From: Uros Bizjak <uros@gcc.gnu.org>
Date: Thu, 4 Jun 2015 10:06:11 +0000 (+0200)
Subject: re PR target/66369 (gcc 4.8.3/5.1.0 miss optimisation with vpmovmskb)
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=e7f40208e1a14281ef5b2ec50f947335e761648f;p=gcc.git

re PR target/66369 (gcc 4.8.3/5.1.0 miss optimisation with vpmovmskb)

	PR target/66369
	* config/i386/sse.md (<sse2_avx2>_pmovmsk): Merge from avx2_pmovmskb
	and sse2_pmovmskb using VI1_AVX2 mode iterator.
	(*<sse>_movmsk<ssemodesuffix><avxsizesuffix>_zext): New insn pattern.
	(*<sse2_avx2>_pmovmskb_zext): Ditto.

From-SVN: r224120
---

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 53d75b1c69f..78d46132672 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1125,7 +1125,7 @@
 	ACX_PROG_CXX_WARNING_ALMOST_PEDANTIC, and
 	ACX_PROG_CXX_WARNINGS_ARE_ERRORS.
 
-2015-05-22 Aditya Kumar <hiraditya@msn.com>
+2015-05-22  Aditya Kumar  <hiraditya@msn.com>
 
 	* auto-profile.c (afdo_calculate_branch_prob): Break once has_sample
 	is true.
@@ -3398,7 +3398,7 @@
 	(aarch_macro_fusion_pair_p): Update uses of current_tune.
 	* arm.h (LOGICAL_OP_NON_SHORT_CIRCUIT): Likewise.
 
-2015-05-12 Sandra Loosemore <sandra@codesourcery.com>
+2015-05-12  Sandra Loosemore  <sandra@codesourcery.com>
 
 	* config/nios2/nios2.md (trap, ctrapsi4): Use "trap" instead of
 	"break".
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 21c6c6cd5be..e44ba9a6d36 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -13112,24 +13112,48 @@
    (set_attr "prefix" "maybe_vex")
    (set_attr "mode" "<MODE>")])
 
-(define_insn "avx2_pmovmskb"
-  [(set (match_operand:SI 0 "register_operand" "=r")
-	(unspec:SI [(match_operand:V32QI 1 "register_operand" "x")]
-		   UNSPEC_MOVMSK))]
-  "TARGET_AVX2"
-  "vpmovmskb\t{%1, %0|%0, %1}"
+(define_insn "*<sse>_movmsk<ssemodesuffix><avxsizesuffix>_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+	(zero_extend:DI
+	  (unspec:SI
+	    [(match_operand:VF_128_256 1 "register_operand" "x")]
+	    UNSPEC_MOVMSK)))]
+  "TARGET_64BIT && TARGET_SSE"
+  "%vmovmsk<ssemodesuffix>\t{%1, %k0|%k0, %1}"
   [(set_attr "type" "ssemov")
-   (set_attr "prefix" "vex")
-   (set_attr "mode" "DI")])
+   (set_attr "prefix" "maybe_vex")
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "sse2_pmovmskb"
+(define_insn "<sse2_avx2>_pmovmskb"
   [(set (match_operand:SI 0 "register_operand" "=r")
-	(unspec:SI [(match_operand:V16QI 1 "register_operand" "x")]
-		   UNSPEC_MOVMSK))]
+	(unspec:SI
+	  [(match_operand:VI1_AVX2 1 "register_operand" "x")]
+	  UNSPEC_MOVMSK))]
   "TARGET_SSE2"
   "%vpmovmskb\t{%1, %0|%0, %1}"
   [(set_attr "type" "ssemov")
-   (set_attr "prefix_data16" "1")
+   (set (attr "prefix_data16")
+     (if_then_else
+       (match_test "TARGET_AVX")
+     (const_string "*")
+     (const_string "1")))
+   (set_attr "prefix" "maybe_vex")
+   (set_attr "mode" "SI")])
+
+(define_insn "*<sse2_avx2>_pmovmskb_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+	(zero_extend:DI
+	  (unspec:SI
+	    [(match_operand:VI1_AVX2 1 "register_operand" "x")]
+	    UNSPEC_MOVMSK)))]
+  "TARGET_64BIT && TARGET_SSE2"
+  "%vpmovmskb\t{%1, %k0|%k0, %1}"
+  [(set_attr "type" "ssemov")
+   (set (attr "prefix_data16")
+     (if_then_else
+       (match_test "TARGET_AVX")
+     (const_string "*")
+     (const_string "1")))
    (set_attr "prefix" "maybe_vex")
    (set_attr "mode" "SI")])