i386: Use lock prefixed insn instead of MFENCE [PR95750]

Currently, __atomic_thread_fence(seq_cst) on x86 and x86-64 generates
mfence instruction. A dummy atomic instruction (a lock-prefixed instruction
or xchg with a memory operand) would provide the same sequential consistency
guarantees while being more efficient on most current CPUs. The mfence
instruction additionally orders non-temporal stores, which is not relevant
for atomic operations and are not ordered by seq_cst atomic operations anyway.

2020-07-20  Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog:
	PR target/95750
	* config/i386/i386.h (TARGET_AVOID_MFENCE):
	Rename from TARGET_USE_XCHG_FOR_ATOMIC_STORE.
	* config/i386/sync.md (mfence_sse2): Disable for TARGET_AVOID_MFENCE.
	(mfence_nosse): Enable also for TARGET_AVOID_MFENCE. Emit stack
	referred memory in word_mode.
	(mem_thread_fence): Do not generate mfence_sse2 pattern when
	TARGET_AVOID_MFENCE is true.
	(atomic_store<mode>): Update for rename.
	* config/i386/x86-tune.def (X86_TUNE_AVOID_MFENCE):
	Rename from X86_TUNE_USE_XCHG_FOR_ATOMIC_STORE.

gcc/testsuite/ChangeLog:
	PR target/95750
	* gcc.target/i386/pr95750.c: New test.

4 files changed, 36 insertions, 11 deletions

diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h
index f4a8f13..114967e 100644
--- a/gcc/config/i386/i386.h
+++ b/gcc/config/i386/i386.h
@@ -598,8 +598,7 @@ extern unsigned char ix86_tune_features[X86_TUNE_LAST];
 	ix86_tune_features[X86_TUNE_AVOID_FALSE_DEP_FOR_BMI]
 #define TARGET_ONE_IF_CONV_INSN \
 	ix86_tune_features[X86_TUNE_ONE_IF_CONV_INSN]
-#define TARGET_USE_XCHG_FOR_ATOMIC_STORE \
-	ix86_tune_features[X86_TUNE_USE_XCHG_FOR_ATOMIC_STORE]
+#define TARGET_AVOID_MFENCE ix86_tune_features[X86_TUNE_AVOID_MFENCE]
 #define TARGET_EMIT_VZEROUPPER \
 	ix86_tune_features[X86_TUNE_EMIT_VZEROUPPER]
 #define TARGET_EXPAND_ABS \
diff --git a/gcc/config/i386/sync.md b/gcc/config/i386/sync.md
index e221090..c682703 100644
--- a/gcc/config/i386/sync.md
+++ b/gcc/config/i386/sync.md
@@ -89,7 +89,8 @@
 (define_insn "mfence_sse2"
   [(set (match_operand:BLK 0)
 	(unspec:BLK [(match_dup 0)] UNSPEC_MFENCE))]
-  "TARGET_64BIT || TARGET_SSE2"
+  "(TARGET_64BIT || TARGET_SSE2)
+   && !TARGET_AVOID_MFENCE"
   "mfence"
   [(set_attr "type" "sse")
    (set_attr "length_address" "0")
@@ -100,8 +101,14 @@
   [(set (match_operand:BLK 0)
 	(unspec:BLK [(match_dup 0)] UNSPEC_MFENCE))
    (clobber (reg:CC FLAGS_REG))]
-  "!(TARGET_64BIT || TARGET_SSE2)"
-  "lock{%;} or{l}\t{$0, (%%esp)|DWORD PTR [esp], 0}"
+  "!(TARGET_64BIT || TARGET_SSE2)
+   || TARGET_AVOID_MFENCE"
+{
+  rtx mem = gen_rtx_MEM (word_mode, stack_pointer_rtx);
+
+  output_asm_insn ("lock{%;} or%z0\t{$0, %0|%0, 0}", &mem);
+  return "";
+}
   [(set_attr "memory" "unknown")])
 
 (define_expand "mem_thread_fence"
@@ -117,7 +124,8 @@
       rtx (*mfence_insn)(rtx);
       rtx mem;
 
-      if (TARGET_64BIT || TARGET_SSE2)
+      if ((TARGET_64BIT || TARGET_SSE2)
+	  && !TARGET_AVOID_MFENCE)
 	mfence_insn = gen_mfence_sse2;
       else
 	mfence_insn = gen_mfence_nosse;
@@ -306,11 +314,10 @@
     {
       operands[1] = force_reg (<MODE>mode, operands[1]);
 
-      /* For seq-cst stores, use XCHG when we lack MFENCE
-      	 or when target prefers XCHG.  */
+      /* For seq-cst stores, use XCHG when we lack MFENCE.  */
       if (is_mm_seq_cst (model)
 	  && (!(TARGET_64BIT || TARGET_SSE2)
-	      || TARGET_USE_XCHG_FOR_ATOMIC_STORE))
+	      || TARGET_AVOID_MFENCE))
 	{
 	  emit_insn (gen_atomic_exchange<mode> (gen_reg_rtx (<MODE>mode),
 						operands[0], operands[1],
diff --git a/gcc/config/i386/x86-tune.def b/gcc/config/i386/x86-tune.def
index 1776aba..6eff825 100644
--- a/gcc/config/i386/x86-tune.def
+++ b/gcc/config/i386/x86-tune.def
@@ -313,8 +313,8 @@ DEF_TUNE (X86_TUNE_ONE_IF_CONV_INSN, "one_if_conv_insn",
 	  m_SILVERMONT | m_KNL | m_KNM | m_INTEL | m_CORE_ALL | m_GOLDMONT
 	  | m_GOLDMONT_PLUS | m_TREMONT | m_GENERIC)
 
-/* X86_TUNE_USE_XCHG_FOR_ATOMIC_STORE: Use xchg instead of mov+mfence.  */
-DEF_TUNE (X86_TUNE_USE_XCHG_FOR_ATOMIC_STORE, "use_xchg_for_atomic_store",
+/* X86_TUNE_AVOID_MFENCE: Use lock prefixed instructions instead of mfence.  */
+DEF_TUNE (X86_TUNE_AVOID_MFENCE, "avoid_mfence",
 	 m_CORE_ALL | m_BDVER | m_ZNVER | m_GENERIC)
 
 /* X86_TUNE_EXPAND_ABS: This enables a new abs pattern by
diff --git a/gcc/testsuite/gcc.target/i386/pr95750.c b/gcc/testsuite/gcc.target/i386/pr95750.c
new file mode 100644
index 0000000..c47108f
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95750.c
@@ -0,0 +1,19 @@
+/* PR target/95750 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -march=core2" } */
+
+void
+foo (void)
+{
+  __atomic_thread_fence (__ATOMIC_SEQ_CST);
+}
+
+int x;
+
+void
+bar (void)
+{
+  __atomic_store_n (&x, -1, __ATOMIC_SEQ_CST);
+}
+
+/* { dg-final { scan-assembler-not "mfence" } } */