[PATCH] [i386][avx512]Add combine splitter to transform vpternlogd/vpcmpeqd/vpxor/vblendvps to vblendvps for ~op0

[Haochen Jiang]

Hi all,

This patch fix the regression previously reported on the combine splitter under '-m32 -march=cascadelake' options.

Regtested on x86_64-pc-linux-gnu.

BRs,
Haochen

gcc/ChangeLog:

	PR target/100738
	* config/i386/sse.md (*avx_cmp<mode>3_lt, *avx_cmp<mode>3_ltint):
	Remove MEM_P restriction and add force_reg for operands[2].
	(*avx_cmp<mode>3_ltint_not): Add new define_insn_and_split.

gcc/testsuite/ChangeLog:

	PR target/100738
	* g++.target/i386/avx512vl-pr100738-1.C: New test.
---
 gcc/config/i386/sse.md                        | 44 +++++++++++++++++--
 .../g++.target/i386/avx512vl-pr100738-1.C     |  8 ++++
 2 files changed, 48 insertions(+), 4 deletions(-)
 create mode 100755 gcc/testsuite/g++.target/i386/avx512vl-pr100738-1.C

diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 5421fb51684..8ec9fb075d0 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -3528,8 +3528,7 @@
 	     UNSPEC_PCMP)))]
   "TARGET_AVX512VL && ix86_pre_reload_split ()
   /* LT or GE 0 */
-  && ((INTVAL (operands[5]) == 1 && !MEM_P (operands[2]))
-      || (INTVAL (operands[5]) == 5 && !MEM_P (operands[1])))"
+  && ((INTVAL (operands[5]) == 1) || (INTVAL (operands[5]) == 5))"
   "#"
   "&& 1"
   [(set (match_dup 0)
@@ -3543,6 +3542,7 @@
 {
   if (INTVAL (operands[5]) == 5)
     std::swap (operands[1], operands[2]);
+  operands[2] = force_reg (<MODE>mode, operands[2]);
 })
 
 (define_insn_and_split "*avx_cmp<mode>3_ltint"
@@ -3557,8 +3557,7 @@
 	     UNSPEC_PCMP)))]
   "TARGET_AVX512VL && ix86_pre_reload_split ()
   /* LT or GE 0 */
-  && ((INTVAL (operands[5]) == 1 && !MEM_P (operands[2]))
-      || (INTVAL (operands[5]) == 5 && !MEM_P (operands[1])))"
+  && ((INTVAL (operands[5]) == 1) || (INTVAL (operands[5]) == 5))"
   "#"
   "&& 1"
   [(set (match_dup 0)
@@ -3575,7 +3574,44 @@
     std::swap (operands[1], operands[2]);
   operands[0] = gen_lowpart (<ssebytemode>mode, operands[0]);
   operands[1] = gen_lowpart (<ssebytemode>mode, operands[1]);
+  operands[2] = force_reg (<ssebytemode>mode,
+			  gen_lowpart (<ssebytemode>mode, operands[2]));
+})
+
+(define_insn_and_split "*avx_cmp<mode>3_ltint_not"
+ [(set (match_operand:VI48_AVX  0 "register_operand")
+       (vec_merge:VI48_AVX
+	 (match_operand:VI48_AVX 1 "vector_operand")
+	 (match_operand:VI48_AVX 2 "vector_operand")
+	 (unspec:<avx512fmaskmode>
+	   [(subreg:VI48_AVX
+	    (not:<ssebytemode>
+	      (match_operand:<ssebytemode> 3 "vector_operand")) 0)
+	    (match_operand:VI48_AVX 4 "const0_operand")
+	    (match_operand:SI 5 "const_0_to_7_operand")]
+	    UNSPEC_PCMP)))]
+  "TARGET_AVX512VL && ix86_pre_reload_split ()
+  /* not LT or GE 0 */
+  && ((INTVAL (operands[5]) == 1) || (INTVAL (operands[5]) == 5))"
+  "#"
+  "&& 1"
+  [(set (match_dup 0)
+	(unspec:<ssebytemode>
+	  [(match_dup 1)
+	   (match_dup 2)
+	   (subreg:<ssebytemode>
+	     (lt:VI48_AVX
+	      (match_dup 3)
+	      (match_dup 4)) 0)]
+	    UNSPEC_BLENDV))]
+{
+  if (INTVAL (operands[5]) == 5)
+    std::swap (operands[1], operands[2]);
+  operands[0] = gen_lowpart (<ssebytemode>mode, operands[0]);
+  operands[1] = force_reg (<ssebytemode>mode,
+			  gen_lowpart (<ssebytemode>mode, operands[1]));
   operands[2] = gen_lowpart (<ssebytemode>mode, operands[2]);
+  operands[3] = lowpart_subreg (<MODE>mode, operands[3], <ssebytemode>mode);
 })
 
 (define_insn "avx_vmcmp<mode>3"
diff --git a/gcc/testsuite/g++.target/i386/avx512vl-pr100738-1.C b/gcc/testsuite/g++.target/i386/avx512vl-pr100738-1.C
new file mode 100755
index 00000000000..ac4d62b94d1
--- /dev/null
+++ b/gcc/testsuite/g++.target/i386/avx512vl-pr100738-1.C
@@ -0,0 +1,8 @@
+/* { dg-do compile } */
+/* { dg-options "-Ofast -march=cascadelake" } */
+/* { dg-final {scan-assembler-times "vblendvps\[ \\t\]" 2 } } */
+/* { dg-final {scan-assembler-not "vpcmpeqd\[ \\t\]" } } */
+/* { dg-final {scan-assembler-not "vpxor\[ \\t\]" } } */
+/* { dg-final {scan-assembler-not "vpternlogd\[ \\t\]" } } */
+
+#include "pr100738-1.C"
-- 
2.18.1