LoongArch: Remove redundant barrier instructions before LL-SC loops

This is isomorphic to the LLVM changes [1-2].

On LoongArch, the LL and SC instructions has memory barrier semantics:

- LL: <memory-barrier> + <load-exclusive>
- SC: <store-conditional> + <memory-barrier>

But the compare and swap operation is allowed to fail, and if it fails
the SC instruction is not executed, thus the guarantee of acquiring
semantics cannot be ensured. Therefore, an acquire barrier needs to be
generated when failure_memorder includes an acquire operation.

On CPUs implementing LoongArch v1.10 or later, "dbar 0b10100" is an
acquire barrier; on CPUs implementing LoongArch v1.00, it is a full
barrier.  So it's always enough for acquire semantics.  OTOH if an
acquire semantic is not needed, we still needs the "dbar 0x700" as the
load-load barrier like all LL-SC loops.

[1]:llvm/llvm-project#67391
[2]:llvm/llvm-project#69339

gcc/ChangeLog:

	* config/loongarch/loongarch.cc
	(loongarch_memmodel_needs_release_fence): Remove.
	(loongarch_cas_failure_memorder_needs_acquire): New static
	function.
	(loongarch_print_operand): Redefine 'G' for the barrier on CAS
	failure.
	* config/loongarch/sync.md (atomic_cas_value_strong<mode>):
	Remove the redundant barrier before the LL instruction, and
	emit an acquire barrier on failure if needed by
	failure_memorder.
	(atomic_cas_value_cmp_and_7_<mode>): Likewise.
	(atomic_cas_value_add_7_<mode>): Remove the unnecessary barrier
	before the LL instruction.
	(atomic_cas_value_sub_7_<mode>): Likewise.
	(atomic_cas_value_and_7_<mode>): Likewise.
	(atomic_cas_value_xor_7_<mode>): Likewise.
	(atomic_cas_value_or_7_<mode>): Likewise.
	(atomic_cas_value_nand_7_<mode>): Likewise.
	(atomic_cas_value_exchange_7_<mode>): Likewise.

gcc/testsuite/ChangeLog:

	* gcc.target/loongarch/cas-acquire.c: New test.

Author: xry111

gcc/config/loongarch/loongarch.cc

@@ -4876,27 +4876,27 @@ loongarch_memmodel_needs_rel_acq_fence (enum memmodel model)
     }
 }
 
-/* Return true if a FENCE should be emitted to before a memory access to
-   implement the release portion of memory model MODEL.  */
+/* Return true if a FENCE should be emitted after a failed CAS to
+   implement the acquire semantic of failure_memorder.  */
 
 static bool
-loongarch_memmodel_needs_release_fence (enum memmodel model)
+loongarch_cas_failure_memorder_needs_acquire (enum memmodel model)
 {
-  switch (model)
+  switch (memmodel_base (model))
     {
+    case MEMMODEL_ACQUIRE:
     case MEMMODEL_ACQ_REL:
     case MEMMODEL_SEQ_CST:
-    case MEMMODEL_SYNC_SEQ_CST:
-    case MEMMODEL_RELEASE:
-    case MEMMODEL_SYNC_RELEASE:
       return true;
 
-    case MEMMODEL_ACQUIRE:
-    case MEMMODEL_CONSUME:
-    case MEMMODEL_SYNC_ACQUIRE:
     case MEMMODEL_RELAXED:
+    case MEMMODEL_RELEASE:
       return false;
 
+    /* MEMMODEL_CONSUME is deliberately not handled because it's always
+       replaced by MEMMODEL_ACQUIRE as at now.  If you see an ICE caused by
+       MEMMODEL_CONSUME, read the change (re)introducing it carefully and
+       decide what to do.  See PR 59448 and get_memmodel in builtins.cc.  */
     default:
       gcc_unreachable ();
     }
@@ -5006,7 +5006,8 @@ loongarch_print_operand_reloc (FILE *file, rtx op, bool hi64_part,
    'C'	Print the integer branch condition for comparison OP.
    'd'	Print CONST_INT OP in decimal.
    'F'	Print the FPU branch condition for comparison OP.
-   'G'	Print a DBAR insn if the memory model requires a release.
+   'G'	Print a DBAR insn for CAS failure (with an acquire semantic if
+	needed, otherwise a simple load-load barrier).
    'H'  Print address 52-61bit relocation associated with OP.
    'h'  Print the high-part relocation associated with OP.
    'i'	Print i if the operand is not a register.
@@ -5073,8 +5074,11 @@ loongarch_print_operand (FILE *file, rtx op, int letter)
       break;
 
     case 'G':
-      if (loongarch_memmodel_needs_release_fence ((enum memmodel) INTVAL (op)))
-	fputs ("dbar\t0", file);
+      if (loongarch_cas_failure_memorder_needs_acquire (
+	    memmodel_from_int (INTVAL (op))))
+	fputs ("dbar\t0b10100", file);
+      else
+	fputs ("dbar\t0x700", file);
       break;
 
     case 'h':

gcc/config/loongarch/sync.md

@@ -129,19 +129,18 @@
    (clobber (match_scratch:GPR 6 "=&r"))]
   ""
 {
-  return "%G5\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "bne\\t%0,%z2,2f\\n\\t"
 	 "or%i3\\t%6,$zero,%3\\n\\t"
 	 "sc.<amo>\\t%6,%1\\n\\t"
-	 "beq\\t$zero,%6,1b\\n\\t"
+	 "beqz\\t%6,1b\\n\\t"
 	 "b\\t3f\\n\\t"
 	 "2:\\n\\t"
-	 "dbar\\t0x700\\n\\t"
+	 "%G5\\n\\t"
 	 "3:\\n\\t";
 }
-  [(set (attr "length") (const_int 32))])
+  [(set (attr "length") (const_int 28))])
 
 (define_expand "atomic_compare_and_swap<mode>"
   [(match_operand:SI 0 "register_operand" "")   ;; bool output
@@ -234,8 +233,7 @@
    (clobber (match_scratch:GPR 7 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%2\\n\\t"
 	 "bne\\t%7,%z4,2f\\n\\t"
@@ -245,10 +243,10 @@
 	 "beq\\t$zero,%7,1b\\n\\t"
 	 "b\\t3f\\n\\t"
 	 "2:\\n\\t"
-	 "dbar\\t0x700\\n\\t"
+	 "%G6\\n\\t"
 	 "3:\\n\\t";
 }
-  [(set (attr "length") (const_int 40))])
+  [(set (attr "length") (const_int 36))])
 
 (define_expand "atomic_compare_and_swap<mode>"
   [(match_operand:SI 0 "register_operand" "")   ;; bool output
@@ -303,8 +301,7 @@
    (clobber (match_scratch:GPR 8 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%3\\n\\t"
 	 "add.w\\t%8,%0,%z5\\n\\t"
@@ -314,7 +311,7 @@
 	 "beq\\t$zero,%7,1b";
 }
 
-  [(set (attr "length") (const_int 32))])
+  [(set (attr "length") (const_int 28))])
 
 (define_insn "atomic_cas_value_sub_7_<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")				;; res
@@ -330,8 +327,7 @@
    (clobber (match_scratch:GPR 8 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%3\\n\\t"
 	 "sub.w\\t%8,%0,%z5\\n\\t"
@@ -340,7 +336,7 @@
 	 "sc.<amo>\\t%7,%1\\n\\t"
 	 "beq\\t$zero,%7,1b";
 }
-  [(set (attr "length") (const_int 32))])
+  [(set (attr "length") (const_int 28))])
 
 (define_insn "atomic_cas_value_and_7_<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")				;; res
@@ -356,8 +352,7 @@
    (clobber (match_scratch:GPR 8 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%3\\n\\t"
 	 "and\\t%8,%0,%z5\\n\\t"
@@ -366,7 +361,7 @@
 	 "sc.<amo>\\t%7,%1\\n\\t"
 	 "beq\\t$zero,%7,1b";
 }
-  [(set (attr "length") (const_int 32))])
+  [(set (attr "length") (const_int 28))])
 
 (define_insn "atomic_cas_value_xor_7_<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")				;; res
@@ -382,8 +377,7 @@
    (clobber (match_scratch:GPR 8 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%3\\n\\t"
 	 "xor\\t%8,%0,%z5\\n\\t"
@@ -393,7 +387,7 @@
 	 "beq\\t$zero,%7,1b";
 }
 
-  [(set (attr "length") (const_int 32))])
+  [(set (attr "length") (const_int 28))])
 
 (define_insn "atomic_cas_value_or_7_<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")				;; res
@@ -409,8 +403,7 @@
    (clobber (match_scratch:GPR 8 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%3\\n\\t"
 	 "or\\t%8,%0,%z5\\n\\t"
@@ -420,7 +413,7 @@
 	 "beq\\t$zero,%7,1b";
 }
 
-  [(set (attr "length") (const_int 32))])
+  [(set (attr "length") (const_int 28))])
 
 (define_insn "atomic_cas_value_nand_7_<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")				;; res
@@ -436,8 +429,7 @@
    (clobber (match_scratch:GPR 8 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%3\\n\\t"
 	 "and\\t%8,%0,%z5\\n\\t"
@@ -446,7 +438,7 @@
 	 "sc.<amo>\\t%7,%1\\n\\t"
 	 "beq\\t$zero,%7,1b";
 }
-  [(set (attr "length") (const_int 32))])
+  [(set (attr "length") (const_int 28))])
 
 (define_insn "atomic_cas_value_exchange_7_<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")
@@ -461,8 +453,7 @@
    (clobber (match_scratch:GPR 7 "=&r"))]
   ""
 {
-  return "%G6\\n\\t"
-	 "1:\\n\\t"
+  return "1:\\n\\t"
 	 "ll.<amo>\\t%0,%1\\n\\t"
 	 "and\\t%7,%0,%z3\\n\\t"
 	 "or%i5\\t%7,%7,%5\\n\\t"

gcc/testsuite/gcc.target/loongarch/cas-acquire.c

@@ -0,0 +1,82 @@
+/* { dg-do run } */
+/* { dg-require-effective-target c99_runtime } */
+/* { dg-require-effective-target pthread } */
+/* { dg-options "-std=c99 -pthread" } */
+
+/* https://github.com/llvm/llvm-project/pull/67391#issuecomment-1752403934
+   reported that this had failed with GCC and 3A6000.  */
+
+#include <pthread.h>
+#include <stdatomic.h>
+#include <stdbool.h>
+#include <stdio.h>
+
+static unsigned int tags[32];
+static unsigned int vals[32];
+
+static void *
+writer_entry (void *data)
+{
+  atomic_uint *pt = (atomic_uint *)tags;
+  atomic_uint *pv = (atomic_uint *)vals;
+
+  for (unsigned int n = 1; n < 10000; n++)
+    {
+      atomic_store_explicit (&pv[n & 31], n, memory_order_release);
+      atomic_store_explicit (&pt[n & 31], n, memory_order_release);
+    }
+
+  return NULL;
+}
+
+static void *
+reader_entry (void *data)
+{
+  atomic_uint *pt = (atomic_uint *)tags;
+  atomic_uint *pv = (atomic_uint *)vals;
+  int i;
+
+  for (;;)
+    {
+      for (i = 0; i < 32; i++)
+        {
+          unsigned int tag = 0;
+          bool res;
+
+          res = atomic_compare_exchange_weak_explicit (
+              &pt[i], &tag, 0, memory_order_acquire, memory_order_acquire);
+          if (!res)
+            {
+              unsigned int val;
+
+              val = atomic_load_explicit (&pv[i], memory_order_relaxed);
+              if (val < tag)
+                __builtin_trap ();
+            }
+        }
+    }
+
+  return NULL;
+}
+
+int
+main (int argc, char *argv[])
+{
+  pthread_t writer;
+  pthread_t reader;
+  int res;
+
+  res = pthread_create (&writer, NULL, writer_entry, NULL);
+  if (res < 0)
+    __builtin_trap ();
+
+  res = pthread_create (&reader, NULL, reader_entry, NULL);
+  if (res < 0)
+    __builtin_trap ();
+
+  res = pthread_join (writer, NULL);
+  if (res < 0)
+    __builtin_trap ();
+
+  return 0;
+}