[PATCH] Optimize 128-bit vector insertion into zero 512-bit vector (PR target/85480)

[Jakub Jelinek]

Hi!

As mentioned in the PR, vmov{aps,apd,dqa{,32,64}} 128-bit instructions
zero the rest of 512-bit register, so we can optimize insertion into zero
vectors using those instructions.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for stage1?

2018-04-23  Jakub Jelinek  <jakub@redhat.com>

	PR target/85480
	* config/i386/sse.md (ssequaterinsnmode): New mode attribute.
	(*<extract_type>_vinsert<shuffletype><extract_suf>_0): New pattern.

	* gcc.target/i386/avx512dq-pr85480-1.c: New test.
	* gcc.target/i386/avx512dq-pr85480-2.c: New test.

--- gcc/config/i386/sse.md.jj	2018-04-17 19:17:54.659879717 +0200
+++ gcc/config/i386/sse.md	2018-04-23 16:56:31.129076790 +0200
@@ -581,6 +581,9 @@ (define_mode_attr shuffletype
 (define_mode_attr ssequartermode
   [(V16SF "V4SF") (V8DF "V2DF") (V16SI "V4SI") (V8DI "V2DI")])
 
+(define_mode_attr ssequarterinsnmode
+  [(V16SF "V4SF") (V8DF "V2DF") (V16SI "TI") (V8DI "TI")])
+
 (define_mode_attr ssedoublemodelower
   [(V16QI "v16hi") (V32QI "v32hi") (V64QI "v64hi")
    (V8HI "v8si")   (V16HI "v16si") (V32HI "v32si")
@@ -12711,6 +12714,40 @@ (define_expand "<extract_type>_vinsert<s
   DONE;
 })
 
+(define_insn "*<extract_type>_vinsert<shuffletype><extract_suf>_0"
+  [(set (match_operand:AVX512_VEC 0 "register_operand" "=v,x,Yv")
+	(vec_merge:AVX512_VEC
+	  (match_operand:AVX512_VEC 1 "reg_or_0_operand" "v,C,C")
+	  (vec_duplicate:AVX512_VEC
+		(match_operand:<ssequartermode> 2 "nonimmediate_operand" "vm,xm,vm"))
+	  (match_operand:SI 3 "const_int_operand" "n,n,n")))]
+  "TARGET_AVX512F
+   && (INTVAL (operands[3])
+       == (GET_MODE_UNIT_SIZE (<MODE>mode) == 4 ? 0xFFF0 : 0xFC))"
+{
+  if (which_alternative == 0)
+    return "vinsert<shuffletype><extract_suf>\t{$0, %2, %1, %0|%0, %1, %2, 0}";
+  switch (<MODE>mode)
+    {
+    case E_V8DFmode:
+      return "vmovapd\t{%2, %x0|%x0, %2}";
+    case E_V16SFmode:
+      return "vmovaps\t{%2, %x0|%x0, %2}";
+    case E_V8DImode:
+      return which_alternative == 2 ? "vmovdqa64\t{%2, %x0|%x0, %2}"
+				    : "vmovdqa\t{%2, %x0|%x0, %2}";
+    case E_V16SImode:
+      return which_alternative == 2 ? "vmovdqa32\t{%2, %x0|%x0, %2}"
+				    : "vmovdqa\t{%2, %x0|%x0, %2}";
+    default:
+      gcc_unreachable ();
+    }
+}
+  [(set_attr "type" "sselog,ssemov,ssemov")
+   (set_attr "length_immediate" "1,0,0")
+   (set_attr "prefix" "evex,vex,evex")
+   (set_attr "mode" "<sseinsnmode>,<ssequarterinsnmode>,<ssequarterinsnmode>")])
+
 (define_insn "<mask_codefor><extract_type>_vinsert<shuffletype><extract_suf>_1<mask_name>"
   [(set (match_operand:AVX512_VEC 0 "register_operand" "=v")
 	(vec_merge:AVX512_VEC
--- gcc/testsuite/gcc.target/i386/avx512dq-pr85480-1.c.jj	2018-04-23 16:53:26.684907642 +0200
+++ gcc/testsuite/gcc.target/i386/avx512dq-pr85480-1.c	2018-04-23 16:55:05.649998402 +0200
@@ -0,0 +1,26 @@
+/* PR target/85480 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx512dq" } */
+/* { dg-final { scan-assembler-times "vmovaps\[^\n\r]*xmm0\[^\n\r]*xmm0" 1 } } */
+/* { dg-final { scan-assembler-times "vmovapd\[^\n\r]*xmm0\[^\n\r]*xmm0" 1 } } */
+/* { dg-final { scan-assembler-times "vmovdqa\[^\n\r]*xmm0\[^\n\r]*xmm0" 1 } } */
+
+#include <x86intrin.h>
+
+__m512
+f1 (__m128 a)
+{
+  return _mm512_insertf32x4 (_mm512_set1_ps (0.0f), a, 0);
+}
+
+__m512d
+f2 (__m128d a)
+{
+  return _mm512_insertf64x2 (_mm512_set1_pd (0.0), a, 0);
+}
+
+__m512i
+f3 (__m128i a)
+{
+  return _mm512_inserti32x4 (_mm512_set1_epi32 (0), a, 0);
+}
--- gcc/testsuite/gcc.target/i386/avx512dq-pr85480-2.c.jj	2018-04-23 16:55:22.151013534 +0200
+++ gcc/testsuite/gcc.target/i386/avx512dq-pr85480-2.c	2018-04-23 17:07:40.655693375 +0200
@@ -0,0 +1,38 @@
+/* PR target/85480 */
+/* { dg-do compile { target lp64 } } */
+/* { dg-options "-O2 -mavx512dq -mavx512vl" } */
+/* { dg-final { scan-assembler-times "vmovaps\[^\n\r]*xmm1\[67]\[^\n\r]*xmm1\[67]" 1 } } */
+/* { dg-final { scan-assembler-times "vmovapd\[^\n\r]*xmm1\[67]\[^\n\r]*xmm1\[67]" 1 } } */
+/* { dg-final { scan-assembler-times "vmovdqa32\[^\n\r]*xmm1\[67]\[^\n\r]*xmm1\[67]" 1 } } */
+
+#include <x86intrin.h>
+
+__m512
+f1 (__m128 a)
+{
+  register __m128 b __asm ("xmm16");
+  asm ("" : "=v" (b) : "0" (a));
+  register __m512 c __asm ("xmm17") = _mm512_insertf32x4 (_mm512_set1_ps (0.0f), b, 0);
+  asm ("" : "+v" (c));
+  return c;
+}
+
+__m512d
+f2 (__m128d a)
+{
+  register __m128d b __asm ("xmm16");
+  asm ("" : "=v" (b) : "0" (a));
+  register __m512d c __asm ("xmm17") = _mm512_insertf64x2 (_mm512_set1_pd (0.0), b, 0);
+  asm ("" : "+v" (c));
+  return c;
+}
+
+__m512i
+f3 (__m128i a)
+{
+  register __m128i b __asm ("xmm16");
+  asm ("" : "=v" (b) : "0" (a));
+  register __m512i c __asm ("xmm17") = _mm512_inserti32x4 (_mm512_set1_epi32 (0), b, 0);
+  asm ("" : "+v" (c));
+  return c;
+}

	Jakub