AVX-512: Truncating store for i1 vectors

Implemented truncstore for KNL and skylake-avx512.
Covered vectors from v2i1 to v64i1. We save the value in bits (not in bytes) - v32i1 is saved in 4 bytes.

Differential Revision: http://reviews.llvm.org/D18740

llvm-svn: 265283

2 files changed, 170 insertions, 416 deletions

diff --git a/llvm/test/CodeGen/X86/avx512-mask-op.ll b/llvm/test/CodeGen/X86/avx512-mask-op.ll
index f6d5ceb5f2f..3b7263ba39a 100644
--- a/llvm/test/CodeGen/X86/avx512-mask-op.ll
+++ b/llvm/test/CodeGen/X86/avx512-mask-op.ll
@@ -521,17 +521,9 @@ define <32 x i16> @test21(<32 x i16> %x , <32 x i1> %mask) nounwind readnone {
 define void @test22(<4 x i1> %a, <4 x i1>* %addr) {
 ; KNL-LABEL: test22:
 ; KNL:       ## BB#0:
-; KNL-NEXT:    vpextrd $3, %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vpextrd $2, %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vpextrd $1, %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vmovd %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
+; KNL-NEXT:    vpslld $31, %ymm0, %ymm0
+; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, %eax
 ; KNL-NEXT:    movb %al, (%rdi)
 ; KNL-NEXT:    retq
 ;
@@ -548,11 +540,9 @@ define void @test22(<4 x i1> %a, <4 x i1>* %addr) {
 define void @test23(<2 x i1> %a, <2 x i1>* %addr) {
 ; KNL-LABEL: test23:
 ; KNL:       ## BB#0:
-; KNL-NEXT:    vpextrq $1, %xmm0, %rax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vmovq %xmm0, %rax
-; KNL-NEXT:    andl $1, %eax
+; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
+; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, %eax
 ; KNL-NEXT:    movb %al, (%rdi)
 ; KNL-NEXT:    retq
 ;
@@ -596,11 +586,9 @@ define void @store_v2i1(<2 x i1> %c , <2 x i1>* %ptr) {
 ; KNL-LABEL: store_v2i1:
 ; KNL:       ## BB#0:
 ; KNL-NEXT:    vpxor {{.*}}(%rip), %xmm0, %xmm0
-; KNL-NEXT:    vpextrq $1, %xmm0, %rax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vmovq %xmm0, %rax
-; KNL-NEXT:    andl $1, %eax
+; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
+; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, %eax
 ; KNL-NEXT:    movb %al, (%rdi)
 ; KNL-NEXT:    retq
 ;
@@ -621,17 +609,9 @@ define void @store_v4i1(<4 x i1> %c , <4 x i1>* %ptr) {
 ; KNL:       ## BB#0:
 ; KNL-NEXT:    vpbroadcastd {{.*}}(%rip), %xmm1
 ; KNL-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; KNL-NEXT:    vpextrd $3, %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vpextrd $2, %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vpextrd $1, %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
-; KNL-NEXT:    movb %al, (%rdi)
-; KNL-NEXT:    vmovd %xmm0, %eax
-; KNL-NEXT:    andl $1, %eax
+; KNL-NEXT:    vpslld $31, %ymm0, %ymm0
+; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, %eax
 ; KNL-NEXT:    movb %al, (%rdi)
 ; KNL-NEXT:    retq
 ;
@@ -1038,3 +1018,128 @@ define <64 x i8> @load_64i1(<64 x i1>* %a) {
   %c = sext <64 x i1> %b to <64 x i8>
   ret <64 x i8> %c
 }
+
+define void @store_8i1(<8 x i1>* %a, <8 x i1> %v) {
+; KNL-LABEL: store_8i1:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
+; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, %eax
+; KNL-NEXT:    movb %al, (%rdi)
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: store_8i1:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpsllw $15, %xmm0, %xmm0
+; SKX-NEXT:    vpmovw2m %xmm0, %k0
+; SKX-NEXT:    kmovb %k0, (%rdi)
+; SKX-NEXT:    retq
+  store <8 x i1> %v, <8 x i1>* %a
+  ret void
+}
+
+define void @store_8i1_1(<8 x i1>* %a, <8 x i16> %v) {
+; KNL-LABEL: store_8i1_1:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
+; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, %eax
+; KNL-NEXT:    movb %al, (%rdi)
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: store_8i1_1:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpsllw $15, %xmm0, %xmm0
+; SKX-NEXT:    vpmovw2m %xmm0, %k0
+; SKX-NEXT:    kmovb %k0, (%rdi)
+; SKX-NEXT:    retq
+  %v1 = trunc <8 x i16> %v to <8 x i1>
+  store <8 x i1> %v1, <8 x i1>* %a
+  ret void
+}
+
+define void @store_16i1(<16 x i1>* %a, <16 x i1> %v) {
+; KNL-LABEL: store_16i1:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
+; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, (%rdi)
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: store_16i1:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpsllw $7, %xmm0, %xmm0
+; SKX-NEXT:    vpmovb2m %xmm0, %k0
+; SKX-NEXT:    kmovw %k0, (%rdi)
+; SKX-NEXT:    retq
+  store <16 x i1> %v, <16 x i1>* %a
+  ret void
+}
+
+define void @store_32i1(<32 x i1>* %a, <32 x i1> %v) {
+; KNL-LABEL: store_32i1:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; KNL-NEXT:    vpmovsxbd %xmm1, %zmm1
+; KNL-NEXT:    vpslld $31, %zmm1, %zmm1
+; KNL-NEXT:    vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT:    kmovw %k0, 2(%rdi)
+; KNL-NEXT:    vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
+; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, (%rdi)
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: store_32i1:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpsllw $7, %ymm0, %ymm0
+; SKX-NEXT:    vpmovb2m %ymm0, %k0
+; SKX-NEXT:    kmovd %k0, (%rdi)
+; SKX-NEXT:    retq
+  store <32 x i1> %v, <32 x i1>* %a
+  ret void
+}
+
+define void @store_32i1_1(<32 x i1>* %a, <32 x i16> %v) {
+; KNL-LABEL: store_32i1_1:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT:    vpmovdb %zmm0, %xmm0
+; KNL-NEXT:    vpmovsxwd %ymm1, %zmm1
+; KNL-NEXT:    vpmovdb %zmm1, %xmm1
+; KNL-NEXT:    vpmovsxbd %xmm1, %zmm1
+; KNL-NEXT:    vpslld $31, %zmm1, %zmm1
+; KNL-NEXT:    vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT:    kmovw %k0, 2(%rdi)
+; KNL-NEXT:    vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
+; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, (%rdi)
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: store_32i1_1:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpsllw $15, %zmm0, %zmm0
+; SKX-NEXT:    vpmovw2m %zmm0, %k0
+; SKX-NEXT:    kmovd %k0, (%rdi)
+; SKX-NEXT:    retq
+  %v1 = trunc <32 x i16> %v to <32 x i1>
+  store <32 x i1> %v1, <32 x i1>* %a
+  ret void
+}
+
+
+define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
+;
+; SKX-LABEL: store_64i1:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpsllw $7, %zmm0, %zmm0
+; SKX-NEXT:    vpmovb2m %zmm0, %k0
+; SKX-NEXT:    kmovq %k0, (%rdi)
+; SKX-NEXT:    retq
+  store <64 x i1> %v, <64 x i1>* %a
+  ret void
+}
diff --git a/llvm/test/CodeGen/X86/vector-compare-results.ll b/llvm/test/CodeGen/X86/vector-compare-results.ll
index 1ac7f5155ff..bfe1a65eb54 100644
--- a/llvm/test/CodeGen/X86/vector-compare-results.ll
+++ b/llvm/test/CodeGen/X86/vector-compare-results.ll
@@ -1,3 +1,4 @@
+; NOTE: Assertions have been autogenerated by update_llc_test_checks.py
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE42
@@ -6569,393 +6570,41 @@ define <128 x i1> @test_cmp_v128i8(<128 x i8> %a0, <128 x i8> %a1) nounwind {
 ; AVX512-NEXT:    vpcmpgtb %ymm6, %ymm2, %ymm2
 ; AVX512-NEXT:    vpcmpgtb %ymm7, %ymm3, %ymm3
 ; AVX512-NEXT:    vextracti128 $1, %ymm3, %xmm4
-; AVX512-NEXT:    vpextrb $15, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm4, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $15, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 12(%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm4, %zmm4
+; AVX512-NEXT:    vpslld $31, %zmm4, %zmm4
+; AVX512-NEXT:    vptestmd %zmm4, %zmm4, %k0
+; AVX512-NEXT:    kmovw %k0, 14(%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm3, %zmm3
+; AVX512-NEXT:    vpslld $31, %zmm3, %zmm3
+; AVX512-NEXT:    vptestmd %zmm3, %zmm3, %k0
+; AVX512-NEXT:    kmovw %k0, 12(%rdi)
 ; AVX512-NEXT:    vextracti128 $1, %ymm2, %xmm3
-; AVX512-NEXT:    vpextrb $15, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm3, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $15, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 8(%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm3, %zmm3
+; AVX512-NEXT:    vpslld $31, %zmm3, %zmm3
+; AVX512-NEXT:    vptestmd %zmm3, %zmm3, %k0
+; AVX512-NEXT:    kmovw %k0, 10(%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm2, %zmm2
+; AVX512-NEXT:    vpslld $31, %zmm2, %zmm2
+; AVX512-NEXT:    vptestmd %zmm2, %zmm2, %k0
+; AVX512-NEXT:    kmovw %k0, 8(%rdi)
 ; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512-NEXT:    vpextrb $15, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm2, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $15, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, 4(%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm2, %zmm2
+; AVX512-NEXT:    vpslld $31, %zmm2, %zmm2
+; AVX512-NEXT:    vptestmd %zmm2, %zmm2, %k0
+; AVX512-NEXT:    kmovw %k0, 6(%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm1, %zmm1
+; AVX512-NEXT:    vpslld $31, %zmm1, %zmm1
+; AVX512-NEXT:    vptestmd %zmm1, %zmm1, %k0
+; AVX512-NEXT:    kmovw %k0, 4(%rdi)
 ; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX512-NEXT:    vpextrb $15, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm1, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $15, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $14, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $13, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $12, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $11, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $10, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $9, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $8, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $7, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $6, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $5, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $4, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $3, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $2, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $1, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
-; AVX512-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX512-NEXT:    andb $1, %al
-; AVX512-NEXT:    movb %al, (%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm1, %zmm1
+; AVX512-NEXT:    vpslld $31, %zmm1, %zmm1
+; AVX512-NEXT:    vptestmd %zmm1, %zmm1, %k0
+; AVX512-NEXT:    kmovw %k0, 2(%rdi)
+; AVX512-NEXT:    vpmovsxbd %xmm0, %zmm0
+; AVX512-NEXT:    vpslld $31, %zmm0, %zmm0
+; AVX512-NEXT:    vptestmd %zmm0, %zmm0, %k0
+; AVX512-NEXT:    kmovw %k0, (%rdi)
 ; AVX512-NEXT:    movq %rdi, %rax
 ; AVX512-NEXT:    retq
   %1 = icmp sgt <128 x i8> %a0, %a1