From 61a70e6b6b44bf420eae559d998e109b70e5a9b6 Mon Sep 17 00:00:00 2001
From: Lulu Cheng <chenglulu@loongson.cn>
Date: Fri, 17 Nov 2023 16:04:45 +0800
Subject: [PATCH 034/188] LoongArch: atomic_load and atomic_store are
 implemented using dbar grading.

Because the la464 memory model design allows the same address load out of order,
so in the following test example, the Load of 23 lines may be executed first over
the load of 21 lines, resulting in an error.
So when memmodel is MEMMODEL_RELAXED, the load instruction will be followed by
"dbar 0x700" when implementing _atomic_load.

  1 void *
  2 gomp_ptrlock_get_slow (gomp_ptrlock_t *ptrlock)
  3 {
  4   int *intptr;
  5   uintptr_t oldval = 1;
  6
  7   __atomic_compare_exchange_n (ptrlock, &oldval, 2, false,
  8                                MEMMODEL_RELAXED, MEMMODEL_RELAXED);
  9
 10   /* futex works on ints, not pointers.
 11      But a valid work share pointer will be at least
 12      8 byte aligned, so it is safe to assume the low
 13      32-bits of the pointer won't contain values 1 or 2.  */
 14   __asm volatile ("" : "=r" (intptr) : "0" (ptrlock));
 15 #if __BYTE_ORDER == __BIG_ENDIAN
 16   if (sizeof (*ptrlock) > sizeof (int))
 17     intptr += (sizeof (*ptrlock) / sizeof (int)) - 1;
 18 #endif
 19   do
 20     do_wait (intptr, 2);
 21   while (__atomic_load_n (intptr, MEMMODEL_RELAXED) == 2);
 22   __asm volatile ("" : : : "memory");
 23   return (void *) __atomic_load_n (ptrlock, MEMMODEL_ACQUIRE);
 24 }

gcc/ChangeLog:

	* config/loongarch/sync.md (atomic_load<mode>): New template.
---
 gcc/config/loongarch/sync.md | 70 +++++++++++++++++++++++++++++++++---
 1 file changed, 65 insertions(+), 5 deletions(-)

diff --git a/gcc/config/loongarch/sync.md b/gcc/config/loongarch/sync.md
index 1eabaec04..f4673c856 100644
--- a/gcc/config/loongarch/sync.md
+++ b/gcc/config/loongarch/sync.md
@@ -30,6 +30,7 @@
   UNSPEC_SYNC_OLD_OP
   UNSPEC_SYNC_EXCHANGE
   UNSPEC_ATOMIC_STORE
+  UNSPEC_ATOMIC_LOAD
   UNSPEC_MEMORY_BARRIER
 ])
 
@@ -103,16 +104,75 @@
 
 ;; Atomic memory operations.
 
+(define_insn "atomic_load<mode>"
+  [(set (match_operand:QHWD 0 "register_operand" "=r")
+    (unspec_volatile:QHWD
+      [(match_operand:QHWD 1 "memory_operand" "+m")
+       (match_operand:SI 2 "const_int_operand")]                        ;; model
+      UNSPEC_ATOMIC_LOAD))]
+  ""
+{
+  enum memmodel model = memmodel_base (INTVAL (operands[2]));
+
+  switch (model)
+    {
+    case MEMMODEL_SEQ_CST:
+      return "dbar\t0x11\\n\\t"
+	     "ld.<size>\t%0,%1\\n\\t"
+	     "dbar\t0x14\\n\\t";
+    case MEMMODEL_ACQUIRE:
+      return "ld.<size>\t%0,%1\\n\\t"
+	     "dbar\t0x14\\n\\t";
+    case MEMMODEL_RELAXED:
+      return "ld.<size>\t%0,%1\\n\\t"
+	     "dbar\t0x700\\n\\t";
+
+    default:
+      /* The valid memory order variants are __ATOMIC_RELAXED, __ATOMIC_SEQ_CST,
+	 __ATOMIC_CONSUME and __ATOMIC_ACQUIRE.
+	 The expand_builtin_atomic_store function converts all invalid memmodels
+	 to MEMMODEL_SEQ_CST.
+
+	 __atomic builtins doc: "Consume is implemented using the
+	 stronger acquire memory order because of a deficiency in C++11's
+	 semantics."  See PR 59448 and get_memmodel in builtins.cc.  */
+      gcc_unreachable ();
+    }
+}
+  [(set (attr "length") (const_int 12))])
+
 ;; Implement atomic stores with amoswap.  Fall back to fences for atomic loads.
 (define_insn "atomic_store<mode>"
-  [(set (match_operand:GPR 0 "memory_operand" "+ZB")
-    (unspec_volatile:GPR
-      [(match_operand:GPR 1 "reg_or_0_operand" "rJ")
+  [(set (match_operand:QHWD 0 "memory_operand" "+m")
+    (unspec_volatile:QHWD
+      [(match_operand:QHWD 1 "reg_or_0_operand" "rJ")
        (match_operand:SI 2 "const_int_operand")]      ;; model
       UNSPEC_ATOMIC_STORE))]
   ""
-  "amswap%A2.<amo>\t$zero,%z1,%0"
-  [(set (attr "length") (const_int 8))])
+{
+  enum memmodel model = memmodel_base (INTVAL (operands[2]));
+
+  switch (model)
+    {
+    case MEMMODEL_SEQ_CST:
+      return "dbar\t0x12\\n\\t"
+	     "st.<size>\t%z1,%0\\n\\t"
+	     "dbar\t0x18\\n\\t";
+    case MEMMODEL_RELEASE:
+      return "dbar\t0x12\\n\\t"
+	     "st.<size>\t%z1,%0\\n\\t";
+    case MEMMODEL_RELAXED:
+      return "st.<size>\t%z1,%0";
+
+    default:
+      /* The valid memory order variants are __ATOMIC_RELAXED, __ATOMIC_SEQ_CST,
+	 and __ATOMIC_RELEASE.
+	 The expand_builtin_atomic_store function converts all invalid memmodels
+	 to MEMMODEL_SEQ_CST.  */
+      gcc_unreachable ();
+    }
+}
+  [(set (attr "length") (const_int 12))])
 
 (define_insn "atomic_<atomic_optab><mode>"
   [(set (match_operand:GPR 0 "memory_operand" "+ZB")
-- 
2.43.0