perform the conversion of origin_loop to new_loop. */ ++static void ++convert_to_new_loop (class loop *loop) ++{ ++ create_new_loops (origin_loop.entry_edge); ++ remove_origin_loop (loop); ++ update_loop_dominator (CDI_DOMINATORS); ++ update_ssa (TODO_update_ssa); ++} ++ + /* The main entry of loop crc optimizes. */ + static unsigned int + tree_ssa_loop_crc () + { ++ if(TARGET_CRC32 == false){ ++ warning (OPT____,"The loop-crc optimization is not working."\ ++ "You should make sure that the specified architecture supports"\ ++ " crc:-march=armv8.1-a"); ++ return 0; ++ } + unsigned int todo = 0; + class loop *loop; + +@@ -553,28 +1230,28 @@ tree_ssa_loop_crc () + } + + FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "======================================\n"); ++ fprintf (dump_file, "Processing loop %d:\n", loop->num); ++ fprintf (dump_file, "======================================\n"); ++ flow_loop_dump (loop, dump_file, NULL, 1); ++ fprintf (dump_file, "\n\n"); ++ } ++ ++ if (match_crc_loop (loop)) + { + if (dump_file && (dump_flags & TDF_DETAILS)) +- { +- fprintf (dump_file, "======================================\n"); +- fprintf (dump_file, "Processing loop %d:\n", loop->num); +- fprintf (dump_file, "======================================\n"); +- flow_loop_dump (loop, dump_file, NULL, 1); +- fprintf (dump_file, "\n\n"); +- } +- +- if (match_crc_loop (loop)) +- { +- if (dump_file && (dump_flags & TDF_DETAILS)) +- { +- fprintf (dump_file, "The %dth loop form is success matched," +- "and the loop can be optimized.\n", +- loop->num); +- } +- +- convert_to_new_loop (loop); +- } ++ { ++ fprintf (dump_file, "The %dth loop form is success matched," ++ "and the loop can be optimized.\n", ++ loop->num); ++ } ++ ++ convert_to_new_loop (loop); + } ++ } + + todo |= (TODO_update_ssa); + return todo; +@@ -641,4 +1318,4 @@ gimple_opt_pass * + make_pass_loop_crc (gcc::context *ctxt) + { + return new pass_loop_crc (ctxt); +-} +\ No newline at end of file ++} +-- +2.33.0 + diff --git a/0147-add-insn-defs-and-correct-costs-for-cmlt-generation.patch b/0147-add-insn-defs-and-correct-costs-for-cmlt-generation.patch new file mode 100644 index 0000000..b03ad48 --- /dev/null +++ b/0147-add-insn-defs-and-correct-costs-for-cmlt-generation.patch @@ -0,0 +1,194 @@ +From 80b7de670da46d8921118799904cba4a0753bb72 Mon Sep 17 00:00:00 2001 +From: Diachkov Ilia WX1215920 +Date: Wed, 23 Aug 2023 15:03:00 +0300 +Subject: [PATCH 09/13] add insn defs and correct costs for cmlt generation + +--- + gcc/config/aarch64/aarch64-simd.md | 48 +++++++++++++++++++++++++++++ + gcc/config/aarch64/aarch64.c | 15 +++++++++ + gcc/config/aarch64/aarch64.opt | 4 +++ + gcc/config/aarch64/iterators.md | 3 +- + gcc/config/aarch64/predicates.md | 25 +++++++++++++++ + gcc/testsuite/gcc.dg/combine-cmlt.c | 20 ++++++++++++ + 6 files changed, 114 insertions(+), 1 deletion(-) + create mode 100755 gcc/testsuite/gcc.dg/combine-cmlt.c + +diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md +index 6049adc3f..f4213fd62 100644 +--- a/gcc/config/aarch64/aarch64-simd.md ++++ b/gcc/config/aarch64/aarch64-simd.md +@@ -4719,6 +4719,54 @@ + [(set_attr "type" "neon_compare, neon_compare_zero")] + ) + ++;; Use cmlt to replace vector arithmetic operations like this (SImode example): ++;; B = (((A >> 15) & 0x00010001) << 16) - ((A >> 15) & 0x00010001) ++;; TODO: maybe extend to scalar operations or other cm** instructions. ++ ++(define_insn "*aarch64_cmlt_as_arith" ++ [(set (match_operand: 0 "register_operand" "=w") ++ (minus: ++ (ashift: ++ (and: ++ (lshiftrt: ++ (match_operand:VDQHSD 1 "register_operand" "w") ++ (match_operand:VDQHSD 2 "half_size_minus_one_operand")) ++ (match_operand:VDQHSD 3 "cmlt_arith_mask_operand")) ++ (match_operand:VDQHSD 4 "half_size_operand")) ++ (and: ++ (lshiftrt: ++ (match_dup 1) ++ (match_dup 2)) ++ (match_dup 3))))] ++ "TARGET_SIMD && flag_cmlt_arith" ++ "cmlt\t%0., %1., #0" ++ [(set_attr "type" "neon_compare_zero")] ++) ++ ++;; The helper definition that allows combiner to use the previous pattern. ++ ++(define_insn_and_split "*aarch64_cmlt_tmp" ++ [(set (match_operand: 0 "register_operand" "=w") ++ (and: ++ (lshiftrt: ++ (match_operand:VDQHSD 1 "register_operand" "w") ++ (match_operand:VDQHSD 2 "half_size_minus_one_operand")) ++ (match_operand:VDQHSD 3 "cmlt_arith_mask_operand")))] ++ "TARGET_SIMD && flag_cmlt_arith" ++ "#" ++ "&& reload_completed" ++ [(set (match_operand: 0 "register_operand") ++ (lshiftrt: ++ (match_operand:VDQHSD 1 "register_operand") ++ (match_operand:VDQHSD 2 "half_size_minus_one_operand"))) ++ (set (match_dup 0) ++ (and: ++ (match_dup 0) ++ (match_operand:VDQHSD 3 "cmlt_arith_mask_operand")))] ++ "" ++ [(set_attr "type" "neon_compare_zero")] ++) ++ + (define_insn_and_split "aarch64_cmdi" + [(set (match_operand:DI 0 "register_operand" "=w,w,r") + (neg:DI +diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c +index cbdde11b0..7a00a0817 100644 +--- a/gcc/config/aarch64/aarch64.c ++++ b/gcc/config/aarch64/aarch64.c +@@ -12659,6 +12659,21 @@ cost_minus: + return true; + } + ++ /* Detect aarch64_cmlt_as_arith instruction. Now only this pattern ++ matches the condition. The costs of cmlt and sub instructions ++ are comparable, so we are not increasing the cost here. */ ++ if (flag_cmlt_arith && GET_CODE (op0) == ASHIFT ++ && GET_CODE (op1) == AND) ++ { ++ rtx op0_subop0 = XEXP (op0, 0); ++ if (rtx_equal_p (op0_subop0, op1)) ++ { ++ rtx lshrt_op = XEXP (op0_subop0, 0); ++ if (GET_CODE (lshrt_op) == LSHIFTRT) ++ return true; ++ } ++ } ++ + /* Look for SUB (extended register). */ + if (is_a (mode, &int_mode) + && aarch64_rtx_arith_op_extract_p (op1, int_mode)) +diff --git a/gcc/config/aarch64/aarch64.opt b/gcc/config/aarch64/aarch64.opt +index bb888461a..c42494036 100644 +--- a/gcc/config/aarch64/aarch64.opt ++++ b/gcc/config/aarch64/aarch64.opt +@@ -273,6 +273,10 @@ Use an immediate to offset from the stack protector guard register, sp_el0. + This option is for use with fstack-protector-strong and not for use in + user-land code. + ++mcmlt-arith ++Target Report Var(flag_cmlt_arith) Optimization Init(0) ++Use SIMD cmlt instruction to perform some arithmetic/logic calculations. ++ + TargetVariable + long aarch64_stack_protector_guard_offset = 0 + +diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md +index 0a7145281..d3be06c6f 100644 +--- a/gcc/config/aarch64/iterators.md ++++ b/gcc/config/aarch64/iterators.md +@@ -1228,7 +1228,8 @@ + (V2DI "2s")]) + + ;; Register suffix narrowed modes for VQN. +-(define_mode_attr V2ntype [(V8HI "16b") (V4SI "8h") ++(define_mode_attr V2ntype [(V4HI "8b") (V2SI "4h") ++ (V8HI "16b") (V4SI "8h") + (V2DI "4s")]) + + ;; Widened modes of vector modes. +diff --git a/gcc/config/aarch64/predicates.md b/gcc/config/aarch64/predicates.md +index 1754b1eff..de58562a7 100644 +--- a/gcc/config/aarch64/predicates.md ++++ b/gcc/config/aarch64/predicates.md +@@ -47,6 +47,31 @@ + return CONST_INT_P (op) && IN_RANGE (INTVAL (op), 1, 3); + }) + ++(define_predicate "half_size_minus_one_operand" ++ (match_code "const_vector") ++{ ++ op = unwrap_const_vec_duplicate (op); ++ unsigned int size = GET_MODE_UNIT_BITSIZE (mode) / 2; ++ return CONST_INT_P (op) && (UINTVAL (op) == size - 1); ++}) ++ ++(define_predicate "half_size_operand" ++ (match_code "const_vector") ++{ ++ op = unwrap_const_vec_duplicate (op); ++ unsigned int size = GET_MODE_UNIT_BITSIZE (mode) / 2; ++ return CONST_INT_P (op) && (UINTVAL (op) == size); ++}) ++ ++(define_predicate "cmlt_arith_mask_operand" ++ (match_code "const_vector") ++{ ++ op = unwrap_const_vec_duplicate (op); ++ unsigned int size = GET_MODE_UNIT_BITSIZE (mode) / 2; ++ unsigned long long mask = ((unsigned long long) 1 << size) | 1; ++ return CONST_INT_P (op) && (UINTVAL (op) == mask); ++}) ++ + (define_predicate "subreg_lowpart_operator" + (ior (match_code "truncate") + (and (match_code "subreg") +diff --git a/gcc/testsuite/gcc.dg/combine-cmlt.c b/gcc/testsuite/gcc.dg/combine-cmlt.c +new file mode 100755 +index 000000000..b4c9a37ff +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/combine-cmlt.c +@@ -0,0 +1,20 @@ ++/* { dg-do compile { target aarch64-*-* } } */ ++/* { dg-options "-O3 -mcmlt-arith" } */ ++ ++/* The test checks usage of cmlt insns for arithmetic/logic calculations ++ * in foo (). It's inspired by sources of x264 codec. */ ++ ++typedef unsigned short int uint16_t; ++typedef unsigned int uint32_t; ++ ++void foo( uint32_t *a, uint32_t *b) ++{ ++ for (unsigned i = 0; i < 4; i++) ++ { ++ uint32_t s = ((a[i]>>((8 * sizeof(uint16_t))-1)) ++ &(((uint32_t)1<<(8 * sizeof(uint16_t)))+1))*((uint16_t)-1); ++ b[i] = (a[i]+s)^s; ++ } ++} ++ ++/* { dg-final { scan-assembler-times {cmlt\t} 1 } } */ +-- +2.33.0 + diff --git a/0148-Introduce-RTL-ifcvt-enhancements.patch b/0148-Introduce-RTL-ifcvt-enhancements.patch new file mode 100644 index 0000000..1bc1aca --- /dev/null +++ b/0148-Introduce-RTL-ifcvt-enhancements.patch @@ -0,0 +1,502 @@ +From df68d120a049049671e44f6cda51e96a9a82c613 Mon Sep 17 00:00:00 2001 +From: Chernonog Vyacheslav 00812786 +Date: Mon, 28 Nov 2022 14:16:48 +0300 +Subject: [PATCH 10/13] Introduce RTL ifcvt enhancements + +It is controlled by option -fifcvt-allow-complicated-cmps, allowing +ifcvt to deal with complicated cmps like + if (cmp) + X = reg1 + else + X = reg2 + reg3 +and + if (cmp) + X = reg1 + reg3 + Y = reg2 + reg4 + Z = reg3 + +Parameter -param=ifcvt-allow-register-renaming=[0,1,2] allows ifcvt to +aggressively rename registers in basic blocks. +* 0: does not allow ifcvt to rename registers +* 1: allows ifcvt to rename registers in then and else bb +* 2: allows to rename registers in condition and else/then bb +--- + gcc/ifcvt.c | 298 ++++++++++++++++++++++++++++++++++++++----------- + gcc/params.opt | 8 ++ + 2 files changed, 240 insertions(+), 66 deletions(-) + +diff --git a/gcc/ifcvt.c b/gcc/ifcvt.c +index 2452f231c..50a73a7ca 100644 +--- a/gcc/ifcvt.c ++++ b/gcc/ifcvt.c +@@ -1,5 +1,5 @@ + /* If-conversion support. +- Copyright (C) 2000-2020 Free Software Foundation, Inc. ++ Copyright (C) 2000-2022 Free Software Foundation, Inc. + + This file is part of GCC. + +@@ -876,7 +876,9 @@ noce_emit_store_flag (struct noce_if_info *if_info, rtx x, int reversep, + } + + /* Don't even try if the comparison operands or the mode of X are weird. */ +- if (cond_complex || !SCALAR_INT_MODE_P (GET_MODE (x))) ++ if (!param_ifcvt_allow_complicated_cmps ++ && (cond_complex ++ || !SCALAR_INT_MODE_P (GET_MODE (x)))) + return NULL_RTX; + + return emit_store_flag (x, code, XEXP (cond, 0), +@@ -1743,8 +1745,9 @@ noce_emit_cmove (struct noce_if_info *if_info, rtx x, enum rtx_code code, + + /* Don't even try if the comparison operands are weird + except that the target supports cbranchcc4. */ +- if (! general_operand (cmp_a, GET_MODE (cmp_a)) +- || ! general_operand (cmp_b, GET_MODE (cmp_b))) ++ if (! param_ifcvt_allow_complicated_cmps ++ && (! general_operand (cmp_a, GET_MODE (cmp_a)) ++ || ! general_operand (cmp_b, GET_MODE (cmp_b)))) + { + if (!have_cbranchcc4 + || GET_MODE_CLASS (GET_MODE (cmp_a)) != MODE_CC +@@ -1915,19 +1918,6 @@ noce_try_cmove (struct noce_if_info *if_info) + return FALSE; + } + +-/* Return true if X contains a conditional code mode rtx. */ +- +-static bool +-contains_ccmode_rtx_p (rtx x) +-{ +- subrtx_iterator::array_type array; +- FOR_EACH_SUBRTX (iter, array, x, ALL) +- if (GET_MODE_CLASS (GET_MODE (*iter)) == MODE_CC) +- return true; +- +- return false; +-} +- + /* Helper for bb_valid_for_noce_process_p. Validate that + the rtx insn INSN is a single set that does not set + the conditional register CC and is in general valid for +@@ -1946,7 +1936,6 @@ insn_valid_noce_process_p (rtx_insn *insn, rtx cc) + /* Currently support only simple single sets in test_bb. */ + if (!sset + || !noce_operand_ok (SET_DEST (sset)) +- || contains_ccmode_rtx_p (SET_DEST (sset)) + || !noce_operand_ok (SET_SRC (sset))) + return false; + +@@ -1960,13 +1949,17 @@ insn_valid_noce_process_p (rtx_insn *insn, rtx cc) + in this function. */ + + static bool +-bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename) ++bbs_ok_for_cmove_arith (basic_block bb_a, ++ basic_block bb_b, ++ rtx to_rename, ++ bitmap conflict_regs) + { + rtx_insn *a_insn; + bitmap bba_sets = BITMAP_ALLOC (®_obstack); +- ++ bitmap intersections = BITMAP_ALLOC (®_obstack); + df_ref def; + df_ref use; ++ rtx_insn *last_a = last_active_insn (bb_a, FALSE); + + FOR_BB_INSNS (bb_a, a_insn) + { +@@ -1976,30 +1969,25 @@ bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename) + rtx sset_a = single_set (a_insn); + + if (!sset_a) +- { +- BITMAP_FREE (bba_sets); +- return false; +- } ++ goto end_cmove_arith_check_and_fail; ++ if (a_insn == last_a) ++ continue; + /* Record all registers that BB_A sets. */ + FOR_EACH_INSN_DEF (def, a_insn) + if (!(to_rename && DF_REF_REG (def) == to_rename)) + bitmap_set_bit (bba_sets, DF_REF_REGNO (def)); + } + ++ bitmap_and (intersections, df_get_live_in (bb_b), bba_sets); + rtx_insn *b_insn; +- + FOR_BB_INSNS (bb_b, b_insn) + { + if (!active_insn_p (b_insn)) + continue; +- + rtx sset_b = single_set (b_insn); + + if (!sset_b) +- { +- BITMAP_FREE (bba_sets); +- return false; +- } ++ goto end_cmove_arith_check_and_fail; + + /* Make sure this is a REG and not some instance + of ZERO_EXTRACT or SUBREG or other dangerous stuff. +@@ -2011,25 +1999,34 @@ bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename) + if (MEM_P (SET_DEST (sset_b))) + gcc_assert (rtx_equal_p (SET_DEST (sset_b), to_rename)); + else if (!REG_P (SET_DEST (sset_b))) +- { +- BITMAP_FREE (bba_sets); +- return false; +- } ++ goto end_cmove_arith_check_and_fail; + +- /* If the insn uses a reg set in BB_A return false. */ ++ /* If the insn uses a reg set in BB_A return false ++ or try to collect register list for renaming. */ + FOR_EACH_INSN_USE (use, b_insn) + { +- if (bitmap_bit_p (bba_sets, DF_REF_REGNO (use))) ++ if (bitmap_bit_p (intersections, DF_REF_REGNO (use))) + { +- BITMAP_FREE (bba_sets); +- return false; ++ if (param_ifcvt_allow_register_renaming < 1) ++ goto end_cmove_arith_check_and_fail; ++ ++ /* Those regs should be renamed. We can't rename CC reg, but ++ possibly we can provide combined comparison in the future. */ ++ if (GET_MODE_CLASS (GET_MODE (DF_REF_REG (use))) == MODE_CC) ++ goto end_cmove_arith_check_and_fail; ++ bitmap_set_bit (conflict_regs, DF_REF_REGNO (use)); + } + } +- + } + + BITMAP_FREE (bba_sets); ++ BITMAP_FREE (intersections); + return true; ++ ++end_cmove_arith_check_and_fail: ++ BITMAP_FREE (bba_sets); ++ BITMAP_FREE (intersections); ++ return false; + } + + /* Emit copies of all the active instructions in BB except the last. +@@ -2084,6 +2081,134 @@ noce_emit_bb (rtx last_insn, basic_block bb, bool simple) + return true; + } + ++/* This function tries to rename regs that intersect with considered bb. */ ++ ++static bool ++noce_rename_regs_in_cond (struct noce_if_info *if_info, bitmap cond_rename_regs) ++{ ++ bool success = true; ++ if (bitmap_empty_p (cond_rename_regs)) ++ return true; ++ if (param_ifcvt_allow_register_renaming < 2) ++ return false; ++ df_ref use; ++ rtx_insn* cmp_insn = if_info->cond_earliest; ++ /* Jump instruction as a condion currently unsupported. */ ++ if (JUMP_P (cmp_insn)) ++ return false; ++ rtx_insn* before_cmp = PREV_INSN (cmp_insn); ++ start_sequence (); ++ rtx_insn *copy_of_cmp = as_a (copy_rtx (cmp_insn)); ++ basic_block cmp_block = BLOCK_FOR_INSN (cmp_insn); ++ FOR_EACH_INSN_USE (use, cmp_insn) ++ { ++ if (bitmap_bit_p (cond_rename_regs, DF_REF_REGNO (use))) ++ { ++ rtx use_reg = DF_REF_REG (use); ++ rtx tmp = gen_reg_rtx (GET_MODE (use_reg)); ++ if (!validate_replace_rtx (use_reg, tmp, copy_of_cmp)) ++ { ++ end_sequence (); ++ return false; ++ } ++ noce_emit_move_insn (tmp, use_reg); ++ } ++ } ++ ++ emit_insn (PATTERN (copy_of_cmp)); ++ rtx_insn *seq = get_insns (); ++ unshare_all_rtl_in_chain (seq); ++ end_sequence (); ++ ++ emit_insn_after_setloc (seq, before_cmp, INSN_LOCATION (cmp_insn)); ++ delete_insn_and_edges (cmp_insn); ++ rtx_insn* insn; ++ FOR_BB_INSNS (cmp_block, insn) ++ df_insn_rescan (insn); ++ ++ if_info->cond = noce_get_condition (if_info->jump, ++ ©_of_cmp, ++ if_info->then_else_reversed); ++ if_info->cond_earliest = copy_of_cmp; ++ if_info->rev_cond = NULL_RTX; ++ ++ return success; ++} ++ ++/* This function tries to rename regs that intersect with considered bb. */ ++static bool ++noce_rename_regs_in_bb (basic_block test_bb, bitmap rename_regs) ++{ ++ if (bitmap_empty_p (rename_regs)) ++ return true; ++ rtx_insn* insn; ++ rtx_insn *last_insn = last_active_insn (test_bb, FALSE); ++ bool res = true; ++ start_sequence (); ++ FOR_BB_INSNS (test_bb, insn) ++ { ++ if (!active_insn_p (insn)) ++ continue; ++ /* Only ssets are supported for now. */ ++ rtx sset = single_set (insn); ++ gcc_assert (sset); ++ rtx x = SET_DEST (sset); ++ if (!REG_P (x) || bitmap_bit_p (rename_regs, REGNO (x))) ++ continue; ++ ++ machine_mode mode = GET_MODE (x); ++ rtx tmp = gen_reg_rtx (mode); ++ if (!validate_replace_rtx_part (x, tmp, &SET_DEST (sset), insn)) ++ { ++ gcc_assert (insn != last_insn); ++ /* We can generate additional move for such case, ++ but it will increase register preasure. ++ For now just stop transformation. */ ++ rtx result_rtx = SET_DEST (single_set (last_insn)); ++ if (REG_P (result_rtx) && (x != result_rtx)) ++ { ++ res = false; ++ break; ++ } ++ if (!validate_replace_rtx (x, tmp, insn)) ++ gcc_unreachable (); ++ noce_emit_move_insn (tmp,x); ++ } ++ set_used_flags (insn); ++ rtx_insn* rename_candidate; ++ for (rename_candidate = NEXT_INSN (insn); ++ rename_candidate && rename_candidate!= NEXT_INSN (BB_END (test_bb)); ++ rename_candidate = NEXT_INSN (rename_candidate)) ++ { ++ if (!reg_overlap_mentioned_p (x, rename_candidate)) ++ continue; ++ ++ int replace_res = TRUE; ++ if (rename_candidate == last_insn) ++ { ++ validate_replace_src_group (x, tmp, rename_candidate); ++ replace_res = apply_change_group (); ++ } ++ else ++ replace_res = validate_replace_rtx (x, tmp, rename_candidate); ++ gcc_assert (replace_res); ++ set_used_flags (rename_candidate); ++ ++ } ++ set_used_flags (x); ++ set_used_flags (tmp); ++ ++ } ++ rtx_insn *seq = get_insns (); ++ unshare_all_rtl_in_chain (seq); ++ end_sequence (); ++ emit_insn_before_setloc (seq, first_active_insn (test_bb), ++ INSN_LOCATION (first_active_insn (test_bb))); ++ FOR_BB_INSNS (test_bb, insn) ++ df_insn_rescan (insn); ++ return res; ++} ++ + /* Try more complex cases involving conditional_move. */ + + static int +@@ -2166,11 +2291,29 @@ noce_try_cmove_arith (struct noce_if_info *if_info) + std::swap (then_bb, else_bb); + } + } +- ++ bitmap else_bb_rename_regs = BITMAP_ALLOC (®_obstack); ++ bitmap then_bb_rename_regs = BITMAP_ALLOC (®_obstack); + if (then_bb && else_bb +- && (!bbs_ok_for_cmove_arith (then_bb, else_bb, if_info->orig_x) +- || !bbs_ok_for_cmove_arith (else_bb, then_bb, if_info->orig_x))) +- return FALSE; ++ && (!bbs_ok_for_cmove_arith (then_bb, else_bb, ++ if_info->orig_x, ++ then_bb_rename_regs) ++ || !bbs_ok_for_cmove_arith (else_bb, then_bb, ++ if_info->orig_x, ++ else_bb_rename_regs))) ++ { ++ BITMAP_FREE (then_bb_rename_regs); ++ BITMAP_FREE (else_bb_rename_regs); ++ return FALSE; ++ } ++ bool prepass_renaming = true; ++ prepass_renaming |= noce_rename_regs_in_bb (then_bb, then_bb_rename_regs); ++ prepass_renaming |= noce_rename_regs_in_bb (else_bb, else_bb_rename_regs); ++ ++ BITMAP_FREE (then_bb_rename_regs); ++ BITMAP_FREE (else_bb_rename_regs); ++ ++ if (!prepass_renaming) ++ return FALSE; + + start_sequence (); + +@@ -2178,7 +2321,6 @@ noce_try_cmove_arith (struct noce_if_info *if_info) + came from the test block. The non-empty complex block that we will + emit might clobber the register used by B or A, so move it to a pseudo + first. */ +- + rtx tmp_a = NULL_RTX; + rtx tmp_b = NULL_RTX; + +@@ -3052,7 +3194,8 @@ noce_operand_ok (const_rtx op) + + static bool + bb_valid_for_noce_process_p (basic_block test_bb, rtx cond, +- unsigned int *cost, bool *simple_p) ++ unsigned int *cost, bool *simple_p, ++ bitmap cond_rename_regs) + { + if (!test_bb) + return false; +@@ -3086,10 +3229,10 @@ bb_valid_for_noce_process_p (basic_block test_bb, rtx cond, + rtx_insn *prev_last_insn = PREV_INSN (last_insn); + gcc_assert (prev_last_insn); + +- /* For now, disallow setting x multiple times in test_bb. */ +- if (REG_P (x) && reg_set_between_p (x, first_insn, prev_last_insn)) ++ if (REG_P (x) ++ && reg_set_between_p (x, first_insn, prev_last_insn) ++ && param_ifcvt_allow_register_renaming < 1) + return false; +- + bitmap test_bb_temps = BITMAP_ALLOC (®_obstack); + + /* The regs that are live out of test_bb. */ +@@ -3099,25 +3242,35 @@ bb_valid_for_noce_process_p (basic_block test_bb, rtx cond, + rtx_insn *insn; + FOR_BB_INSNS (test_bb, insn) + { +- if (insn != last_insn) +- { +- if (!active_insn_p (insn)) +- continue; ++ if (insn == last_insn) ++ continue; ++ if (!active_insn_p (insn)) ++ continue; + +- if (!insn_valid_noce_process_p (insn, cc)) +- goto free_bitmap_and_fail; ++ if (!insn_valid_noce_process_p (insn, cc)) ++ goto free_bitmap_and_fail; + +- rtx sset = single_set (insn); +- gcc_assert (sset); ++ rtx sset = single_set (insn); ++ gcc_assert (sset); + +- if (contains_mem_rtx_p (SET_SRC (sset)) +- || !REG_P (SET_DEST (sset)) +- || reg_overlap_mentioned_p (SET_DEST (sset), cond)) +- goto free_bitmap_and_fail; ++ if (contains_mem_rtx_p (SET_SRC (sset)) ++ || !REG_P (SET_DEST (sset))) ++ goto free_bitmap_and_fail; + +- potential_cost += pattern_cost (sset, speed_p); +- bitmap_set_bit (test_bb_temps, REGNO (SET_DEST (sset))); ++ if (reg_overlap_mentioned_p (SET_DEST (sset), cond)) ++ { ++ if (param_ifcvt_allow_register_renaming < 1) ++ goto free_bitmap_and_fail; ++ rtx sset_dest = SET_DEST (sset); ++ if (REG_P (sset_dest) ++ && (GET_MODE_CLASS (GET_MODE (sset_dest)) != MODE_CC)) ++ bitmap_set_bit (cond_rename_regs, REGNO (sset_dest)); ++ else ++ goto free_bitmap_and_fail; + } ++ potential_cost += pattern_cost (sset, speed_p); ++ if (SET_DEST (sset) != SET_DEST (last_set)) ++ bitmap_set_bit (test_bb_temps, REGNO (SET_DEST (sset))); + } + + /* If any of the intermediate results in test_bb are live after test_bb +@@ -3475,14 +3628,27 @@ noce_process_if_block (struct noce_if_info *if_info) + + bool speed_p = optimize_bb_for_speed_p (test_bb); + unsigned int then_cost = 0, else_cost = 0; ++ bitmap cond_rename_regs = BITMAP_ALLOC (®_obstack); + if (!bb_valid_for_noce_process_p (then_bb, cond, &then_cost, +- &if_info->then_simple)) +- return false; ++ &if_info->then_simple, cond_rename_regs)) ++ { ++ BITMAP_FREE (cond_rename_regs); ++ return false; ++ } + + if (else_bb + && !bb_valid_for_noce_process_p (else_bb, cond, &else_cost, +- &if_info->else_simple)) ++ &if_info->else_simple, cond_rename_regs)) ++ { ++ BITMAP_FREE (cond_rename_regs); ++ return false; ++ } ++ ++ if (!noce_rename_regs_in_cond (if_info, cond_rename_regs)) + return false; ++ cond = if_info->cond; ++ ++ BITMAP_FREE (cond_rename_regs); + + if (speed_p) + if_info->original_cost += average_cost (then_cost, else_cost, +@@ -5426,7 +5592,7 @@ if_convert (bool after_combine) + { + basic_block bb; + int pass; +- ++ cleanup_cfg (CLEANUP_EXPENSIVE); + if (optimize == 1) + { + df_live_add_problem (); +diff --git a/gcc/params.opt b/gcc/params.opt +index 83fd705ee..345f9b3ff 100644 +--- a/gcc/params.opt ++++ b/gcc/params.opt +@@ -574,6 +574,14 @@ Maximum permissible cost for the sequence that would be generated by the RTL if- + Common Joined UInteger Var(param_max_rtl_if_conversion_unpredictable_cost) Init(40) IntegerRange(0, 200) Param Optimization + Maximum permissible cost for the sequence that would be generated by the RTL if-conversion pass for a branch that is considered unpredictable. + ++-param=ifcvt-allow-complicated-cmps= ++Common Joined UInteger Var(param_ifcvt_allow_complicated_cmps) IntegerRange(0, 1) Param Optimization ++Allow RTL if-conversion pass to deal with complicated cmps (can increase compilation time). ++ ++-param=ifcvt-allow-register-renaming= ++Common Joined UInteger Var(param_ifcvt_allow_register_renaming) IntegerRange(0, 2) Param Optimization ++Allow RTL if-conversion pass to aggressively rename registers in basic blocks. Sometimes additional moves will be created. ++ + -param=max-sched-extend-regions-iters= + Common Joined UInteger Var(param_max_sched_extend_regions_iters) Param Optimization + The maximum number of iterations through CFG to extend regions. +-- +2.33.0 + diff --git a/0149-Add-more-flexible-check-for-pointer-aliasing-during-.patch b/0149-Add-more-flexible-check-for-pointer-aliasing-during-.patch new file mode 100644 index 0000000..2d53e18 --- /dev/null +++ b/0149-Add-more-flexible-check-for-pointer-aliasing-during-.patch @@ -0,0 +1,239 @@ +From f43bdfbdcfdeb425a0bd303f4787a13323fd2934 Mon Sep 17 00:00:00 2001 +From: vchernon +Date: Wed, 27 Sep 2023 11:07:29 +0800 +Subject: [PATCH 11/13] Add more flexible check for pointer aliasing during + vectorization + +It takes minimum between number of iteration and segment length and helps to +speed up loops with small number of iterations when only tail can be vectorized. +--- + gcc/params.opt | 5 ++ + .../sve/var_stride_flexible_segment_len_1.c | 23 +++++++ + gcc/tree-data-ref.c | 68 +++++++++++++------ + gcc/tree-data-ref.h | 11 ++- + gcc/tree-vect-data-refs.c | 14 +++- + 5 files changed, 95 insertions(+), 26 deletions(-) + create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/var_stride_flexible_segment_len_1.c + +diff --git a/gcc/params.opt b/gcc/params.opt +index 83fd705ee..7f335a94b 100644 +--- a/gcc/params.opt ++++ b/gcc/params.opt +@@ -964,6 +964,11 @@ Maximum number of loop peels to enhance alignment of data references in a loop. + Common Joined UInteger Var(param_vect_max_version_for_alias_checks) Init(10) Param Optimization + Bound on number of runtime checks inserted by the vectorizer's loop versioning for alias check. + ++-param=vect-alias-flexible-segment-len= ++Common Joined UInteger Var(param_flexible_seg_len) Init(0) IntegerRange(0, 1) Param Optimization ++Use a minimum length of different segments. Currently the minimum between ++iteration number and vectorization length is chosen by this param. ++ + -param=vect-max-version-for-alignment-checks= + Common Joined UInteger Var(param_vect_max_version_for_alignment_checks) Init(6) Param Optimization + Bound on number of runtime checks inserted by the vectorizer's loop versioning for alignment check. +diff --git a/gcc/testsuite/gcc.target/aarch64/sve/var_stride_flexible_segment_len_1.c b/gcc/testsuite/gcc.target/aarch64/sve/var_stride_flexible_segment_len_1.c +new file mode 100644 +index 000000000..894f075f3 +--- /dev/null ++++ b/gcc/testsuite/gcc.target/aarch64/sve/var_stride_flexible_segment_len_1.c +@@ -0,0 +1,23 @@ ++/* { dg-do compile } */ ++/* { dg-options "-O2 -ftree-vectorize --param=vect-alias-flexible-segment-len=1" } */ ++ ++#define TYPE int ++#define SIZE 257 ++ ++void __attribute__ ((weak)) ++f (TYPE *x, TYPE *y, unsigned short n, long m __attribute__((unused))) ++{ ++ for (int i = 0; i < SIZE; ++i) ++ x[i * n] += y[i * n]; ++} ++ ++/* { dg-final { scan-assembler {\tld1w\tz[0-9]+} } } */ ++/* { dg-final { scan-assembler {\tst1w\tz[0-9]+} } } */ ++/* { dg-final { scan-assembler {\tldr\tw[0-9]+} } } */ ++/* { dg-final { scan-assembler {\tstr\tw[0-9]+} } } */ ++/* Should use a WAR check that multiplies by (VF-2)*4 rather than ++ an overlap check that multiplies by (257-1)*4. */ ++/* { dg-final { scan-assembler {\tcntb\t(x[0-9]+)\n.*\tsub\tx[0-9]+, \1, #8\n.*\tmul\tx[0-9]+,[^\n]*\1} } } */ ++/* One range check and a check for n being zero. */ ++/* { dg-final { scan-assembler-times {\t(?:cmp|tst)\t} 2 } } */ ++/* { dg-final { scan-assembler-times {\tccmp\t} 1 } } */ +diff --git a/gcc/tree-data-ref.c b/gcc/tree-data-ref.c +index 2cb54def8..8c5f1048c 100644 +--- a/gcc/tree-data-ref.c ++++ b/gcc/tree-data-ref.c +@@ -2071,31 +2071,14 @@ create_intersect_range_checks_index (class loop *loop, tree *cond_expr, + same arguments. Try to optimize cases in which the second access + is a write and in which some overlap is valid. */ + +-static bool +-create_waw_or_war_checks (tree *cond_expr, ++static void ++create_waw_or_war_checks2 (tree *cond_expr, tree seg_len_a, + const dr_with_seg_len_pair_t &alias_pair) + { + const dr_with_seg_len& dr_a = alias_pair.first; + const dr_with_seg_len& dr_b = alias_pair.second; + +- /* Check for cases in which: +- +- (a) DR_B is always a write; +- (b) the accesses are well-ordered in both the original and new code +- (see the comment above the DR_ALIAS_* flags for details); and +- (c) the DR_STEPs describe all access pairs covered by ALIAS_PAIR. */ +- if (alias_pair.flags & ~(DR_ALIAS_WAR | DR_ALIAS_WAW)) +- return false; +- +- /* Check for equal (but possibly variable) steps. */ + tree step = DR_STEP (dr_a.dr); +- if (!operand_equal_p (step, DR_STEP (dr_b.dr))) +- return false; +- +- /* Make sure that we can operate on sizetype without loss of precision. */ +- tree addr_type = TREE_TYPE (DR_BASE_ADDRESS (dr_a.dr)); +- if (TYPE_PRECISION (addr_type) != TYPE_PRECISION (sizetype)) +- return false; + + /* All addresses involved are known to have a common alignment ALIGN. + We can therefore subtract ALIGN from an exclusive endpoint to get +@@ -2112,9 +2095,6 @@ create_waw_or_war_checks (tree *cond_expr, + fold_convert (ssizetype, indicator), + ssize_int (0)); + +- /* Get lengths in sizetype. */ +- tree seg_len_a +- = fold_convert (sizetype, rewrite_to_non_trapping_overflow (dr_a.seg_len)); + step = fold_convert (sizetype, rewrite_to_non_trapping_overflow (step)); + + /* Each access has the following pattern: +@@ -2221,6 +2201,50 @@ create_waw_or_war_checks (tree *cond_expr, + *cond_expr = fold_build2 (GT_EXPR, boolean_type_node, subject, limit); + if (dump_enabled_p ()) + dump_printf (MSG_NOTE, "using an address-based WAR/WAW test\n"); ++} ++ ++/* This is a wrapper function for create_waw_or_war_checks2. */ ++static bool ++create_waw_or_war_checks (tree *cond_expr, ++ const dr_with_seg_len_pair_t &alias_pair) ++{ ++ const dr_with_seg_len& dr_a = alias_pair.first; ++ const dr_with_seg_len& dr_b = alias_pair.second; ++ ++ /* Check for cases in which: ++ ++ (a) DR_B is always a write; ++ (b) the accesses are well-ordered in both the original and new code ++ (see the comment above the DR_ALIAS_* flags for details); and ++ (c) the DR_STEPs describe all access pairs covered by ALIAS_PAIR. */ ++ if (alias_pair.flags & ~(DR_ALIAS_WAR | DR_ALIAS_WAW)) ++ return false; ++ ++ /* Check for equal (but possibly variable) steps. */ ++ tree step = DR_STEP (dr_a.dr); ++ if (!operand_equal_p (step, DR_STEP (dr_b.dr))) ++ return false; ++ ++ /* Make sure that we can operate on sizetype without loss of precision. */ ++ tree addr_type = TREE_TYPE (DR_BASE_ADDRESS (dr_a.dr)); ++ if (TYPE_PRECISION (addr_type) != TYPE_PRECISION (sizetype)) ++ return false; ++ ++ /* Get lengths in sizetype. */ ++ tree seg_len_a ++ = fold_convert (sizetype, ++ rewrite_to_non_trapping_overflow (dr_a.seg_len)); ++ create_waw_or_war_checks2 (cond_expr, seg_len_a, alias_pair); ++ if (param_flexible_seg_len && dr_a.seg_len != dr_a.seg_len2) ++ { ++ tree seg_len2_a ++ = fold_convert (sizetype, ++ rewrite_to_non_trapping_overflow (dr_a.seg_len2)); ++ tree cond_expr2; ++ create_waw_or_war_checks2 (&cond_expr2, seg_len2_a, alias_pair); ++ *cond_expr = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, ++ *cond_expr, cond_expr2); ++ } + return true; + } + +diff --git a/gcc/tree-data-ref.h b/gcc/tree-data-ref.h +index 771d20fbb..5903ce66a 100644 +--- a/gcc/tree-data-ref.h ++++ b/gcc/tree-data-ref.h +@@ -208,12 +208,19 @@ class dr_with_seg_len + public: + dr_with_seg_len (data_reference_p d, tree len, unsigned HOST_WIDE_INT size, + unsigned int a) +- : dr (d), seg_len (len), access_size (size), align (a) {} +- ++ : dr (d), seg_len (len), seg_len2 (len), access_size (size), align (a) ++ {} ++ dr_with_seg_len (data_reference_p d, tree len, tree len2, ++ unsigned HOST_WIDE_INT size, unsigned int a) ++ : dr (d), seg_len (len), seg_len2 (len2), access_size (size), align (a) ++ {} + data_reference_p dr; + /* The offset of the last access that needs to be checked minus + the offset of the first. */ + tree seg_len; ++ /* The second version of segment length. Currently this is used to ++ soften checks for a small number of iterations. */ ++ tree seg_len2; + /* A value that, when added to abs (SEG_LEN), gives the total number of + bytes in the segment. */ + poly_uint64 access_size; +diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c +index e4466a4f3..1b8a03c9c 100644 +--- a/gcc/tree-vect-data-refs.c ++++ b/gcc/tree-vect-data-refs.c +@@ -3498,6 +3498,7 @@ vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo) + { + poly_uint64 lower_bound; + tree segment_length_a, segment_length_b; ++ tree segment_length2_a, segment_length2_b; + unsigned HOST_WIDE_INT access_size_a, access_size_b; + unsigned int align_a, align_b; + +@@ -3598,6 +3599,8 @@ vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo) + { + segment_length_a = size_zero_node; + segment_length_b = size_zero_node; ++ segment_length2_a = size_zero_node; ++ segment_length2_b = size_zero_node; + } + else + { +@@ -3606,8 +3609,15 @@ vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo) + length_factor = scalar_loop_iters; + else + length_factor = size_int (vect_factor); ++ /* In any case we should rememeber scalar_loop_iters ++ this helps to create flexible aliasing check ++ for small number of iterations. */ + segment_length_a = vect_vfa_segment_size (dr_info_a, length_factor); + segment_length_b = vect_vfa_segment_size (dr_info_b, length_factor); ++ segment_length2_a ++ = vect_vfa_segment_size (dr_info_a, scalar_loop_iters); ++ segment_length2_b ++ = vect_vfa_segment_size (dr_info_b, scalar_loop_iters); + } + access_size_a = vect_vfa_access_size (dr_info_a); + access_size_b = vect_vfa_access_size (dr_info_b); +@@ -3652,9 +3662,9 @@ vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo) + } + + dr_with_seg_len dr_a (dr_info_a->dr, segment_length_a, +- access_size_a, align_a); ++ segment_length2_a, access_size_a, align_a); + dr_with_seg_len dr_b (dr_info_b->dr, segment_length_b, +- access_size_b, align_b); ++ segment_length2_b, access_size_b, align_b); + /* Canonicalize the order to be the one that's needed for accurate + RAW, WAR and WAW flags, in cases where the data references are + well-ordered. The order doesn't really matter otherwise, +-- +2.33.0 + diff --git a/0150-Implement-propagation-of-permutations-in-fwprop.patch b/0150-Implement-propagation-of-permutations-in-fwprop.patch new file mode 100644 index 0000000..005730e --- /dev/null +++ b/0150-Implement-propagation-of-permutations-in-fwprop.patch @@ -0,0 +1,1050 @@ +From 07aa5f889dc8bc3e642affe21dcfc197ad7d8b3b Mon Sep 17 00:00:00 2001 +From: Diachkov Ilia +Date: Sun, 3 Sep 2023 05:52:32 +0800 +Subject: [PATCH 12/13] Implement propagation of permutations in fwprop + +It is an implementation of permutation forward propagation, which is a +transformation designed to decrease the number of vector permutation +instructions in vectorized code, moving the permutations over arithmetic +operations. +--- + gcc/config/aarch64/aarch64-simd.md | 26 + + gcc/params.opt | 4 + + gcc/testsuite/gcc.dg/vect/transpose-9.c | 56 ++ + gcc/tree-ssa-forwprop.c | 891 ++++++++++++++++++++++++ + 4 files changed, 977 insertions(+) + create mode 100755 gcc/testsuite/gcc.dg/vect/transpose-9.c + +diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md +index 6049adc3f..af6d3ebf6 100644 +--- a/gcc/config/aarch64/aarch64-simd.md ++++ b/gcc/config/aarch64/aarch64-simd.md +@@ -4615,6 +4615,19 @@ + [(set_attr "type" "neon_shift_imm_long")] + ) + ++(define_insn "*aarch64_simd_vec_unpacks_lo_shiftsi" ++ [(set (match_operand:V4SI 0 "register_operand" "=w") ++ (ashift:V4SI ++ (sign_extend:V4SI ++ (vec_select:V4HI ++ (match_operand:V8HI 1 "register_operand" "w") ++ (match_operand:V8HI 2 "vect_par_cnst_lo_half" ""))) ++ (match_operand:V4SI 3 "aarch64_simd_rshift_imm" "Dr")))] ++ "TARGET_SIMD" ++ "shll\t%0.4s, %1.4h, #%3" ++ [(set_attr "type" "neon_compare_zero")] ++) ++ + ;; vshll_high_n + + (define_insn "aarch64_shll2_n" +@@ -4632,6 +4645,19 @@ + [(set_attr "type" "neon_shift_imm_long")] + ) + ++(define_insn "*aarch64_simd_vec_unpacks_hi_shiftsi" ++ [(set (match_operand:V4SI 0 "register_operand" "=w") ++ (ashift:V4SI ++ (sign_extend:V4SI ++ (vec_select:V4HI ++ (match_operand:V8HI 1 "register_operand" "w") ++ (match_operand:V8HI 2 "vect_par_cnst_hi_half" ""))) ++ (match_operand:V4SI 3 "aarch64_simd_rshift_imm" "Dr")))] ++ "TARGET_SIMD" ++ "shll2\t%0.4s, %1.8h, #%3" ++ [(set_attr "type" "neon_compare_zero")] ++) ++ + ;; vrshr_n + + (define_insn "aarch64_shr_n" +diff --git a/gcc/params.opt b/gcc/params.opt +index 83fd705ee..a87f6f00a 100644 +--- a/gcc/params.opt ++++ b/gcc/params.opt +@@ -852,6 +852,10 @@ Maximum size, in storage units, of an aggregate which should be considered for s + Common Joined UInteger Var(param_sra_max_propagations) Param Optimization Init(32) + Maximum number of artificial accesses to enable forward propagation that Scalar Replacement of Aggregates will keep for one local variable. + ++-param=tree-forwprop-perm= ++Common Joined UInteger Var(param_tree_forwprop_perm) Param Optimization Init(0) ++Propagate permutations in vectorized code on tree forward propagation. ++ + -param=ssa-name-def-chain-limit= + Common Joined UInteger Var(param_ssa_name_def_chain_limit) Init(512) Param Optimization + The maximum number of SSA_NAME assignments to follow in determining a value. +diff --git a/gcc/testsuite/gcc.dg/vect/transpose-9.c b/gcc/testsuite/gcc.dg/vect/transpose-9.c +new file mode 100755 +index 000000000..f20a67c6e +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/vect/transpose-9.c +@@ -0,0 +1,56 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-additional-options "-mtune=tsv110 --param=tree-forwprop-perm=1 -fdump-tree-forwprop-details" } */ ++/* { dg-require-effective-target vect_int } */ ++#include ++#include ++#include "tree-vect.h" ++ ++typedef unsigned short int sum_t; ++typedef unsigned int sum2_t; ++typedef long int intptr_t; ++typedef unsigned char data; ++#define BITS_PER_SUM (8 * sizeof(sum_t)) ++ ++static sum2_t bar(sum2_t a ) ++{ ++ sum2_t s = ((a>>(BITS_PER_SUM-1))&(((sum2_t)1<>BITS_PER_SUM)) >> 1; ++} ++/* { dg-final { scan-tree-dump "Initial permutations were reduced:" "forwprop4" } } */ ++/* { dg-final { scan-tree-dump "Permutations were moved through binary operations:" "forwprop4" } } */ ++ +diff --git a/gcc/tree-ssa-forwprop.c b/gcc/tree-ssa-forwprop.c +index ba0b55f4a..92ef5d036 100644 +--- a/gcc/tree-ssa-forwprop.c ++++ b/gcc/tree-ssa-forwprop.c +@@ -2225,6 +2225,893 @@ simplify_permutation (gimple_stmt_iterator *gsi) + return 0; + } + ++/* Compare the UID of two gimple stmts for sorting in ascending order. */ ++ ++static int ++gimple_uid_cmp (const void *ptr0, const void *ptr1) ++{ ++ const gimple *stmt0 = *(gimple * const *) ptr0; ++ const gimple *stmt1 = *(gimple * const *) ptr1; ++ ++ if (gimple_uid (stmt0) < gimple_uid (stmt1)) ++ return -1; ++ else if (gimple_uid (stmt0) > gimple_uid (stmt1)) ++ return 1; ++ return 0; ++} ++ ++/* Find a source permutation statement in backward direction through a chain of ++ unary, single or binary operations. In the last case only one variable ++ operand is allowed. If it's found, return true and save the statement in ++ perm_stmts, otherwise return false. */ ++ ++static bool ++find_src_perm_stmt (tree op, auto_vec &perm_stmts) ++{ ++ gimple *stmt; ++ while ((stmt = get_prop_source_stmt (op, false, NULL))) ++ { ++ if (!can_propagate_from (stmt)) ++ return false; ++ ++ if (gimple_assign_rhs_code (stmt) == VEC_PERM_EXPR) ++ { ++ perm_stmts.safe_push (stmt); ++ return true; ++ } ++ ++ /* TODO: check vector length and element size. */ ++ enum tree_code code = gimple_assign_rhs_code (stmt); ++ switch (get_gimple_rhs_class (code)) ++ { ++ case GIMPLE_TERNARY_RHS: ++ return false; ++ case GIMPLE_BINARY_RHS: ++ { ++ tree op1 = gimple_assign_rhs1 (stmt); ++ tree op2 = gimple_assign_rhs2 (stmt); ++ bool is_cst_op1 = is_gimple_constant (op1); ++ bool is_cst_op2 = is_gimple_constant (op2); ++ if ((is_cst_op1 && is_cst_op2) || (!is_cst_op1 && !is_cst_op2)) ++ return false; ++ op = !is_cst_op1 && is_cst_op2 ? op1 : op2; ++ break; ++ } ++ case GIMPLE_UNARY_RHS: ++ case GIMPLE_SINGLE_RHS: ++ op = gimple_assign_rhs1 (stmt); ++ break; ++ default: ++ gcc_unreachable (); ++ } ++ if (TREE_CODE (op) != SSA_NAME) ++ return false; ++ } ++ return false; ++} ++ ++/* Check the stmt is binary operation and find initial permutations for both ++ of its sources. */ ++ ++static bool ++find_initial_permutations (gimple_stmt_iterator *gsi, tree &type, ++ auto_vec &perm_stmts) ++{ ++ gimple *stmt = gsi_stmt (*gsi); ++ enum tree_code code = gimple_assign_rhs_code (stmt); ++ ++ // TODO: support other initial binary operations. ++ gcc_checking_assert (code == PLUS_EXPR || code == MINUS_EXPR); ++ ++ type = TREE_TYPE (gimple_assign_lhs (stmt)); ++ if (!VECTOR_TYPE_P (type)) ++ return false; ++ tree op1 = gimple_assign_rhs1 (stmt); ++ tree op2 = gimple_assign_rhs2 (stmt); ++ if (TREE_CODE (op1) != SSA_NAME || TREE_CODE (op2) != SSA_NAME ++ || TREE_TYPE (op1) != type || TREE_TYPE (op2) != type || op1 == op2) ++ return false; ++ ++ if (find_src_perm_stmt (op1, perm_stmts) ++ && find_src_perm_stmt (op2, perm_stmts)) ++ return true; ++ return false; ++} ++ ++/* Check if the permutation statement is suitable for the transformation. */ ++ ++static bool ++check_perm_stmt (gimple *stmt, tree type, vec *perm_stmts, ++ vec *src_vects) ++{ ++ if (!stmt || !can_propagate_from (stmt)) ++ return false; ++ ++ enum tree_code code = gimple_assign_rhs_code (stmt); ++ if (code != VEC_PERM_EXPR) ++ return false; ++ ++ tree op3 = gimple_assign_rhs3 (stmt); ++ tree op1 = gimple_assign_rhs1 (stmt); ++ tree op2 = gimple_assign_rhs2 (stmt); ++ if (TREE_CODE (op1) != SSA_NAME || TREE_CODE (op2) != SSA_NAME ++ || TREE_CODE (op3) != VECTOR_CST) ++ return false; ++ if (type != NULL_TREE && (TREE_TYPE (op1) != type ++ || TREE_TYPE (op2) != type)) ++ return false; ++ if (perm_stmts) ++ perm_stmts->safe_push (stmt); ++ if (src_vects) ++ { ++ src_vects->safe_push (op1); ++ src_vects->safe_push (op2); ++ } ++ return true; ++} ++ ++/* Collect permutation stmts preceding the given stmt. */ ++ ++static bool ++find_perm_set (gimple *stmt, tree type, vec &perm_stmts, ++ vec &src_vects) ++{ ++ auto_vec ops; ++ if (!check_perm_stmt (stmt, NULL, NULL, &ops)) ++ return false; ++ ++ unsigned i; ++ tree op; ++ bool single_use_op = false; ++ FOR_EACH_VEC_ELT (ops, i, op) ++ { ++ /* Skip if we already processed the same operand. */ ++ if (i > 0 && ops[i] == ops[i - 1]) ++ continue; ++ /* Find one permutation stmt. */ ++ gimple *def_stmt = get_prop_source_stmt (op, false, &single_use_op); ++ if (!check_perm_stmt (def_stmt, type, &perm_stmts, &src_vects)) ++ return false; ++ if (single_use_op || src_vects.length () <= 1) ++ return false; ++ unsigned last_i = src_vects.length () - 1; ++ unsigned before_last_i = src_vects.length () - 2; ++ ++ /* Find one more permutation stmt. */ ++ gimple *use_stmt; ++ imm_use_iterator iter; ++ FOR_EACH_IMM_USE_STMT (use_stmt, iter, src_vects[before_last_i]) ++ if (use_stmt != def_stmt) ++ BREAK_FROM_IMM_USE_STMT (iter); ++ if (!use_stmt || use_stmt == def_stmt ++ || gimple_assign_rhs_code (use_stmt) != VEC_PERM_EXPR ++ || src_vects[before_last_i] != gimple_assign_rhs1 (use_stmt) ++ || src_vects[last_i] != gimple_assign_rhs2 (use_stmt)) ++ return false; ++ perm_stmts.safe_push (use_stmt); ++ } ++ return true; ++} ++ ++/* Walk permutation pattern and make a vector of permutation indices. */ ++ ++static bool ++make_vec_of_indices (vec &perm_pattern, vec &perm_indices) ++{ ++ unsigned i, j; ++ tree tree_it; ++ FOR_EACH_VEC_ELT (perm_pattern, i, tree_it) ++ { ++ unsigned HOST_WIDE_INT nelts; ++ if (!VECTOR_CST_NELTS (tree_it).is_constant (&nelts)) ++ return false; ++ for (j = 0; j < nelts; j++) ++ { ++ tree val = VECTOR_CST_ELT (tree_it, j); ++ gcc_checking_assert (TREE_CODE (val) == INTEGER_CST); ++ perm_indices.safe_push (TREE_INT_CST_LOW (val)); ++ } ++ } ++ return true; ++} ++ ++/* Check or collect a permutation pattern in the provided perm_stmts depending ++ on the passed parameters. If collect_pattern is true, collect permutation ++ vectors to pattern. In other case, check the pattern suits perm_stmts. */ ++ ++static bool ++check_or_collect_perm_pattern (vec &perm_stmts, vec &pattern, ++ bool collect_pattern) ++{ ++ unsigned i, j; ++ gimple *stmt_it; ++ tree tree_it; ++ FOR_EACH_VEC_ELT (perm_stmts, i, stmt_it) ++ { ++ gcc_assert (gimple_assign_rhs_code (stmt_it) == VEC_PERM_EXPR); ++ tree perm_vec = gimple_assign_rhs3 (stmt_it); ++ bool found = false; ++ FOR_EACH_VEC_ELT (pattern, j, tree_it) ++ if (operand_equal_p (tree_it, perm_vec)) ++ { ++ found = true; ++ break; ++ } ++ if (collect_pattern && !found) ++ pattern.safe_push (perm_vec); ++ else ++ gcc_assert (found); ++ if (i % pattern.length () != j) ++ return false; ++ } ++ return true; ++} ++ ++/* Identify the permutation pattern and check it. For now, we are checking ++ only transposition permutations with no more than 2 lines in their patterns. ++ Collect permutation const vectors and the second permutation stmts. */ ++ ++static bool ++check_perm_pattern (vec &first_perm_stmts, vec &perm_pattern, ++ vec &second_perm_stmts) ++{ ++ unsigned i, j; ++ gimple *stmt_it; ++ if (!check_or_collect_perm_pattern (first_perm_stmts, perm_pattern, true)) ++ return false; ++ ++ if (perm_pattern.length () == 0 || perm_pattern.length () > 2) ++ return false; ++ ++ /* Find the second permutation stmts. */ ++ hash_set visited; ++ FOR_EACH_VEC_ELT (first_perm_stmts, i, stmt_it) ++ { ++ tree dst = gimple_assign_lhs (stmt_it); ++ use_operand_p use_p; ++ imm_use_iterator iter; ++ FOR_EACH_IMM_USE_FAST (use_p, iter, dst) ++ { ++ gimple *stmt_it2 = USE_STMT (use_p); ++ if (visited.contains (stmt_it2)) ++ continue; ++ second_perm_stmts.safe_push (stmt_it2); ++ visited.add (stmt_it2); ++ } ++ } ++ second_perm_stmts.qsort (gimple_uid_cmp); ++ ++ if (first_perm_stmts.length () != second_perm_stmts.length ()) ++ return false; ++ ++ /* Check that all second_perm_stmts are VEC_PERM_EXPR. */ ++ FOR_EACH_VEC_ELT (second_perm_stmts, i, stmt_it) ++ if (gimple_assign_rhs_code (stmt_it) != VEC_PERM_EXPR) ++ return false; ++ ++ /* Check permutation pattern on the second permutation stmts. */ ++ if (!check_or_collect_perm_pattern (second_perm_stmts, perm_pattern, false)) ++ return false; ++ ++ /* Check values of permutation indices. */ ++ auto_vec perm_indices (vector_cst_encoded_nelts (perm_pattern[0]) ++ * perm_pattern.length ()); ++ if (!make_vec_of_indices (perm_pattern, perm_indices)) ++ return false; ++ ++ unsigned val, half_len = perm_indices.length () / 2; ++ FOR_EACH_VEC_ELT (perm_indices, j, val) ++ if (val != (j % 2 ? half_len + j / 2 : j / 2)) ++ return false; ++ ++ /* Check the correspondence of defs in first_perm_stmts and uses in ++ second_perm_stmts. */ ++ tree type1 = TREE_TYPE (gimple_assign_lhs (first_perm_stmts[0])); ++ tree type2 = TREE_TYPE (gimple_assign_lhs (second_perm_stmts[0])); ++ if (type1 != type2) ++ return false; ++ ++ unsigned HOST_WIDE_INT len = TYPE_VECTOR_SUBPARTS (type1).to_constant (); ++ FOR_EACH_VEC_ELT (second_perm_stmts, i, stmt_it) ++ { ++ /* Vectors of first/second perm stmts consist of blocks, each block ++ transposes its own set of input vectors. J corresponds to the number ++ of such block in the vector. */ ++ unsigned j = (i / len) * len; ++ gimple *src_stmt1 = first_perm_stmts[j + (i - j) / 2]; ++ gimple *src_stmt2 = first_perm_stmts[j + (i - j) / 2 + len / 2]; ++ if (gimple_assign_rhs1 (stmt_it) != gimple_assign_lhs (src_stmt1) ++ || gimple_assign_rhs2 (stmt_it) != gimple_assign_lhs (src_stmt2)) ++ return false; ++ } ++ return true; ++} ++ ++/* For the given vector of stmts find all immediate def or use stmts. ++ It uses SSA and don't go trough loads/stores. */ ++ ++static bool ++find_next_stmts (auto_vec &stmts, auto_vec &next_stmts, ++ bool is_forward, bool skip_perms) ++{ ++ unsigned i; ++ gimple *stmt_it; ++ hash_set new_stmt_set; ++ FOR_EACH_VEC_ELT (stmts, i, stmt_it) ++ { ++ if (is_forward) ++ { ++ tree lhs = gimple_assign_lhs (stmt_it); ++ if (!lhs || TREE_CODE (lhs) != SSA_NAME) ++ continue; ++ imm_use_iterator iter; ++ gimple *use_stmt; ++ FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) ++ if (!new_stmt_set.contains (use_stmt)) ++ { ++ new_stmt_set.add (use_stmt); ++ if (!skip_perms ++ || gimple_assign_rhs_code (use_stmt) != VEC_PERM_EXPR) ++ next_stmts.safe_push (use_stmt); ++ } ++ } ++ else ++ { ++ tree rhs; ++ auto_vec rhs_vec (3); ++ if ((rhs = gimple_assign_rhs1 (stmt_it))) ++ rhs_vec.quick_push (rhs); ++ if ((rhs = gimple_assign_rhs2 (stmt_it))) ++ rhs_vec.quick_push (rhs); ++ if ((rhs = gimple_assign_rhs3 (stmt_it))) ++ rhs_vec.quick_push (rhs); ++ unsigned j; ++ FOR_EACH_VEC_ELT (rhs_vec, j, rhs) ++ { ++ if (TREE_CODE (rhs) == VIEW_CONVERT_EXPR) ++ rhs = TREE_OPERAND (rhs, 0); ++ if (TREE_CODE (rhs) != SSA_NAME) ++ continue; ++ gimple *def_stmt = get_prop_source_stmt (rhs, false, NULL); ++ if (!def_stmt) ++ return false; ++ if (new_stmt_set.contains (def_stmt)) ++ continue; ++ new_stmt_set.add (def_stmt); ++ if (!skip_perms ++ || gimple_assign_rhs_code (def_stmt) != VEC_PERM_EXPR) ++ next_stmts.safe_push (def_stmt); ++ } ++ } ++ } ++ return true; ++} ++ ++/* Check if stmts in the vector have similar code and type. Process only ++ assign stmts. */ ++ ++static bool ++check_stmts_similarity (auto_vec &stmts, enum tree_code &code) ++{ ++ code = NOP_EXPR; ++ tree type = NULL_TREE; ++ unsigned i; ++ gimple *stmt_it; ++ FOR_EACH_VEC_ELT (stmts, i, stmt_it) ++ { ++ if (!is_gimple_assign (stmt_it)) ++ return false; ++ tree lhs = gimple_assign_lhs (stmt_it); ++ enum tree_code code2 = gimple_assign_rhs_code (stmt_it); ++ if (type != NULL_TREE) ++ { ++ /* Unpack lo/hi are the same for the analysis. */ ++ if (((code2 != VEC_UNPACK_LO_EXPR && code2 != VEC_UNPACK_HI_EXPR) ++ || (code != VEC_UNPACK_LO_EXPR && code != VEC_UNPACK_HI_EXPR)) ++ && (!lhs || type != TREE_TYPE (lhs) ++ || (code != NOP_EXPR && code != code2))) ++ return false; ++ } ++ else if (lhs) ++ type = TREE_TYPE (lhs); ++ if (code == NOP_EXPR) ++ code = code2; ++ } ++ return true; ++} ++ ++/* Check that the order of definitions of first_stmts and uses of second_stmts ++ is the same. */ ++ ++static bool ++check_def_use_order (vec &first_stmts, vec &second_stmts) ++{ ++ first_stmts.qsort (gimple_uid_cmp); ++ second_stmts.qsort (gimple_uid_cmp); ++ unsigned len1 = first_stmts.length (); ++ unsigned len2 = second_stmts.length (); ++ ++ /* Skip if one of the blocks is empty or the second block is permutaions. */ ++ if (!len1 || !len2 ++ || gimple_assign_rhs_code (second_stmts[0]) == VEC_PERM_EXPR) ++ return true; ++ ++ unsigned i; ++ gimple *stmt_it; ++ FOR_EACH_VEC_ELT (first_stmts, i, stmt_it) ++ { ++ tree op = gimple_assign_lhs (stmt_it); ++ imm_use_iterator iter; ++ gimple *stmt; ++ FOR_EACH_IMM_USE_STMT (stmt, iter, op) ++ { ++ if ((len1 == len2 && stmt != second_stmts[i]) ++ || (len1 == len2 * 2 && stmt != second_stmts[i % len2])) ++ RETURN_FROM_IMM_USE_STMT (iter, false); ++ enum tree_code code = gimple_assign_rhs_code (stmt); ++ if ((len1 * 2 == len2) ++ && ((code == VEC_UNPACK_LO_EXPR && stmt != second_stmts[2 * i]) ++ || (code == VEC_UNPACK_HI_EXPR ++ && stmt != second_stmts[2 * i + 1]))) ++ RETURN_FROM_IMM_USE_STMT (iter, false); ++ } ++ } ++ return true; ++} ++ ++/* Check similarity of stmts in the block of arithmetic operations. */ ++ ++static bool ++check_arithmetic_block (vec &initial_perm_stmts, unsigned nstmts) ++{ ++ auto_vec next_stmts (nstmts); ++ auto_vec prev_stmts (nstmts); ++ ++ enum tree_code code; ++ unsigned i; ++ gimple *stmt_it; ++ FOR_EACH_VEC_ELT (initial_perm_stmts, i, stmt_it) ++ prev_stmts.quick_push (stmt_it); ++ ++ do ++ { ++ next_stmts.block_remove (0, next_stmts.length ()); ++ if (!find_next_stmts (prev_stmts, next_stmts, false, true)) ++ return false; ++ ++ /* Check that types and codes of all stmts in the list are the same. */ ++ if (!check_stmts_similarity (next_stmts, code)) ++ return false; ++ /* Check that the order of all operands is the same. */ ++ if (!check_def_use_order (next_stmts, prev_stmts)) ++ return false; ++ prev_stmts.block_remove (0, prev_stmts.length ()); ++ ++ FOR_EACH_VEC_ELT (next_stmts, i, stmt_it) ++ prev_stmts.safe_push (stmt_it); ++ } ++ while (code != NOP_EXPR); ++ ++ return true; ++} ++ ++/* Find two blocks of permutations on two sets of input vectors which are ++ used in the same vectorized arithmetic operations after the permutaion: ++ Va1...VaN = PERM{P1} (Sa1...SaN) ++ Vb1...VbN = PERM{P1} (Sb1...SbN) ++ Vc1...VcN = binops (Va1...VaN, Vb1...VbN) ++ The goal of the transformation is to execute the block of permutations ++ only once on the result of the arithmetic operations: ++ Va1...VaN = binops (Sa1...SaN, Sb1...SbN) ++ Vc1...VcN = PERM{P1} (Va1...VaN) ++ ++ Currently the analysis looks for transposition permutations that consist ++ of two layers of statements e.g.: ++ Vt1 = PERM { 0, 4, 1, 5 } Sa1, Sa2 // the first ++ Vt2 = PERM { 2, 6, 3, 7 } Sa1, Sa2 ++ Vt3 = PERM { 0, 4, 1, 5 } Sa3, Sa4 ++ Vt4 = PERM { 2, 6, 3, 7 } Sa3, Sa4 ++ Va1 = PERM { 0, 4, 1, 5 } Vt1, Vt3 // the second ++ Va2 = PERM { 2, 6, 3, 7 } Vt1, Vt3 ++ Va3 = PERM { 0, 4, 1, 5 } Vt2, Vt4 ++ Va4 = PERM { 2, 6, 3, 7 } Vt2, Vt4 ++ Permutation stmts are collected in first_perm_stmts and second_perm_stmts ++ vectors correspondinglys. ++ ++ Arithmetic operations may contain several stmts for one pair of input source ++ vectors e.g.: ++ Vtmp1 = unop (Va1) ++ Vtmp2 = binop (Vb1, const) ++ Vc1 = binop (Vtmp1, Vtmp2) ++ The last stmts of each sequence in the arithmetic block are collected ++ in final_arith_stmts. */ ++ ++static bool ++analyze_perm_fwprop (tree type, unsigned HOST_WIDE_INT nelts, ++ vec &stmts, auto_vec &src_vects, ++ auto_vec &perm_pattern, ++ auto_vec &final_arith_stmts, ++ auto_vec &second_perm_stmts) ++{ ++ gcc_checking_assert (stmts.length () == 2); ++ auto_vec first_perm_stmts (nelts * 2); ++ if (!find_perm_set (stmts[0], type, first_perm_stmts, src_vects) ++ || !find_perm_set (stmts[1], type, first_perm_stmts, src_vects)) ++ return false; ++ first_perm_stmts.qsort (gimple_uid_cmp); ++ ++ /* Determine permutation pattern. */ ++ if (!check_perm_pattern (first_perm_stmts, perm_pattern, second_perm_stmts)) ++ return false; ++ ++ /* Find all arithmetic stmts. */ ++ unsigned i; ++ gimple *stmt_it; ++ auto_vec all_arith_stmts (nelts * 2); ++ hash_set visited; ++ FOR_EACH_VEC_ELT (second_perm_stmts, i, stmt_it) ++ { ++ tree dst = gimple_assign_lhs (stmt_it); ++ use_operand_p use_p; ++ gimple *use_stmt; ++ if (!single_imm_use (dst, &use_p, &use_stmt)) ++ return false; ++ all_arith_stmts.quick_push (use_stmt); ++ visited.add (use_stmt); ++ } ++ ++ /* Select final arithmetic stmts. */ ++ FOR_EACH_VEC_ELT (all_arith_stmts, i, stmt_it) ++ { ++ tree dst = gimple_assign_lhs (stmt_it); ++ use_operand_p use_p; ++ imm_use_iterator iter; ++ bool use_only_outside_arith_stmts = true; ++ FOR_EACH_IMM_USE_FAST (use_p, iter, dst) ++ if (visited.contains (USE_STMT (use_p))) ++ { ++ use_only_outside_arith_stmts = false; ++ break; ++ } ++ if (use_only_outside_arith_stmts) ++ final_arith_stmts.quick_push (stmt_it); ++ } ++ ++ /* Check that all results has the same arithmetic patterns. */ ++ if (!check_arithmetic_block (final_arith_stmts, nelts)) ++ return false; ++ ++ if (final_arith_stmts.length () < nelts) ++ return false; ++ return true; ++} ++ ++/* Substitute uses of stmts' results by new_uses. */ ++ ++static void ++substitute_uses (vec &stmts, vec &new_uses) ++{ ++ gcc_checking_assert (stmts.length () == new_uses.length ()); ++ unsigned i; ++ gimple *stmt_it; ++ FOR_EACH_VEC_ELT (stmts, i, stmt_it) ++ { ++ tree op = gimple_assign_lhs (stmt_it); ++ imm_use_iterator iter; ++ gimple *use_stmt; ++ FOR_EACH_IMM_USE_STMT (use_stmt, iter, op) ++ { ++ use_operand_p use_p; ++ FOR_EACH_IMM_USE_ON_STMT (use_p, iter) ++ SET_USE (use_p, new_uses[i]); ++ update_stmt (use_stmt); ++ } ++ } ++} ++ ++/* Propagate permutations through the block of arithmetic operations. */ ++ ++static void ++fwprop_perms (tree type, auto_vec &src_vects, ++ auto_vec &perm_pattern, ++ auto_vec &final_arith_stmts, ++ auto_vec &second_perm_stmts) ++{ ++ /* Build new permutation stmts after the block of arithmetic stmts. */ ++ gimple_seq new_stmts = NULL; ++ unsigned perm_block_size = final_arith_stmts.length (); ++ auto_vec new_first_perm_vals (perm_block_size); ++ hash_set new_stmts_set; ++ unsigned i, perm_pattern_size = perm_pattern.length (); ++ for (i = 0; i < perm_block_size; i++) ++ { ++ tree op0 = gimple_assign_lhs (final_arith_stmts[i / 2]); ++ unsigned idx = i / 2 + perm_block_size / 2; ++ tree op1 = gimple_assign_lhs (final_arith_stmts[idx]); ++ tree res = gimple_build (&new_stmts, VEC_PERM_EXPR, type, op0, op1, ++ perm_pattern[i % perm_pattern_size]); ++ new_first_perm_vals.quick_push (res); ++ new_stmts_set.add (gimple_seq_last (new_stmts)); ++ } ++ auto_vec new_second_perm_vals (perm_block_size); ++ for (i = 0; i < perm_block_size; i++) ++ { ++ tree op0 = new_first_perm_vals[i / 2]; ++ tree op1 = new_first_perm_vals[i / 2 + perm_block_size/ 2]; ++ tree res = gimple_build (&new_stmts, VEC_PERM_EXPR, type, op0, op1, ++ perm_pattern[i % perm_pattern_size]); ++ new_second_perm_vals.quick_push (res); ++ new_stmts_set.add (gimple_seq_last (new_stmts)); ++ } ++ ++ gimple_stmt_iterator g = gsi_for_stmt (final_arith_stmts.last ()); ++ gsi_insert_seq_after (&g, new_stmts, GSI_SAME_STMT); ++ ++ /* Replace old uses of the arithmetic block results by destinations of ++ the new permutation block. */ ++ gimple *stmt_it; ++ FOR_EACH_VEC_ELT (final_arith_stmts, i, stmt_it) ++ { ++ tree op0 = gimple_assign_lhs (final_arith_stmts[i]); ++ imm_use_iterator iter; ++ gimple *use_stmt; ++ use_operand_p use_p; ++ FOR_EACH_IMM_USE_STMT (use_stmt, iter, op0) ++ { ++ if (new_stmts_set.contains (use_stmt)) ++ continue; ++ FOR_EACH_IMM_USE_ON_STMT (use_p, iter) ++ SET_USE (use_p, new_second_perm_vals[i]); ++ update_stmt (use_stmt); ++ } ++ } ++ ++ /* Disconnect the old permutation stmts. */ ++ substitute_uses (second_perm_stmts, src_vects); ++} ++ ++/* Find the permutation stmts in the forward or backward direction (in terms of ++ def/use graph) starting from the vector of initial stmts. Count reduction ++ stmts (i.e. binary operations) if they can change the number of processed ++ elements. */ ++ ++static bool ++find_perm_stmts (vec &initial_stmts, unsigned nstmts, ++ vec &final_perm_stmts, bool is_forward, ++ unsigned &nreduct) ++{ ++ auto_vec next_stmts (nstmts); ++ auto_vec prev_stmts (nstmts); ++ ++ nreduct = 0; ++ enum tree_code code; ++ unsigned i; ++ gimple *stmt_it; ++ FOR_EACH_VEC_ELT (initial_stmts, i, stmt_it) ++ prev_stmts.quick_push (stmt_it); ++ ++ do ++ { ++ next_stmts.block_remove (0, next_stmts.length ()); ++ if (!find_next_stmts (prev_stmts, next_stmts, is_forward, false)) ++ return false; ++ ++ /* Check that types and codes of all stmts in the list are the same. */ ++ if (!check_stmts_similarity (next_stmts, code)) ++ return false; ++ ++ /* TODO: don't take into account binary operations with constants. */ ++ if (TREE_CODE_CLASS (code) == tcc_binary) ++ nreduct += 1; ++ ++ if (is_forward ? !check_def_use_order (prev_stmts, next_stmts) ++ : !check_def_use_order (next_stmts, prev_stmts)) ++ return false; ++ ++ prev_stmts.block_remove (0, prev_stmts.length ()); ++ ++ FOR_EACH_VEC_ELT (next_stmts, i, stmt_it) ++ prev_stmts.safe_push (stmt_it); ++ } ++ while (code != NOP_EXPR && code != VEC_PERM_EXPR); ++ ++ if (code != VEC_PERM_EXPR) ++ return false; ++ ++ FOR_EACH_VEC_ELT (next_stmts, i, stmt_it) ++ final_perm_stmts.safe_push (stmt_it); ++ final_perm_stmts.qsort (gimple_uid_cmp); ++ return true; ++} ++ ++/* Check if the initial and the final permutations can be optimized i.e. ++ the initial permutation can be removed with the modification of ++ the final one. */ ++ ++static bool ++can_reduce_permutations (unsigned init_nelts, vec &perm_pattern, ++ vec &init_perm_stmts) ++{ ++ auto_vec perm_indices (init_nelts); ++ if (!make_vec_of_indices (perm_pattern, perm_indices)) ++ return false; ++ unsigned i, j; ++ gimple *stmt_it; ++ unsigned perm_vec_size = perm_indices.length (); ++ FOR_EACH_VEC_ELT (init_perm_stmts, i, stmt_it) ++ { ++ gcc_assert (gimple_assign_rhs_code (stmt_it) == VEC_PERM_EXPR); ++ tree perm_vec2 = gimple_assign_rhs3 (stmt_it); ++ unsigned HOST_WIDE_INT mask_elts; ++ if (!VECTOR_CST_NELTS (perm_vec2).is_constant (&mask_elts)) ++ return false; ++ for (j = 0; j < mask_elts; j++) ++ { ++ tree val = VECTOR_CST_ELT (perm_vec2, j); ++ gcc_assert (TREE_CODE (val) == INTEGER_CST); ++ unsigned HOST_WIDE_INT int_val = TREE_INT_CST_LOW (val); ++ if (int_val != perm_indices[j % perm_vec_size] ++ + (j / perm_vec_size) * perm_vec_size) ++ return false; ++ } ++ } ++ return true; ++} ++ ++/* Find permutation blocks before and after arithmetic operations and decide ++ if the number of permutations can be reduced, e.g: ++ Va1...VaN = PERM{P1} (Sa1...SaN) ++ Vb1...VbM = some operations (Va1...VaN) ++ Vb1...VbM = PERM{P2} (Sb1...SbM) ++ can be transformed to: ++ Vb1...VbM = some operations (Va1...VaN) ++ Vb1...VbM = PERM{P3} (Sb1...SbM) ++ ++ Currently it supports initial permutations like this: ++ Va1 = PERM { 0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15} Sa1 ++ and transposition permutations with two layers of permutation stmts as ++ final permutaions. ++ ++ Operations between permutations can include unary and binary arithmetic, ++ element conversions and vector packing/unpacking. */ ++ ++static bool ++analyze_perm_reduction (unsigned HOST_WIDE_INT nelts, ++ vec &perm_stmts, ++ vec &init_perm_stmts, ++ vec &second_perm_stmts) ++{ ++ auto_vec first_perm_stmts (nelts * 2); ++ if (!check_perm_stmt (perm_stmts[0], NULL_TREE, &first_perm_stmts, NULL) ++ || !check_perm_stmt (perm_stmts[1], NULL_TREE, &first_perm_stmts, NULL)) ++ return false; ++ ++ unsigned nreduct; ++ auto_vec final_perm_stmts (nelts * 2); ++ if (!find_perm_stmts (first_perm_stmts, nelts, final_perm_stmts, true, ++ nreduct)) ++ return false; ++ ++ if (!find_perm_stmts (final_perm_stmts, nelts, init_perm_stmts, false, ++ nreduct)) ++ return false; ++ ++ /* Check number of elemetns in the inital and final data block. */ ++ tree init_elem_type = TREE_TYPE (gimple_assign_lhs (init_perm_stmts[0])); ++ unsigned init_nelts = TYPE_VECTOR_SUBPARTS (init_elem_type).to_constant () ++ * init_perm_stmts.length (); ++ tree final_elem_type = TREE_TYPE (gimple_assign_lhs (final_perm_stmts[0])); ++ unsigned final_nelts = TYPE_VECTOR_SUBPARTS (final_elem_type).to_constant () ++ * final_perm_stmts.length (); ++ if (init_nelts != final_nelts * (1 + nreduct)) ++ return false; ++ ++ /* Check the final permutations and detect its pattern. */ ++ auto_vec perm_pattern (nelts); ++ if (!check_perm_pattern (final_perm_stmts, perm_pattern, second_perm_stmts)) ++ return false; ++ ++ return can_reduce_permutations (init_nelts, perm_pattern, init_perm_stmts); ++} ++ ++/* Do the optimization: skip the initial permutation and change the order ++ of destinations after the second layer of permutation statements in ++ the final permutation block. */ ++ ++static void ++reduce_perms (vec &init_perm_stmts, vec &second_perm_stmts) ++{ ++ unsigned i; ++ gimple *stmt_it; ++ auto_vec new_srcs (init_perm_stmts.length ()); ++ FOR_EACH_VEC_ELT (init_perm_stmts, i, stmt_it) ++ new_srcs.quick_push (gimple_assign_rhs1 (stmt_it)); ++ substitute_uses (init_perm_stmts, new_srcs); ++ ++ unsigned half = second_perm_stmts.length () / 2; ++ auto_vec new_dsts (second_perm_stmts.length ()); ++ FOR_EACH_VEC_ELT (second_perm_stmts, i, stmt_it) ++ { ++ unsigned idx = i < half ? i << 1 : ((i - half) << 1) + 1; ++ new_dsts.quick_push (gimple_assign_lhs (second_perm_stmts[idx])); ++ } ++ ++ FOR_EACH_VEC_ELT (second_perm_stmts, i, stmt_it) ++ { ++ gimple_assign_set_lhs (stmt_it, new_dsts[i]); ++ update_stmt (stmt_it); ++ } ++} ++ ++/* Optimize permutations in the following two cases: ++ 1. Recognize the same permutations of two sets of vectors with subsequent ++ binary arithmetic operations on them: ++ V1 = PERM{1} (S1); ++ V2 = PERM{1} (S2); ++ V3 = V1 binop V2; ++ then move the permutation after the operations: ++ V0 = S1 binop S2; ++ V3 = PERM{1} V0; ++ 2. Detect the first permutation before some operations on a set of vectors ++ and the second one after the operations: ++ V1 = PERM{1} (S1) ++ V2 = set of operations (V1) ++ V3 = PERM{2} (V2) ++ try to reduce them: ++ V2 = set of operations (S1) ++ V3 = PERM{3} (V2) ++ Return true if the optimization is successful. */ ++ ++static bool ++propagate_permutations (gimple_stmt_iterator *gsi) ++{ ++ tree type; ++ auto_vec perm_stmts (2); ++ ++ if (!find_initial_permutations (gsi, type, perm_stmts)) ++ return false; ++ ++ unsigned HOST_WIDE_INT nelts = TYPE_VECTOR_SUBPARTS (type).to_constant (); ++ auto_vec final_arith_stmts (nelts * 2); ++ auto_vec second_perm_stmts (nelts * 2); ++ auto_vec src_vects (nelts * 2); ++ auto_vec perm_pattern (nelts); ++ if (analyze_perm_fwprop (type, nelts, perm_stmts, src_vects, perm_pattern, ++ final_arith_stmts, second_perm_stmts)) ++ { ++ fwprop_perms (type, src_vects, perm_pattern, final_arith_stmts, ++ second_perm_stmts); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ unsigned i; ++ gimple *stmt_it; ++ fprintf (dump_file, "Permutations were moved through " ++ "binary operations:\n"); ++ FOR_EACH_VEC_ELT (second_perm_stmts, i, stmt_it) ++ print_gimple_stmt (dump_file, stmt_it, 0); ++ } ++ return true; ++ } ++ ++ auto_vec init_perm_stmts (nelts * 2); ++ auto_vec final_perm_stmts (nelts * 2); ++ if (analyze_perm_reduction (nelts, perm_stmts, init_perm_stmts, ++ final_perm_stmts)) ++ { ++ reduce_perms (init_perm_stmts, final_perm_stmts); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ unsigned i; ++ gimple *stmt_it; ++ fprintf (dump_file, "Initial permutations were reduced:\n"); ++ FOR_EACH_VEC_ELT (init_perm_stmts, i, stmt_it) ++ print_gimple_stmt (dump_file, stmt_it, 0); ++ } ++ return true; ++ } ++ return false; ++} ++ + /* Get the BIT_FIELD_REF definition of VAL, if any, looking through + conversions with code CONV_CODE or update it if still ERROR_MARK. + Return NULL_TREE if no such matching def was found. */ +@@ -3155,6 +4042,10 @@ pass_forwprop::execute (function *fun) + || code == BIT_XOR_EXPR) + && simplify_rotate (&gsi)) + changed = true; ++ else if ((code == PLUS_EXPR || code == MINUS_EXPR) ++ && param_tree_forwprop_perm ++ && propagate_permutations (&gsi)) ++ changed = true; + else if (code == VEC_PERM_EXPR) + { + int did_something = simplify_permutation (&gsi); +-- +2.33.0 + diff --git a/0151-Fix-bugs-and-add-tests-for-RTL-ifcvt.patch b/0151-Fix-bugs-and-add-tests-for-RTL-ifcvt.patch new file mode 100644 index 0000000..b228e2e --- /dev/null +++ b/0151-Fix-bugs-and-add-tests-for-RTL-ifcvt.patch @@ -0,0 +1,381 @@ +From 4bcb19923cdcb042d66057766d661ef68bf70e92 Mon Sep 17 00:00:00 2001 +From: Chernonog Vyacheslav 00812786 +Date: Wed, 29 Mar 2023 05:22:17 +0300 +Subject: [PATCH 13/13] Fix bugs and add tests for RTL ifcvt + +1. Fix bug in rtl ifcvt that run pass despite renaming failure. +2. Fix bug that prevent final set register to be renamed. +3. Clean up dominance info before runnig cleanup_cfg to avoid fixup + invalid dominance info. +4. Remove duplicated cleanup_cfg. +5. Add tests. +--- + gcc/common.opt | 4 + + gcc/ifcvt.c | 88 ++++++++++++------- + gcc/params.opt | 4 - + .../gcc.c-torture/execute/ifcvt-renaming-1.c | 38 ++++++++ + gcc/testsuite/gcc.dg/ifcvt-6.c | 29 ++++++ + 5 files changed, 128 insertions(+), 35 deletions(-) + create mode 100644 gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c + create mode 100644 gcc/testsuite/gcc.dg/ifcvt-6.c + +diff --git a/gcc/common.opt b/gcc/common.opt +index 6f0ed7cea..92d3a1986 100644 +--- a/gcc/common.opt ++++ b/gcc/common.opt +@@ -3534,4 +3534,8 @@ fipa-ra + Common Report Var(flag_ipa_ra) Optimization + Use caller save register across calls if possible. + ++fifcvt-allow-complicated-cmps ++Common Report Var(flag_ifcvt_allow_complicated_cmps) Optimization ++Allow RTL if-conversion pass to deal with complicated cmps (can increase compilation time). ++ + ; This comment is to ensure we retain the blank line above. +diff --git a/gcc/ifcvt.c b/gcc/ifcvt.c +index 50a73a7ca..209987ebc 100644 +--- a/gcc/ifcvt.c ++++ b/gcc/ifcvt.c +@@ -876,7 +876,7 @@ noce_emit_store_flag (struct noce_if_info *if_info, rtx x, int reversep, + } + + /* Don't even try if the comparison operands or the mode of X are weird. */ +- if (!param_ifcvt_allow_complicated_cmps ++ if (!flag_ifcvt_allow_complicated_cmps + && (cond_complex + || !SCALAR_INT_MODE_P (GET_MODE (x)))) + return NULL_RTX; +@@ -1745,7 +1745,7 @@ noce_emit_cmove (struct noce_if_info *if_info, rtx x, enum rtx_code code, + + /* Don't even try if the comparison operands are weird + except that the target supports cbranchcc4. */ +- if (! param_ifcvt_allow_complicated_cmps ++ if (! flag_ifcvt_allow_complicated_cmps + && (! general_operand (cmp_a, GET_MODE (cmp_a)) + || ! general_operand (cmp_b, GET_MODE (cmp_b)))) + { +@@ -1918,6 +1918,19 @@ noce_try_cmove (struct noce_if_info *if_info) + return FALSE; + } + ++/* Return true if X contains a conditional code mode rtx. */ ++ ++static bool ++contains_ccmode_rtx_p (rtx x) ++{ ++ subrtx_iterator::array_type array; ++ FOR_EACH_SUBRTX (iter, array, x, ALL) ++ if (GET_MODE_CLASS (GET_MODE (*iter)) == MODE_CC) ++ return true; ++ ++ return false; ++} ++ + /* Helper for bb_valid_for_noce_process_p. Validate that + the rtx insn INSN is a single set that does not set + the conditional register CC and is in general valid for +@@ -1936,6 +1949,8 @@ insn_valid_noce_process_p (rtx_insn *insn, rtx cc) + /* Currently support only simple single sets in test_bb. */ + if (!sset + || !noce_operand_ok (SET_DEST (sset)) ++ || (!flag_ifcvt_allow_complicated_cmps ++ && contains_ccmode_rtx_p (SET_DEST (sset))) + || !noce_operand_ok (SET_SRC (sset))) + return false; + +@@ -1974,8 +1989,7 @@ bbs_ok_for_cmove_arith (basic_block bb_a, + continue; + /* Record all registers that BB_A sets. */ + FOR_EACH_INSN_DEF (def, a_insn) +- if (!(to_rename && DF_REF_REG (def) == to_rename)) +- bitmap_set_bit (bba_sets, DF_REF_REGNO (def)); ++ bitmap_set_bit (bba_sets, DF_REF_REGNO (def)); + } + + bitmap_and (intersections, df_get_live_in (bb_b), bba_sets); +@@ -1984,6 +1998,7 @@ bbs_ok_for_cmove_arith (basic_block bb_a, + { + if (!active_insn_p (b_insn)) + continue; ++ + rtx sset_b = single_set (b_insn); + + if (!sset_b) +@@ -2081,7 +2096,12 @@ noce_emit_bb (rtx last_insn, basic_block bb, bool simple) + return true; + } + +-/* This function tries to rename regs that intersect with considered bb. */ ++/* This function tries to rename regs that intersect with considered bb ++ inside condition expression. Condition expression will be moved down ++ if the optimization will be applied, so it is essential to be sure that ++ all intersected registers will be renamed otherwise transformation ++ can't be applied. Function returns true if renaming was successful ++ and optimization can proceed futher. */ + + static bool + noce_rename_regs_in_cond (struct noce_if_info *if_info, bitmap cond_rename_regs) +@@ -2092,11 +2112,11 @@ noce_rename_regs_in_cond (struct noce_if_info *if_info, bitmap cond_rename_regs) + if (param_ifcvt_allow_register_renaming < 2) + return false; + df_ref use; +- rtx_insn* cmp_insn = if_info->cond_earliest; ++ rtx_insn *cmp_insn = if_info->cond_earliest; + /* Jump instruction as a condion currently unsupported. */ + if (JUMP_P (cmp_insn)) + return false; +- rtx_insn* before_cmp = PREV_INSN (cmp_insn); ++ rtx_insn *before_cmp = PREV_INSN (cmp_insn); + start_sequence (); + rtx_insn *copy_of_cmp = as_a (copy_rtx (cmp_insn)); + basic_block cmp_block = BLOCK_FOR_INSN (cmp_insn); +@@ -2122,7 +2142,7 @@ noce_rename_regs_in_cond (struct noce_if_info *if_info, bitmap cond_rename_regs) + + emit_insn_after_setloc (seq, before_cmp, INSN_LOCATION (cmp_insn)); + delete_insn_and_edges (cmp_insn); +- rtx_insn* insn; ++ rtx_insn *insn; + FOR_BB_INSNS (cmp_block, insn) + df_insn_rescan (insn); + +@@ -2135,13 +2155,15 @@ noce_rename_regs_in_cond (struct noce_if_info *if_info, bitmap cond_rename_regs) + return success; + } + +-/* This function tries to rename regs that intersect with considered bb. */ ++/* This function tries to rename regs that intersect with considered bb. ++ return true if the renaming was successful and optimization can ++ proceed futher, false otherwise. */ + static bool + noce_rename_regs_in_bb (basic_block test_bb, bitmap rename_regs) + { + if (bitmap_empty_p (rename_regs)) + return true; +- rtx_insn* insn; ++ rtx_insn *insn; + rtx_insn *last_insn = last_active_insn (test_bb, FALSE); + bool res = true; + start_sequence (); +@@ -2153,7 +2175,7 @@ noce_rename_regs_in_bb (basic_block test_bb, bitmap rename_regs) + rtx sset = single_set (insn); + gcc_assert (sset); + rtx x = SET_DEST (sset); +- if (!REG_P (x) || bitmap_bit_p (rename_regs, REGNO (x))) ++ if (!REG_P (x) || !bitmap_bit_p (rename_regs, REGNO (x))) + continue; + + machine_mode mode = GET_MODE (x); +@@ -2175,7 +2197,7 @@ noce_rename_regs_in_bb (basic_block test_bb, bitmap rename_regs) + noce_emit_move_insn (tmp,x); + } + set_used_flags (insn); +- rtx_insn* rename_candidate; ++ rtx_insn *rename_candidate; + for (rename_candidate = NEXT_INSN (insn); + rename_candidate && rename_candidate!= NEXT_INSN (BB_END (test_bb)); + rename_candidate = NEXT_INSN (rename_candidate)) +@@ -2193,17 +2215,16 @@ noce_rename_regs_in_bb (basic_block test_bb, bitmap rename_regs) + replace_res = validate_replace_rtx (x, tmp, rename_candidate); + gcc_assert (replace_res); + set_used_flags (rename_candidate); +- + } + set_used_flags (x); + set_used_flags (tmp); +- + } +- rtx_insn *seq = get_insns (); +- unshare_all_rtl_in_chain (seq); +- end_sequence (); +- emit_insn_before_setloc (seq, first_active_insn (test_bb), +- INSN_LOCATION (first_active_insn (test_bb))); ++ ++ rtx_insn *seq = get_insns (); ++ unshare_all_rtl_in_chain (seq); ++ end_sequence (); ++ emit_insn_before_setloc (seq, first_active_insn (test_bb), ++ INSN_LOCATION (first_active_insn (test_bb))); + FOR_BB_INSNS (test_bb, insn) + df_insn_rescan (insn); + return res; +@@ -2305,9 +2326,10 @@ noce_try_cmove_arith (struct noce_if_info *if_info) + BITMAP_FREE (else_bb_rename_regs); + return FALSE; + } +- bool prepass_renaming = true; +- prepass_renaming |= noce_rename_regs_in_bb (then_bb, then_bb_rename_regs); +- prepass_renaming |= noce_rename_regs_in_bb (else_bb, else_bb_rename_regs); ++ bool prepass_renaming = noce_rename_regs_in_bb (then_bb, ++ then_bb_rename_regs) ++ && noce_rename_regs_in_bb (else_bb, ++ else_bb_rename_regs); + + BITMAP_FREE (then_bb_rename_regs); + BITMAP_FREE (else_bb_rename_regs); +@@ -2321,6 +2343,7 @@ noce_try_cmove_arith (struct noce_if_info *if_info) + came from the test block. The non-empty complex block that we will + emit might clobber the register used by B or A, so move it to a pseudo + first. */ ++ + rtx tmp_a = NULL_RTX; + rtx tmp_b = NULL_RTX; + +@@ -3233,6 +3256,7 @@ bb_valid_for_noce_process_p (basic_block test_bb, rtx cond, + && reg_set_between_p (x, first_insn, prev_last_insn) + && param_ifcvt_allow_register_renaming < 1) + return false; ++ + bitmap test_bb_temps = BITMAP_ALLOC (®_obstack); + + /* The regs that are live out of test_bb. */ +@@ -3268,9 +3292,10 @@ bb_valid_for_noce_process_p (basic_block test_bb, rtx cond, + else + goto free_bitmap_and_fail; + } +- potential_cost += pattern_cost (sset, speed_p); +- if (SET_DEST (sset) != SET_DEST (last_set)) +- bitmap_set_bit (test_bb_temps, REGNO (SET_DEST (sset))); ++ ++ potential_cost += pattern_cost (sset, speed_p); ++ if (SET_DEST (sset) != SET_DEST (last_set)) ++ bitmap_set_bit (test_bb_temps, REGNO (SET_DEST (sset))); + } + + /* If any of the intermediate results in test_bb are live after test_bb +@@ -3645,11 +3670,12 @@ noce_process_if_block (struct noce_if_info *if_info) + } + + if (!noce_rename_regs_in_cond (if_info, cond_rename_regs)) +- return false; +- cond = if_info->cond; +- ++ { ++ BITMAP_FREE (cond_rename_regs); ++ return false; ++ } + BITMAP_FREE (cond_rename_regs); +- ++ cond = if_info->cond; + if (speed_p) + if_info->original_cost += average_cost (then_cost, else_cost, + find_edge (test_bb, then_bb)); +@@ -5592,12 +5618,13 @@ if_convert (bool after_combine) + { + basic_block bb; + int pass; +- cleanup_cfg (CLEANUP_EXPENSIVE); ++ + if (optimize == 1) + { + df_live_add_problem (); + df_live_set_all_dirty (); + } ++ cleanup_cfg (CLEANUP_EXPENSIVE); + + /* Record whether we are after combine pass. */ + ifcvt_after_combine = after_combine; +@@ -5702,7 +5729,6 @@ rest_of_handle_if_conversion (void) + dump_reg_info (dump_file); + dump_flow_info (dump_file, dump_flags); + } +- cleanup_cfg (CLEANUP_EXPENSIVE); + if_convert (false); + if (num_updated_if_blocks) + /* Get rid of any dead CC-related instructions. */ +diff --git a/gcc/params.opt b/gcc/params.opt +index 345f9b3ff..272a0eb2b 100644 +--- a/gcc/params.opt ++++ b/gcc/params.opt +@@ -574,10 +574,6 @@ Maximum permissible cost for the sequence that would be generated by the RTL if- + Common Joined UInteger Var(param_max_rtl_if_conversion_unpredictable_cost) Init(40) IntegerRange(0, 200) Param Optimization + Maximum permissible cost for the sequence that would be generated by the RTL if-conversion pass for a branch that is considered unpredictable. + +--param=ifcvt-allow-complicated-cmps= +-Common Joined UInteger Var(param_ifcvt_allow_complicated_cmps) IntegerRange(0, 1) Param Optimization +-Allow RTL if-conversion pass to deal with complicated cmps (can increase compilation time). +- + -param=ifcvt-allow-register-renaming= + Common Joined UInteger Var(param_ifcvt_allow_register_renaming) IntegerRange(0, 2) Param Optimization + Allow RTL if-conversion pass to aggressively rename registers in basic blocks. Sometimes additional moves will be created. +diff --git a/gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c b/gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c +new file mode 100644 +index 000000000..761c8ab7e +--- /dev/null ++++ b/gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c +@@ -0,0 +1,38 @@ ++ ++extern void abort(void); ++ ++__attribute__ ((noinline)) ++int foo (int x, int y, int z, int a, int b) ++{ ++ if (a < 2) ++ { ++ if (a == 0) ++ { ++ if (x - y < 0) ++ x = x - y + z; ++ else ++ x = x - y; ++ } ++ else ++ { ++ if (x + y >= z) ++ x = x + y - z; ++ else ++ x = x + y; ++ } ++ } ++ return x; ++} ++ ++int main(void) ++{ ++ if (foo (5,10,7,0,1) != 2) // x - y + z = -5 + 7 = 2 ++ abort (); ++ if (foo (50,10,7,0,1) != 40) // x - y = 40 ++ abort (); ++ if (foo (5,10,7,1,1) != 8) // x + y - z = 5 + 10 - 7 = 8 ++ abort (); ++ if (foo (5,10,70,1,1) != 15) // x + y = 15 ++ abort (); ++ return 0; ++} +diff --git a/gcc/testsuite/gcc.dg/ifcvt-6.c b/gcc/testsuite/gcc.dg/ifcvt-6.c +new file mode 100644 +index 000000000..7d2a8d58b +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/ifcvt-6.c +@@ -0,0 +1,29 @@ ++/* { dg-do compile { target { aarch64*-*-* } } } */ ++/* { dg-options "-fdump-rtl-ce1 -O2 -fifcvt-allow-complicated-cmps --param max-rtl-if-conversion-unpredictable-cost=100 --param max-rtl-if-conversion-predictable-cost=100 --param=ifcvt-allow-register-renaming=2 " } */ ++ ++typedef unsigned int uint16_t; ++ ++uint16_t ++foo (uint16_t x, uint16_t y, uint16_t z, uint16_t a, ++ uint16_t b, uint16_t c, uint16_t d) ++{ ++ int i = 1; ++ int j = 1; ++ if (a > b) ++ { ++ j = x; ++ if (b > c) ++ i = y; ++ else ++ i = z; ++ } ++ else ++ { ++ j = y; ++ if (c > d) ++ i = z; ++ } ++ return i * j; ++} ++ ++/* { dg-final { scan-rtl-dump "7 true changes made" "ce1" } } */ +-- +2.33.0 + diff --git a/0152-Add-LLC-Allocation-Pass.patch b/0152-Add-LLC-Allocation-Pass.patch new file mode 100644 index 0000000..0a19214 --- /dev/null +++ b/0152-Add-LLC-Allocation-Pass.patch @@ -0,0 +1,4905 @@ +From e0e139bf642398d1e1b8cfd803ee6ce276404991 Mon Sep 17 00:00:00 2001 +From: huangxiaoquan +Date: Wed, 6 Dec 2023 17:51:11 +0800 +Subject: [PATCH] Add LLC-Allocation Pass LLC allocation allows the compiler to + identify frequently-used data in the program and strengthens the ability to + prefetch and distribute it to the last level cache (LLC) through memory + accesses of the corresponding data variables. Add flag -fllc-allocate to + enable LLC allocation. + +--- + gcc/Makefile.in | 1 + + gcc/cfgloop.h | 3 + + gcc/common.opt | 4 + + gcc/config/aarch64/aarch64-sve.md | 48 +- + gcc/config/aarch64/aarch64.c | 18 + + gcc/doc/tm.texi | 21 + + gcc/doc/tm.texi.in | 6 + + gcc/internal-fn.c | 115 + + gcc/internal-fn.def | 4 + + gcc/optabs.def | 2 + + gcc/params.opt | 53 + + gcc/passes.def | 1 + + gcc/target.def | 31 + + gcc/testsuite/gcc.dg/llc-allocate/llc-1.c | 61 + + gcc/testsuite/gcc.dg/llc-allocate/llc-2.c | 54 + + .../gcc.dg/llc-allocate/llc-allocate.exp | 27 + + .../llc-allocate/llc-issue-builtin-prefetch.c | 48 + + .../gcc.dg/llc-allocate/llc-nonzero-offset.c | 50 + + .../gcc.dg/llc-allocate/llc-ref-trace.c | 62 + + .../llc-allocate/llc-tool-insertion-1.c | 48 + + .../llc-allocate/llc-tool-insertion-2.c | 48 + + .../llc-allocate/llc-tool-insertion-3.c | 48 + + .../llc-allocate/llc-tool-insertion-4.c | 47 + + .../llc-allocate/llc-tool-insertion-5.c | 48 + + .../llc-allocate/llc-tool-insertion-6.c | 47 + + .../llc-tool-insertion-7-null-var-name.c | 52 + + .../llc-tool-insertion-8-tmp-var-name.c | 54 + + .../gfortran.dg/llc-allocate/llc-3.f90 | 213 ++ + .../gfortran.dg/llc-allocate/llc-allocate.exp | 29 + + .../llc-trace-multiple-base-var.f90 | 63 + + .../llc-unknown-type-size-unit.f90 | 58 + + gcc/timevar.def | 1 + + gcc/tree-cfg.c | 11 + + gcc/tree-cfg.h | 1 + + gcc/tree-pass.h | 1 + + gcc/tree-scalar-evolution.c | 8 +- + gcc/tree-scalar-evolution.h | 3 +- + gcc/tree-ssa-llc-allocate.c | 2898 +++++++++++++++++ + gcc/tree-ssa-loop-niter.c | 38 +- + gcc/tree-ssa-loop-niter.h | 3 +- + 40 files changed, 4297 insertions(+), 31 deletions(-) + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-1.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-2.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-nonzero-offset.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-ref-trace.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-2.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-3.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-4.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-5.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-6.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-7-null-var-name.c + create mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c + create mode 100644 gcc/testsuite/gfortran.dg/llc-allocate/llc-3.f90 + create mode 100644 gcc/testsuite/gfortran.dg/llc-allocate/llc-allocate.exp + create mode 100644 gcc/testsuite/gfortran.dg/llc-allocate/llc-trace-multiple-base-var.f90 + create mode 100644 gcc/testsuite/gfortran.dg/llc-allocate/llc-unknown-type-size-unit.f90 + create mode 100644 gcc/tree-ssa-llc-allocate.c + +diff --git a/gcc/Makefile.in b/gcc/Makefile.in +index 2a59acfbe..31bf2cde2 100644 +--- a/gcc/Makefile.in ++++ b/gcc/Makefile.in +@@ -1594,6 +1594,7 @@ OBJS = \ + tree-ssa-loop-array-widen-compare.o \ + tree-ssa-loop-crc.o \ + tree-ssa-loop-prefetch.o \ ++ tree-ssa-llc-allocate.o \ + tree-ssa-loop-split.o \ + tree-ssa-loop-unswitch.o \ + tree-ssa-loop.o \ +diff --git a/gcc/cfgloop.h b/gcc/cfgloop.h +index 18b404e29..e3ecf5076 100644 +--- a/gcc/cfgloop.h ++++ b/gcc/cfgloop.h +@@ -272,6 +272,9 @@ public: + the basic-block from being collected but its index can still be + reused. */ + basic_block former_header; ++ ++ /* Number of latch executions from vectorization. */ ++ tree vec_nb_iterations; + }; + + /* Set if the loop is known to be infinite. */ +diff --git a/gcc/common.opt b/gcc/common.opt +index 4db061b44..2dde0f673 100644 +--- a/gcc/common.opt ++++ b/gcc/common.opt +@@ -2233,6 +2233,10 @@ Common Joined RejectNegative UInteger Var(prefetch_level) Init(0) IntegerRange(0 + Generate prefetch instructions, if available, for arrays in loops. The prefetch + level can control the optimize level to array prefetch. + ++fllc-allocate ++Common Report Var(flag_llc_allocate) Init(-1) Optimization ++Generate LLC hint instructions. ++ + fprofile + Common Report Var(profile_flag) + Enable basic program profiling code. +diff --git a/gcc/config/aarch64/aarch64-sve.md b/gcc/config/aarch64/aarch64-sve.md +index d17a77706..c5b99b6c4 100644 +--- a/gcc/config/aarch64/aarch64-sve.md ++++ b/gcc/config/aarch64/aarch64-sve.md +@@ -1940,7 +1940,7 @@ + (define_insn "@aarch64_sve_prefetch" + [(prefetch (unspec:DI + [(match_operand: 0 "register_operand" "Upl") +- (match_operand:SVE_FULL_I 1 "aarch64_sve_prefetch_operand" "UP") ++ (match_operand:SVE_FULL 1 "aarch64_sve_prefetch_operand" "UP") + (match_operand:DI 2 "const_int_operand")] + UNSPEC_SVE_PREFETCH) + (match_operand:DI 3 "const_int_operand") +@@ -1973,14 +1973,14 @@ + ;; 6: the prefetch operator (an svprfop) + ;; 7: the normal RTL prefetch rw flag + ;; 8: the normal RTL prefetch locality value +-(define_insn "@aarch64_sve_gather_prefetch" ++(define_insn "@aarch64_sve_gather_prefetch" + [(prefetch (unspec:DI + [(match_operand:VNx4BI 0 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl") +- (match_operand:DI 1 "aarch64_sve_gather_offset_" "Z, vg, rk, rk, rk, rk") ++ (match_operand:DI 1 "aarch64_sve_gather_offset_" "Z, vg, rk, rk, rk, rk") + (match_operand:VNx4SI_ONLY 2 "register_operand" "w, w, w, w, w, w") + (match_operand:DI 3 "const_int_operand" "i, i, Z, Ui1, Z, Ui1") +- (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, Ui1, Ui1, Ui1, i, i") +- (match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero") ++ (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, Ui1, Ui1, Ui1, i, i") ++ (match_operand:SVE_FULL 5 "aarch64_simd_imm_zero") + (match_operand:DI 6 "const_int_operand")] + UNSPEC_SVE_PREFETCH_GATHER) + (match_operand:DI 7 "const_int_operand") +@@ -1988,12 +1988,12 @@ + "TARGET_SVE" + { + static const char *const insns[][2] = { +- "prf", "%0, [%2.s]", +- "prf", "%0, [%2.s, #%1]", ++ "prf", "%0, [%2.s]", ++ "prf", "%0, [%2.s, #%1]", + "prfb", "%0, [%1, %2.s, sxtw]", + "prfb", "%0, [%1, %2.s, uxtw]", +- "prf", "%0, [%1, %2.s, sxtw %p4]", +- "prf", "%0, [%1, %2.s, uxtw %p4]" ++ "prf", "%0, [%1, %2.s, sxtw %p4]", ++ "prf", "%0, [%1, %2.s, uxtw %p4]" + }; + const char *const *parts = insns[which_alternative]; + return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]); +@@ -2002,14 +2002,14 @@ + + ;; Predicated gather prefetches for 64-bit elements. The value of operand 3 + ;; doesn't matter in this case. +-(define_insn "@aarch64_sve_gather_prefetch" ++(define_insn "@aarch64_sve_gather_prefetch" + [(prefetch (unspec:DI + [(match_operand:VNx2BI 0 "register_operand" "Upl, Upl, Upl, Upl") +- (match_operand:DI 1 "aarch64_sve_gather_offset_" "Z, vg, rk, rk") ++ (match_operand:DI 1 "aarch64_sve_gather_offset_" "Z, vg, rk, rk") + (match_operand:VNx2DI_ONLY 2 "register_operand" "w, w, w, w") + (match_operand:DI 3 "const_int_operand") +- (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, Ui1, Ui1, i") +- (match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero") ++ (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, Ui1, Ui1, i") ++ (match_operand:SVE_FULL 5 "aarch64_simd_imm_zero") + (match_operand:DI 6 "const_int_operand")] + UNSPEC_SVE_PREFETCH_GATHER) + (match_operand:DI 7 "const_int_operand") +@@ -2017,10 +2017,10 @@ + "TARGET_SVE" + { + static const char *const insns[][2] = { +- "prf", "%0, [%2.d]", +- "prf", "%0, [%2.d, #%1]", ++ "prf", "%0, [%2.d]", ++ "prf", "%0, [%2.d, #%1]", + "prfb", "%0, [%1, %2.d]", +- "prf", "%0, [%1, %2.d, lsl %p4]" ++ "prf", "%0, [%1, %2.d, lsl %p4]" + }; + const char *const *parts = insns[which_alternative]; + return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]); +@@ -2028,7 +2028,7 @@ + ) + + ;; Likewise, but with the offset being sign-extended from 32 bits. +-(define_insn_and_rewrite "*aarch64_sve_gather_prefetch_sxtw" ++(define_insn_and_rewrite "*aarch64_sve_gather_prefetch_sxtw" + [(prefetch (unspec:DI + [(match_operand:VNx2BI 0 "register_operand" "Upl, Upl") + (match_operand:DI 1 "register_operand" "rk, rk") +@@ -2039,8 +2039,8 @@ + (match_operand:VNx2DI 2 "register_operand" "w, w")))] + UNSPEC_PRED_X) + (match_operand:DI 3 "const_int_operand") +- (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, i") +- (match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero") ++ (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, i") ++ (match_operand:SVE_FULL 5 "aarch64_simd_imm_zero") + (match_operand:DI 6 "const_int_operand")] + UNSPEC_SVE_PREFETCH_GATHER) + (match_operand:DI 7 "const_int_operand") +@@ -2049,7 +2049,7 @@ + { + static const char *const insns[][2] = { + "prfb", "%0, [%1, %2.d, sxtw]", +- "prf", "%0, [%1, %2.d, sxtw %p4]" ++ "prf", "%0, [%1, %2.d, sxtw %p4]" + }; + const char *const *parts = insns[which_alternative]; + return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]); +@@ -2061,7 +2061,7 @@ + ) + + ;; Likewise, but with the offset being zero-extended from 32 bits. +-(define_insn "*aarch64_sve_gather_prefetch_uxtw" ++(define_insn "*aarch64_sve_gather_prefetch_uxtw" + [(prefetch (unspec:DI + [(match_operand:VNx2BI 0 "register_operand" "Upl, Upl") + (match_operand:DI 1 "register_operand" "rk, rk") +@@ -2069,8 +2069,8 @@ + (match_operand:VNx2DI 2 "register_operand" "w, w") + (match_operand:VNx2DI 9 "aarch64_sve_uxtw_immediate")) + (match_operand:DI 3 "const_int_operand") +- (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, i") +- (match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero") ++ (match_operand:DI 4 "aarch64_gather_scale_operand_" "Ui1, i") ++ (match_operand:SVE_FULL 5 "aarch64_simd_imm_zero") + (match_operand:DI 6 "const_int_operand")] + UNSPEC_SVE_PREFETCH_GATHER) + (match_operand:DI 7 "const_int_operand") +@@ -2079,7 +2079,7 @@ + { + static const char *const insns[][2] = { + "prfb", "%0, [%1, %2.d, uxtw]", +- "prf", "%0, [%1, %2.d, uxtw %p4]" ++ "prf", "%0, [%1, %2.d, uxtw %p4]" + }; + const char *const *parts = insns[which_alternative]; + return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]); +diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c +index dbdc6dffb..aa077ec0a 100644 +--- a/gcc/config/aarch64/aarch64.c ++++ b/gcc/config/aarch64/aarch64.c +@@ -2367,6 +2367,13 @@ aarch64_sve_data_mode_p (machine_mode mode) + return aarch64_classify_vector_mode (mode) & VEC_SVE_DATA; + } + ++/* Return true if MODE is an full SVE data vector mode. */ ++static bool ++aarch64_full_sve_data_mode_p (machine_mode mode) ++{ ++ return aarch64_classify_vector_mode (mode) == VEC_SVE_DATA; ++} ++ + /* Return the number of defined bytes in one constituent vector of + SVE mode MODE, which has vector flags VEC_FLAGS. */ + static poly_int64 +@@ -24370,6 +24377,17 @@ aarch64_libgcc_floating_mode_supported_p + #undef TARGET_ASM_FUNCTION_EPILOGUE + #define TARGET_ASM_FUNCTION_EPILOGUE aarch64_sls_emit_blr_function_thunks + ++#undef TARGET_VECTORIZE_CODE_FOR_PREFETCH ++#define TARGET_VECTORIZE_CODE_FOR_PREFETCH code_for_aarch64_sve_prefetch ++ ++#undef TARGET_VECTORIZE_CODE_FOR_GATHER_PREFETCH ++#define TARGET_VECTORIZE_CODE_FOR_GATHER_PREFETCH \ ++ code_for_aarch64_sve_gather_prefetch ++ ++#undef TARGET_VECTORIZE_PREFETCH_HANDLEABLE_MODE_P ++#define TARGET_VECTORIZE_PREFETCH_HANDLEABLE_MODE_P \ ++ aarch64_full_sve_data_mode_p ++ + struct gcc_target targetm = TARGET_INITIALIZER; + + #include "gt-aarch64.h" +diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi +index b46418d0b..ef3566510 100644 +--- a/gcc/doc/tm.texi ++++ b/gcc/doc/tm.texi +@@ -6122,6 +6122,27 @@ The default is @code{NULL_TREE} which means to not vectorize scatter + stores. + @end deftypefn + ++@deftypefn {Target Hook} insn_code TARGET_VECTORIZE_CODE_FOR_PREFETCH (machine_mode @var{arg}) ++This hook should return the decl of a function that implements the ++vectorized variant of the function with the @code{combined_fn} code ++@var{code} or @code{NULL_TREE} if such a function is not available. ++The return type of the vectorized function shall be of vector type ++@var{vec_type_out} and the argument types should be @var{vec_type_in}. ++@end deftypefn ++ ++@deftypefn {Target Hook} insn_code TARGET_VECTORIZE_CODE_FOR_GATHER_PREFETCH (machine_mode @var{mode_to}, machine_mode @var{mode_form}) ++This hook should return the decl of a function that implements the ++vectorized variant of the function with the @code{combined_fn} code ++@var{code} or @code{NULL_TREE} if such a function is not available. ++The return type of the vectorized function shall be of vector type ++@var{vec_type_out} and the argument types should be @var{vec_type_in}. ++@end deftypefn ++ ++@deftypefn {Target Hook} bool TARGET_VECTORIZE_PREFETCH_HANDLEABLE_MODE_P (machine_mode @var{arg}) ++This hook should return true if the target hardware architecture ++supports a full SVE data vector mode. ++@end deftypefn ++ + @deftypefn {Target Hook} int TARGET_SIMD_CLONE_COMPUTE_VECSIZE_AND_SIMDLEN (struct cgraph_node *@var{}, struct cgraph_simd_clone *@var{}, @var{tree}, @var{int}) + This hook should set @var{vecsize_mangle}, @var{vecsize_int}, @var{vecsize_float} + fields in @var{simd_clone} structure pointed by @var{clone_info} argument and also +diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in +index 2663547c7..945d0f696 100644 +--- a/gcc/doc/tm.texi.in ++++ b/gcc/doc/tm.texi.in +@@ -4195,6 +4195,12 @@ address; but often a machine-dependent strategy can generate better code. + + @hook TARGET_VECTORIZE_BUILTIN_SCATTER + ++@hook TARGET_VECTORIZE_CODE_FOR_PREFETCH ++ ++@hook TARGET_VECTORIZE_CODE_FOR_GATHER_PREFETCH ++ ++@hook TARGET_VECTORIZE_PREFETCH_HANDLEABLE_MODE_P ++ + @hook TARGET_SIMD_CLONE_COMPUTE_VECSIZE_AND_SIMDLEN + + @hook TARGET_SIMD_CLONE_ADJUST +diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c +index 644f234e0..e8a3bb654 100644 +--- a/gcc/internal-fn.c ++++ b/gcc/internal-fn.c +@@ -102,10 +102,12 @@ init_internal_fns () + direct_internal_fn. */ + #define not_direct { -2, -2, false } + #define mask_load_direct { -1, 2, false } ++#define mask_prefetch_direct { -1, 2, false } + #define load_lanes_direct { -1, -1, false } + #define mask_load_lanes_direct { -1, -1, false } + #define gather_load_direct { 3, 1, false } + #define mask_store_direct { 3, 2, false } ++#define gather_prefetch_direct { 3, 1, false } + #define store_lanes_direct { 0, 0, false } + #define mask_store_lanes_direct { 0, 0, false } + #define vec_cond_mask_direct { 0, 0, false } +@@ -2520,6 +2522,53 @@ expand_mask_load_optab_fn (internal_fn, gcall *stmt, convert_optab optab) + + #define expand_mask_load_lanes_optab_fn expand_mask_load_optab_fn + ++/* Expand MASK_PREFETCH call STMT using optab OPTAB. ++ .MASK_STORE (_5, 64B, loop_mask_98, vect__8.10_102); ++ .MASK_PREFETCH (_68, 64B, loop_mask_98, vect__8.10_102, 4); ++*/ ++ ++static void ++expand_mask_prefetch_optab_fn (internal_fn, gcall *stmt, direct_optab optab) ++{ ++ if (targetm.vectorize.code_for_prefetch == NULL ++ || targetm.vectorize.prefetch_handleable_mode_p == NULL) ++ return; ++ ++ tree base = gimple_call_arg (stmt, 0); ++ if (base == NULL_TREE) ++ return; ++ ++ tree maskt = gimple_call_arg (stmt, 2); ++ tree target = gimple_call_arg (stmt, 3); ++ tree prfop = gimple_call_arg (stmt, 4); ++ HOST_WIDE_INT prfop_int = tree_to_uhwi (prfop); ++ /* Bit 3 of the prfop selects stores over loads. */ ++ HOST_WIDE_INT access = prfop_int & 8; ++ /* Bits 1 and 2 specify the locality; 0-based for svprfop but ++ 1-based for PREFETCH. */ ++ HOST_WIDE_INT locality = ((prfop_int >> 1) & 3) + 1; ++ ++ machine_mode m_mode = TYPE_MODE (TREE_TYPE (target)); ++ if (!targetm.vectorize.prefetch_handleable_mode_p (m_mode)) ++ return; ++ insn_code icode = targetm.vectorize.code_for_prefetch (m_mode); ++ ++ rtx mask = expand_normal (maskt); ++ rtx base_rtx = expand_normal (base); ++ /* Convert ptr_mode value X to Pmode. */ ++ if (ptr_mode == SImode) ++ base_rtx = simplify_gen_unary (ZERO_EXTEND, DImode, base_rtx, SImode); ++ ++ unsigned i = 0; ++ class expand_operand ops[5]; ++ create_input_operand (&ops[i++], mask, TYPE_MODE (TREE_TYPE (maskt))); ++ create_address_operand (&ops[i++], base_rtx); ++ create_integer_operand (&ops[i++], prfop_int); ++ create_integer_operand (&ops[i++], access); ++ create_integer_operand (&ops[i++], locality); ++ expand_insn (icode, i, ops); ++} ++ + /* Expand MASK_STORE{,_LANES} call STMT using optab OPTAB. */ + + static void +@@ -2920,6 +2969,70 @@ expand_gather_load_optab_fn (internal_fn, gcall *stmt, direct_optab optab) + emit_move_insn (lhs_rtx, ops[0].value); + } + ++/* Expand {MASK_,}GATHER_PREFETCH call CALL using optab OPTAB. ++ vect_patt_97.14_77 = .MASK_GATHER_LOAD (_78, vect__14.13_79, 8, { 0.0, ... }, loop_mask_87); ++ .MASK_GATHER_PREFETCH (_45, vect__14.13_79, 8, { 0.0, ... }, loop_mask_87, vect_patt_97.14_77, 4); ++*/ ++ ++static void ++expand_gather_prefetch_optab_fn (internal_fn, gcall *stmt, direct_optab optab) ++{ ++ if (targetm.vectorize.code_for_gather_prefetch == NULL ++ || targetm.vectorize.prefetch_handleable_mode_p == NULL) ++ return; ++ ++ /* Extracting tree nodes, only expand for scalar base and vector index. */ ++ tree base = gimple_call_arg (stmt, 0); ++ if (VECTOR_TYPE_P (TREE_TYPE (base))) ++ return; ++ tree offset = gimple_call_arg (stmt, 1); ++ if (VECTOR_TYPE_P (TREE_TYPE (offset)) == false) ++ return; ++ ++ tree scale = gimple_call_arg (stmt, 2); ++ tree mask = gimple_call_arg (stmt, 4); ++ tree target = gimple_call_arg (stmt, 5); ++ tree prfop = gimple_call_arg (stmt, 6); ++ ++ /* Convert to the rtx node. */ ++ rtx base_rtx = expand_normal (base); ++ /* Convert ptr_mode value X to Pmode. */ ++ if (ptr_mode == SImode) ++ base_rtx = simplify_gen_unary (ZERO_EXTEND, DImode, base_rtx, SImode); ++ rtx offset_rtx = expand_normal (offset); ++ rtx const_rtx = CONST0_RTX (TYPE_MODE (TREE_TYPE (target))); ++ rtx mask_rtx = expand_normal (mask); ++ HOST_WIDE_INT scale_int = tree_to_shwi (scale); ++ HOST_WIDE_INT prfop_int = tree_to_uhwi (prfop); ++ /* Bit 3 of the prfop selects stores over loads. */ ++ HOST_WIDE_INT access = prfop_int & 8; ++ /* Bits 1 and 2 specify the locality; 0-based for svprfop but ++ 1-based for PREFETCH. */ ++ HOST_WIDE_INT locality = ((prfop_int >> 1) & 3) + 1; ++ ++ /* add operand. */ ++ unsigned int i = 0; ++ class expand_operand ops[9]; ++ create_input_operand (&ops[i++], mask_rtx, TYPE_MODE (TREE_TYPE (mask))); ++ create_address_operand (&ops[i++], base_rtx); ++ create_input_operand (&ops[i++], offset_rtx, TYPE_MODE (TREE_TYPE (offset))); ++ /* Check whether the index has unsigned. */ ++ create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); ++ create_integer_operand (&ops[i++], scale_int); ++ create_input_operand (&ops[i++], const_rtx, GET_MODE (const_rtx)); ++ create_integer_operand (&ops[i++], prfop_int); ++ create_integer_operand (&ops[i++], access); ++ create_integer_operand (&ops[i++], locality); ++ ++ machine_mode reg_mode = GET_MODE (offset_rtx); ++ machine_mode m_mode = TYPE_MODE (TREE_TYPE (target)); ++ if (!targetm.vectorize.prefetch_handleable_mode_p (m_mode)) ++ return; ++ insn_code icode = targetm.vectorize.code_for_gather_prefetch ++ (m_mode, reg_mode); ++ expand_insn (icode, i, ops); ++} ++ + /* Expand DIVMOD() using: + a) optab handler for udivmod/sdivmod if it is available. + b) If optab_handler doesn't exist, generate call to +@@ -3210,9 +3323,11 @@ multi_vector_optab_supported_p (convert_optab optab, tree_pair types, + #define direct_cond_binary_optab_supported_p direct_optab_supported_p + #define direct_cond_ternary_optab_supported_p direct_optab_supported_p + #define direct_mask_load_optab_supported_p direct_optab_supported_p ++#define direct_mask_prefetch_optab_supported_p direct_optab_supported_p + #define direct_load_lanes_optab_supported_p multi_vector_optab_supported_p + #define direct_mask_load_lanes_optab_supported_p multi_vector_optab_supported_p + #define direct_gather_load_optab_supported_p convert_optab_supported_p ++#define direct_gather_prefetch_optab_supported_p direct_optab_supported_p + #define direct_mask_store_optab_supported_p direct_optab_supported_p + #define direct_store_lanes_optab_supported_p multi_vector_optab_supported_p + #define direct_mask_store_lanes_optab_supported_p multi_vector_optab_supported_p +diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def +index 0c6fc3711..cc0f42b98 100644 +--- a/gcc/internal-fn.def ++++ b/gcc/internal-fn.def +@@ -119,6 +119,8 @@ along with GCC; see the file COPYING3. If not see + #endif + + DEF_INTERNAL_OPTAB_FN (MASK_LOAD, ECF_PURE, maskload, mask_load) ++DEF_INTERNAL_OPTAB_FN (MASK_PREFETCH, ECF_NOVOPS | ECF_LEAF, ++ maskprefetch, mask_prefetch) + DEF_INTERNAL_OPTAB_FN (LOAD_LANES, ECF_CONST, vec_load_lanes, load_lanes) + DEF_INTERNAL_OPTAB_FN (MASK_LOAD_LANES, ECF_PURE, + vec_mask_load_lanes, mask_load_lanes) +@@ -126,6 +128,8 @@ DEF_INTERNAL_OPTAB_FN (MASK_LOAD_LANES, ECF_PURE, + DEF_INTERNAL_OPTAB_FN (GATHER_LOAD, ECF_PURE, gather_load, gather_load) + DEF_INTERNAL_OPTAB_FN (MASK_GATHER_LOAD, ECF_PURE, + mask_gather_load, gather_load) ++DEF_INTERNAL_OPTAB_FN (MASK_GATHER_PREFETCH, ECF_NOVOPS | ECF_LEAF, ++ mask_gather_prefetch, gather_prefetch) + + DEF_INTERNAL_OPTAB_FN (SCATTER_STORE, 0, scatter_store, scatter_store) + DEF_INTERNAL_OPTAB_FN (MASK_SCATTER_STORE, 0, +diff --git a/gcc/optabs.def b/gcc/optabs.def +index 0c64eb52a..ee25bc3f7 100644 +--- a/gcc/optabs.def ++++ b/gcc/optabs.def +@@ -90,9 +90,11 @@ OPTAB_CD(vec_cmp_optab, "vec_cmp$a$b") + OPTAB_CD(vec_cmpu_optab, "vec_cmpu$a$b") + OPTAB_CD(vec_cmpeq_optab, "vec_cmpeq$a$b") + OPTAB_CD(maskload_optab, "maskload$a$b") ++OPTAB_CD(maskprefetch_optab, "maskprefetch$a$b") + OPTAB_CD(maskstore_optab, "maskstore$a$b") + OPTAB_CD(gather_load_optab, "gather_load$a$b") + OPTAB_CD(mask_gather_load_optab, "mask_gather_load$a$b") ++OPTAB_CD(mask_gather_prefetch_optab, "mask_gather_prefetch$a$b") + OPTAB_CD(scatter_store_optab, "scatter_store$a$b") + OPTAB_CD(mask_scatter_store_optab, "mask_scatter_store$a$b") + OPTAB_CD(vec_extract_optab, "vec_extract$a$b") +diff --git a/gcc/params.opt b/gcc/params.opt +index 2044524a3..c429359e3 100644 +--- a/gcc/params.opt ++++ b/gcc/params.opt +@@ -1005,4 +1005,57 @@ Target size of compressed pointer, which should be 8, 16 or 32. + Common Joined UInteger Var(semi_relayout_level) Init(13) IntegerRange(11, 15) Param Optimization + Set capacity of each bucket to semi-relayout to (1 << semi-relayout-level) / 8 . + ++-param=mem-access-ratio= ++Common Joined UInteger Var(param_mem_access_ratio) Init(20) IntegerRange(0, 100) Param Optimization ++Memory access ratio (in percent). ++ ++-param=mem-access-num= ++Common Joined UInteger Var(param_mem_access_num) Init(3) Param Optimization ++Memory access num. ++ ++-param=prefetch-offset= ++Common Joined UInteger Var(param_prefetch_offset) Init(1024) ++IntegerRange(1, 999999) Param Optimization ++Prefetch Offset, which is usually a power of two due to cache line size. ++ ++-param=branch-prob-threshold= ++Common Joined UInteger Var(param_branch_prob_threshold) Init(80) IntegerRange(50, 100) ++Param Optimization ++High Execution Rate Branch Threshold. ++ ++-param=issue-topn= ++Common Joined UInteger Var(param_issue_topn) Init(1) Param Optimization ++Issue topn LLC mem_ref hint. ++ ++-param=force-issue= ++Common Joined UInteger Var(param_force_issue) Init(0) IntegerRange(0, 1) Param ++Force issue the topn LLC mem_ref hint, without generating dynamic multi-branches. ++ ++-param=llc-capacity-per-core= ++Common Joined UInteger Var(param_llc_capacity_per_core) Init(114) IntegerRange(0, 999999) Param ++LLC capacity per core. ++ ++-param=target-variables= ++Common Joined Var(param_target_variables) Init("") Param Optimization ++--param=target-variables=[,,...] Target variables for prefetching, separated by comma, ++without space. The representation of a variable can be complex and containing space, please surround ++it by quotation marks and escape special characters in Linux. The input length should be no more ++than 512 characters. ++ ++-param=use-ref-group-index= ++Common Joined UInteger Var(param_use_ref_group_index) Init(0) IntegerRange(0, 1) Param Optimization ++Prefetch the target variables by their indices in sorted ref_groups, use together with parameter ++target-variables. ++ ++-param=mem-ref-index= ++Common Joined Var(param_mem_ref_index) Init("") Param Optimization ++--param=mem-ref-index=[,,...] Prefetch the target variable at the memory reference ++location with the index of customized order, separated by comma, without space. The input length ++should be no more than 512 characters. ++ ++-param=filter-kernels= ++Common Joined UInteger Var(param_filter_kernels) Init(1) IntegerRange(0, 1) Param ++Allow LLC allocate pass to greedily filter kernels by traversing the corresponding basic blocks ++through edges with branch probability no less than param_branch_prob_threshold. ++ + ; This comment is to ensure we retain the blank line above. +diff --git a/gcc/passes.def b/gcc/passes.def +index df7d65733..ea59fc8ca 100644 +--- a/gcc/passes.def ++++ b/gcc/passes.def +@@ -303,6 +303,7 @@ along with GCC; see the file COPYING3. If not see + /* Run IVOPTs after the last pass that uses data-reference analysis + as that doesn't handle TARGET_MEM_REFs. */ + NEXT_PASS (pass_iv_optimize); ++ NEXT_PASS (pass_llc_allocate); + NEXT_PASS (pass_lim); + NEXT_PASS (pass_tree_loop_done); + POP_INSERT_PASSES () +diff --git a/gcc/target.def b/gcc/target.def +index 34d3561bd..351c94c37 100644 +--- a/gcc/target.def ++++ b/gcc/target.def +@@ -2072,6 +2072,37 @@ DEFHOOK + (void *data), + default_destroy_cost_data) + ++/* Function for vector prefetch operation. */ ++DEFHOOK ++(code_for_prefetch, ++ "This hook should return the decl of a function that implements the\n\ ++vectorized variant of the function with the @code{combined_fn} code\n\ ++@var{code} or @code{NULL_TREE} if such a function is not available.\n\ ++The return type of the vectorized function shall be of vector type\n\ ++@var{vec_type_out} and the argument types should be @var{vec_type_in}.", ++ insn_code, (machine_mode arg), ++ NULL) ++ ++/* Function for vector gather prefetch operation. */ ++DEFHOOK ++(code_for_gather_prefetch, ++ "This hook should return the decl of a function that implements the\n\ ++vectorized variant of the function with the @code{combined_fn} code\n\ ++@var{code} or @code{NULL_TREE} if such a function is not available.\n\ ++The return type of the vectorized function shall be of vector type\n\ ++@var{vec_type_out} and the argument types should be @var{vec_type_in}.", ++ insn_code, (machine_mode mode_to, machine_mode mode_form), ++ NULL) ++ ++/* Function to check whether the target hardware architecture supports ++ a full SVE data vector mode. */ ++DEFHOOK ++(prefetch_handleable_mode_p, ++ "This hook should return true if the target hardware architecture\n\ ++supports a full SVE data vector mode.", ++ bool, (machine_mode arg), ++ NULL) ++ + HOOK_VECTOR_END (vectorize) + + #undef HOOK_PREFIX +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-1.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-1.c +new file mode 100644 +index 000000000..a4828eaab +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-1.c +@@ -0,0 +1,61 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param issue-topn=2" } */ ++ ++#include ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cell ++ ++#define N 100000 ++ ++int A_i[N]; ++int A_j[N]; ++double A_data[N]; ++double x_data[N]; ++double y_data[N]; ++int num_rows = N; ++ ++void ++MatMult (int *A_i, int *A_j, double *A_data, double *x_data, ++ int num_rows, double *y_data) ++{ ++ int i = 0; ++ int j = 0; ++ double temp = 0; ++ for (i = 0; i < num_rows; i++) ++ { ++ temp = y_data[i]; ++ for (j = A_i[i]; j < A_i[i+1]; j++) ++ temp += A_data[j] * x_data[A_j[j]]; ++ y_data[i] = temp; ++ } ++} ++ ++int ++main (int argc, char *argv[]) ++{ ++ int testIter = 2; ++ ++ for (int i = 0; i < testIter; i++) ++ MatMult (A_i, A_j, A_data, x_data, num_rows, y_data); ++ ++ return 0; ++} ++ ++/* { dg-final { scan-tree-dump-times "ref_count = (?:\[3-9\]|\[1-9\]\\d{1,}), ninsns = \[1-9\]\\d*, mem_to_insn_ratio = 0.\[2-9\]\\d*" 2 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "Tracing succeeded" 6 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-not "Tracing failed" "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-not "static_data_size:" "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "\{ (?:\\d+\\(\\d+\\) ){1}\}" 4 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-not ", size: (?!(0\.000000))" "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times ", size: 0\.000000" 6 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "\\d x_data \\(0.000000, 1, 0\\) : 3" 2 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "\\d A_j \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "\\d A_data \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "static issue" 2 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "insert svprfd_gather" 2 "llc_allocate" } } */ +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp b/gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp +new file mode 100644 +index 000000000..4f34e722f +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp +@@ -0,0 +1,27 @@ ++# Copyright (C) 2022-2023 Free Software Foundation, Inc. ++ ++# This program is free software; you can redistribute it and/or modify ++# it under the terms of the GNU General Public License as published by ++# the Free Software Foundation; either version 3 of the License, or ++# (at your option) any later version. ++# ++# This program is distributed in the hope that it will be useful, ++# but WITHOUT ANY WARRANTY; without even the implied warranty of ++# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the ++# GNU General Public License for more details. ++# ++# You should have received a copy of the GNU General Public License ++# along with GCC; see the file COPYING3. If not see ++# . ++ ++load_lib gcc-dg.exp ++load_lib target-supports.exp ++ ++# Initialize `dg'. ++dg-init ++ ++dg-runtest [lsort [glob -nocomplain $srcdir/$subdir/*.c]] \ ++ "" "-fllc-allocate" ++ ++# All done. ++dg-finish +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c +new file mode 100644 +index 000000000..2a58c501f +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c +@@ -0,0 +1,48 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=uPtr" } */ ++ ++#include ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cell ++ ++typedef struct stack_def ++{ ++ int top; /* index to top stack element */ ++ unsigned long reg_set; /* set of live registers */ ++ unsigned char reg[128]; /* register - stack mapping */ ++} *stack; ++ ++typedef struct block_info_def ++{ ++ struct stack_def stack_in; /* Input stack configuration. */ ++ struct stack_def stack_out; /* Output stack configuration. */ ++ unsigned long out_reg_set; /* Stack regs live on output. */ ++ int done; /* True if block already converted. */ ++ int predecessors; /* Number of predecessors that need ++ to be visited. */ ++} *block_info; ++ ++typedef struct basic_block_def ++{ ++ void *aux; ++} *basic_block; ++ ++unsigned char ++convert_regs_exit (basic_block bb, int value_reg_low, int value_reg_high) ++{ ++ stack output_stack; ++ ++ output_stack = &(((block_info) bb->aux)->stack_in); ++ if (value_reg_low == -1) ++ output_stack->top = -1; ++ else ++ { ++ int reg; ++ output_stack->top = value_reg_high - value_reg_low; ++ for (reg = value_reg_low; reg <= value_reg_high; ++reg) ++ { ++ (output_stack->reg + 16)[value_reg_high - reg] = reg; ++ output_stack->reg_set |= (unsigned long) 1 << reg; ++ } ++ } ++ return output_stack->reg[0]; ++} ++ ++/* { dg-final { scan-tree-dump-times "runtime issue" 1 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "static issue" 1 "llc_allocate" } } */ +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-ref-trace.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-ref-trace.c +new file mode 100644 +index 000000000..27cd574cf +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-ref-trace.c +@@ -0,0 +1,62 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-options "-O3 -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param filter-kernels=0" } */ ++ ++#include ++#include ++ ++#define N 1000 ++ ++long a[N] = {0}; ++long b[N] = {0}; ++long c[N] = {0}; ++ ++double ++referenceTrace (double *psiPtr, int *lPtr, int *uPtr, int nCells) ++{ ++ double sum; ++ for (int cell = 0; cell < nCells; cell++) ++ { ++ // Multi-layer pointer ++ sum += psiPtr[lPtr[cell]]; ++ psiPtr[uPtr[cell]] = sum; ++ ++ // Outer pointer, inner array ++ sum += psiPtr[b[cell]]; ++ psiPtr[a[cell]] = sum; ++ ++ // Multi-layer array, currently failed tracing at b[cell] and a[cell] ++ sum += a[b[cell]]; ++ c[a[cell]] = sum; ++ ++ // Outer array, inner pointer, currently failed tracing at lPtr[cell] ++ sum += a[lPtr[cell]]; ++ c[lPtr[cell]] = sum; ++ } ++ return sum; ++} ++ ++int ++main (int argc, char *argv[]) ++{ ++ int testIter = 2; ++ ++ double *psiPtr = NULL; ++ int *lPtr = NULL; ++ int *uPtr = NULL; ++ psiPtr = (double *) calloc (N, sizeof(double)); ++ lPtr = (int *) calloc (N, sizeof(int)); ++ uPtr = (int *) calloc (N, sizeof(int)); ++ ++ for (int i = 0; i < testIter; i++) ++ referenceTrace (psiPtr, lPtr, uPtr, N); ++ ++ free (psiPtr); ++ free (lPtr); ++ free (uPtr); ++ ++ return 0; ++} ++ ++/* { dg-final { scan-tree-dump-times "Tracing succeeded" 16 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "Tracing failed" 8 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "unhandled issue scene" 2 "llc_allocate" } } */ +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c +new file mode 100644 +index 000000000..276781c4f +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c +@@ -0,0 +1,48 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=lPtr" } */ ++ ++#include ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cell ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cell ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cell ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cell ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cell ++ ++#define N 131590 ++#define F 384477 ++ ++double diagPtr[N]; ++double psiPtr[N]; ++double ApsiPtr[N]; ++int lPtr[F]; ++int uPtr[F]; ++double lowerPtr[F]; ++double upperPtr[F]; ++ ++void ++AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr, ++ int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces) ++{ ++ for (int cell=0; cellaux\"" } */ ++ ++#include ++ ++typedef struct stack_def ++{ ++ int top; /* index to top stack element */ ++ unsigned long reg_set; /* set of live registers */ ++ unsigned char reg[128]; /* register - stack mapping */ ++} *stack; ++ ++typedef struct block_info_def ++{ ++ struct stack_def stack_in; /* Input stack configuration. */ ++ struct stack_def stack_out; /* Output stack configuration. */ ++ unsigned long out_reg_set; /* Stack regs live on output. */ ++ int done; /* True if block already converted. */ ++ int predecessors; /* Number of predecessors that need ++ to be visited. */ ++} *block_info; ++ ++typedef struct basic_block_def ++{ ++ void *aux; ++} *basic_block; ++ ++unsigned char ++convert_regs_exit (basic_block bb, int value_reg_low, int value_reg_high) ++{ ++ stack output_stack; ++ ++ output_stack = &(((block_info) bb->aux)->stack_in); ++ if (value_reg_low == -1) ++ output_stack->top = -1; ++ else ++ { ++ int reg; ++ output_stack->top = value_reg_high - value_reg_low; ++ for (reg = value_reg_low; reg <= value_reg_high; ++reg) ++ { ++ (output_stack->reg + 16)[value_reg_high - reg] = reg; ++ output_stack->reg_set |= (unsigned long) 1 << reg; ++ } ++ } ++ return output_stack->reg[0]; ++} ++ ++/* { dg-final { scan-tree-dump-not "Unrecognizable variable name" ++ "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-not "static issue" "llc_allocate" } } */ +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c +new file mode 100644 +index 000000000..09a525ce1 +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c +@@ -0,0 +1,54 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-options "-O3 -c -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param filter-kernels=0 --param target-variables=tmp_var_0" } */ ++ ++#include ++ ++typedef struct stack_def ++{ ++ int top; /* index to top stack element */ ++ unsigned long reg_set; /* set of live registers */ ++ unsigned char reg[128]; /* register - stack mapping */ ++} *stack; ++ ++typedef struct block_info_def ++{ ++ struct stack_def stack_in; /* Input stack configuration. */ ++ struct stack_def stack_out; /* Output stack configuration. */ ++ unsigned long out_reg_set; /* Stack regs live on output. */ ++ int done; /* True if block already converted. */ ++ int predecessors; /* Number of predecessors that need ++ to be visited. */ ++} *block_info; ++ ++typedef struct basic_block_def ++{ ++ void *aux; ++} *basic_block; ++ ++unsigned char ++convert_regs_exit (basic_block bb, int value_reg_low, int value_reg_high) ++{ ++ stack output_stack; ++ ++ output_stack = &(((block_info) bb->aux)->stack_in); ++ if (value_reg_low == -1) ++ output_stack->top = -1; ++ else ++ { ++ int reg; ++ output_stack->top = value_reg_high - value_reg_low; ++ for (reg = value_reg_low; reg <= value_reg_high; ++reg) ++ { ++ (output_stack->reg + 16)[value_reg_high - reg] = reg; ++ output_stack->reg_set |= (unsigned long) 1 << reg; ++ } ++ } ++ return output_stack->reg[0]; ++} ++ ++/* { dg-final { scan-tree-dump-not "Unrecognizable variable name" ++ "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "NOTICE: Prefetching target variable \"" ++ " bb_16(D)->aux \"" 1 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "static issue" 1 "llc_allocate" } } */ +diff --git a/gcc/testsuite/gfortran.dg/llc-allocate/llc-3.f90 b/gcc/testsuite/gfortran.dg/llc-allocate/llc-3.f90 +new file mode 100644 +index 000000000..ec918e144 +--- /dev/null ++++ b/gcc/testsuite/gfortran.dg/llc-allocate/llc-3.f90 +@@ -0,0 +1,213 @@ ++! { dg-do compile { target { aarch64*-*-linux* } } } ++! { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param branch-prob-threshold=50" } ++ ++program main ++ ++ IMPLICIT NONE ++ INTEGER :: ids,ide, jds,jde, kds,kde ++ INTEGER,parameter :: ims=-4,kms=1,jms=-4 ++ INTEGER,parameter :: ime=210,kme=36,jme=192 ++ INTEGER :: its,ite, jts,jte, kts,kte ++ INTEGER :: number_of_small_timesteps,rk_step, rk_order, step ++ ++ REAL, DIMENSION(ims:ime, kms:kme, jms:jme) :: t_1, t_2, c2a, p, ph, pm1, al, alt ++ ++ ++ REAL, DIMENSION(ims:ime, jms:jme) :: mu, muts ++ ++ REAL, DIMENSION(kms:kme) :: dnw, rdnw, znu ++ ++ REAL :: rdx,rdy ++ REAL :: dts, t0, smdiv ++ REAL :: random1,time_begin,time_end,total_time ++ ++ INTEGER :: i, j, k ++ INTEGER :: i_start, i_end, j_start, j_end, k_start, k_end ++ INTEGER :: i_endu, j_endv ++ INTEGER :: interval=1 ++ INTEGER :: epoch,iter ++ ++ LOGICAL :: non_hydrostatic ++ ++ data ids, jds, kds, its, jts, kts /6*1/ ++ data ide, ite /2*205/ ++ data jde, jte /2*187/ ++ data kde, kte /2*36/ ++ ++ number_of_small_timesteps = 1 ++ rk_step = 3 ++ rk_order = 1 ++ dts = 1. ++ ++ rdx = 1. ++ rdy = 1. ++ ++ t0 = 0. ++ smdiv = 1. ++ step = 1 ++ non_hydrostatic = .true. ++ ++ call random_number(random1) ++ interval = random1*100 ++ interval=1 ++ ++ call random_seed(put=(/(i,i=1,10000,interval)/)) ++ ++ call random_number(alt) ++ call random_number(c2a) ++ call random_number(ph) ++ call random_number(pm1) ++ call random_number(mu) ++ call random_number(muts) ++ call random_number(dnw) ++ call random_number(rdnw) ++ call random_number(znu) ++ ++ do iter=1,2 ++ call calc_p_rho( al, p, ph, & ++ alt, t_2, t_1, c2a, pm1, & ++ mu, muts, znu, t0, & ++ rdnw, dnw, smdiv, & ++ non_hydrostatic, step, & ++ ids, ide, jds, jde, kds, kde, & ++ ims, ime, jms, jme, kms, kme, & ++ its,ite, jts,jte, kts,kte ) ++ ++ enddo ++ ++end program ++ ++ ++SUBROUTINE calc_p_rho( al, p, ph, & ++ alt, t_2, t_1, c2a, pm1, & ++ mu, muts, znu, t0, & ++ rdnw, dnw, smdiv, & ++ non_hydrostatic, step, & ++ ids, ide, jds, jde, kds, kde, & ++ ims, ime, jms, jme, kms, kme, & ++ its,ite, jts,jte, kts,kte ) ++ ++ IMPLICIT NONE ! religion first ++ !asb ++! declarations for the stuff coming in ++ ++ INTEGER, INTENT(IN ) :: ids,ide, jds,jde, kds,kde ++ INTEGER, INTENT(IN ) :: ims,ime, jms,jme, kms,kme ++ INTEGER, INTENT(IN ) :: its,ite, jts,jte, kts,kte ++ ++ INTEGER, INTENT(IN ) :: step ++ ++ REAL, DIMENSION(ims:ime, kms:kme, jms:jme),INTENT( OUT) :: al, & ++ p ++ ++ REAL, DIMENSION(ims:ime, kms:kme, jms:jme),INTENT(IN ) :: alt, & ++ t_2, & ++ t_1, & ++ c2a ++ ++ REAL, DIMENSION(ims:ime, kms:kme, jms:jme),INTENT(INOUT) :: ph, pm1 ++ ++ REAL, DIMENSION(ims:ime, jms:jme) , INTENT(IN ) :: mu, & ++ muts ++ ++ REAL, DIMENSION(kms:kme) , INTENT(IN ) :: dnw, & ++ rdnw, & ++ znu ++ ++ REAL, INTENT(IN ) :: t0, smdiv ++ ++ LOGICAL, INTENT(IN ) :: non_hydrostatic ++ ++! local variables ++ ++ INTEGER :: i, j, k ++ INTEGER :: i_start, i_end, j_start, j_end, k_start, k_end ++ REAL :: ptmp ++ ++ i_start = its ++ i_end = min(ite,ide-1) ++ j_start = jts ++ j_end = min(jte,jde-1) ++ k_start = kts ++ k_end = min(kte,kde-1) ++ ++ IF (non_hydrostatic) THEN ++ DO j=j_start, j_end ++ DO k=k_start, k_end ++ DO i=i_start, i_end ++ ++! al computation is all dry, so ok with moisture ++ ++ al(i,k,j)=-1./muts(i,j)*(alt(i,k,j)*mu(i,j) & ++ +rdnw(k)*(ph(i,k+1,j)-ph(i,k,j))) ++ ++! this is temporally linearized p, no moisture correction needed ++ ++ p(i,k,j)=c2a(i,k,j)*(alt(i,k,j)*(t_2(i,k,j)-mu(i,j)*t_1(i,k,j)) & ++ /(muts(i,j)*(t0+t_1(i,k,j)))-al (i,k,j)) ++ ++ ENDDO ++ ENDDO ++ ENDDO ++ ++ ELSE ! hydrostatic calculation ++ ++ DO j=j_start, j_end ++ DO k=k_start, k_end ++ DO i=i_start, i_end ++ p(i,k,j)=mu(i,j)*znu(k) ++ al(i,k,j)=alt(i,k,j)*(t_2(i,k,j)-mu(i,j)*t_1(i,k,j)) & ++ /(muts(i,j)*(t0+t_1(i,k,j)))-p(i,k,j)/c2a(i,k,j) ++ ph(i,k+1,j)=ph(i,k,j)-dnw(k)*(muts(i,j)*al (i,k,j) & ++ +mu(i,j)*alt(i,k,j)) ++ ENDDO ++ ENDDO ++ ENDDO ++ ++ END IF ++ ++! divergence damping setup ++ ++ IF (step == 0) then ! we're initializing small timesteps ++ DO j=j_start, j_end ++ DO k=k_start, k_end ++ DO i=i_start, i_end ++ pm1(i,k,j)=p(i,k,j) ++ ENDDO ++ ENDDO ++ ENDDO ++ ELSE ! we're in the small timesteps ++ DO j=j_start, j_end ! and adding div damping component ++ DO k=k_start, k_end ++ DO i=i_start, i_end ++ ptmp = p(i,k,j) ++ p(i,k,j) = p(i,k,j) + smdiv*(p(i,k,j)-pm1(i,k,j)) ++ pm1(i,k,j) = ptmp ++ ENDDO ++ ENDDO ++ ENDDO ++ END IF ++ ++END SUBROUTINE calc_p_rho ++ ++! { dg-final { scan-tree-dump-times "ref_count = (?:\[3-9\]|\[1-9\]\\d{1,}), ninsns = \[1-9\]\\d*, mem_to_insn_ratio = 0.\[2-9\]\\d*" 6 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "Tracing succeeded" 48 "llc_allocate" } } ++! { dg-final { scan-tree-dump-not "Tracing failed" "llc_allocate" } } ++! { dg-final { scan-tree-dump-not "static_data_size:" "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\{ (?:\\d+\\(\\d+\\) ){2}\}" 3 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\{ (?:\\d+\\(\\d+\\) ){4}\}" 1 "llc_allocate" } } ++! { dg-final { scan-tree-dump-not ", size: (?!(0\.000000))" "llc_allocate" } } ++! { dg-final { scan-tree-dump-times ", size: 0\.000000" 28 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d p \\(0.000000, 3, 0\\) : 8" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d pm1 \\(0.000000, 2, 0\\) : 5" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d ph \\(0.000000, 2, 0\\) : 4" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d al \\(0.000000, 1, 0\\) : 3" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d alt \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d t_1 \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d t_2 \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d c2a \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d mu \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "\\d muts \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "runtime issue" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "static issue" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "insert svprfd" 2 "llc_allocate" } } +diff --git a/gcc/testsuite/gfortran.dg/llc-allocate/llc-allocate.exp b/gcc/testsuite/gfortran.dg/llc-allocate/llc-allocate.exp +new file mode 100644 +index 000000000..068341784 +--- /dev/null ++++ b/gcc/testsuite/gfortran.dg/llc-allocate/llc-allocate.exp +@@ -0,0 +1,29 @@ ++# Copyright (C) 2022-2023 Free Software Foundation, Inc. ++ ++# This program is free software; you can redistribute it and/or modify ++# it under the terms of the GNU General Public License as published by ++# the Free Software Foundation; either version 3 of the License, or ++# (at your option) any later version. ++# ++# This program is distributed in the hope that it will be useful, ++# but WITHOUT ANY WARRANTY; without even the implied warranty of ++# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the ++# GNU General Public License for more details. ++# ++# You should have received a copy of the GNU General Public License ++# along with GCC; see the file COPYING3. If not see ++# . ++ ++# GCC testsuite that uses the `dg.exp' driver. ++ ++load_lib gfortran-dg.exp ++ ++# Initialize `dg'. ++dg-init ++ ++# Main loop. ++gfortran-dg-runtest [lsort \ ++ [glob -nocomplain $srcdir/$subdir/*.\[fF\]{,90,95,03,08} ] ] "" "" ++ ++# All done. ++dg-finish +diff --git a/gcc/testsuite/gfortran.dg/llc-allocate/llc-trace-multiple-base-var.f90 b/gcc/testsuite/gfortran.dg/llc-allocate/llc-trace-multiple-base-var.f90 +new file mode 100644 +index 000000000..23e360540 +--- /dev/null ++++ b/gcc/testsuite/gfortran.dg/llc-allocate/llc-trace-multiple-base-var.f90 +@@ -0,0 +1,63 @@ ++! { dg-do compile { target { aarch64*-*-linux* } } } ++! { dg-options "-O3 -march=armv8.2-a+sve -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno" } ++ ++MODULE INPUT ++ IMPLICIT NONE ++ ++ INTEGER, PARAMETER :: wp = 8, jpi = 25, jpj = 39, jpk = 31, kjpt = 2 ++ ++ INTEGER :: kt = 1, jpkm1 = 30, jpjm1 = 38, fs_jpim1 = 24, fs_2 = 2 ++ REAL(wp), DIMENSION(jpi, jpj) :: e12t ++ REAL(wp), DIMENSION(jpi, jpj, jpk) :: fse3t_n ++ REAL(wp), DIMENSION(jpi, jpj, jpk, kjpt) :: pta ++ ++END MODULE INPUT ++ ++PROGRAM MAIN ++ USE INPUT ++ ++ IMPLICIT NONE ++ ++ INTEGER :: EPOCH ++ ++! Initialize arrays ++ ++ e12t = 1 ++ fse3t_n = 1 ++ pta = 1 ++! ++ ++ DO EPOCH=1,2 ++ CALL tra_ldf_iso ++ ENDDO ++ ++END PROGRAM MAIN ++ ++SUBROUTINE tra_ldf_iso ++ USE INPUT ++ ++ IMPLICIT NONE ++ ! ++ INTEGER :: ji, jj, jk, jn ! dummy loop indices ++ REAL(wp) :: zbtr, ztra ! - - ++ REAL(wp), DIMENSION(jpi, jpj, jpk) :: ztfw ++ ++ DO jn = 1, kjpt ++ ztfw(:, :, 1) = 0.e0; ztfw(:, :, jpk) = 0.e0 ++ ++ DO jk = 1, jpkm1 ++ DO jj = 2, jpjm1 ++ DO ji = fs_2, fs_jpim1 ! vector opt. ++ zbtr = 1.0/(e12t(ji, jj)*fse3t_n(ji, jj, jk)) ++ ztra = (ztfw(ji, jj, jk) - ztfw(ji, jj, jk + 1))*zbtr ++ pta(ji, jj, jk, jn) = pta(ji, jj, jk, jn) + ztra ++ END DO ++ END DO ++ END DO ++ ! ++ END DO ++ ! ++END SUBROUTINE tra_ldf_iso ++ ++! { dg-final { scan-tree-dump-times "Traced variables at vectp_ztfw" 2 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "Tracing unusual number or occurrences of base variables. Choose ztfw." 2 "llc_allocate" } } +diff --git a/gcc/testsuite/gfortran.dg/llc-allocate/llc-unknown-type-size-unit.f90 b/gcc/testsuite/gfortran.dg/llc-allocate/llc-unknown-type-size-unit.f90 +new file mode 100644 +index 000000000..d76c75b5b +--- /dev/null ++++ b/gcc/testsuite/gfortran.dg/llc-allocate/llc-unknown-type-size-unit.f90 +@@ -0,0 +1,58 @@ ++! { dg-do compile { target { aarch64*-*-linux* } } } ++! { dg-options "-c -O3 -march=armv8.2-a+sve -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param filter-kernels=0 --param issue-topn=1 --param mem-access-ratio=5 --param mem-access-num=1" } ++ ++Module module_domain ++ IMPLICIT NONE ++ ++ REAL, PARAMETER :: g = 9.8 ++ TYPE :: grid_type ++ REAL, POINTER :: phb(:,:,:), ph_2(:,:,:), p(:,:,:), pb(:,:,:) ++ REAL, POINTER :: fnm(:), fnp(:) ++ END TYPE ++END Module ++ ++SUBROUTINE calc_p8w(p8w, ix, iy, k_start, k_end) ++ ++ USE module_domain ++ !USE module_model_constants ++ ++ IMPLICIT NONE ++ ++ ++ !TYPE (domain), INTENT(IN) :: grid ++ INTEGER, INTENT(IN) :: k_start, k_end, ix, iy ++ REAL, DIMENSION(k_start:k_end), INTENT(OUT) :: p8w ++ ++ ++ INTEGER :: k ++ REAL :: z0, z1, z2, w1, w2 ++ REAL, DIMENSION(k_start:k_end) :: z_at_w ++ REAL, DIMENSION(k_start:k_end-1) :: z ++ TYPE (grid_type), POINTER :: grid ++ ++ ++ DO k = k_start, k_end ++ z_at_w(k) = (grid%phb(ix,k,iy)+grid%ph_2(ix,k,iy))/g ++ END DO ++ ++ DO k = k_start, k_end-1 ++ z(k) = 0.5*(z_at_w(k) + z_at_w(k+1)) ++ END DO ++ ++ DO k = k_start+1, k_end-1 ++ p8w(k) = grid%fnm(k)*(grid%p(ix,k,iy)+grid%pb(ix,k,iy)) + & ++ grid%fnp(k)*(grid%p(ix,k-1,iy)+grid%pb(ix,k-1,iy)) ++ END DO ++ ++ z0 = z_at_w(k_start) ++ z1 = z(k_start) ++ z2 = z(k_start+1) ++ w1 = (z0 - z2)/(z1 - z2) ++ w2 = 1. - w1 ++ p8w(k_start) = w1*(grid%p(ix,k_start,iy)+grid%pb(ix,k_start,iy)) + & ++ w2*(grid%p(ix,k_start+1,iy)+grid%pb(ix,k_start+1,iy)) ++ ++END SUBROUTINE calc_p8w ++ ++! { dg-final { scan-tree-dump-times "runtime issue" 1 "llc_allocate" } } ++! { dg-final { scan-tree-dump-times "static issue" 1 "llc_allocate" } } +diff --git a/gcc/timevar.def b/gcc/timevar.def +index ba86a1b7b..4b643538f 100644 +--- a/gcc/timevar.def ++++ b/gcc/timevar.def +@@ -207,6 +207,7 @@ DEFTIMEVAR (TV_TREE_LOOP_DISTRIBUTION, "tree loop distribution") + DEFTIMEVAR (TV_CHECK_DATA_DEPS , "tree check data dependences") + DEFTIMEVAR (TV_TREE_PREFETCH , "tree prefetching") + DEFTIMEVAR (TV_TREE_LOOP_IVOPTS , "tree iv optimization") ++DEFTIMEVAR (TV_TREE_LLC_ALLOCATE , "tree llc allocation") + DEFTIMEVAR (TV_PREDCOM , "predictive commoning") + DEFTIMEVAR (TV_TREE_CH , "tree copy headers") + DEFTIMEVAR (TV_TREE_SSA_UNCPROP , "tree SSA uncprop") +diff --git a/gcc/tree-cfg.c b/gcc/tree-cfg.c +index d82fe23d8..9eb173d69 100644 +--- a/gcc/tree-cfg.c ++++ b/gcc/tree-cfg.c +@@ -8365,6 +8365,17 @@ print_loops (FILE *file, int verbosity) + print_loop_and_siblings (file, bb->loop_father, 0, verbosity); + } + ++/* Dump a loop to file. */ ++ ++void ++loop_dump (FILE *file, class loop *loop) ++{ ++ print_loop (file, loop, 0, 0); ++ fprintf (file, "vec_niter = "); ++ print_generic_expr (file, loop->vec_nb_iterations); ++ fprintf (file, "\n"); ++} ++ + /* Dump a loop. */ + + DEBUG_FUNCTION void +diff --git a/gcc/tree-cfg.h b/gcc/tree-cfg.h +index beb4997a6..dad0ca0a6 100644 +--- a/gcc/tree-cfg.h ++++ b/gcc/tree-cfg.h +@@ -83,6 +83,7 @@ extern void dump_function_to_file (tree, FILE *, dump_flags_t); + extern void debug_function (tree, dump_flags_t); + extern void print_loops_bb (FILE *, basic_block, int, int); + extern void print_loops (FILE *, int); ++extern void loop_dump (FILE *file, class loop *loop); + extern void debug (class loop &ref); + extern void debug (class loop *ptr); + extern void debug_verbose (class loop &ref); +diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h +index 027f8992d..a1e215901 100644 +--- a/gcc/tree-pass.h ++++ b/gcc/tree-pass.h +@@ -383,6 +383,7 @@ extern gimple_opt_pass *make_pass_complete_unrolli (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_parallelize_loops (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_loop_prefetch (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_iv_optimize (gcc::context *ctxt); ++extern gimple_opt_pass *make_pass_llc_allocate (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_tree_loop_done (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_ch (gcc::context *ctxt); + extern gimple_opt_pass *make_pass_ch_vect (gcc::context *ctxt); +diff --git a/gcc/tree-scalar-evolution.c b/gcc/tree-scalar-evolution.c +index edab77827..73ffa0759 100644 +--- a/gcc/tree-scalar-evolution.c ++++ b/gcc/tree-scalar-evolution.c +@@ -2789,7 +2789,7 @@ resolve_mixers (class loop *loop, tree chrec, bool *folded_casts) + the loop body has been executed 6 times. */ + + tree +-number_of_latch_executions (class loop *loop) ++number_of_latch_executions (class loop *loop, bool guarantee) + { + edge exit; + class tree_niter_desc niter_desc; +@@ -2810,7 +2810,8 @@ number_of_latch_executions (class loop *loop) + res = chrec_dont_know; + exit = single_exit (loop); + +- if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false)) ++ if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false, ++ true, NULL, guarantee)) + { + may_be_zero = niter_desc.may_be_zero; + res = niter_desc.niter; +@@ -2836,7 +2837,8 @@ number_of_latch_executions (class loop *loop) + fprintf (dump_file, "))\n"); + } + +- loop->nb_iterations = res; ++ if (guarantee) ++ loop->nb_iterations = res; + return res; + } + +diff --git a/gcc/tree-scalar-evolution.h b/gcc/tree-scalar-evolution.h +index e2fbfb55b..218155650 100644 +--- a/gcc/tree-scalar-evolution.h ++++ b/gcc/tree-scalar-evolution.h +@@ -21,7 +21,8 @@ along with GCC; see the file COPYING3. If not see + #ifndef GCC_TREE_SCALAR_EVOLUTION_H + #define GCC_TREE_SCALAR_EVOLUTION_H + +-extern tree number_of_latch_executions (class loop *); ++extern tree number_of_latch_executions (class loop *, ++ bool guarantee = true); + extern gcond *get_loop_exit_condition (const class loop *); + + extern void scev_initialize (void); +diff --git a/gcc/tree-ssa-llc-allocate.c b/gcc/tree-ssa-llc-allocate.c +new file mode 100644 +index 000000000..746a1cf95 +--- /dev/null ++++ b/gcc/tree-ssa-llc-allocate.c +@@ -0,0 +1,2898 @@ ++/* LLC allocate. ++ Copyright (C) 2022-2023 Free Software Foundation, Inc. ++ ++This file is part of GCC. ++ ++GCC is free software; you can redistribute it and/or modify it ++under the terms of the GNU General Public License as published by the ++Free Software Foundation; either version 3, or (at your option) any ++later version. ++ ++GCC is distributed in the hope that it will be useful, but WITHOUT ++ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or ++FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License ++for more details. ++ ++You should have received a copy of the GNU General Public License ++along with GCC; see the file COPYING3. If not see ++. */ ++ ++#include "config.h" ++#define INCLUDE_MAP ++#define INCLUDE_SET ++#define INCLUDE_VECTOR ++#define INCLUDE_LIST ++#define INCLUDE_ALGORITHM ++#define INCLUDE_STRING ++#include "system.h" ++#include "coretypes.h" ++#include "backend.h" ++#include "target.h" ++#include "rtl.h" ++#include "tree.h" ++#include "gimple.h" ++#include "predict.h" ++#include "tree-pass.h" ++#include "gimple-ssa.h" ++#include "optabs-query.h" ++#include "tree-pretty-print.h" ++#include "fold-const.h" ++#include "stor-layout.h" ++#include "gimplify.h" ++#include "gimple-iterator.h" ++#include "gimplify-me.h" ++#include "tree-ssa-loop-ivopts.h" ++#include "tree-ssa-loop-manip.h" ++#include "tree-ssa-loop-niter.h" ++#include "tree-ssa-loop.h" ++#include "ssa.h" ++#include "tree-into-ssa.h" ++#include "cfgloop.h" ++#include "tree-scalar-evolution.h" ++#include "langhooks.h" ++#include "tree-inline.h" ++#include "tree-data-ref.h" ++#include "diagnostic-core.h" ++#include "dbgcnt.h" ++#include "gimple-pretty-print.h" ++#include "internal-fn.h" ++#include "tree-cfg.h" ++#include "profile-count.h" ++ ++/* Number of parallel cores. */ ++const unsigned int PARALLEL_NUM = 288; ++ ++/* Indirect access weight. */ ++const unsigned int INDIRECT_ACCESS_VALUE = 2; ++ ++/* Write memory weight. */ ++const unsigned int WRITE_COST = 2; ++ ++/* Prefetch tool input max length. */ ++#ifndef PREFETCH_TOOL_INPUT_MAX_LEN ++#define PREFETCH_TOOL_INPUT_MAX_LEN 512 ++#endif ++ ++/* Prefetch tool number max length. */ ++#ifndef PREFETCH_TOOL_NUM_MAX_LEN ++#define PREFETCH_TOOL_NUM_MAX_LEN 9 ++#endif ++ ++namespace { ++ ++using namespace std; ++ ++/* loop bound info of the memory reference located. */ ++struct loop_bound ++{ ++ /* iv tree_node. */ ++ tree iv; ++ ++ /* define stmt of iv. */ ++ gimple *def_stmt; ++ ++ /* loop where stmt is located. */ ++ class loop *loop; ++ ++ /* loop unroll factor. */ ++ unsigned int unroll; ++ ++ /* Number of iterations of loop. */ ++ tree niters; ++ ++ loop_bound (tree t, gimple *stmt) ++ { ++ iv = t; ++ def_stmt = stmt; ++ loop = loop_containing_stmt (stmt); ++ unroll = 1; ++ niters = chrec_dont_know; ++ } ++}; ++ ++/* method of calculating the data size. */ ++ ++enum calc_type ++{ ++ UNHANDLE_CALC = 0, ++ RUNTIME_CALC, ++ STATIC_CALC ++}; ++ ++/* Describes a info of a memory reference. */ ++ ++struct data_ref ++{ ++ /* The memory reference. */ ++ tree ref; ++ ++ /* Statement where the ref is located. */ ++ gimple *stmt; ++ ++ /* var_decl or param_decl, used for the ref_group. */ ++ tree var; ++ ++ /* Base of the reference. */ ++ tree base; ++ ++ /* Constant offset of the reference. */ ++ tree offset; ++ ++ /* index of the reference. */ ++ tree index; ++ ++ /* Constant step of the reference. */ ++ tree step; ++ ++ /* loop boundary info of each dimension. */ ++ vector loop_bounds; ++ ++ /* memory data size, Unit: MB. */ ++ double data_size; ++ ++ /* method of calculating the data size. */ ++ calc_type calc_by; ++ ++ /* True if the info of ref is traced, and then record it. */ ++ unsigned int trace_status_p : 1; ++ ++ /* True if the loop is vectorized. */ ++ unsigned int vectorize_p : 1; ++ ++ /* True if the memory reference is shared. */ ++ unsigned int parallel_p : 1; ++ ++ /* True if the memory reference is regular. */ ++ unsigned int regular_p : 1; ++ ++ /* True if the memory reference is read. */ ++ unsigned int read_p : 1; ++ ++ data_ref () ++ { ++ ref = NULL_TREE; ++ stmt = NULL; ++ var = NULL_TREE; ++ base = NULL_TREE; ++ offset = NULL_TREE; ++ index = NULL_TREE; ++ step = NULL_TREE; ++ data_size = 0; ++ calc_by = UNHANDLE_CALC; ++ trace_status_p = false; ++ vectorize_p = false; ++ parallel_p = false; ++ regular_p = true; ++ read_p = true; ++ } ++}; ++ ++/* ================ phase 1 get_dense_memory_kernels ================ */ ++ ++/* Add ref node and print. */ ++ ++void ++add_ref (vector &references, tree op, gimple *stmt, ++ bool vectorize_p, bool read_p) ++{ ++ data_ref ref; ++ ref.ref = op; ++ ref.stmt = stmt; ++ ref.vectorize_p = vectorize_p; ++ ref.read_p = read_p; ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ print_generic_expr (dump_file, ref.ref, TDF_LINENO); ++ fprintf (dump_file, "\n"); ++ } ++ references.push_back (ref); ++} ++ ++/* Get the references from the simple call (vectorization type). */ ++ ++void ++get_references_in_gimple_call (gimple *stmt, vector &references) ++{ ++ if (gimple_code (stmt) != GIMPLE_CALL) ++ return; ++ ++ if (gimple_call_internal_p (stmt)) ++ { ++ bool read_p = false; ++ switch (gimple_call_internal_fn (stmt)) ++ { ++ case IFN_MASK_GATHER_LOAD: ++ case IFN_MASK_LOAD: ++ { ++ if (gimple_call_lhs (stmt) == NULL_TREE) ++ return; ++ read_p = true; ++ // FALLTHRU ++ } ++ case IFN_MASK_STORE: ++ { ++ /* _1 = &MEM[base: a_2(D), index: ivtmp_3, step: 8, offset: 0B]; ++ vect__1.1 = .MASK_LOAD (_1, 64B, loop_mask_4); ++ ++ _1 = &MEM[base: a_2(D), index: ivtmp_3, step: 8, offset: 0B]; ++ .MASK_STORE (_1, 64B, loop_mask_4, vect__1.2); ++ ++ _1 = (sizetype) a_2(D); ++ vect_patt_3.3 = .MASK_GATHER_LOAD (_1, vect__4.4, 8, ++ { 0.0, ... }, loop_mask_5); ++ */ ++ tree op1 = gimple_call_arg (stmt, 0); ++ if (TREE_CODE (op1) != SSA_NAME) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "get_references_in_gimple_call: "); ++ fprintf (dump_file, "find base that not ssa_name: "); ++ print_generic_expr (dump_file, op1, TDF_LINENO); ++ fprintf (dump_file, "\n"); ++ } ++ return; ++ } ++ gimple *op1_def = SSA_NAME_DEF_STMT (op1); ++ if (op1_def != NULL && gimple_code (op1_def) == GIMPLE_ASSIGN) ++ { ++ /* &MEM[base: xx] */ ++ tree rhs1 = gimple_assign_rhs1 (op1_def); ++ /* If the definition stmt of the operation is memory ++ reference type, read it directly. */ ++ if (TREE_CODE (rhs1) == ADDR_EXPR ++ && TREE_CODE (TREE_OPERAND (rhs1, 0)) == TARGET_MEM_REF) ++ op1 = TREE_OPERAND (rhs1, 0); /* MEM[base: xx] */ ++ } ++ ++ add_ref (references, op1, stmt, true, read_p); ++ return; ++ } ++ default: ++ return; ++ } ++ } ++} ++ ++/* Stores the locations of memory references in STMT to REFERENCES. */ ++ ++void ++get_references_in_stmt (gimple *stmt, vector &references) ++{ ++ if (!gimple_vuse (stmt)) ++ return; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "gimple_vuse: "); ++ print_gimple_stmt (dump_file, stmt, 0, TDF_LINENO); ++ } ++ ++ if (gimple_code (stmt) == GIMPLE_ASSIGN) ++ { ++ tree op0 = gimple_assign_lhs (stmt); ++ tree op1 = gimple_assign_rhs1 (stmt); ++ tree base = NULL_TREE; ++ ++ /* _1 = MEM[base: a, index: i, step: 8, offset: 0B]; */ ++ if (REFERENCE_CLASS_P (op1) && (base = get_base_address (op1)) ++ && TREE_CODE (base) != SSA_NAME && !is_gimple_min_invariant (base)) ++ add_ref (references, op1, stmt, false, true); ++ ++ if (REFERENCE_CLASS_P (op0) && get_base_address (op0)) ++ add_ref (references, op0, stmt, false, false); ++ } ++ else if (gimple_code (stmt) == GIMPLE_CALL) ++ get_references_in_gimple_call (stmt, references); ++ ++ return; ++} ++ ++/* flag of loop filter out. */ ++ ++struct loop_filter_out_flag ++{ ++ /* Use external gimple. */ ++ bool use_ext_gimple; ++ ++ /* Use external call. */ ++ bool use_ext_call; ++ ++ /* Use external node. */ ++ bool use_ext_node; ++ ++ /* Use loop defined in macros. */ ++ bool use_macro_loop; ++ ++ /* Use external node. */ ++ bool use_cond_func; ++}; ++ ++/* Check whether an external node is used. */ ++ ++bool use_ext_node_p (const vector &references, ++ unsigned int &start) ++{ ++ expanded_location cfun_xloc ++ = expand_location (DECL_SOURCE_LOCATION (current_function_decl)); ++ ++ unsigned i = start; ++ start = references.size (); ++ for (; i < references.size (); i++) ++ { ++ data_ref ref = references[i]; ++ expanded_location xloc = expand_location (ref.stmt->location); ++ if (xloc.file && filename_cmp (cfun_xloc.file, xloc.file)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "use_ext_node\n\n"); ++ return true; ++ } ++ } ++ return false; ++} ++ ++/* Determine whether to filter out loops by stmt. */ ++ ++bool ++filter_out_loop_by_stmt_p (loop_filter_out_flag &loop_filter, gimple *stmt, ++ const vector &references, unsigned int &start) ++{ ++ /* check use_ext_gimple. */ ++ expanded_location cfun_xloc ++ = expand_location (DECL_SOURCE_LOCATION (current_function_decl)); ++ expanded_location xloc = expand_location (stmt->location); ++ if (xloc.file && filename_cmp (cfun_xloc.file, xloc.file)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "use_ext_gimple: "); ++ print_gimple_stmt (dump_file, stmt, 0, TDF_LINENO); ++ } ++ loop_filter.use_ext_gimple = true; ++ return true; ++ } ++ ++ /* check use_ext_call. */ ++ if (gimple_code (stmt) == GIMPLE_CALL && !gimple_call_internal_p (stmt)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "use_ext_call: "); ++ print_gimple_stmt (dump_file, stmt, 0, TDF_LINENO); ++ } ++ loop_filter.use_ext_call = true; ++ return true; ++ } ++ ++ /* check use_macro_loop. */ ++ if (xloc.file && xloc.column != 1) ++ loop_filter.use_macro_loop = false; ++ ++ /* checke use_cond_func, VEC_COND_EXPR/MIN_EXPR/MAX_EXPR. */ ++ if (gimple_code (stmt) == GIMPLE_ASSIGN) ++ { ++ enum tree_code rhs_code = gimple_assign_rhs_code (stmt); ++ if (rhs_code == VEC_COND_EXPR || rhs_code == MIN_EXPR ++ || rhs_code == MAX_EXPR) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "use_cond_func: "); ++ print_gimple_stmt (dump_file, stmt, 0, TDF_LINENO); ++ } ++ loop_filter.use_cond_func = true; ++ return true; ++ } ++ } ++ ++ /* check use_ext_node. */ ++ if (use_ext_node_p (references, start)) ++ { ++ loop_filter.use_ext_node = true; ++ return true; ++ } ++ ++ return false; ++} ++ ++/* Dump the flag type of the loop is filtered out. */ ++ ++void ++dump_loop_filter_out_flag (loop_filter_out_flag &loop_filter) ++{ ++ if (loop_filter.use_ext_gimple) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "non-dense mem access: use_ext_gimple\n"); ++ } ++ if (loop_filter.use_ext_call) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "non-dense mem access: use_ext_call\n"); ++ } ++ ++ if (loop_filter.use_ext_node) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "non-dense mem access: use_ext_node\n"); ++ } ++ ++ if (loop_filter.use_macro_loop) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "non-dense mem access: use_macro_loop\n"); ++ } ++ ++ if (loop_filter.use_cond_func) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "non-dense mem access: use_cond_func\n"); ++ } ++} ++ ++/* Get references in loop. */ ++ ++bool ++get_references_in_loop (vector &references, ++ loop_filter_out_flag &loop_filter, ++ class loop *loop) ++{ ++ unsigned int start = 0; ++ bool filter_out_loop = true; ++ ++ /* Analyze each bb in the loop. */ ++ basic_block *body = get_loop_body_in_dom_order (loop); ++ for (unsigned i = 0; i < loop->num_nodes; i++) ++ { ++ basic_block bb = body[i]; ++ if (bb->loop_father != loop) ++ continue; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\n==== the %dth loop bb body ====\n", i); ++ gimple_dump_bb (dump_file, bb, 0, dump_flags); ++ fprintf (dump_file, "\n"); ++ } ++ ++ gimple_stmt_iterator bsi; ++ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) ++ { ++ gimple *stmt = gsi_stmt (bsi); ++ get_references_in_stmt (stmt, references); ++ filter_out_loop = filter_out_loop_by_stmt_p (loop_filter, stmt, ++ references, start); ++ if (filter_out_loop) ++ break; ++ } ++ if (filter_out_loop) ++ break; ++ } ++ free (body); ++ return !filter_out_loop; ++} ++ ++/* Determine whether the loop is a single path. */ ++ ++bool ++single_path_p (class loop *loop, basic_block bb) ++{ ++ if (bb == NULL) ++ return false; ++ if (bb == loop->latch) ++ return true; ++ ++ gimple *stmt = last_stmt (bb); ++ bool res = false; ++ ++ if (stmt && gimple_code (stmt) == GIMPLE_COND) ++ { ++ gcc_assert (EDGE_COUNT (bb->succs) == 2); ++ edge true_edge = NULL; ++ edge false_edge = NULL; ++ extract_true_false_edges_from_block (bb, &true_edge, &false_edge); ++ ++ /* Returns false, if a branch occurs. */ ++ if (true_edge->dest->loop_father == loop ++ && false_edge->dest->loop_father == loop) ++ return false; ++ ++ if (true_edge->dest->loop_father == loop) ++ res = single_path_p (loop, true_edge->dest); ++ else ++ res = single_path_p (loop, false_edge->dest); ++ } ++ else ++ { ++ edge e = find_fallthru_edge (bb->succs); ++ if (e) ++ res = single_path_p (loop, e->dest); ++ } ++ return res; ++} ++ ++/* Computes an estimated number of insns in LOOP, weighted by WEIGHTS. ++ Assume that the HPC data reading and calculation process does not involve ++ adding branches in loops. Therefore, all bbs of loops are directly used for ++ calculation (excluding embedded loops) without considering branch weighting. ++*/ ++ ++unsigned ++estimate_loop_insns (class loop *loop, eni_weights *weights) ++{ ++ basic_block *body = get_loop_body (loop); ++ gimple_stmt_iterator gsi; ++ unsigned size = 0, i; ++ ++ for (i = 0; i < loop->num_nodes; i++) ++ { ++ basic_block bb = body[i]; ++ if (bb->loop_father != loop) ++ { ++ continue; ++ } ++ for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi); gsi_next (&gsi)) ++ size += estimate_num_insns (gsi_stmt (gsi), weights); ++ } ++ free (body); ++ ++ return size; ++} ++ ++/* Check whether the memory access is dense. */ ++ ++bool ++dense_memory_p (const vector &references, class loop *loop) ++{ ++ int ref_count = references.size (); ++ unsigned int ninsns = estimate_loop_insns (loop, &eni_size_weights); ++ float mem_to_insn_ratio = (float)ref_count / (float)ninsns; ++ ++ /* The number of cores to be run and DDR bandwidth information can be ++ transferred to flexibly adjust the threshold. */ ++ bool dense_mem = (mem_to_insn_ratio >= (param_mem_access_ratio / 100.0) ++ && ref_count >= param_mem_access_num); ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ const char *fn_name = IDENTIFIER_POINTER (DECL_NAME (cfun->decl)); ++ ++ /* Dump dense memory source code location. */ ++ if (ref_count && references[0].stmt->location) ++ { ++ expanded_location xloc = expand_location ++ (references[0].stmt->location); ++ int fn_start = 0; ++ if (DECL_SOURCE_LOCATION (current_function_decl)) ++ fn_start = expand_location ( ++ DECL_SOURCE_LOCATION (current_function_decl)).line; ++ int fn_end = fn_start; ++ if (cfun->function_end_locus) ++ fn_end = expand_location (cfun->function_end_locus).line; ++ if (xloc.file) ++ fprintf (dump_file, "[%s:%s(%d-%d):%d:%d] ", ++ xloc.file, fn_name, fn_start, fn_end, ++ xloc.line, xloc.column); ++ } ++ ++ /* Dump memory dense information. */ ++ if (dense_mem) ++ fprintf (dump_file, "dense memory access: "); ++ else ++ fprintf (dump_file, "non-dense mem access: "); ++ fprintf (dump_file, ++ "ref_count = %d, ninsns = %d, mem_to_insn_ratio = %f\n\n", ++ ref_count, ninsns, mem_to_insn_ratio); ++ } ++ ++ return dense_mem; ++} ++ ++/* Analyze the inner loop and get the loop with dense memory access. */ ++ ++bool ++get_dense_memory_kernels (vector &kernels, ++ map > &kernels_refs) ++{ ++ if (dump_file) ++ fprintf (dump_file, "\nPhase 1: get_dense_memory_kernels\n\n"); ++ class loop *loop = NULL; ++ FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) ++ { ++ number_of_latch_executions (loop); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\n========== Processing loop %d: ==========\n", ++ loop->num); ++ loop_dump (dump_file, loop); ++ flow_loop_dump (loop, dump_file, NULL, 1); ++ fprintf (dump_file, "loop unroll: %d\n", loop->unroll); ++ } ++ ++ if (get_loop_exit_edges (loop).length () != 1 ++ || !single_path_p (loop, loop->header)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "non-dense mem access: loop_branching\n"); ++ continue; ++ } ++ ++ vector references; ++ loop_filter_out_flag loop_filter = {false, false, false, true, false}; ++ ++ if (!get_references_in_loop (references, loop_filter, loop)) ++ { ++ dump_loop_filter_out_flag (loop_filter); ++ continue; ++ } ++ ++ if (dense_memory_p (references, loop)) ++ { ++ kernels_refs[loop] = references; ++ kernels.push_back (loop); ++ } ++ } ++ return kernels.size () > 0; ++} ++ ++/* ================ phase 2 trace_data_refs_info ================ */ ++ ++/* Determine whether the declaration is a non-vectorized. */ ++ ++bool ++generic_decl_p (tree expr) ++{ ++ if (expr == NULL_TREE) ++ return false; ++ enum tree_code expr_code = TREE_CODE (expr); ++ if (expr_code != VAR_DECL && expr_code != PARM_DECL ++ && expr_code != COMPONENT_REF) ++ return false; ++ ++ tree type = TREE_TYPE (expr); ++ while (type) ++ { ++ if (TREE_CODE (type) != VECTOR_TYPE) ++ /* TREE_TYPE (NODE) ( ++ CONTAINS_STRUCT_CHECK (NODE, TS_TYPED)->typed.type) */ ++ type = CONTAINS_STRUCT_CHECK (type, TS_TYPED) ? TREE_TYPE (type) : NULL; ++ else ++ return false; ++ } ++ return true; ++} ++ ++/* Initial worklist preparation for source variable tracing. ++ Add different initial node based on different gimple statements. */ ++ ++void ++add_worklist (vector &worklist, set &walked, gimple *def_stmt) ++{ ++ if (gimple_code (def_stmt) == GIMPLE_PHI) ++ { ++ for (unsigned i = 0; i < gimple_phi_num_args (def_stmt); i++) ++ { ++ tree node = gimple_phi_arg_def (def_stmt, i); ++ if (!walked.count (node)) ++ { ++ worklist.push_back (node); ++ walked.insert (node); ++ } ++ } ++ } ++ else if (is_gimple_assign (def_stmt)) ++ { ++ tree_code rhs_code = gimple_assign_rhs_code (def_stmt); ++ if (rhs_code == POINTER_PLUS_EXPR || rhs_code == NEGATE_EXPR ++ || rhs_code == NOP_EXPR || rhs_code == SSA_NAME ++ || rhs_code == COMPONENT_REF) ++ { ++ tree node = gimple_assign_rhs1 (def_stmt); ++ if (!walked.count (node)) ++ { ++ worklist.push_back (node); ++ walked.insert (node); ++ } ++ } ++ else if (rhs_code == PLUS_EXPR || rhs_code == MINUS_EXPR) ++ { ++ tree node = gimple_assign_rhs1 (def_stmt); ++ if (!walked.count (node)) ++ { ++ worklist.push_back (node); ++ walked.insert (node); ++ } ++ node = gimple_assign_rhs2 (def_stmt); ++ if (!walked.count (node)) ++ { ++ worklist.push_back (node); ++ walked.insert (node); ++ } ++ } ++ else ++ { ++ /* unhandled assign rhs_code: _219 = _17 * _70; ++ _17 = *grid_56(D).sst.span; ++ _70 = *grid_56(D).sst.dim[0].stride; ++ */ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "unhandled assign rhs_code: "); ++ print_gimple_stmt (dump_file, def_stmt, 0, TDF_LINENO); ++ fprintf (dump_file, "\n"); ++ } ++ } ++ } ++ else ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "unsupported tracing stmt: "); ++ print_gimple_stmt (dump_file, def_stmt, 0, TDF_LINENO); ++ fprintf (dump_file, "\n"); ++ } ++ } ++} ++ ++ ++/* Tracing source variables: ++ vectp.1 = a_2(D) + _3; ++ _4 = &MEM[base: vectp.1, index: ivtmp_5, step: 8, offset: 0B]; ++ vect__1.6 = .MASK_LOAD (_4, 64B, loop_mask_7); ++ ++ _1 = (sizetype) b_2(D); ++ vect_patt_3.3 = .MASK_GATHER_LOAD (_1, vect__4.4, 8, { 0.0, ... }, ++ loop_mask_5); ++ ... ++ Due to previous pass optimizations, the current tracing method can find ++ several source variable candidates. We decide to record them in a map and ++ later filter out the true base variable by some criteria. ++*/ ++ ++void ++trace_base_var_helper (tree arg, set &walked, ++ map& base_var_candid) ++{ ++ if (arg == NULL) ++ return; ++ ++ /* Array type. */ ++ tree op0 = NULL; ++ if (TREE_CODE (arg) == ADDR_EXPR ++ && (op0 = TREE_OPERAND (arg, 0)) && generic_decl_p (op0)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "array type\n"); ++ base_var_candid[op0] += 1; ++ return; ++ } ++ ++ /* Pointer type. */ ++ if (TREE_CODE (TREE_TYPE (arg)) == POINTER_TYPE && generic_decl_p (arg)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "pointer type\n"); ++ base_var_candid[arg] += 1; ++ return; ++ } ++ ++ /* SSA_NAME type. */ ++ if (TREE_CODE (arg) != SSA_NAME) ++ return; ++ ++ tree tmp_var = SSA_NAME_VAR (arg); ++ if (tmp_var && generic_decl_p (tmp_var) ++ && TREE_CODE (TREE_TYPE (arg)) == POINTER_TYPE) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "ssa pointer type\n"); ++ base_var_candid[tmp_var] += 1; ++ return; ++ } ++ ++ gimple *def_stmt = SSA_NAME_DEF_STMT (arg); ++ if (def_stmt == NULL) ++ return; ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ print_generic_expr (dump_file, arg, TDF_SLIM); ++ fprintf (dump_file, "\t\t: "); ++ print_gimple_stmt (dump_file, def_stmt, 0, TDF_SLIM); ++ } ++ ++ vector worklist; ++ add_worklist (worklist, walked, def_stmt); ++ for (unsigned i = 0; i < worklist.size (); ++i) ++ trace_base_var_helper (worklist[i], walked, base_var_candid); ++} ++ ++/* Identify the base variable traced from base address of memory reference. ++ We recognize that current method could detect several base variable ++ candidates and the temporary criteria for base variable determination ++ is that either one of the following statement is true: ++ 1. The number of base variable candidates is 1; ++ 2. The number of detected gimple statements for some variable is 1. ++ We may use other criteria or relax the current criteria ++ (e.g., criterion 2: 1 -> any odd number). */ ++ ++bool ++trace_base_var (tree &var, tree arg, set &walked) ++{ ++ map base_var_candid; ++ trace_base_var_helper (arg, walked, base_var_candid); ++ bool is_tracing_unusual = false; ++ if (base_var_candid.size () == 1) ++ var = base_var_candid.begin ()->first; ++ else ++ { ++ is_tracing_unusual = true; ++ for (const pair& base_var_count : base_var_candid) ++ if (base_var_count.second == 1) ++ var = base_var_count.first; ++ } ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "Traced variables at "); ++ print_generic_expr (dump_file, arg, TDF_SLIM); ++ fprintf (dump_file, ":\n"); ++ for (const pair& base_var_count : base_var_candid) ++ fprintf (dump_file, "%s:%d, ", get_name (base_var_count.first), ++ base_var_count.second); ++ fprintf (dump_file, "\n"); ++ ++ if (var == NULL_TREE) ++ fprintf (dump_file, "Unhandled scenario for tracing base variable.\n"); ++ else if (is_tracing_unusual && var != NULL_TREE) ++ fprintf (dump_file, "Tracing unusual number or occurrences of base " ++ "variables. Choose %s.\n", get_name (var)); ++ } ++ return var != NULL_TREE; ++} ++ ++/* Tracing direct memory reference information. */ ++ ++bool ++trace_direct_mem_ref (data_ref &mem_ref, set &traced_ref_stmt) ++{ ++ if (TREE_CODE (mem_ref.ref) != TARGET_MEM_REF) ++ return false; ++ ++ /* Direct memory access, regardless of whether it is in vectorized form, ++ can be determined through TARGET_MEM_REF. */ ++ mem_ref.base = TREE_OPERAND (mem_ref.ref, 0); ++ mem_ref.offset = TREE_OPERAND (mem_ref.ref, 1); ++ mem_ref.index = TREE_OPERAND (mem_ref.ref, 2); ++ mem_ref.step = TREE_OPERAND (mem_ref.ref, 3); ++ ++ set walked; ++ if (mem_ref.var == NULL_TREE ++ && !trace_base_var (mem_ref.var, mem_ref.base, walked)) ++ return false; ++ ++ traced_ref_stmt.insert (mem_ref.stmt); ++ return true; ++} ++ ++/* Recursively trace and check whether the definition stmt of the ++ index operand is a recorded stmt in direct access tracing. ++ If true, it is an indirect access. */ ++ ++bool ++trace_indirect_operand (tree arg, set &traced_ref_stmt) ++{ ++ if (TREE_CODE (arg) != SSA_NAME) ++ return false; ++ ++ gimple *def_stmt = SSA_NAME_DEF_STMT (arg); ++ ++ if (traced_ref_stmt.count (def_stmt)) ++ return true; ++ ++ if (!def_stmt || !is_gimple_assign (def_stmt)) ++ return false; ++ ++ tree_code rhs_code = gimple_assign_rhs_code (def_stmt); ++ /* Collect a whitelist of gimple_assign_rhs_code for tracing pointer/array ++ type indirect memory access. Please check examples before function ++ trace_indirect_ptr and trace_indirect_array. */ ++ if (rhs_code != MULT_EXPR && rhs_code != NOP_EXPR ++ && rhs_code != CONVERT_EXPR && rhs_code != PLUS_EXPR ++ && rhs_code != ARRAY_REF) ++ return false; ++ ++ tree op = NULL_TREE; ++ ssa_op_iter iter; ++ FOR_EACH_SSA_TREE_OPERAND (op, def_stmt, iter, SSA_OP_USE) ++ { ++ if (trace_indirect_operand (op, traced_ref_stmt)) ++ return true; ++ } ++ return false; ++} ++ ++/* Trace the pointer of the indirect memory access: ++ 1) obtain the base address of the indirect memory access. ++ 2) ensure that the index has been traced in the direct memory access. ++ ++ _1 = MEM[base: a_2(D), index: ivtmp.3_3, step: 4, offset: 0B]; // Traced in ++ direct access ++ _4 = (long unsigned int) _1; ++ _5 = _4 * 8; ++ _6 = p(D) + _5; // get base ++ _7 = *_6; // start tracing ++*/ ++ ++bool ++trace_indirect_ptr (tree &base, tree &index, tree arg, ++ set traced_ref_stmt) ++{ ++ gimple *def_stmt = SSA_NAME_DEF_STMT (arg); ++ ++ if (!def_stmt || !is_gimple_assign (def_stmt)) ++ return false; ++ ++ tree_code rhs_code = gimple_assign_rhs_code (def_stmt); ++ if (rhs_code != POINTER_PLUS_EXPR) ++ return false; ++ ++ /* POINTER_PLUS_EXPR, The first operand is always a pointer/reference type. ++ The second operand is always an unsigned integer type compatible with ++ sizetype. */ ++ base = gimple_assign_rhs1 (def_stmt); ++ index = gimple_assign_rhs2 (def_stmt); ++ ++ return trace_indirect_operand (index, traced_ref_stmt); ++} ++ ++/* Trace the array of the indirect memory access: ++ 1) obtain the base address of the indirect memory access. ++ 2) ensure that the index has been traced in the direct memory access. ++ ++ _1 = MEM[base: a_2(D), index: ivtmp.3_3, step: 4, offset: 0B]; // Traced in ++ direct access ++ _4 = (integer(kind=8)) _1; ++ _5 = _4 + 135; ++ _6 = p[_5]; // start tracing ++*/ ++ ++bool ++trace_indirect_array (tree &base, tree &index, ++ set traced_ref_stmt, tree ref) ++{ ++ if (TREE_CODE (ref) != ARRAY_REF) ++ return false; ++ base = TREE_OPERAND (ref, 0); ++ index = TREE_OPERAND (ref, 1); ++ return trace_indirect_operand (index, traced_ref_stmt); ++} ++ ++/* Tracing indirect memory reference information. ++ Include tracing of base addresses and source variable. ++ _x(ssa name) -> a_2(base addr) -> a(src var) */ ++ ++bool ++trace_indirect_mem_ref (data_ref &mem_ref, ++ set &traced_ref_stmt) ++{ ++ /* Processing of vectorization types. */ ++ if (mem_ref.vectorize_p) ++ { ++ tree op = gimple_call_arg (mem_ref.stmt, 1); ++ if (trace_indirect_operand (op, traced_ref_stmt)) ++ { ++ mem_ref.base = gimple_call_arg (mem_ref.stmt, 0); ++ mem_ref.regular_p = false; ++ set walked; ++ if (mem_ref.var == NULL_TREE ++ && !trace_base_var (mem_ref.var, mem_ref.base, walked)) ++ return false; ++ return true; ++ } ++ return false; ++ } ++ ++ /* Processing of non-vectorized types. */ ++ tree op = NULL_TREE; ++ ssa_op_iter iter; ++ FOR_EACH_SSA_TREE_OPERAND (op, mem_ref.stmt, iter, SSA_OP_USE) ++ { ++ ++ /* Array type: ++ _1 = MEM[base: a_2(D), index: ivtmp.3_3, step: 4, offset: 0B]; ++ _4 = c[_1]; ++ ++ Pointer type: ++ _1 = MEM[base: a_2(D), index: ivtmp.3_3, step: 4, offset: 0B]; ++ _4 = (long unsigned int) _1; ++ _5 = _4 * 8; ++ _6 = p(D) + _5; ++ _7 = *_6; ++ */ ++ tree base = NULL_TREE; ++ tree index = NULL_TREE; ++ if (trace_indirect_array (base, index, traced_ref_stmt, mem_ref.ref) ++ || trace_indirect_ptr (base, index, op, traced_ref_stmt)) ++ { ++ /* ARRAY_REF, The first operand is the array; ++ the second is the index. */ ++ mem_ref.base = base; ++ mem_ref.index = index; ++ mem_ref.regular_p = false; ++ set walked; ++ if (mem_ref.var == NULL_TREE ++ && !trace_base_var (mem_ref.var, mem_ref.base, walked)) ++ return false; ++ return true; ++ } ++ } ++ ++ return false; ++} ++ ++/* Trace references base info: ++ 1) Parallel analysis ++ 2) Memory access rule analysis ++ 3) Tracing base address and source variable of memory references ++ We will extend parallel analysis later. ++*/ ++ ++void ++trace_ref_info (data_ref &mem_ref, set &traced_ref_stmt) ++{ ++ enum tree_code ref_code = TREE_CODE (mem_ref.ref); ++ if (/* Vectorized and non-vectorized direct access. */ ++ ref_code != TARGET_MEM_REF ++ /* non-vectorized indirect memory access. */ ++ && ref_code != MEM_REF && ref_code != ARRAY_REF ++ /* vectorized indirect memory access. */ ++ && ref_code != SSA_NAME) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "ref is another tree-code: "); ++ fprintf (dump_file, "stmt: "); ++ print_gimple_stmt (dump_file, mem_ref.stmt, 0, TDF_LINENO); ++ fprintf (dump_file, "ref: "); ++ print_generic_expr (dump_file, mem_ref.ref, TDF_LINENO); ++ fprintf (dump_file, "\n"); ++ } ++ return; ++ } ++ ++ /* 1) Direct and indirect access traces and traces source variables. */ ++ if (!trace_direct_mem_ref (mem_ref, traced_ref_stmt) ++ && !trace_indirect_mem_ref (mem_ref, traced_ref_stmt)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Tracing failed.\n\n"); ++ return; ++ } ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Tracing succeeded.\n\n"); ++ mem_ref.trace_status_p = true; ++} ++ ++/* Tracing and sorting reference groups. */ ++ ++void ++trace_data_refs_info (vector &kernels, ++ map > &loop_refs) ++{ ++ if (dump_file) ++ fprintf (dump_file, "\nPhase 2: trace_all_references_info\n\n"); ++ ++ set traced_ref_stmt; ++ ++ for (unsigned i = 0; i < kernels.size (); ++i) ++ { ++ class loop* loop = kernels[i]; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "loop header %d:\n", loop->header->index); ++ for (unsigned j = 0; j < loop_refs[loop].size (); ++j) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "trace_references_base_info %d:\n", j); ++ print_generic_expr (dump_file, loop_refs[loop][j].ref, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ trace_ref_info (loop_refs[loop][j], traced_ref_stmt); ++ } ++ } ++} ++ ++/* ================ phase 3 analyze_nested_kernels ================ */ ++ ++/* Return the inner most type for arrays and pointers of TYPE. */ ++ ++tree ++inner_type (tree type) ++{ ++ while (POINTER_TYPE_P (type) ++ || TREE_CODE (type) == ARRAY_TYPE) ++ type = TREE_TYPE (type); ++ return type; ++} ++ ++/* Check whether the input iv is the loop dimension boundary. */ ++ ++bool ++loop_bound_iv_p (tree t, tree &outer_loop_t) ++{ ++ if (t == NULL || TREE_CODE (t) != SSA_NAME ++ || TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE) ++ return false; ++ ++ gimple *def_stmt = SSA_NAME_DEF_STMT (t); ++ if (gimple_code (def_stmt) != GIMPLE_PHI) ++ return false; ++ ++ /* Filter scenarios with only two phi inputs. */ ++ if (gimple_phi_num_args (def_stmt) != 2) ++ return false; ++ ++ gphi *phi_stmt = as_a (def_stmt); ++ basic_block src0 = gimple_phi_arg_edge (phi_stmt, 0)->src; ++ basic_block src1 = gimple_phi_arg_edge (phi_stmt, 1)->src; ++ ++ class loop *loop = loop_containing_stmt (def_stmt); ++ bool res = false; ++ /* Two phi inputs, one from the current loop and one from the outer loop. */ ++ if ((src0->loop_father == loop) && (src1->loop_father == loop_outer (loop))) ++ { ++ outer_loop_t = gimple_phi_arg_def (def_stmt, 1); ++ res = true; ++ } ++ else if ((src1->loop_father == loop) ++ && (src0->loop_father == loop_outer (loop))) ++ { ++ outer_loop_t = gimple_phi_arg_def (def_stmt, 0); ++ res = true; ++ } ++ ++ if (res) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "===> "); ++ print_gimple_stmt (dump_file, def_stmt, 0, TDF_SLIM); ++ } ++ return true; ++ } ++ return false; ++} ++ ++/* add worklist and walked list. */ ++ ++void ++add_worklist_walked (vector &worklist, set &walked, tree node) ++{ ++ if (!walked.count (node)) ++ { ++ worklist.push_back (node); ++ /* Avoid phi node cycle introduction, which makes the worklist unable ++ to end. */ ++ walked.insert (node); ++ } ++} ++ ++/* check bound iv and add worklist. */ ++ ++void ++check_bound_iv_and_add_worklist (vector &worklist, set &walked, ++ tree t, data_ref &mem_ref) ++{ ++ if (TREE_CODE (t) != SSA_NAME) ++ return; ++ ++ gimple *def_stmt = SSA_NAME_DEF_STMT (t); ++ if (def_stmt == NULL) ++ return; ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ print_generic_expr (dump_file, t, TDF_SLIM); ++ fprintf (dump_file, "\t\t: "); ++ print_gimple_stmt (dump_file, def_stmt, 0, TDF_SLIM); ++ } ++ ++ if (gimple_code (def_stmt) == GIMPLE_PHI) ++ { ++ tree out_loop_t = NULL_TREE; ++ if (loop_bound_iv_p (t, out_loop_t)) ++ { ++ mem_ref.loop_bounds.push_back (loop_bound (t, def_stmt)); ++ add_worklist_walked (worklist, walked, out_loop_t); ++ } ++ } ++ else if (is_gimple_assign (def_stmt)) ++ { ++ tree_code rhs_code = gimple_assign_rhs_code (def_stmt); ++ ++ /* unary. */ ++ if (rhs_code == SSA_NAME || rhs_code == NOP_EXPR) ++ add_worklist_walked (worklist, walked, gimple_assign_rhs1 (def_stmt)); ++ else if (rhs_code == POINTER_PLUS_EXPR) ++ add_worklist_walked (worklist, walked, gimple_assign_rhs2 (def_stmt)); ++ ++ /* binary. */ ++ else if (rhs_code == PLUS_EXPR || rhs_code == MINUS_EXPR ++ || rhs_code == MULT_EXPR) ++ { ++ add_worklist_walked (worklist, walked, gimple_assign_rhs1 (def_stmt)); ++ add_worklist_walked (worklist, walked, gimple_assign_rhs2 (def_stmt)); ++ } ++ } ++} ++ ++/* DFS trace the loop bound of iv. */ ++ ++bool ++trace_loop_bound_iv (data_ref &mem_ref) ++{ ++ /* Indirect memory access, the size cannot be determined based on the loop ++ boundary. */ ++ if (!mem_ref.regular_p) ++ return false; ++ ++ /* Determine and record the boundary iv of the current index, ++ but do not trace it. */ ++ tree outer_loop_t = NULL_TREE; ++ if (loop_bound_iv_p (mem_ref.index, outer_loop_t)) ++ mem_ref.loop_bounds.push_back ( ++ loop_bound (mem_ref.index, SSA_NAME_DEF_STMT (mem_ref.index))); ++ ++ vector worklist; ++ worklist.push_back (mem_ref.base); ++ set walked; ++ ++ while (worklist.size ()) ++ { ++ tree t = worklist.back (); ++ worklist.pop_back (); ++ ++ /* add worklist. */ ++ check_bound_iv_and_add_worklist (worklist, walked, t, mem_ref); ++ } ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nmem_ref access dimension: %ld\n", ++ mem_ref.loop_bounds.size ()); ++ ++ return mem_ref.loop_bounds.size () > 0; ++} ++ ++/* dump loop bound. */ ++ ++void ++loop_bound_dump (FILE *file, loop_bound &lb) ++{ ++ class loop *loop = lb.loop; ++ fprintf (file, "loop_bound: loop_%d (", loop->num); ++ if (loop->header) ++ fprintf (file, "header = %d", loop->header->index); ++ else ++ { ++ fprintf (file, "deleted)\n"); ++ return; ++ } ++ if (loop->latch) ++ fprintf (file, ", latch = %d", loop->latch->index); ++ fprintf (file, ", lb_niters = "); ++ print_generic_expr (file, lb.niters); ++ fprintf (file, ")\n"); ++} ++ ++/* static calculate data size. */ ++ ++void ++static_calculate_data_size (data_ref &mem_ref) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nstatic_calculate_data_size\n"); ++ ++ tree size_unit = TYPE_SIZE_UNIT (inner_type (TREE_TYPE (mem_ref.var))); ++ HOST_WIDE_INT type_size = size_unit ? tree_to_uhwi (size_unit) : 0; ++ for (unsigned i = 0; i < mem_ref.loop_bounds.size (); ++i) ++ { ++ HOST_WIDE_INT est_niter = tree_to_uhwi (mem_ref.loop_bounds[i].niters); ++ unsigned int unroll = mem_ref.loop_bounds[i].unroll; ++ if (i == 0) ++ { ++ /* The unit conversion between byte, kilobytes, and megabytes is ++ 1024. */ ++ mem_ref.data_size = double (type_size ++ * est_niter * unroll) / 1024 / 1024; ++ } ++ else ++ mem_ref.data_size *= est_niter * unroll; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "static_data_size: %lf\n", mem_ref.data_size); ++ } ++} ++ ++/* Recursive tracing and creating of dominant nodes. */ ++ ++tree ++trace_and_create_dominate_expr (tree expr, class loop *outermost) ++{ ++ if (expr == NULL_TREE || is_gimple_constant (expr)) ++ return expr; ++ ++ if (TREE_CODE (expr) != SSA_NAME) ++ return NULL_TREE; ++ ++ if (SSA_NAME_IS_DEFAULT_DEF (expr)) ++ return expr; ++ ++ gimple *stmt = SSA_NAME_DEF_STMT (expr); ++ basic_block def_bb = gimple_bb (stmt); ++ if (def_bb == NULL || def_bb->loop_father == NULL) ++ return NULL_TREE; ++ ++ if (dominated_by_p (CDI_DOMINATORS, outermost->header, def_bb)) ++ return expr; ++ ++ if (gimple_code (stmt) != GIMPLE_ASSIGN) ++ return NULL_TREE; ++ ++ enum tree_code rhs_code = gimple_assign_rhs_code (stmt); ++ tree_code_class code_class = TREE_CODE_CLASS (rhs_code); ++ tree type = TREE_TYPE (gimple_assign_lhs (stmt)); ++ tree rhs1 = trace_and_create_dominate_expr ++ (gimple_assign_rhs1 (stmt), outermost); ++ if (rhs1 == NULL_TREE) ++ return NULL_TREE; ++ ++ if (code_class == tcc_unary) ++ { ++ tree expr_new = build1 (rhs_code, type, rhs1); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "expr_new = "); ++ print_generic_expr (dump_file, expr_new, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ return expr_new; ++ } ++ else if (code_class == tcc_binary) ++ { ++ tree rhs2 = trace_and_create_dominate_expr ++ (gimple_assign_rhs2 (stmt), outermost); ++ if (rhs2 == NULL_TREE) ++ return NULL_TREE; ++ ++ tree expr_new = fold_build2 (rhs_code, type, rhs1, rhs2); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "expr_new = "); ++ print_generic_expr (dump_file, expr_new, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ return expr_new; ++ } ++ ++ return NULL_TREE; ++} ++ ++/* Recursive parsing and craating of nodes in expr expressions. */ ++ ++tree ++parse_and_create_expr (tree expr, class loop *outermost) ++{ ++ if (expr == NULL_TREE || expr == chrec_dont_know ++ || is_gimple_constant (expr) || TREE_CODE (expr) == ADDR_EXPR) ++ { ++ /* tcc_expression (e.g., &q) situation combined with tcc_unary. */ ++ if (TREE_CODE (expr) == ADDR_EXPR && dump_file ++ && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "tcc_expression case in ADDR_EXPR: "); ++ print_generic_expr (dump_file, expr, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ return expr; ++ } ++ ++ if (TREE_CODE (expr) == SSA_NAME) ++ return trace_and_create_dominate_expr (expr, outermost); ++ else if (EXPR_P (expr)) ++ { ++ enum tree_code tree_code = TREE_CODE (expr); ++ tree_code_class code_class = TREE_CODE_CLASS (tree_code); ++ tree type = TREE_TYPE (expr); ++ tree op1 = parse_and_create_expr (TREE_OPERAND (expr, 0), outermost); ++ if (op1 == NULL_TREE) ++ return NULL_TREE; ++ ++ if (code_class == tcc_unary) ++ { ++ tree expr_new = build1 (tree_code, type, op1); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "expr_new = "); ++ print_generic_expr (dump_file, expr_new, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ return expr_new; ++ } ++ else if (code_class == tcc_binary) ++ { ++ tree op2 = parse_and_create_expr (TREE_OPERAND (expr, 1), outermost); ++ if (op2 == NULL_TREE) ++ return NULL_TREE; ++ ++ tree expr_new = fold_build2 (tree_code, type, op1, op2); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "expr_new = "); ++ print_generic_expr (dump_file, expr_new, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ return expr_new; ++ } ++ } ++ return NULL_TREE; ++} ++ ++/* Trace and creat dominate loop bounds. */ ++ ++void ++trace_and_create_dominate_loop_bounds (data_ref &mem_ref) ++{ ++ /* Check whether the niters is a loop dominant. ++ If not, trace and determine whether the result is dominant. If yes, create ++ the expr of the dominant node. ++ */ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\ntrace_and_create_dominate_loop_bounds\n"); ++ ++ /* Determine the relationship between the boundary of the innermost loop and ++ the dominant of the outer loop and the processing. */ ++ loop_bound &outermost = mem_ref.loop_bounds.back (); ++ for (unsigned i = 0; i < mem_ref.loop_bounds.size (); ++i) ++ { ++ loop_bound ¤t = mem_ref.loop_bounds[i]; ++ tree &niters = current.niters; ++ if (TREE_CODE (niters) == COND_EXPR) ++ niters = TREE_OPERAND (niters, 1); ++ ++ niters = parse_and_create_expr (niters, outermost.loop); ++ ++ if (niters == NULL_TREE) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ print_generic_expr (dump_file, mem_ref.ref, TDF_SLIM); ++ fprintf (dump_file, "Tracing loop bound failed at dimension %d", ++ i); ++ } ++ mem_ref.calc_by = UNHANDLE_CALC; ++ break; ++ } ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ loop_bound_dump (dump_file, mem_ref.loop_bounds[i]); ++ } ++} ++ ++/* trace the dimension and corresponding loop bounds of mem_ref. ++ This function is used to supplement the information of mem_ref.loop_bounds. ++*/ ++ ++void ++trace_ref_dimension_and_loop_bounds (data_ref &mem_ref) ++{ ++ /* In the same loop, some memory access dimensions are different. Remove ++ variables with fewer dimensions. ++ Previous cyclic filtering conditions and memory access node records and ++ tracing. ++ The false result is also processed. ++ */ ++ if (dump_file) ++ fprintf (dump_file, "\ncalculate_data_size\n"); ++ ++ /* Trace the loop bound iv of ref to determine the dimension. */ ++ /* Record data from the loop perspective to avoid repeated tracing. */ ++ if (!trace_loop_bound_iv (mem_ref)) ++ return; ++ ++ /* The traced mem_ref may have multiple dimensions, which corresponds to ++ multiple loops. */ ++ /* And in the dimension-by-dimensional analysis, the computable way is ++ continuously reduced. */ ++ mem_ref.calc_by = STATIC_CALC; ++ for (unsigned i = 0; i < mem_ref.loop_bounds.size (); ++i) ++ { ++ class loop *loop = mem_ref.loop_bounds[i].loop; ++ tree &niters = mem_ref.loop_bounds[i].niters; ++ ++ /* Set NULL_TREE to ensure that nb_iterations are retraced and ++ vec_nb_iterations are also extracted. */ ++ loop->nb_iterations = NULL_TREE; ++ niters = number_of_latch_executions (loop, false); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ loop_dump (dump_file, loop); ++ ++ if (loop->unroll) ++ { ++ if (loop->unroll == USHRT_MAX && dump_file ++ && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "loop->unroll = USHRT_MAX = %d", USHRT_MAX); ++ mem_ref.loop_bounds[i].unroll = loop->unroll; ++ } ++ ++ if ((niters == chrec_dont_know) && loop->vec_nb_iterations ++ && (loop->vec_nb_iterations != chrec_dont_know)) ++ niters = loop->vec_nb_iterations; ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ loop_bound_dump (dump_file, mem_ref.loop_bounds[i]); ++ ++ if (niters == NULL_TREE || niters == chrec_dont_know) ++ mem_ref.calc_by = min (mem_ref.calc_by, UNHANDLE_CALC); ++ else if (TREE_CODE (niters) != INTEGER_CST) ++ mem_ref.calc_by = min (mem_ref.calc_by, RUNTIME_CALC); ++ else ++ mem_ref.calc_by = min (mem_ref.calc_by, STATIC_CALC); ++ } ++ ++ if (mem_ref.calc_by == RUNTIME_CALC) ++ trace_and_create_dominate_loop_bounds (mem_ref); ++ else if (mem_ref.calc_by == STATIC_CALC) ++ static_calculate_data_size (mem_ref); ++} ++ ++/* analyze nested kernels. ++ 1. multidimension loop analyze. ++ 2. extended outer loop analyze. ++ Later we will extend outer loop analysis. ++*/ ++ ++bool ++analyze_nested_kernels (vector &kernels, ++ map > &loop_refs) ++{ ++ if (dump_file) ++ fprintf (dump_file, "\nPhase 3: analyze_nested_kernels\n\n"); ++ ++ for (unsigned i = 0; i < kernels.size (); ++i) ++ { ++ class loop* loop = kernels[i]; ++ if (loop_refs.count (loop) == 0) ++ continue; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\n\nloop header %d:\n", loop->header->index); ++ for (unsigned j = 0; j < loop_refs[loop].size (); ++j) ++ { ++ if (loop_refs[loop][j].trace_status_p == false) ++ continue; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\ntrace_reference_dimension at mem_ref " ++ "index %d in loop %d:\n", j, loop->num); ++ print_generic_expr (dump_file, loop_refs[loop][j].ref, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ trace_ref_dimension_and_loop_bounds (loop_refs[loop][j]); ++ } ++ ++ } ++ return true; ++} ++ ++/* ================ phase 4 filter_and_sort_kernels ================ */ ++ ++/* Get the edge probability information of each basic block in the loop. */ ++ ++float ++get_edge_prob (edge e, float minimum) ++{ ++ float fvalue = 0; ++ ++ profile_probability probability = e->probability; ++ if (probability.initialized_p ()) ++ { ++ fvalue = probability.to_reg_br_prob_base () / float (REG_BR_PROB_BASE); ++ if (fvalue < minimum && probability.to_reg_br_prob_base ()) ++ fvalue = minimum; ++ } ++ return fvalue; ++} ++ ++/* Get the next bb with a high branch probability. */ ++ ++basic_block ++next_high_probability_bb (basic_block bb) ++{ ++ if (bb == NULL) ++ return NULL; ++ ++ /* Limit the minimum probability value. */ ++ const float MINNUM_PROB = 0.00001f; ++ float minimum = MINNUM_PROB; ++ ++ gimple *stmt = last_stmt (bb); ++ if (stmt && gimple_code (stmt) == GIMPLE_COND) ++ { ++ edge true_edge = NULL; ++ edge false_edge = NULL; ++ extract_true_false_edges_from_block (bb, &true_edge, &false_edge); ++ ++ float true_edge_prob = get_edge_prob (true_edge, minimum); ++ float false_edge_prob = get_edge_prob (false_edge, minimum); ++ /* If the content of the branch does not include the candidate ++ kernel, the branch probability may not be limited. */ ++ /* The edge_prob may have precision error during static prediction, ++ so we need to relax the limit before comparison. */ ++ if ((true_edge_prob >= (param_branch_prob_threshold / 100.0) - minimum) ++ && flow_bb_inside_loop_p (bb->loop_father, true_edge->dest)) ++ return true_edge->dest; ++ else if ((false_edge_prob >= (param_branch_prob_threshold / 100.0) ++ - minimum) && flow_bb_inside_loop_p (bb->loop_father, ++ false_edge->dest)) ++ return false_edge->dest; ++ else ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "No high probability bb:"); ++ fprintf (dump_file, "current bb: %d, true: %f, false: %f\n", ++ bb->index, true_edge_prob, false_edge_prob); ++ } ++ return NULL; ++ } ++ } ++ else ++ { ++ edge e = find_fallthru_edge (bb->succs); ++ if (e) ++ return e->dest; ++ } ++ return NULL; ++} ++ ++ ++/* Dump loop header bb. */ ++ ++void ++dump_loop_headers (const char *name, vector &loops) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\n\n%s:\n", name); ++ fprintf (dump_file, "{ "); ++ for (unsigned int i = 0; i < loops.size (); i++) ++ fprintf (dump_file, "%d(%d) ", loops[i]->num, loops[i]->header->index); ++ fprintf (dump_file, "}\n\n"); ++ } ++} ++ ++/* Combine and sort candidate loops. */ ++ ++bool ++filter_and_sort_kernels (vector &sorted_kernels, ++ vector &kernels) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nPhase 4: filter_and_sort_kernels:\n\n"); ++ ++ set end_bb; ++ list walked_header_bb; /* Used to record nested loops. */ ++ ++ for (unsigned i = 0; i < kernels.size (); ++i) ++ end_bb.insert (kernels[i]->header); ++ ++ dump_loop_headers ("kernels", kernels); ++ ++ if (!param_filter_kernels) ++ { ++ for (vector::iterator it = kernels.begin (); ++ it != kernels.end (); ++it) ++ sorted_kernels.push_back (*it); ++ } ++ else ++ { ++ basic_block bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); ++ ++ while (bb) ++ { ++ if (bb == NULL) ++ return false; ++ if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) ++ break; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "%d ", bb->index); ++ ++ /* bb is not the head of the loop, go to the next. */ ++ if (bb != bb->loop_father->header) ++ { ++ bb = next_high_probability_bb (bb); ++ continue; ++ } ++ ++ /* bb is the head of the loop. */ ++ if (bb != walked_header_bb.back ()) ++ { ++ if (end_bb.count (bb)) ++ { ++ sorted_kernels.push_back (bb->loop_father); ++ bb = single_exit (bb->loop_father)->dest; ++ continue; ++ } ++ if (loop_outer (bb->loop_father) != NULL ++ && get_loop_exit_edges (bb->loop_father).length () != 1) ++ return false; ++ walked_header_bb.push_back (bb); ++ bb = next_high_probability_bb (bb); ++ continue; ++ } ++ else ++ { ++ walked_header_bb.pop_back (); ++ bb = single_exit (bb->loop_father)->dest; ++ continue; ++ } ++ } ++ } ++ ++ dump_loop_headers ("sorted_kernels", sorted_kernels); ++ return true; ++} ++ ++/* ================ phase 5 record_and_sort_ref_groups ================ */ ++/* Memory reference score, different aspects of one memory reference. */ ++ ++struct ref_score ++{ ++ /* certain memory reference. */ ++ data_ref d_ref; ++ ++ /* local count for bb where memory reference is located. */ ++ gcov_type bb_count; ++ ++ /* line-location of memory reference. */ ++ int line; ++}; ++ ++ ++/* Memory reference group, different reference of the same variable. */ ++ ++struct ref_group ++{ ++ /* source variables. */ ++ tree var; ++ ++ /* variable size, Unit: MB. */ ++ double var_size; ++ ++ /* first ref for insert hint. */ ++ data_ref first_use; ++ ++ /* reuse scores of variables. */ ++ unsigned int reuse_level; ++ ++ /* method of calculating the var size. */ ++ calc_type calc_by; ++ ++ /* memory reference index for specific variable. */ ++ unsigned int mem_ref_index; ++ ++ /* Accessing Reference Records in Different Modes (key_index): ++ 000: write, random, non-parallel ++ 001: write, random, parallel ++ 010: write, regular, non-parallel ++ 011: write, regular, parallel ++ 100: read, random, non-parallel ++ 101: read, random, parallel ++ 110: read, regular, non-parallel ++ 111: read, regular, parallel ++ */ ++ map > ref_use; ++ ++ /* scores for different memory references. */ ++ vector ref_scores; ++ ++ ref_group () ++ { ++ var = NULL_TREE; ++ var_size = 0; ++ reuse_level = 0; ++ calc_by = UNHANDLE_CALC; ++ mem_ref_index = 0; ++ } ++}; ++ ++/* calculate reuse level. */ ++ ++unsigned int ++calculate_reuse_level (map > &var_use) ++{ ++ unsigned int level = 0; ++ for (map >::iterator it = var_use.begin (); ++ it != var_use.end (); ++it) ++ { ++ unsigned int parallel = 1; ++ unsigned int regular = 1; ++ unsigned int cost = 1; ++ ++ if ((*it).second[0].parallel_p) ++ parallel = PARALLEL_NUM; ++ if (!(*it).second[0].regular_p) ++ regular = INDIRECT_ACCESS_VALUE; ++ if (!(*it).second[0].read_p) ++ cost = WRITE_COST; ++ ++ /* In serial reuse, we will later check whether they are in the ++ same cacheline. If yes, delete the reuse. For details, see the ++ reuse analysis of prefetching and eliminate redundancy. */ ++ unsigned int add = parallel * ((*it).second.size () * (cost + regular)); ++ level += add; ++ if (add && dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "%d : %d * (%ld * (%d + %d)) = %d\n", ++ (*it).first, parallel, (*it).second.size (), cost, regular, add); ++ } ++ return level; ++} ++ ++/* Comparison of reference reuse level. */ ++ ++bool ++ref_group_reuse_cmp (const ref_group &a, const ref_group &b) ++{ ++ return a.reuse_level > b.reuse_level; ++} ++ ++/* Sort reference groups. */ ++ ++void ++sort_ref_groups (vector &ref_groups, ++ map &ref_groups_map) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nsort_ref_groups_by_reuse_level\n"); ++ ++ for (map::iterator it = ref_groups_map.begin (); ++ it != ref_groups_map.end (); ++it) ++ { ++ (*it).second.reuse_level = calculate_reuse_level ((*it).second.ref_use); ++ ref_groups.push_back ((*it).second); ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ print_generic_expr (dump_file, (*it).second.var, TDF_SLIM); ++ fprintf (dump_file, " : %d\n", (*it).second.reuse_level); ++ } ++ } ++ ++ sort (ref_groups.begin (), ref_groups.end (), ref_group_reuse_cmp); ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\nsorted ref_groups:\n"); ++ fprintf (dump_file, "rank var (data_size, num_of_mem_ref, need_tmp_name):" ++ " reuse_level_score\n"); ++ for (unsigned int i = 0; i < ref_groups.size (); ++i) ++ { ++ fprintf (dump_file, "%d ", i); ++ print_generic_expr (dump_file, ref_groups[i].var, TDF_SLIM); ++ int need_tmp_name = !get_name (ref_groups[i].var) ? 1 : 0; ++ fprintf (dump_file, " (%lf, %lu, %d)", ref_groups[i].var_size, ++ ref_groups[i].ref_scores.size (), need_tmp_name); ++ fprintf (dump_file, " : %d\n", ref_groups[i].reuse_level); ++ } ++ fprintf (dump_file, "\n"); ++ ++ fprintf (dump_file, "first_use:\n"); ++ for (unsigned int i = 0; i < ref_groups.size (); ++i) ++ { ++ fprintf (dump_file, "%d ", i); ++ print_generic_expr (dump_file, ref_groups[i].var, TDF_SLIM); ++ fprintf (dump_file, " : "); ++ if (!ref_groups[i].first_use.vectorize_p) ++ print_generic_expr (dump_file, ref_groups[i].first_use.ref, ++ TDF_SLIM); ++ else ++ print_gimple_stmt (dump_file, ref_groups[i].first_use.stmt, ++ TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ fprintf (dump_file, "\n"); ++ } ++} ++ ++/* Attributes of variable data. */ ++ ++enum data_attribute ++{ ++ DA_PARALLEL = 0, ++ DA_REGULAR, ++ DA_READ ++}; ++ ++/* Record memory reference by use mode. ++ If the reference group is not found, create a group. */ ++ ++void ++record_mem_ref (map &ref_groups, data_ref &mem_ref) ++{ ++ unsigned int index = (mem_ref.parallel_p << DA_PARALLEL) ++ + (mem_ref.regular_p << DA_REGULAR) + (mem_ref.read_p << DA_READ); ++ ++ if (!ref_groups.count (mem_ref.var)) ++ { ++ ref_group ref_group; ++ ref_group.var = mem_ref.var; ++ ref_group.first_use = mem_ref; ++ ref_groups[mem_ref.var] = ref_group; ++ } ++ ++ /* Ref_groups' calc_by depends on the inserted mem_ref's calc_by. ++ Runtime issue requires the specified mem_ref's calc_by to be >= 1. ++ Temporarily modified ref_group's first_use after sorting mem_refs. */ ++ ref_groups[mem_ref.var].calc_by = max (ref_groups[mem_ref.var].calc_by, ++ mem_ref.calc_by); ++ ref_groups[mem_ref.var].var_size = max (ref_groups[mem_ref.var].var_size, ++ mem_ref.data_size); ++ ref_groups[mem_ref.var].ref_use[index].push_back (mem_ref); ++ ++ ref_score ref_level{ mem_ref, ((mem_ref.stmt)->bb->count).to_gcov_type (), ++ expand_location (mem_ref.stmt->location).line }; ++ ref_groups[mem_ref.var].ref_scores.push_back (ref_level); ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "recorded in: "); ++ print_generic_expr (dump_file, mem_ref.var, TDF_SLIM); ++ fprintf (dump_file, ":%d:%ld\n", index, ++ ref_groups[mem_ref.var].ref_use[index].size () - 1); ++ ++ fprintf (dump_file, "base: "); ++ print_generic_expr (dump_file, mem_ref.base, TDF_SLIM); ++ ++ fprintf (dump_file, ", index: "); ++ print_generic_expr (dump_file, mem_ref.index, TDF_SLIM); ++ ++ fprintf (dump_file, ", step: "); ++ if (mem_ref.step && cst_and_fits_in_hwi (mem_ref.step)) ++ fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, ++ int_cst_value (mem_ref.step)); ++ else ++ print_generic_expr (dump_file, mem_ref.step, TDF_SLIM); ++ ++ fprintf (dump_file, ", offset: "); ++ if (mem_ref.offset && cst_and_fits_in_hwi (mem_ref.offset)) ++ fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, ++ int_cst_value (mem_ref.offset)); ++ else ++ print_generic_expr (dump_file, mem_ref.offset, TDF_SLIM); ++ fprintf (dump_file, ", %s", mem_ref.read_p ? "read" : "write"); ++ ++ fprintf (dump_file, ", size: %lf", mem_ref.data_size); ++ fprintf (dump_file, "\n\n"); ++ } ++} ++ ++/* Rank data reference index level by the scheme of source code line number. */ ++ ++bool ++data_ref_reuse_cmp (const ref_score &a, const ref_score &b) ++{ ++ return a.line < b.line; ++} ++ ++/* Sort data reference index level within one reference group in non-decreasing ++ order of the customized sorting scheme. */ ++ ++void ++sort_mem_ref_in_ref_group (map &ref_groups_map) ++{ ++ if (dump_file) ++ fprintf (dump_file, "\nsorted data_references:\n"); ++ for (map::iterator it = ref_groups_map.begin (); ++ it != ref_groups_map.end (); ++it) ++ { ++ vector &ref_scores = (*it).second.ref_scores; ++ stable_sort (ref_scores.begin (), ref_scores.end (), data_ref_reuse_cmp); ++ /* Update ref_group's first_use and calc_by with the first mem_ref after ++ sorting. */ ++ (*it).second.first_use = (*it).second.ref_scores[0].d_ref; ++ (*it).second.calc_by = (*it).second.first_use.calc_by; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ print_generic_expr (dump_file, (*it).first, TDF_SLIM); ++ fprintf (dump_file, " : %lu\n", ref_scores.size ()); ++ for (unsigned int i = 0; i < ref_scores.size (); ++i) ++ { ++ fprintf (dump_file, "mem_ref_index %u: ", i); ++ print_gimple_stmt (dump_file, ref_scores[i].d_ref.stmt, 0, ++ TDF_LINENO); ++ } ++ fprintf (dump_file, "\n\n"); ++ } ++ } ++} ++ ++/* Tracing and sorting reference groups. */ ++ ++bool ++record_and_sort_ref_groups (vector &ref_groups, ++ vector &kernels, ++ map > &loop_refs) ++{ ++ if (dump_file) ++ fprintf (dump_file, "\nPhase 5: trace_all_references_details\n\n"); ++ ++ map ref_groups_map; ++ ++ for (unsigned i = 0; i < kernels.size (); ++i) ++ { ++ class loop* loop = kernels[i]; ++ if (loop_refs.count (loop) == 0) ++ continue; ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "loop header %d:\n", loop->header->index); ++ for (unsigned j = 0; j < loop_refs[loop].size (); ++j) ++ { ++ if (loop_refs[loop][j].trace_status_p) ++ record_mem_ref (ref_groups_map, loop_refs[loop][j]); ++ } ++ } ++ ++ /* Sort mem_ref within ref_group by local count and update first_use's ++ data_ref, stable sort. */ ++ sort_mem_ref_in_ref_group (ref_groups_map); ++ sort_ref_groups (ref_groups, ref_groups_map); ++ ++ return ref_groups.size () > 0; ++} ++ ++/* ================ phase 6 issue_llc_hint ================ */ ++ ++/* Issue vectorized mask prefetch gimple. */ ++ ++void ++issue_mask_prefetch (gimple *stmt) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "insert svprfd.\n"); ++ ++ /* vect__1.1 = .MASK_LOAD (_2, 32B, loop_mask_3); ++ .MASK_STORE (_4, 32B, loop_mask_5, vect__6.6); ++ */ ++ tree dataref_ptr = gimple_call_arg (stmt, 0); ++ tree scale = gimple_call_arg (stmt, 1); ++ tree final_mask = gimple_call_arg (stmt, 2); ++ tree target = NULL_TREE; ++ if (gimple_call_internal_fn (stmt) == IFN_MASK_STORE) ++ target = gimple_call_arg (stmt, 3); ++ else if (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD) ++ target = gimple_call_lhs (stmt); ++ /* 4: PLDL3KEEP. */ ++ tree prfop = build_int_cst (TREE_TYPE (integer_zero_node), 4); ++ ++ /* add offset. */ ++ gimple_stmt_iterator si = gsi_for_stmt (stmt); ++ /* target: vector_type - XXX_type. */ ++ if (target == NULL_TREE) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "unhandled scene: target vect is null"); ++ return; ++ } ++ HOST_WIDE_INT distance = param_prefetch_offset * tree_to_uhwi ++ (TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (target)))); ++ tree addr = fold_build_pointer_plus_hwi (dataref_ptr, distance); ++ addr = force_gimple_operand_gsi (&si, unshare_expr (addr), true, ++ NULL, true, GSI_SAME_STMT); ++ ++ gcall *call = gimple_build_call_internal (IFN_MASK_PREFETCH, ++ 5, addr, scale, final_mask, target, prfop); ++ gsi_insert_after (&si, call, GSI_SAME_STMT); ++ update_ssa (TODO_update_ssa_only_virtuals); ++} ++ ++/* Issue vectorized mask gather prefetch gimple. */ ++ ++void ++issue_mask_gather_prefetch (gimple *stmt) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "insert svprfd_gather_uxindex.\n"); ++ ++ /* vect_patt_1.1 = .MASK_GATHER_LOAD (_2, vect__3.3, 8, { 0.0, ... }, ++ loop_mask_4); */ ++ tree dataref_ptr = gimple_call_arg (stmt, 0); ++ tree vec_offset = gimple_call_arg (stmt, 1); ++ tree scale = gimple_call_arg (stmt, 2); ++ tree zero = gimple_call_arg (stmt, 3); ++ tree final_mask = gimple_call_arg (stmt, 4); ++ tree prfop = build_int_cst (TREE_TYPE (integer_zero_node), 4); ++ tree target = gimple_call_lhs (stmt); ++ ++ /* add offset. */ ++ gimple_stmt_iterator si = gsi_for_stmt (stmt); ++ if (target == NULL_TREE) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "unhandled scene: target vect is null"); ++ return; ++ } ++ HOST_WIDE_INT distance = param_prefetch_offset * tree_to_uhwi ++ (TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (target)))); ++ tree addr = fold_build_pointer_plus_hwi (dataref_ptr, distance); ++ addr = force_gimple_operand_gsi (&si, unshare_expr (addr), true, ++ NULL, true, GSI_SAME_STMT); ++ ++ gcall *call = gimple_build_call_internal ++ (IFN_MASK_GATHER_PREFETCH, 7, addr, ++ vec_offset, scale, zero, final_mask, target, prfop); ++ gsi_insert_after (&si, call, GSI_SAME_STMT); ++ update_ssa (TODO_update_ssa_only_virtuals); ++} ++ ++/* Issue builtin prefetch gimple. */ ++ ++void ++issue_builtin_prefetch (data_ref &mem_ref) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "insert prfm.\n"); ++ /* MEM[symbol: diagPtr, index: ivtmp_102, step: 8, offset: 0B] */ ++ gimple* stmt = mem_ref.stmt; ++ tree dataref_ptr = mem_ref.base; ++ tree data_idx = mem_ref.index; ++ tree scale = mem_ref.step; ++ tree offset = mem_ref.offset; ++ /* add offset. */ ++ gimple_stmt_iterator si = gsi_for_stmt (stmt); ++ if (scale == NULL_TREE) ++ { ++ /* _190 = (void *) ivtmp.444_221; ++ Cannot detect size unit at (void *). */ ++ scale = TYPE_SIZE_UNIT (inner_type (TREE_TYPE (mem_ref.var))); ++ if (scale == NULL_TREE) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "ERROR: Unknown size unit for the prefetching " ++ "variable. Stop builtin_prefetch.\n\n"); ++ return; ++ } ++ } ++ ++ data_idx = data_idx ? data_idx : size_zero_node; ++ data_idx = build1 (NOP_EXPR, TREE_TYPE (scale), data_idx); ++ tree displacement = fold_build2 (MULT_EXPR, TREE_TYPE (scale), data_idx, ++ scale); ++ if (offset != NULL_TREE && TREE_CODE (offset) != TREE_CODE (size_zero_node)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "WARNING: offset's TREE_TYPE is not integer_cst: " ++ "%s\nStop builtin_prefetch.\n", ++ get_tree_code_name (TREE_CODE (offset))); ++ return; ++ } ++ offset = offset ? offset : size_zero_node; ++ offset = build1 (NOP_EXPR, TREE_TYPE (scale), offset); ++ dataref_ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (dataref_ptr), ++ dataref_ptr, offset); ++ tree addr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (dataref_ptr), ++ dataref_ptr, displacement); ++ HOST_WIDE_INT distance = param_prefetch_offset * tree_to_uhwi (scale); ++ ++ addr = fold_build_pointer_plus_hwi (addr, distance); ++ addr = force_gimple_operand_gsi (&si, unshare_expr (addr), true, ++ NULL, true, GSI_SAME_STMT); ++ /* __builtin_prefetch (_68, 0, 1); ++ 1st param: *addr, 2nd param: write/read (1/0), 3rd param: temporal locality ++ (high means strong locality) */ ++ gcall *call = gimple_build_call (builtin_decl_explicit (BUILT_IN_PREFETCH), ++ 3, addr, integer_zero_node, integer_one_node); ++ gsi_insert_after (&si, call, GSI_SAME_STMT); ++ update_ssa (TODO_update_ssa_only_virtuals); ++} ++ ++/* Retrieve memory reference at the specific index. */ ++ ++data_ref ++get_data_ref_at_idx (ref_group &var_ref_group) ++{ ++ unsigned int mem_ref_size = static_cast( ++ var_ref_group.ref_scores.size ()); ++ if (strlen (param_mem_ref_index) == 0) ++ return var_ref_group.first_use; ++ else ++ { ++ /* Insert prefetch hint at highly-likely-used location with the given ++ index. */ ++ if (var_ref_group.mem_ref_index >= mem_ref_size) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "WARNING: The target data_ref index is out " ++ "of range. Use top index instead!\n"); ++ return var_ref_group.ref_scores[0].d_ref; ++ } ++ return var_ref_group.ref_scores[var_ref_group.mem_ref_index].d_ref; ++ } ++} ++ ++/* Static form insertion and issue instruction. We may check the ++ determination of the ARM SVE architecture before SVE hint insertion. */ ++ ++void ++static_issue (vector &ref_groups, int num_issue_var) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "static issue\n"); ++ ++ for (int i = 0; i < num_issue_var; ++i) ++ { ++ data_ref mem_ref = get_data_ref_at_idx (ref_groups[i]); ++ if (mem_ref.vectorize_p) ++ { ++ enum internal_fn ifn_code = gimple_call_internal_fn ++ (mem_ref.stmt); ++ if (ifn_code == IFN_MASK_STORE || ifn_code == IFN_MASK_LOAD) ++ issue_mask_prefetch (mem_ref.stmt); ++ else if (ifn_code == IFN_MASK_GATHER_LOAD) ++ issue_mask_gather_prefetch (mem_ref.stmt); ++ else ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "other vectorized internal function\n"); ++ } ++ else ++ issue_builtin_prefetch (mem_ref); ++ } ++} ++ ++/* Generate the stmts for calculating the size. Later we will consider nested ++ multi-branches scenarios and check more information of niters when it is ++ a COND_EXPR. */ ++ ++tree ++calc_stmts_gen (vector &ref_groups, gimple_seq &cond_expr_stmt_list, ++ int num_issue_var) ++{ ++ /* Accumulated keep size. */ ++ tree total_size = build_real_from_int_cst ++ (double_type_node, integer_zero_node); ++ for (int i = 0; i < num_issue_var; ++i) ++ { ++ data_ref &mem_ref = ref_groups[i].first_use; ++ tree var = mem_ref.var; ++ for (unsigned j = 0; j < mem_ref.loop_bounds.size (); ++j) ++ { ++ tree niters = mem_ref.loop_bounds[j].niters; ++ ++ /* COND_EXPR. */ ++ if (TREE_CODE (niters) == COND_EXPR) ++ niters = TREE_OPERAND (niters, 1); ++ tree unit = TYPE_SIZE_UNIT (inner_type (TREE_TYPE (var))); ++ /* _190 = (void *) ivtmp.444_221; ++ Cannot detect size unit at (void *). */ ++ if (unit == NULL_TREE) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "WARNING: Cannot detect size unit " ++ "(use 1 byte) for variable %s: ", get_name (var)); ++ print_generic_expr (dump_file, mem_ref.ref, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ unit = size_one_node; ++ } ++ unit = build1 (NOP_EXPR, TREE_TYPE (niters), unit); ++ tree size = fold_build2 (MULT_EXPR, TREE_TYPE (niters), niters, unit); ++ size = build1 (FLOAT_EXPR, double_type_node, size); ++ total_size = fold_build2 ++ (PLUS_EXPR, double_type_node, total_size, size); ++ } ++ } ++ /* Create a stmt list for size calculation. */ ++ tree div = build_int_cst (TREE_TYPE (integer_zero_node), 1024 * 1024); ++ div = build1 (NOP_EXPR, double_type_node, div); ++ total_size = fold_build2 (RDIV_EXPR, double_type_node, total_size, div); ++ ++ tree threshold = build_int_cst (TREE_TYPE (integer_zero_node), ++ param_llc_capacity_per_core / 2); ++ threshold = build_real_from_int_cst (double_type_node, threshold); ++ tree cond_expr = fold_build2 ++ (LE_EXPR, boolean_type_node, total_size, threshold); ++ ++ /* Convert cond_expr to stmt list. */ ++ cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr), ++ &cond_expr_stmt_list, is_gimple_condexpr, NULL_TREE); ++ return cond_expr; ++} ++ ++/* Runtime form insertion and issue instruction. */ ++ ++void ++runtime_issue (vector &ref_groups, int num_issue_var) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "runtime issue\n"); ++ ++ if (ref_groups.size () == 0) ++ return; ++ data_ref &mem_ref = ref_groups[0].first_use; ++ class loop *loop = mem_ref.loop_bounds.back ().loop; ++ /* Ensure that variables are in the same loop. */ ++ for (int i = 1; i < num_issue_var; ++i) ++ { ++ data_ref &mem_ref = ref_groups[i].first_use; ++ if (loop != mem_ref.loop_bounds.back ().loop) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "topn var are not in the same loop\n"); ++ return; ++ } ++ } ++ if (loop == NULL) ++ return; ++ ++ /* If the exit edge points to bb with multiple inputs, split the exit edge ++ and create a new bb, make the exit edge point to bb only single input. */ ++ edge e = single_exit (loop); ++ if (e == NULL) ++ return; ++ if (!single_pred_p (e->dest)) ++ { ++ split_loop_exit_edge (e, true); ++ if (dump_enabled_p ()) ++ dump_printf (MSG_NOTE, "split exit edge\n"); ++ } ++ ++ gimple_seq cond_expr_stmt_list = NULL; ++ tree cond_expr = calc_stmts_gen (ref_groups, cond_expr_stmt_list, ++ num_issue_var); ++ ++ /* Use the previous cond and generate a new branch and copy loop. */ ++ basic_block condition_bb = NULL; ++ profile_probability prob = profile_probability::likely (); ++ initialize_original_copy_tables (); ++ class loop *nloop = loop_version (loop, cond_expr, &condition_bb, ++ prob, prob.invert (), prob, prob.invert (), true); ++ free_original_copy_tables (); ++ ++ /* Insert the generated stmt list before cond_expr. */ ++ gimple_stmt_iterator cond_exp_gsi; ++ if (cond_expr_stmt_list) ++ { ++ cond_exp_gsi = gsi_last_bb (condition_bb); ++ gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, ++ GSI_SAME_STMT); ++ } ++ update_ssa (TODO_update_ssa); ++ ++ /* Perform hint issue for branches that meet conditions. */ ++ static_issue (ref_groups, num_issue_var); ++} ++ ++/* Issue llc hints through prefetch instructions. */ ++ ++void ++issue_llc_hint (vector &ref_groups) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "issue_llc_hint:\n"); ++ ++ /* 1. If the issue-topn and force-issue options are available, top N var is ++ forcibly allocated and no runtime branch is generated. ++ 2. If the issue-topn option is available and the size of top N var is ++ statically known, top N is statically allocated and no runtime branch ++ is generated. ++ 3. If the issue-topn option is available and the size of the top N var is ++ unknown, but them is dynamically known, the top N is dynamically ++ allocated and generate runtime branches. (also depends on the screening ++ of the innermost variable boundary type) ++ 4. If the dynamic runtime cannot know the size, such as indirect access, ++ optimization is skipped. ++ */ ++ if (ref_groups.size () == 0) ++ return; ++ ++ int num_issue_var = min (param_issue_topn, ++ static_cast(ref_groups.size ())); ++ if (num_issue_var < param_issue_topn ++ && dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "WARNING: Only %u (less than param_issue_topn = %d) " ++ "ref_group(s) is found for llc hint.\n", ++ num_issue_var, param_issue_topn); ++ } ++ if (param_force_issue) ++ { ++ if (strlen (param_target_variables) > 0) ++ static_issue (ref_groups, static_cast(ref_groups.size ())); ++ else ++ static_issue (ref_groups, num_issue_var); ++ return; ++ } ++ calc_type topn_calc_type = STATIC_CALC; ++ for (int i = 0; i < num_issue_var; ++i) ++ topn_calc_type = min (topn_calc_type, ref_groups[i].calc_by); ++ ++ if (topn_calc_type == STATIC_CALC) ++ { ++ /* Before static issue, we still need to collect data size of all target ++ variables and compare the summation with LLC cache size. */ ++ double prefetch_data_size = 0.; ++ for (int i = 0; i < num_issue_var; ++i) ++ prefetch_data_size += ref_groups[i].var_size; ++ if (prefetch_data_size <= (double) param_llc_capacity_per_core * 0.8) ++ static_issue (ref_groups, num_issue_var); ++ else ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "static issue: Prefetch size exceeds LLC cache " ++ "size: %lf > %lf.\n", prefetch_data_size, ++ (double) param_llc_capacity_per_core * 0.8); ++ } ++ else if (topn_calc_type == RUNTIME_CALC) ++ runtime_issue (ref_groups, num_issue_var); ++ else ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "unhandled issue scene\n"); ++ } ++} ++ ++/* ==================== phase entry ==================== */ ++/* Check whether a string can be converted to an unsigned integer. */ ++ ++bool is_unsigned_int (const string &s) ++{ ++ if (s.empty () || s.size () > PREFETCH_TOOL_NUM_MAX_LEN) ++ return false; ++ ++ for (unsigned int i = 0; i < s.size (); ++i) ++ { ++ if (s[i] < '0' || s[i] > '9') ++ return false; ++ } ++ return true; ++} ++ ++/* Parse a substring separated by comma. If the substring is valid and ++ non-empty, store it as a parsed element. */ ++ ++bool ++parse_string_helper (const string &substr, vector& str_elts, ++ bool check_unsigned, size_t start, size_t end) ++{ ++ if (substr == "" && dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "WARNING: The input string from %lu to %lu is " ++ "empty.\n", start, end); ++ else if (check_unsigned && !is_unsigned_int (substr)) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "ERROR: not an unsigned integer: %s\n", ++ substr.c_str ()); ++ str_elts.clear (); ++ return false; ++ } ++ else ++ str_elts.push_back (substr); ++ return true; ++} ++ ++/* Parse a user input string, separated by comma. */ ++ ++void ++parse_string (const string &s, vector& str_elts, ++ bool check_unsigned = false) ++{ ++ string delim = ","; ++ size_t start = 0; ++ size_t end = s.find (delim); ++ string substr = s.substr (start, end - start); ++ while (end != string::npos) ++ { ++ if (!parse_string_helper (substr, str_elts, check_unsigned, start, end)) ++ return; ++ start = end + delim.size (); ++ end = s.find (delim, start); ++ substr = s.substr (start, end - start); ++ } ++ parse_string_helper (substr, str_elts, check_unsigned, start, end); ++} ++ ++/* Parse user input of target variables and memory indices and create a map ++ that assigns a target variable to a memory index. */ ++ ++void ++parse_param_inputs (map &var2mem_idx) ++{ ++ /* The user input length should have an input length limit. */ ++ if ((strlen (param_target_variables) >= PREFETCH_TOOL_INPUT_MAX_LEN ++ || strlen (param_mem_ref_index) >= PREFETCH_TOOL_INPUT_MAX_LEN) ++ && dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "INVALID INPUT: The user inputs for target variables " ++ "and/or memory reference indices are too long for parsing.\n"); ++ ++ vector var_names; ++ string target_variables = param_target_variables; ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Start parsing target variables:\n"); ++ if (param_use_ref_group_index) ++ parse_string (target_variables, var_names, true); ++ else ++ parse_string (target_variables, var_names, false); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Finish parsing target variables.\n\n"); ++ ++ vector var_mem_indices; ++ string mem_indices = param_mem_ref_index; ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Start parsing memory reference indices:\n"); ++ parse_string (mem_indices, var_mem_indices, true); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Finish parsing memory reference indices.\n\n"); ++ ++ /* Construct a map of var_name: var_mem_index. */ ++ if (var_names.size () > 0) ++ { ++ if (var_mem_indices.size () < var_names.size ()) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "WARNING: The number of provided memory " ++ "reference indices is less than that of target " ++ "variables.\nUse the top index for all variables " ++ "instead.\n"); ++ for (string& var_name : var_names) ++ var2mem_idx[var_name] = 0; ++ } ++ else ++ { ++ if (var_mem_indices.size () > var_names.size () ++ && dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "WARNING: The number of target variables is " ++ "less than that of memory reference indices.\n"); ++ for (unsigned int i = 0; i < var_names.size (); ++i) ++ { ++ var2mem_idx[var_names[i]] = static_cast( ++ atoi (var_mem_indices[i].c_str ())); ++ } ++ } ++ } ++} ++ ++/* Filter reference groups by only selecting target variables from the user ++ input. There are two options for prefetching target variables: ++ 1. Specify variable name parsed by the pass, which you can double-check at ++ "sorted ref_groups" section in the dump file. ++ 2. Specify variable rank exhibited at "sorted ref_groups" section in the ++ dump file. ++*/ ++ ++void ++prefetch_variables (const vector& ref_groups, ++ vector& reduced_ref_groups) ++{ ++ map ref_group2mem_idx; ++ ++ map var2mem_idx; /* externally defined. */ ++ parse_param_inputs (var2mem_idx); ++ ++ if (param_use_ref_group_index) ++ { ++ /* Use ref_group index at "sorted ref_groups" section to specify ++ variable. */ ++ /* Collect the variables in "reduced_ref_group" only if their indices ++ show up at "sorted ref_groups" section. */ ++ for (const pair &var_mem_idx : var2mem_idx) ++ { ++ unsigned int var_idx = static_cast(atoi ( ++ var_mem_idx.first.c_str ())); ++ if (var_idx < ref_groups.size ()) ++ ref_group2mem_idx[var_idx] = var_mem_idx.second; ++ else if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "WARNING: The index \"%u\" does not show " ++ "up in the ref_groups.\n", var_idx); ++ } ++ } ++ else ++ { ++ /* Use variable name shown up at "sorted ref_groups" section to specify ++ variable: ++ var2ref_group_idx + var2mem_idx -> ref_group2mem_idx. */ ++ /* Create a map that assigns the variable name to its corresponding ++ ref_group index. */ ++ map var2ref_group_idx; /* internally detected. */ ++ for (unsigned int i = 0; i < ref_groups.size (); ++i) ++ { ++ const ref_group &curr_ref_group = ref_groups[i]; ++ const int UINT_MAX_DIGIT = 10; ++ /* Unrecognizable variable name related to ref_group. */ ++ if (!get_name (curr_ref_group.var)) ++ { ++ /* If the variable name does not have a string representation, ++ we can rename it by "tmp_var_" + . */ ++ char group_idx[UINT_MAX_DIGIT]; ++ sprintf (group_idx, "%u", i); ++ string tmp_var_name = "tmp_var_" + std::string (group_idx); ++ fprintf (dump_file, "Unrecognizable variable name at ref_group " ++ "index %u.\nThe tree expression for variable is: ", i); ++ print_generic_expr (dump_file, curr_ref_group.var, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ var2ref_group_idx[tmp_var_name] = i; ++ } ++ else ++ var2ref_group_idx[std::string (get_name (curr_ref_group.var))] = i; ++ } ++ /* Collect the variables in "reduced_ref_group" only if they show up in ++ the ref_groups. */ ++ for (const pair &var_mem_idx : var2mem_idx) ++ { ++ if (var2ref_group_idx.count (var_mem_idx.first)) ++ { ++ unsigned int ref_group_idx = var2ref_group_idx[var_mem_idx.first]; ++ ref_group2mem_idx[ref_group_idx] = var_mem_idx.second; ++ } ++ else if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "WARNING: Target variable \" %s \" does " ++ "not show up in the ref_groups. Check whether it needs " ++ "temporary variable name.\n", ++ var_mem_idx.first.c_str ()); ++ } ++ } ++ ++ for (const pair &ref_group_mem_idx : ++ ref_group2mem_idx) ++ { ++ ref_group curr_ref_group = ref_groups[ref_group_mem_idx.first]; ++ curr_ref_group.mem_ref_index = ref_group_mem_idx.second; ++ reduced_ref_groups.push_back (curr_ref_group); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\nNOTICE: Prefetching target variable \" "); ++ print_generic_expr (dump_file, curr_ref_group.var, TDF_SLIM); ++ fprintf (dump_file, " \" at ref_group index %u and memory location " ++ "index %u.\n", ref_group_mem_idx.first, ++ ref_group_mem_idx.second); ++ } ++ } ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\n\n"); ++} ++ ++ ++/* The LLC intelligent allocation consists of 6 steps. */ ++ ++void ++llc_allocate (void) ++{ ++ map > kernels_refs; ++ vector kernels; ++ if (!get_dense_memory_kernels (kernels, kernels_refs)) ++ return; ++ ++ trace_data_refs_info (kernels, kernels_refs); ++ ++ if (!analyze_nested_kernels (kernels, kernels_refs)) ++ return; ++ ++ vector sorted_kernels; ++ if (!filter_and_sort_kernels (sorted_kernels, kernels)) ++ return; ++ ++ vector ref_groups; ++ if (!record_and_sort_ref_groups (ref_groups, sorted_kernels, kernels_refs)) ++ return; ++ ++ if (strlen (param_target_variables) > 0) ++ { ++ /* If "param_target_variables" is not empty, we will issue parsed target ++ variables compulsorily. */ ++ param_force_issue = true; ++ vector reduced_ref_groups; ++ prefetch_variables (ref_groups, reduced_ref_groups); ++ issue_llc_hint (reduced_ref_groups); ++ } ++ else ++ issue_llc_hint (ref_groups); ++} ++ ++/* Check whether the function is an operator reloading function. */ ++ ++bool ++operator_func_p (function *fn) ++{ ++ const char *fn_name = IDENTIFIER_POINTER (DECL_NAME (fn->decl)); ++ ++ if (fn_name && strncmp (fn_name, "operator", 8) == 0) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "operator_func: %s ", fn_name); ++ ++ return true; ++ } ++ return false; ++} ++ ++/* Check whether the function file location is known. */ ++ ++bool ++func_location_p (function *fn) ++{ ++ expanded_location fn_decl_xloc ++ = expand_location (DECL_SOURCE_LOCATION (current_function_decl)); ++ expanded_location fn_xloc ++ = expand_location (fn->function_start_locus); ++ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "fn->function_start_locus = %d \n", ++ fn->function_start_locus); ++ fprintf (dump_file, "fn_xloc.file = %s \n", ++ fn_xloc.file ? fn_xloc.file : "NULL"); ++ fprintf (dump_file, "fn_decl_xloc.file = %s \n", ++ fn_decl_xloc.file ? fn_decl_xloc.file : "NULL"); ++ fprintf (dump_file, "LOCATION_FILE (input_location) = %s \n", ++ LOCATION_FILE (input_location) ? LOCATION_FILE (input_location) ++ : "NULL"); ++ } ++ if (fn_decl_xloc.file == NULL) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Function location unknown, skip analysis \n"); ++ return false; ++ } ++ /* Newly generated functions are filtered out, such as function constant ++ propagation func.constprop (). */ ++ if (LOCATION_FILE (input_location) != fn_decl_xloc.file) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "Function location non-local, skip analysis \n"); ++ return false; ++ } ++ return true; ++} ++ ++/* Dump function information. */ ++ ++void ++dump_function_info (function *fn) ++{ ++ const char *fn_name = IDENTIFIER_POINTER (DECL_NAME (fn->decl)); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\nfn_name: %s\n", fn_name); ++ expanded_location cfun_xloc ++ = expand_location (DECL_SOURCE_LOCATION (current_function_decl)); ++ if (cfun_xloc.line) ++ { ++ if (cfun_xloc.file) ++ fprintf (dump_file, "[%s:%d:%d]\n", ++ cfun_xloc.file, cfun_xloc.line, cfun_xloc.column); ++ } ++ fprintf (dump_file, "\n"); ++ flow_loops_dump (dump_file, NULL, 1); ++ fprintf (dump_file, "\n"); ++ } ++} ++ ++/* dump param. */ ++ ++void ++dump_param (void) ++{ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "LLC allocate parameters:\n"); ++ fprintf (dump_file, " block size: %d\n", param_l1_cache_line_size); ++ fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", ++ param_l1_cache_size * 1024 / param_l1_cache_line_size, ++ param_l1_cache_size); ++ fprintf (dump_file, " L1 cache line size: %d\n", ++ param_l1_cache_line_size); ++ fprintf (dump_file, " L2 cache size: %d kB\n", param_l2_cache_size); ++ fprintf (dump_file, " min mem_access_ratio: %d \n", ++ param_mem_access_ratio); ++ fprintf (dump_file, " min mem_access_num: %d \n", ++ param_mem_access_num); ++ fprintf (dump_file, "\n"); ++ } ++} ++ ++const pass_data pass_data_llc_allocate = ++{ ++ GIMPLE_PASS, /* type. */ ++ "llc_allocate", /* name. */ ++ OPTGROUP_LOOP, /* optinfo_flags. */ ++ TV_TREE_PREFETCH, /* tv_id. */ ++ (PROP_cfg | PROP_ssa), /* properties_required. */ ++ 0, /* properties_provided. */ ++ 0, /* properties_destroyed. */ ++ 0, /* todo_flags_start. */ ++ 0, /* todo_flags_finish. */ ++}; ++ ++class pass_llc_allocate : public gimple_opt_pass ++{ ++public: ++ pass_llc_allocate (gcc::context *ctxt) ++ : gimple_opt_pass (pass_data_llc_allocate, ctxt) ++ {} ++ ++ /* opt_pass methods. */ ++ virtual bool gate (function *) ++ { ++ return (optimize >= 2 && flag_llc_allocate > 0); ++ } ++ virtual unsigned int execute (function *); ++ ++}; // class pass_llc_allocate ++ ++unsigned int ++pass_llc_allocate::execute (function *fn) ++{ ++ unsigned int ret = 0; ++ ++ if (!targetm.have_prefetch () ++ || targetm.vectorize.code_for_prefetch == NULL ++ || targetm.vectorize.prefetch_handleable_mode_p == NULL ++ || targetm.vectorize.code_for_gather_prefetch == NULL) ++ return 0; ++ ++ if (!builtin_decl_explicit_p (BUILT_IN_PREFETCH)) ++ { ++ tree type = build_function_type_list (void_type_node, ++ const_ptr_type_node, NULL_TREE); ++ tree decl = add_builtin_function ("__builtin_prefetch", type, ++ BUILT_IN_PREFETCH, BUILT_IN_NORMAL, ++ NULL, NULL_TREE); ++ DECL_IS_NOVOPS (decl) = true; ++ set_builtin_decl (BUILT_IN_PREFETCH, decl, false); ++ } ++ ++ dump_param (); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "llc_allocate: %s\n", ++ IDENTIFIER_POINTER (DECL_NAME (fn->decl))); ++ ++ if (number_of_loops (fn) <= 1 || !func_location_p (fn) ++ || operator_func_p (fn)) ++ return ret; ++ ++ dump_function_info (fn); ++ ++ llc_allocate (); ++ ++ return ret; ++} ++ ++} // anon namespace ++ ++gimple_opt_pass * ++make_pass_llc_allocate (gcc::context *ctxt) ++{ ++ return new pass_llc_allocate (ctxt); ++} +diff --git a/gcc/tree-ssa-loop-niter.c b/gcc/tree-ssa-loop-niter.c +index 7775bc727..c500d5e20 100644 +--- a/gcc/tree-ssa-loop-niter.c ++++ b/gcc/tree-ssa-loop-niter.c +@@ -2384,6 +2384,37 @@ loop_only_exit_p (const class loop *loop, basic_block *body, const_edge exit) + return true; + } + ++/* Returns whether the number of vectorized iterations for the loop can be ++ estimated from the given IR and update the corresponding loop attribute, ++ e.g., next_mask_114 = .WHILE_ULT (_122, niters.5_75, { 0, ... }); */ ++ ++bool ++number_of_iterations_vect (class loop *loop, tree lhs, tree rhs) ++{ ++ loop->vec_nb_iterations = chrec_dont_know; ++ ++ if ((TREE_CODE (lhs) != SSA_NAME && TREE_CODE (rhs) != SSA_NAME) ++ || (TREE_CODE (lhs) == SSA_NAME && TREE_CODE (rhs) == SSA_NAME)) ++ return false; ++ ++ tree ssa = TREE_CODE (lhs) == SSA_NAME ? lhs : rhs; ++ gimple *def_stmt = SSA_NAME_DEF_STMT (ssa); ++ ++ if (gimple_code (def_stmt) != GIMPLE_CALL ++ || !gimple_call_internal_p (def_stmt)) ++ return false; ++ ++ internal_fn ifn = gimple_call_internal_fn (def_stmt); ++ if (ifn != IFN_WHILE_ULT) ++ return false; ++ ++ gcall *call = dyn_cast (def_stmt); ++ tree niters = gimple_call_arg (call, 1); ++ loop->vec_nb_iterations = niters; ++ ++ return true; ++} ++ + /* Stores description of number of iterations of LOOP derived from + EXIT (an exit edge of the LOOP) in NITER. Returns true if some useful + information could be derived (and fields of NITER have meaning described +@@ -2454,6 +2485,9 @@ number_of_iterations_exit_assumptions (class loop *loop, edge exit, + op1 = gimple_cond_rhs (stmt); + type = TREE_TYPE (op0); + ++ if (TREE_CODE (type) == VECTOR_TYPE) ++ number_of_iterations_vect (loop, op0, op1); ++ + if (TREE_CODE (type) != INTEGER_TYPE + && !POINTER_TYPE_P (type)) + return false; +@@ -2730,14 +2764,14 @@ bool + number_of_iterations_exit (class loop *loop, edge exit, + class tree_niter_desc *niter, + bool warn, bool every_iteration, +- basic_block *body) ++ basic_block *body, bool guarantee) + { + gcond *stmt; + if (!number_of_iterations_exit_assumptions (loop, exit, niter, + &stmt, every_iteration, body)) + return false; + +- if (integer_nonzerop (niter->assumptions)) ++ if (integer_nonzerop (niter->assumptions) || guarantee == false) + return true; + + if (warn && dump_enabled_p ()) +diff --git a/gcc/tree-ssa-loop-niter.h b/gcc/tree-ssa-loop-niter.h +index eb8d15794..d38472e52 100644 +--- a/gcc/tree-ssa-loop-niter.h ++++ b/gcc/tree-ssa-loop-niter.h +@@ -27,7 +27,8 @@ extern bool loop_only_exit_p (const class loop *, basic_block *body, + extern bool number_of_iterations_exit (class loop *, edge, + class tree_niter_desc *niter, bool, + bool every_iteration = true, +- basic_block * = NULL); ++ basic_block * = NULL, ++ bool guarantee = true); + extern bool number_of_iterations_exit_assumptions (class loop *, edge, + class tree_niter_desc *, + gcond **, bool = true, +-- +2.33.0 + diff --git a/0153-LLC-add-extending-outer-loop.patch b/0153-LLC-add-extending-outer-loop.patch new file mode 100644 index 0000000..fef87dc --- /dev/null +++ b/0153-LLC-add-extending-outer-loop.patch @@ -0,0 +1,1285 @@ +From 4a365290cd9563385d32a22f7b1532c50b69e063 Mon Sep 17 00:00:00 2001 +From: zhaoshujian +Date: Mon, 11 Dec 2023 15:06:28 +0800 +Subject: [PATCH] LLC add extending outer loop + + +diff --git a/gcc/params.opt b/gcc/params.opt +index c429359e3..227175eef 100644 +--- a/gcc/params.opt ++++ b/gcc/params.opt +@@ -1058,4 +1058,10 @@ Common Joined UInteger Var(param_filter_kernels) Init(1) IntegerRange(0, 1) Para + Allow LLC allocate pass to greedily filter kernels by traversing the corresponding basic blocks + through edges with branch probability no less than param_branch_prob_threshold. + ++-param=outer-loop-nums= ++Common Joined UInteger Var(param_outer_loop_num) Init(1) IntegerRange(1, 10) Param ++Maximum number of outer loops allowed to extend outer loops for loops that ++cannot recognize inner loop boundaries. ++ ++ + ; This comment is to ensure we retain the blank line above. +diff --git a/gcc/testsuite/g++.dg/llc-allocate/llc-allocate.exp b/gcc/testsuite/g++.dg/llc-allocate/llc-allocate.exp +new file mode 100644 +index 000000000..9e98191ed +--- /dev/null ++++ b/gcc/testsuite/g++.dg/llc-allocate/llc-allocate.exp +@@ -0,0 +1,27 @@ ++# Copyright (C) 1997-2022 Free Software Foundation, Inc. ++ ++# This program is free software; you can redistribute it and/or modify ++# it under the terms of the GNU General Public License as published by ++# the Free Software Foundation; either version 3 of the License, or ++# (at your option) any later version. ++# ++# This program is distributed in the hope that it will be useful, ++# but WITHOUT ANY WARRANTY; without even the implied warranty of ++# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the ++# GNU General Public License for more details. ++# ++# You should have received a copy of the GNU General Public License ++# along with GCC; see the file COPYING3. If not see ++# . ++ ++load_lib g++-dg.exp ++load_lib target-supports.exp ++ ++# Initialize `dg'. ++dg-init ++ ++dg-runtest [lsort [glob -nocomplain $srcdir/$subdir/*.C]] \ ++ "" "-fllc-allocate" ++ ++# All done. ++dg-finish +\ No newline at end of file +diff --git a/gcc/testsuite/g++.dg/llc-allocate/llc-relion-expand-kernels.C b/gcc/testsuite/g++.dg/llc-allocate/llc-relion-expand-kernels.C +new file mode 100644 +index 000000000..44a9d7c66 +--- /dev/null ++++ b/gcc/testsuite/g++.dg/llc-allocate/llc-relion-expand-kernels.C +@@ -0,0 +1,53 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-options "-O3 -march=armv8.2-a+sve -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param branch-prob-threshold=50 --param filter-kernels=0 --param mem-access-num=2 --param issue-topn=1 --param force-issue=1" } */ ++#include "multidim_array.h" ++ ++class Input ++{ ++ public: ++ int metadata_offset = 13; ++ int exp_nr_images = 1; ++ MultidimArray exp_Mweight; ++ void convertAllSquaredDifferencesToWeights(); ++}; ++ ++int main() ++{ ++ clock_t start = clock(); ++ Input input; ++ int testIter = 2; ++ ++ for (int i = 0; i < testIter; ++i) ++ { ++ input.convertAllSquaredDifferencesToWeights(); ++ } ++ return 0; ++} ++ ++void Input::convertAllSquaredDifferencesToWeights() ++{ ++ for (int img_id = 0; img_id < exp_nr_images; img_id++) ++ { ++ int my_metadata_offset = metadata_offset + img_id; ++ MultidimArray sorted_weight; ++ ++ exp_Mweight.getRow(img_id, sorted_weight); ++ long int np = 0; ++ FOR_ALL_DIRECT_ELEMENTS_IN_MULTIDIMARRAY(sorted_weight) ++ { ++ if (DIRECT_MULTIDIM_ELEM(sorted_weight, n) > 0.) ++ { ++ DIRECT_MULTIDIM_ELEM(sorted_weight, np) = DIRECT_MULTIDIM_ELEM( \ ++ sorted_weight, n); ++ np++; ++ } ++ } ++ } ++} ++ ++ ++ ++/* { dg-final { scan-tree-dump-times "dense memory access" 1 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "__builtin_prefetch" 1 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "static issue" 1 "llc_allocate" } } */ ++ +diff --git a/gcc/testsuite/g++.dg/llc-allocate/multidim_array.h b/gcc/testsuite/g++.dg/llc-allocate/multidim_array.h +new file mode 100644 +index 000000000..d65066ebf +--- /dev/null ++++ b/gcc/testsuite/g++.dg/llc-allocate/multidim_array.h +@@ -0,0 +1,186 @@ ++#ifndef MULTIDIM_ARRAY_H ++#define MULTIDIM_ARRAY_H ++ ++#include ++ ++#define RELION_ALIGNED_MALLOC malloc ++#define RELION_ALIGNED_FREE free ++ ++#define STARTINGX(v) ((v).xinit) ++#define STARTINGY(v) ((v).yinit) ++#define NZYXSIZE(v) ((v).nzyxdim) ++ ++#define DIRECT_MULTIDIM_ELEM(v,n) ((v).data[(n)]) ++#define FOR_ALL_DIRECT_ELEMENTS_IN_MULTIDIMARRAY(v) \ ++ for (long int n=0; n ++class MultidimArray ++{ ++public: ++ T* data; ++ bool destroyData; ++ long int ndim; ++ long int zdim; ++ long int ydim; ++ long int xdim; ++ long int yxdim; ++ long int zyxdim; ++ long int nzyxdim; ++ long int zinit; ++ long int yinit; ++ long int xinit; ++ long int nzyxdimAlloc; ++ ++public: ++ void clear() ++ { ++ coreDeallocate(); ++ coreInit(); ++ } ++ ++ void coreInit() ++ { ++ xdim=0; ++ yxdim=0; ++ zyxdim=0; ++ nzyxdim=0; ++ ydim=1; ++ zdim=1; ++ ndim=1; ++ zinit=0; ++ yinit=0; ++ xinit=0; ++ data=NULL; ++ nzyxdimAlloc = 0; ++ destroyData=true; ++ } ++ ++ void coreAllocate(long int _ndim, long int _zdim, long int _ydim, long int _xdim) ++ { ++ if (_ndim <= 0 || _zdim <= 0 || _ydim<=0 || _xdim<=0) ++ { ++ clear(); ++ return; ++ } ++ ++ ndim=_ndim; ++ zdim=_zdim; ++ ydim=_ydim; ++ xdim=_xdim; ++ yxdim=ydim*xdim; ++ zyxdim=zdim*yxdim; ++ nzyxdim=ndim*zyxdim; ++ ++ coreAllocate(); ++ } ++ ++ void coreAllocate() ++ { ++ data = (T*)RELION_ALIGNED_MALLOC(sizeof(T) * nzyxdim); ++ nzyxdimAlloc = nzyxdim; ++ } ++ ++ void coreDeallocate() ++ { ++ if (data != NULL && destroyData) ++ { ++ RELION_ALIGNED_FREE(data); ++ } ++ data=NULL; ++ nzyxdimAlloc = 0; ++ } ++ ++ void resize(long int Ndim, long int Zdim, long int Ydim, long int Xdim) ++ { ++ if (Ndim*Zdim*Ydim*Xdim == nzyxdimAlloc && data != NULL) ++ { ++ ndim = Ndim; ++ xdim = Xdim; ++ ydim = Ydim; ++ zdim = Zdim; ++ yxdim = Ydim * Xdim; ++ zyxdim = Zdim * yxdim; ++ nzyxdim = Ndim * zyxdim; ++ nzyxdimAlloc = nzyxdim; ++ return; ++ } ++ ++ if (Xdim <= 0 || Ydim <= 0 || Zdim <= 0 || Ndim <= 0) ++ { ++ clear(); ++ return; ++ } ++ ++ if (NZYXSIZE(*this) > 0 && data == NULL) ++ { ++ coreAllocate(); ++ return; ++ } ++ ++ size_t YXdim=Ydim*Xdim; ++ size_t ZYXdim=Zdim*YXdim; ++ size_t NZYXdim=Ndim*ZYXdim; ++ ++ T * new_data = (T*)RELION_ALIGNED_MALLOC(sizeof(T) * NZYXdim); ++ for (long int l = 0; l < Ndim; l++) ++ for (long int k = 0; k < Zdim; k++) ++ for (long int i = 0; i < Ydim; i++) ++ for (long int j = 0; j < Xdim; j++) ++ { ++ T val; ++ new_data[l*ZYXdim + k*YXdim+i*Xdim+j] = val; ++ } ++ coreDeallocate(); ++ ++ data = new_data; ++ ndim = Ndim; ++ xdim = Xdim; ++ ydim = Ydim; ++ zdim = Zdim; ++ yxdim = Ydim * Xdim; ++ zyxdim = Zdim * yxdim; ++ nzyxdim = Ndim * zyxdim; ++ nzyxdimAlloc = nzyxdim; ++ } ++ ++ void resize(long int Xdim) ++ { ++ resize(1, 1, 1, Xdim); ++ } ++ ++ inline T& operator()(long int i, long int j) const ++ { ++ return A2D_ELEM(*this, i, j); ++ } ++ ++ inline T& operator()(long int i) const ++ { ++ return A1D_ELEM(*this, i); ++ } ++ ++ void getRow(long int i, MultidimArray& v) const ++ { ++ if (xdim == 0 || ydim == 0) ++ { ++ v.clear(); ++ return; ++ } ++ ++ v.resize(xdim); ++ for (long int j = 0; j < xdim; j++) ++ v(j) = (*this)(i, j); ++ } ++}; ++ ++#endif /* MULTIDIM_ARRAY_H */ +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c +index 9bc6cc32b..9f8a5c307 100644 +--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c +@@ -39,13 +39,13 @@ main (int argc, char *argv[]) + return 0; + } + +-/* { dg-final { scan-tree-dump-times "ref_count = (?:\[3-9\]|\[1-9\]\\d{1,}), ninsns = \[1-9\]\\d*, mem_to_insn_ratio = 0.\[2-9\]\\d*" 2 "llc_allocate" } } */ +-/* { dg-final { scan-tree-dump-times "Tracing succeeded" 6 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "ref_count = (?:\[3-9\]|\[1-9\]\\d{1,}), ninsns = \[1-9\]\\d*, mem_to_insn_ratio = 0.\[2-9\]\\d*" 4 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "Tracing succeeded" 14 "llc_allocate" } } */ + /* { dg-final { scan-tree-dump-not "Tracing failed" "llc_allocate" } } */ + /* { dg-final { scan-tree-dump-not "static_data_size:" "llc_allocate" } } */ +-/* { dg-final { scan-tree-dump-times "\{ (?:\\d+\\(\\d+\\) ){1}\}" 4 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times "\{ (?:\\d+\\(\\d+\\) ){1}\}" 2 "llc_allocate" } } */ + /* { dg-final { scan-tree-dump-not ", size: (?!(0\.000000))" "llc_allocate" } } */ +-/* { dg-final { scan-tree-dump-times ", size: 0\.000000" 6 "llc_allocate" } } */ ++/* { dg-final { scan-tree-dump-times ", size: 0\.000000" 8 "llc_allocate" } } */ + /* { dg-final { scan-tree-dump-times "\\d x_data \\(0.000000, 1, 0\\) : 3" 2 "llc_allocate" } } */ + /* { dg-final { scan-tree-dump-times "\\d A_j \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } */ + /* { dg-final { scan-tree-dump-times "\\d A_data \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } */ +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp b/gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp +index 4f34e722f..05a3bf842 100644 +--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-allocate.exp +@@ -24,4 +24,4 @@ dg-runtest [lsort [glob -nocomplain $srcdir/$subdir/*.c]] \ + "" "-fllc-allocate" + + # All done. +-dg-finish ++dg-finish +\ No newline at end of file +diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-extend-outer-loop.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-extend-outer-loop.c +new file mode 100644 +index 000000000..9b2b656fd +--- /dev/null ++++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-extend-outer-loop.c +@@ -0,0 +1,61 @@ ++/* { dg-do compile { target { aarch64*-*-linux* } } } */ ++/* { dg-options "-O3 -march=armv8.2-a+sve -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param=outer-loop-nums=10 --param=issue-topn=4 --param=force-issue=1 --param=filter-kernels=0" } */ ++#include ++#define N 131590 ++#define F 384477 ++ ++int ownStartPtr[F]; ++double bPrimePtr[N]; ++double diagPtr[N]; ++double psiPtr[N]; ++double upperPtr[F]; ++double lowerPtr[F]; ++int uPtr[F]; ++ ++void SMOOTH(int *ownStartPtr, double *bPrimePtr, double *diagPtr, double *psiPtr, int *uPtr, double *lowerPtr, double *upperPtr, int nCells); ++ ++int main(int argc, char *argv[]) ++{ ++ int nCells = N; ++ int nFaces = F; ++ int testIter = 2; ++ for (int i = 0; i < testIter; i++) ++ { ++ SMOOTH(ownStartPtr, bPrimePtr, diagPtr, psiPtr, uPtr, lowerPtr, upperPtr, nCells); ++ } ++ return 0; ++} ++ ++ ++void SMOOTH(int *ownStartPtr, double *bPrimePtr, double *diagPtr, double *psiPtr, int *uPtr, double *lowerPtr, double *upperPtr, int nCells) ++{ ++ double psii; ++ int fStart; ++ int fEnd = ownStartPtr[0]; ++ ++ for (int celli = 0; celli < nCells; celli++) ++ { ++ fStart = fEnd; ++ fEnd = ownStartPtr[celli + 1]; ++ psii = bPrimePtr[celli]; ++ ++ for (int facei = fStart; facei &references) + + struct loop_filter_out_flag + { +- /* Use external gimple. */ +- bool use_ext_gimple; +- + /* Use external call. */ + bool use_ext_call; + +@@ -358,21 +355,7 @@ bool + filter_out_loop_by_stmt_p (loop_filter_out_flag &loop_filter, gimple *stmt, + const vector &references, unsigned int &start) + { +- /* check use_ext_gimple. */ +- expanded_location cfun_xloc +- = expand_location (DECL_SOURCE_LOCATION (current_function_decl)); + expanded_location xloc = expand_location (stmt->location); +- if (xloc.file && filename_cmp (cfun_xloc.file, xloc.file)) +- { +- if (dump_file && (dump_flags & TDF_DETAILS)) +- { +- fprintf (dump_file, "use_ext_gimple: "); +- print_gimple_stmt (dump_file, stmt, 0, TDF_LINENO); +- } +- loop_filter.use_ext_gimple = true; +- return true; +- } +- + /* check use_ext_call. */ + if (gimple_code (stmt) == GIMPLE_CALL && !gimple_call_internal_p (stmt)) + { +@@ -421,11 +404,6 @@ filter_out_loop_by_stmt_p (loop_filter_out_flag &loop_filter, gimple *stmt, + void + dump_loop_filter_out_flag (loop_filter_out_flag &loop_filter) + { +- if (loop_filter.use_ext_gimple) +- { +- if (dump_file && (dump_flags & TDF_DETAILS)) +- fprintf (dump_file, "non-dense mem access: use_ext_gimple\n"); +- } + if (loop_filter.use_ext_call) + { + if (dump_file && (dump_flags & TDF_DETAILS)) +@@ -493,45 +471,6 @@ get_references_in_loop (vector &references, + return !filter_out_loop; + } + +-/* Determine whether the loop is a single path. */ +- +-bool +-single_path_p (class loop *loop, basic_block bb) +-{ +- if (bb == NULL) +- return false; +- if (bb == loop->latch) +- return true; +- +- gimple *stmt = last_stmt (bb); +- bool res = false; +- +- if (stmt && gimple_code (stmt) == GIMPLE_COND) +- { +- gcc_assert (EDGE_COUNT (bb->succs) == 2); +- edge true_edge = NULL; +- edge false_edge = NULL; +- extract_true_false_edges_from_block (bb, &true_edge, &false_edge); +- +- /* Returns false, if a branch occurs. */ +- if (true_edge->dest->loop_father == loop +- && false_edge->dest->loop_father == loop) +- return false; +- +- if (true_edge->dest->loop_father == loop) +- res = single_path_p (loop, true_edge->dest); +- else +- res = single_path_p (loop, false_edge->dest); +- } +- else +- { +- edge e = find_fallthru_edge (bb->succs); +- if (e) +- res = single_path_p (loop, e->dest); +- } +- return res; +-} +- + /* Computes an estimated number of insns in LOOP, weighted by WEIGHTS. + Assume that the HPC data reading and calculation process does not involve + adding branches in loops. Therefore, all bbs of loops are directly used for +@@ -611,6 +550,45 @@ dense_memory_p (const vector &references, class loop *loop) + + /* Analyze the inner loop and get the loop with dense memory access. */ + ++void ++analyze_loop_dense_memory (vector &kernels, ++ map > &kernels_refs, ++ class loop *loop) ++{ ++ vector references; ++ number_of_latch_executions (loop); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "\n========== Processing loop %d: ==========\n", ++ loop->num); ++ loop_dump (dump_file, loop); ++ flow_loop_dump (loop, dump_file, NULL, 1); ++ fprintf (dump_file, "loop unroll: %d\n", loop->unroll); ++ } ++ ++ if (get_loop_exit_edges (loop).length () != 1) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "non-dense mem access: loop_branching\n"); ++ return; ++ } ++ ++ loop_filter_out_flag loop_filter = {false, false, true, false}; ++ ++ if (!get_references_in_loop (references, loop_filter, loop)) ++ { ++ dump_loop_filter_out_flag (loop_filter); ++ return; ++ } ++ ++ if (dense_memory_p (references, loop)) ++ { ++ kernels_refs[loop] = references; ++ kernels.push_back (loop); ++ } ++} ++/* Analyze the inner loop and get the loop with dense memory access. */ ++ + bool + get_dense_memory_kernels (vector &kernels, + map > &kernels_refs) +@@ -619,40 +597,7 @@ get_dense_memory_kernels (vector &kernels, + fprintf (dump_file, "\nPhase 1: get_dense_memory_kernels\n\n"); + class loop *loop = NULL; + FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) +- { +- number_of_latch_executions (loop); +- if (dump_file && (dump_flags & TDF_DETAILS)) +- { +- fprintf (dump_file, "\n========== Processing loop %d: ==========\n", +- loop->num); +- loop_dump (dump_file, loop); +- flow_loop_dump (loop, dump_file, NULL, 1); +- fprintf (dump_file, "loop unroll: %d\n", loop->unroll); +- } +- +- if (get_loop_exit_edges (loop).length () != 1 +- || !single_path_p (loop, loop->header)) +- { +- if (dump_file && (dump_flags & TDF_DETAILS)) +- fprintf (dump_file, "non-dense mem access: loop_branching\n"); +- continue; +- } +- +- vector references; +- loop_filter_out_flag loop_filter = {false, false, false, true, false}; +- +- if (!get_references_in_loop (references, loop_filter, loop)) +- { +- dump_loop_filter_out_flag (loop_filter); +- continue; +- } +- +- if (dense_memory_p (references, loop)) +- { +- kernels_refs[loop] = references; +- kernels.push_back (loop); +- } +- } ++ analyze_loop_dense_memory (kernels, kernels_refs, loop); + return kernels.size () > 0; + } + +@@ -1094,33 +1039,41 @@ trace_ref_info (data_ref &mem_ref, set &traced_ref_stmt) + mem_ref.trace_status_p = true; + } + ++/* Trace all references in the loop. */ ++ ++void ++trace_loop_refs_info (vector &refs, set &traced_ref_stmt) ++{ ++ for (unsigned i = 0; i < refs.size (); ++i) ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "trace_references_base_info %d:\n", i); ++ print_generic_expr (dump_file, refs[i].ref, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ trace_ref_info (refs[i], traced_ref_stmt); ++ } ++} ++ + /* Tracing and sorting reference groups. */ + + void + trace_data_refs_info (vector &kernels, +- map > &loop_refs) ++ map > &loop_refs, ++ set &traced_ref_stmt) + { + if (dump_file) + fprintf (dump_file, "\nPhase 2: trace_all_references_info\n\n"); + +- set traced_ref_stmt; +- + for (unsigned i = 0; i < kernels.size (); ++i) + { +- class loop* loop = kernels[i]; +- ++ class loop *loop = kernels[i]; ++ if (loop_refs.count (loop) == 0) ++ continue; + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "loop header %d:\n", loop->header->index); +- for (unsigned j = 0; j < loop_refs[loop].size (); ++j) +- { +- if (dump_file && (dump_flags & TDF_DETAILS)) +- { +- fprintf (dump_file, "trace_references_base_info %d:\n", j); +- print_generic_expr (dump_file, loop_refs[loop][j].ref, TDF_SLIM); +- fprintf (dump_file, "\n"); +- } +- trace_ref_info (loop_refs[loop][j], traced_ref_stmt); +- } ++ trace_loop_refs_info (loop_refs[loop], traced_ref_stmt); + } + } + +@@ -1205,7 +1158,7 @@ void + check_bound_iv_and_add_worklist (vector &worklist, set &walked, + tree t, data_ref &mem_ref) + { +- if (TREE_CODE (t) != SSA_NAME) ++ if (t == NULL_TREE || TREE_CODE (t) != SSA_NAME) + return; + + gimple *def_stmt = SSA_NAME_DEF_STMT (t); +@@ -1278,8 +1231,13 @@ trace_loop_bound_iv (data_ref &mem_ref) + } + + if (dump_file && (dump_flags & TDF_DETAILS)) +- fprintf (dump_file, "\nmem_ref access dimension: %ld\n", +- mem_ref.loop_bounds.size ()); ++ { ++ fprintf (dump_file, "\nmem_ref access dimension: %ld\n", ++ mem_ref.loop_bounds.size ()); ++ fprintf (dump_file, "Traced variables: "); ++ print_generic_expr (dump_file, mem_ref.base, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } + + return mem_ref.loop_bounds.size () > 0; + } +@@ -1487,7 +1445,7 @@ trace_and_create_dominate_loop_bounds (data_ref &mem_ref) + if (dump_file && (dump_flags & TDF_DETAILS)) + { + print_generic_expr (dump_file, mem_ref.ref, TDF_SLIM); +- fprintf (dump_file, "Tracing loop bound failed at dimension %d", ++ fprintf (dump_file, "Tracing loop bound failed at dimension %d\n", + i); + } + mem_ref.calc_by = UNHANDLE_CALC; +@@ -1565,42 +1523,246 @@ trace_ref_dimension_and_loop_bounds (data_ref &mem_ref) + static_calculate_data_size (mem_ref); + } + +-/* analyze nested kernels. +- 1. multidimension loop analyze. +- 2. extended outer loop analyze. +- Later we will extend outer loop analysis. ++/* Get the loop's niters tree. ++ Return NULL_TREE if not found. */ ++ ++tree ++get_cur_loop_niters (map > &loop_refs, ++ class loop* loop) ++{ ++ if (loop_refs.count (loop) == 0) ++ return NULL_TREE; ++ vector bounds = loop_refs[loop][0].loop_bounds; ++ return bounds.size () ? bounds[0].niters : NULL_TREE; ++} ++ ++/* Trace the sources of the niters tree and return the ++ outermost depth of the loops containing them. ++ Return start_depth if not found. ++ ++ example: ++ niters:(long) (((int) i_end_417 - (int) i_start_452) + 1) ++ operand_num: 1, subtree:(long) (((int) i_end_417 - (int) i_start_452) + 1) ++ operand_num: 2, subtree:((int) i_end_417 - (int) i_start_452) + 1 ++ operand_num: 2, subtree:(int) i_end_417 - (int) i_start_452 ++ operand_num: 1, subtree:(int) i_end_417 ++ SSA_NAME of niters: i_end_417 ++ gimple of SSA: i_end_417 = PHI ++ return gimple depth; ++*/ ++ ++unsigned ++trace_outer_loop_depth (tree niters, unsigned start_depth) ++{ ++ /* If niter does not exist or the type is INTEGER_CST, ++ the loop bound is determined and return start_depth. */ ++ if (niters == NULL_TREE || TREE_CODE (niters) == INTEGER_CST) ++ return start_depth; ++ ++ gimple *def_stmt = NULL; ++ /* niters examples: i_start_452, fEnd_35, fEnd_100. */ ++ enum tree_code niter_code = TREE_CODE (niters); ++ if (niter_code == SSA_NAME) ++ { ++ /* Trace the SSA that define this niter. */ ++ def_stmt = SSA_NAME_DEF_STMT (niters); ++ enum gimple_code stmt_code = gimple_code (def_stmt); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "ssa_name of niters: "); ++ print_generic_expr (dump_file, niters); ++ fprintf (dump_file, "\ngimple of ssa: \n"); ++ print_gimple_stmt (dump_file, def_stmt, 0, TDF_LINENO); ++ fprintf (dump_file, "\n"); ++ } ++ /* Termination condition of dfs. Return the depth of the bb block. */ ++ if (stmt_code == GIMPLE_PHI || stmt_code == GIMPLE_NOP) ++ { ++ basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (niters)); ++ if (def_bb == NULL || def_bb->loop_father == NULL) ++ return start_depth; ++ unsigned ret_depth = loop_depth (def_bb->loop_father); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "Stop tracing the outer loop depth, "); ++ fprintf (dump_file, "current depth: %d, current bb: %d\n", \ ++ ret_depth, def_bb->index); ++ } ++ return ret_depth; ++ } ++ /* 'ASSIGN': Use dfs to trace the rhs of the assignment statement. */ ++ else if (stmt_code == GIMPLE_ASSIGN) ++ { ++ tree rhs = gimple_assign_rhs1 (def_stmt); ++ if (TREE_CODE (rhs) == TARGET_MEM_REF) ++ /* fEnd_35 = MEM[base: _19, index: ivtmp.96, step: 4, ++ offset: 0B] */ ++ return trace_outer_loop_depth (TREE_OPERAND (rhs, 2), start_depth); ++ else ++ { ++ /* M.218_658 = MIN_EXPR <_631, _657> */ ++ unsigned min_depth = start_depth; ++ unsigned operand_num = gimple_num_ops (def_stmt); ++ /* 'ASSIGN': start from 1 because op[0] is the lhs. */ ++ for (unsigned i = 1; i < operand_num; i++) ++ { ++ tree subtree = GIMPLE_CHECK2 ++ (def_stmt)->op[i]; ++ if (subtree == NULL) ++ continue; ++ unsigned depth = trace_outer_loop_depth (subtree, \ ++ start_depth); ++ min_depth = MIN (min_depth, depth); ++ } ++ return min_depth; ++ } ++ } ++ else ++ { ++ /* Adding termination conditions: ++ 1. Niters is MEM variable; ++ 2. Niters is a runtime value (smooth_uPtr), and consider \ ++ finding footprint in other mem_ref; ++ 3. Niters is loop variable (i_start/i_end), and the boundary in \ ++ the outer loop depends on the variable j_start/j_end. */ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "The loop termination condition"); ++ fprintf (dump_file, "is to be extended.\n"); ++ } ++ return start_depth; ++ } ++ } ++ /* The operand nums can be obtained when the tree code is as follows. */ ++ else if (niter_code == NOP_EXPR || niter_code == MEM_REF ++ || niter_code == ARRAY_REF || niter_code == COND_EXPR ++ || niter_code == PLUS_EXPR || niter_code == MINUS_EXPR ++ || niter_code == TARGET_MEM_REF || niter_code == POINTER_PLUS_EXPR) ++ { ++ /* operand_num is the operand in the niters statement. ++ example: In the following niter statement, operand_num = 3. ++ (unsigned int) fEnd_35 - (unsigned int) fEnd_100 + 4294967295. */ ++ unsigned operand_num = TREE_OPERAND_LENGTH (niters); ++ unsigned min_depth = start_depth; ++ for (unsigned i = 0; i < operand_num; i++) ++ { ++ tree subtree = TREE_OPERAND (niters, i); ++ if (subtree == NULL) ++ continue; ++ unsigned depth = trace_outer_loop_depth (subtree, start_depth); ++ min_depth = MIN (min_depth, depth); ++ } ++ return min_depth; ++ } ++ else ++ { ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "niters is another tree code: %s\n", \ ++ get_tree_code_name (niter_code)); ++ print_generic_expr (dump_file, niters, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ return start_depth; ++ } ++} ++ ++/* Traces the ref dimension information in each loop. */ ++ ++void ++analyze_loop_refs_dimension (vector &refs) ++{ ++ for (unsigned i = 0; i < refs.size (); ++i) ++ { ++ if (refs[i].trace_status_p == false) ++ continue; ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "trace_reference_dimension %d:\n", i); ++ print_generic_expr (dump_file, refs[i].ref, TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } ++ trace_ref_dimension_and_loop_bounds (refs[i]); ++ } ++} ++/* analyze nested kernels ++ 1. multidimension loop analyze ++ 2. extended outer loop analyze + */ + + bool + analyze_nested_kernels (vector &kernels, +- map > &loop_refs) ++ map > &loop_refs, ++ set &traced_ref_stmt) + { + if (dump_file) + fprintf (dump_file, "\nPhase 3: analyze_nested_kernels\n\n"); + +- for (unsigned i = 0; i < kernels.size (); ++i) ++ /* `kernels` may be added in during outer loop extension phase, ++ thus using initial size to avoid repeatedly analyzing. */ ++ unsigned init_kernels_size = kernels.size (); ++ for (unsigned i = 0; i < init_kernels_size; ++i) + { + class loop* loop = kernels[i]; + if (loop_refs.count (loop) == 0) + continue; + + if (dump_file && (dump_flags & TDF_DETAILS)) +- fprintf (dump_file, "\n\nloop header %d:\n", loop->header->index); +- for (unsigned j = 0; j < loop_refs[loop].size (); ++j) ++ fprintf (dump_file, "loop header %d:\n", loop->header->index); ++ analyze_loop_refs_dimension (loop_refs[loop]); ++ ++ unsigned depth = loop_depth (loop); ++ unsigned outer_depth = trace_outer_loop_depth (get_cur_loop_niters \ ++ (loop_refs, loop), depth); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "cur_depth: %d, outer_depth: %d\n", \ ++ depth, outer_depth); ++ /* param_outer_loop_num: number of loops of the extended outer loop. ++ Outermost loop should not be extended when outer_depth = 0. ++ `outer_depth == depth` means the current loop is the loop which ++ boundary is known, so there is no need to extend the outer loop. */ ++ if (outer_depth == 0 || outer_depth == depth ++ || depth > outer_depth + param_outer_loop_num) ++ continue; ++ /* Extend outer loop. */ ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ fprintf (dump_file, "\nStart extending outer loop\n"); ++ /* Superloops of the loop, start from the loop closest to the \ ++ current loop in the outermost loop. */ ++ for (unsigned j = 0; j < param_outer_loop_num && --depth; ++j) + { +- if (loop_refs[loop][j].trace_status_p == false) ++ class loop* outer_loop = (*loop->superloops)[depth]; ++ /* The outer loop may be added when analyzing previous inner loops, ++ i.e. the outer loop contains two or more inner loops. */ ++ if (loop_refs.count (outer_loop)) + continue; +- +- if (dump_file && (dump_flags & TDF_DETAILS)) ++ /* phase1~phase3 analysis on the extended outer loop. */ ++ analyze_loop_dense_memory (kernels, loop_refs, outer_loop); ++ if (loop_refs.count (outer_loop) == 0) ++ continue; ++ for (unsigned k = 0; k < loop_refs[outer_loop].size (); ++k) + { +- fprintf (dump_file, "\ntrace_reference_dimension at mem_ref " +- "index %d in loop %d:\n", j, loop->num); +- print_generic_expr (dump_file, loop_refs[loop][j].ref, TDF_SLIM); +- fprintf (dump_file, "\n"); ++ if (dump_file && (dump_flags & TDF_DETAILS)) ++ { ++ fprintf (dump_file, "outer_analyze_nested_kernels %d: ", k); ++ print_generic_expr (dump_file, loop_refs[outer_loop][k].ref,\ ++ TDF_SLIM); ++ fprintf (dump_file, "\n"); ++ } + } +- trace_ref_dimension_and_loop_bounds (loop_refs[loop][j]); ++ trace_loop_refs_info (loop_refs[outer_loop], traced_ref_stmt); ++ analyze_loop_refs_dimension (loop_refs[outer_loop]); ++ outer_depth = trace_outer_loop_depth (get_cur_loop_niters \ ++ (loop_refs, outer_loop), depth); ++ /* `outer_depth == depth` means the current loop is the loop which ++ boundary is known, so there is no need to extend the outer loop. */ ++ if (outer_depth == depth) ++ break; ++ else ++ /* The outer loop cannot find the current loop boundary, ++ Remove the record of outer_loop from the loop_refs. */ ++ loop_refs.erase (outer_loop); + } +- + } + return true; + } +@@ -2694,9 +2856,10 @@ llc_allocate (void) + if (!get_dense_memory_kernels (kernels, kernels_refs)) + return; + +- trace_data_refs_info (kernels, kernels_refs); ++ set traced_ref_stmt; ++ trace_data_refs_info (kernels, kernels_refs, traced_ref_stmt); + +- if (!analyze_nested_kernels (kernels, kernels_refs)) ++ if (!analyze_nested_kernels (kernels, kernels_refs, traced_ref_stmt)) + return; + + vector sorted_kernels; +-- +2.33.0 + diff --git a/gcc.spec b/gcc.spec index 49207d9..6eb6e0a 100644 --- a/gcc.spec +++ b/gcc.spec @@ -46,7 +46,7 @@ %else %global build_libitm 0 %endif -%global build_isl 0 +%global build_isl 1 %global build_libstdcxx_docs 0 %ifarch %{ix86} x86_64 ppc ppc64 ppc64le ppc64p7 s390 s390x %{arm} aarch64 %{mips} %global attr_ifunc 1 @@ -61,7 +61,7 @@ Summary: Various compilers (C, C++, Objective-C, ...) Name: gcc Version: %{gcc_version} -Release: 41 +Release: 42 License: GPLv3+ and GPLv3+ with exceptions and GPLv2+ with exceptions and LGPLv2+ and BSD URL: https://gcc.gnu.org @@ -249,6 +249,18 @@ Patch138: 0138-Fix-ICE-bugs-in-transpose-test-cases-with-vector-ind.patch Patch139: 0139-Fix-errors-on-testsuite-c-c-tests-and-505.mcf_r.patch Patch140: 0140-Fix-an-error-in-memory-allocation-deallocation.patch Patch141: 0141-Fix-warnings-and-errors-with-debug-prints.patch +Patch142: 0142-crc-loop-optimization-initial.patch +Patch143: 0143-Perform-early-if-conversion-of-simple-arithmetic.patch +Patch144: 0144-Add-option-to-allow-matching-uaddsub-overflow-for-wi.patch +Patch145: 0145-Match-double-sized-mul-pattern.patch +Patch146: 0146-LOOP-CRC32-Add-Crc32-Optimization-in-Gzip-For-crc32-.patch +Patch147: 0147-add-insn-defs-and-correct-costs-for-cmlt-generation.patch +Patch148: 0148-Introduce-RTL-ifcvt-enhancements.patch +Patch149: 0149-Add-more-flexible-check-for-pointer-aliasing-during-.patch +Patch150: 0150-Implement-propagation-of-permutations-in-fwprop.patch +Patch151: 0151-Fix-bugs-and-add-tests-for-RTL-ifcvt.patch +Patch152: 0152-Add-LLC-Allocation-Pass.patch +Patch153: 0153-LLC-add-extending-outer-loop.patch %global gcc_target_platform %{_arch}-linux-gnu @@ -843,6 +855,18 @@ not stable, so plugins must be rebuilt any time GCC is updated. %patch139 -p1 %patch140 -p1 %patch141 -p1 +%patch142 -p1 +%patch143 -p1 +%patch144 -p1 +%patch145 -p1 +%patch146 -p1 +%patch147 -p1 +%patch148 -p1 +%patch149 -p1 +%patch150 -p1 +%patch151 -p1 +%patch152 -p1 +%patch153 -p1 %build @@ -908,15 +932,10 @@ CC="$CC" CFLAGS="$OPT_FLAGS" \ --with-system-zlib --enable-__cxa_atexit --disable-libunwind-exceptions \ --enable-gnu-unique-object --enable-linker-build-id --with-linker-hash-style=gnu \ --enable-languages=c,c++,fortran${enablelobjc}${enablelada}${enablelgo}${enableld},lto --enable-plugin \ - --enable-initfini-array --disable-libgcj --without-cloog \ + --enable-initfini-array --disable-libgcj --with-isl --without-cloog \ --enable-gnu-indirect-function --build=%{gcc_target_platform} \ --with-stage1-ldflags="$OPT_LDFLAGS" \ --with-boot-ldflags="$OPT_LDFLAGS" --disable-bootstrap \ -%if %{build_isl} - --with-isl \ -%else - --without-isl \ -%endif %ifarch x86_64 --with-tune=generic \ --with-arch_32=x86-64 \ @@ -930,11 +949,6 @@ CC="$CC" CFLAGS="$OPT_FLAGS" \ --with-arch=rv64g --with-abi=lp64d \ --disable-libquadmath --disable-multilib %endif -%ifarch ppc64le - --disable-multilib \ - --enable-targets=powerpcle-linux \ - --with-cpu-32=power8 --with-tune-32=power8 --with-cpu-64=power8 --with-tune-64=power8 \ -%endif %ifarch sparc sparcv9 sparc64 make %{?_smp_mflags} BOOT_CFLAGS="$OPT_FLAGS" bootstrap @@ -2877,6 +2891,12 @@ end %doc rpm.doc/changelogs/libcc1/ChangeLog* %changelog +* Wed May 29 2024 zhengchenhui - 10.3.1-42 +- Type:Spec +- ID:NA +- SUG:NA +- DESC: Revert last two commits about isl and ppc64le, and Sync patch from openeuler/gcc + * Mon Apr 15 2024 huyubiao - 10.3.1-41 - Type:SPEC - ID:NA -- Gitee