; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE2
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=KNL
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefix=SKX
define i1 @allones_v16i8_sign(<16 x i8> %arg) {
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v16i8_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i8_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i8_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i8_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v32i8_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: cmpl $-1, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v32i8_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v32i8_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v32i8_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v64i8_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; KNL-NEXT: vpcmpgtb %ymm0, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm3
+; KNL-NEXT: vptestmd %zmm3, %zmm3, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: vpcmpgtb %ymm1, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %edx
+; KNL-NEXT: shll $16, %edx
+; KNL-NEXT: orl %eax, %edx
+; KNL-NEXT: shlq $32, %rdx
+; KNL-NEXT: orq %rcx, %rdx
+; KNL-NEXT: cmpq $-1, %rdx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v64i8_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v64i8_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; KNL-NEXT: vpcmpgtb %ymm0, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm3
+; KNL-NEXT: vptestmd %zmm3, %zmm3, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: vpcmpgtb %ymm1, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %edx
+; KNL-NEXT: shll $16, %edx
+; KNL-NEXT: orl %eax, %edx
+; KNL-NEXT: shlq $32, %rdx
+; KNL-NEXT: orq %rcx, %rdx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v64i8_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %zmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v8i16_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
+; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i16_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovw2m %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i16_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
+; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i16_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovw2m %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v16i16_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i16_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovw2m %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i16_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i16_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovw2m %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v32i16_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; KNL-NEXT: vpcmpgtw %ymm0, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vpcmpgtw %ymm1, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: cmpl $-1, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v32i16_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovw2m %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v32i16_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; KNL-NEXT: vpcmpgtw %ymm0, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vpcmpgtw %ymm1, %ymm2, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v32i16_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovw2m %zmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v4i32_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $15, %al
+; KNL-NEXT: cmpb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v4i32_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v4i32_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v4i32_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %xmm0, %k0
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v8i32_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i32_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %ymm0, %k0
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i32_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i32_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v16i32_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i32_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i32_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i32_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %zmm0, %k0
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v4i64_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $15, %al
+; KNL-NEXT: cmpb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v4i64_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovq2m %ymm0, %k0
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v4i64_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v4i64_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovq2m %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v8i64_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i64_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovq2m %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i64_sign:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i64_sign:
; SKX: # %bb.0:
; SKX-NEXT: vpmovq2m %zmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v16i8_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i8_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i8_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i8_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v32i8_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: cmpl $-1, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v32i8_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v32i8_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v32i8_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v64i8_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm3
+; KNL-NEXT: vptestmd %zmm3, %zmm3, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %edx
+; KNL-NEXT: shll $16, %edx
+; KNL-NEXT: orl %eax, %edx
+; KNL-NEXT: shlq $32, %rdx
+; KNL-NEXT: orq %rcx, %rdx
+; KNL-NEXT: cmpq $-1, %rdx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v64i8_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v64i8_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm3
+; KNL-NEXT: vptestmd %zmm3, %zmm3, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %edx
+; KNL-NEXT: shll $16, %edx
+; KNL-NEXT: orl %eax, %edx
+; KNL-NEXT: shlq $32, %rdx
+; KNL-NEXT: orq %rcx, %rdx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v64i8_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %zmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v8i16_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i16_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i16_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i16_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v16i16_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i16_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v32i16_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: cmpl $-1, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v32i16_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v32i16_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v32i16_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i16_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i16_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %ymm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v4i32_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [1,1,1,1]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $15, %al
+; KNL-NEXT: cmpb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v4i32_and1:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} xmm1 = [1,1,1,1]
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v4i32_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [1,1,1,1]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v4i32_and1:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} xmm1 = [1,1,1,1]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v8i32_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i32_and1:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i32_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i32_and1:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v16i32_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vpbroadcastd {{.*#+}} zmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i32_and1:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} zmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i32_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vpbroadcastd {{.*#+}} zmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i32_and1:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} zmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v2i64_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,1]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $3, %al
+; KNL-NEXT: cmpb $3, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v2i64_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip), %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v2i64_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,1]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $3, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v2i64_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip), %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v4i64_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastq {{.*#+}} ymm1 = [1,1,1,1]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $15, %al
+; KNL-NEXT: cmpb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v4i64_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to4}, %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v4i64_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastq {{.*#+}} ymm1 = [1,1,1,1]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v4i64_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to4}, %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v8i64_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i64_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i64_and1:
+; KNL: # %bb.0:
+; KNL-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i64_and1:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v16i8_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i8_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i8_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i8_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v32i8_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: cmpl $-1, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v32i8_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v32i8_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v32i8_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v64i8_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm3
+; KNL-NEXT: vptestmd %zmm3, %zmm3, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %edx
+; KNL-NEXT: shll $16, %edx
+; KNL-NEXT: orl %eax, %edx
+; KNL-NEXT: shlq $32, %rdx
+; KNL-NEXT: orq %rcx, %rdx
+; KNL-NEXT: cmpq $-1, %rdx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v64i8_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v64i8_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm3
+; KNL-NEXT: vptestmd %zmm3, %zmm3, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: vpcmpeqb %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %edx
+; KNL-NEXT: shll $16, %edx
+; KNL-NEXT: orl %eax, %edx
+; KNL-NEXT: shlq $32, %rdx
+; KNL-NEXT: orq %rcx, %rdx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v64i8_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmb {{.*}}(%rip), %zmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v8i16_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i16_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i16_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
+; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
+; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i16_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v16i16_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i16_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v32i16_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: cmpl $-1, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v32i16_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v32i16_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm2 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
+; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vpcmpeqw %ymm2, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: shll $16, %ecx
+; KNL-NEXT: orl %eax, %ecx
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v32i16_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i16_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vpand %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
+; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i16_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmw {{.*}}(%rip), %ymm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v4i32_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $15, %al
+; KNL-NEXT: cmpb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v4i32_and4:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v4i32_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v4i32_and4:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v8i32_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i32_and4:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i32_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastd {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i32_and4:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v16i32_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vpbroadcastd {{.*#+}} zmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: setb %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v16i32_and4:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} zmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v16i32_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vpbroadcastd {{.*#+}} zmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
+; KNL-NEXT: vptestmd %zmm1, %zmm0, %k0
+; KNL-NEXT: kortestw %k0, %k0
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v16i32_and4:
; SKX: # %bb.0:
; SKX-NEXT: vpbroadcastd {{.*#+}} zmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allones_v2i64_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [4,4]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $3, %al
+; KNL-NEXT: cmpb $3, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v2i64_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip), %xmm0, %k0
; AVX-NEXT: sete %al
; AVX-NEXT: retq
;
+; KNL-LABEL: allzeros_v2i64_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vmovdqa {{.*#+}} xmm1 = [4,4]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $3, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v2i64_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip), %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v4i64_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4,4,4,4]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andb $15, %al
+; KNL-NEXT: cmpb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v4i64_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to4}, %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v4i64_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4,4,4,4]
+; KNL-NEXT: vptestmq %zmm1, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb $15, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v4i64_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to4}, %ymm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allones_v8i64_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: cmpb $-1, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allones_v8i64_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: allzeros_v8i64_and4:
+; KNL: # %bb.0:
+; KNL-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: testb %al, %al
+; KNL-NEXT: sete %al
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: allzeros_v8i64_and4:
; SKX: # %bb.0:
; SKX-NEXT: vptestmq {{.*}}(%rip){1to8}, %zmm0, %k0
; AVX-NEXT: vmovmskpd %xmm0, %eax
; AVX-NEXT: retq
;
+; KNL-LABEL: movmskpd:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andl $3, %eax
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: movmskpd:
; SKX: # %bb.0:
; SKX-NEXT: vpmovq2m %xmm0, %k0
; AVX-NEXT: vmovmskps %xmm0, %eax
; AVX-NEXT: retq
;
+; KNL-LABEL: movmskps:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andl $15, %eax
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: movmskps:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %xmm0, %k0
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
+; KNL-LABEL: movmskpd256:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: andl $15, %eax
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: movmskpd256:
; SKX: # %bb.0:
; SKX-NEXT: vpmovq2m %ymm0, %k0
; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
+; KNL-LABEL: movmskps256:
+; KNL: # %bb.0:
+; KNL-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtd %zmm0, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: movzbl %al, %eax
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: movmskps256:
; SKX: # %bb.0:
; SKX-NEXT: vpmovd2m %ymm0, %k0
; AVX-NEXT: vpmovmskb %xmm0, %eax
; AVX-NEXT: retq
;
+; KNL-LABEL: movmskb:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: movmskb:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %xmm0, %k0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
+; KNL-LABEL: movmskb256:
+; KNL: # %bb.0:
+; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; KNL-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
+; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
+; KNL-NEXT: kmovw %k0, %ecx
+; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
+; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
+; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
+; KNL-NEXT: kmovw %k0, %eax
+; KNL-NEXT: shll $16, %eax
+; KNL-NEXT: orl %ecx, %eax
+; KNL-NEXT: vzeroupper
+; KNL-NEXT: retq
+;
; SKX-LABEL: movmskb256:
; SKX: # %bb.0:
; SKX-NEXT: vpmovb2m %ymm0, %k0