public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH, i386]: Fix PR91719, emit XCHG for seq_cst store on big cores
@ 2019-09-16 18:43 Uros Bizjak
  0 siblings, 0 replies; only message in thread
From: Uros Bizjak @ 2019-09-16 18:43 UTC (permalink / raw)
  To: gcc-patches

[-- Attachment #1: Type: text/plain, Size: 609 bytes --]

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  <ubizjak@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.

[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 2106 bytes --]

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                                          */
 /*****************************************************************************/

^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2019-09-16 18:43 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-09-16 18:43 [PATCH, i386]: Fix PR91719, emit XCHG for seq_cst store on big cores Uros Bizjak

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).