Attached patch emits XCHG instead of store+MFENCE on big cores and
generic tuning
m_CORE_ALL | m_BDVER | m_ZNVER | m_GENERIC
(The tune can be added for other targets, too.)
2019-09-16 Uroš Bizjak <ubiz...@gmail.com>
PR target/91719
* config/i386/i386.h (TARGET_USE_XCHG_FOR_ATOMIC_STORE): New macro.
* config/i386/x86-tune.def (X86_TUNE_USE_XCHG_FOR_ATOMIC_STORE): New.
* config/i386/sync.md (atomic_store<mode>): emit XCHG for
TARGET_USE_XCHG_FOR_ATOMIC_STORE.
Bootstrapped and regression tested on x86_64-linux-gnu {,-m32}.
Committed to mainline SVN.
Uros.
Index: config/i386/i386.h
===================================================================
--- config/i386/i386.h (revision 275752)
+++ config/i386/i386.h (working copy)
@@ -590,6 +590,8 @@ extern unsigned char ix86_tune_features[X86_TUNE_L
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_EMIT_VZEROUPPER \
ix86_tune_features[X86_TUNE_EMIT_VZEROUPPER]
Index: config/i386/sync.md
===================================================================
--- config/i386/sync.md (revision 275752)
+++ config/i386/sync.md (working copy)
@@ -306,8 +306,11 @@
{
operands[1] = force_reg (<MODE>mode, operands[1]);
- /* For seq-cst stores, when we lack MFENCE, use XCHG. */
- if (is_mm_seq_cst (model) && !(TARGET_64BIT || TARGET_SSE2))
+ /* For seq-cst stores, use XCHG
+ when we lack MFENCE or when target prefers XCHG. */
+ if (is_mm_seq_cst (model)
+ && (!(TARGET_64BIT || TARGET_SSE2)
+ || TARGET_USE_XCHG_FOR_ATOMIC_STORE))
{
emit_insn (gen_atomic_exchange<mode> (gen_reg_rtx (<MODE>mode),
operands[0], operands[1],
Index: config/i386/x86-tune.def
===================================================================
--- config/i386/x86-tune.def (revision 275752)
+++ config/i386/x86-tune.def (working copy)
@@ -313,6 +313,10 @@ DEF_TUNE (X86_TUNE_ONE_IF_CONV_INSN, "one_if_conv_
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",
+ m_CORE_ALL | m_BDVER | m_ZNVER | m_GENERIC)
+
/*****************************************************************************/
/* 387 instruction selection tuning */
/*****************************************************************************/
