[X86] Add a DAG combine to widen (i4 (bitcast (v4i1))) before type legalization sees the i4 and changes to load/store.

Same for v2i1 and i2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@321602 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 134 insertions, 148 deletions

diff --git a/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll b/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
index bda3a848522..c2bd80b053a 100644
--- a/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
+++ b/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
@@ -5333,8 +5333,9 @@ define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b)
 ; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask:
@@ -5361,8 +5362,9 @@ define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %
 ; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem:
@@ -5391,8 +5393,9 @@ define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask:
@@ -5430,8 +5433,9 @@ define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem:
@@ -5470,8 +5474,9 @@ define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b)
 ; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem_b:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpeqq (%rdi){1to2}, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem_b:
@@ -5502,8 +5507,9 @@ define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b:
@@ -14443,8 +14449,9 @@ define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b)
 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask:
@@ -14471,8 +14478,9 @@ define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>*
 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem:
@@ -14501,8 +14509,9 @@ define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i6
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask:
@@ -14540,8 +14549,9 @@ define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem:
@@ -14580,8 +14590,9 @@ define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b
 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpgtq (%rdi){1to2}, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
@@ -14612,8 +14623,9 @@ define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
@@ -23681,8 +23693,9 @@ define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b)
 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask:
@@ -23711,8 +23724,9 @@ define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>*
 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpnltq (%rdi), %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem:
@@ -23744,8 +23758,9 @@ define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i6
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask:
@@ -23783,8 +23798,9 @@ define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem:
@@ -23824,8 +23840,9 @@ define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b
 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpnltq (%rdi){1to2}, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
@@ -23858,8 +23875,9 @@ define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
@@ -33075,8 +33093,9 @@ define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b)
 ; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask:
@@ -33106,8 +33125,9 @@ define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>*
 ; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpltuq (%rdi), %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem:
@@ -33139,8 +33159,9 @@ define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i6
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask:
@@ -33181,8 +33202,9 @@ define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem:
@@ -33224,8 +33246,9 @@ define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b
 ; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem_b:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vpcmpltuq (%rdi){1to2}, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem_b:
@@ -33259,8 +33282,9 @@ define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b:
@@ -39955,8 +39979,9 @@ define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b)
 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask:
@@ -39983,8 +40008,9 @@ define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>*
 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem:
@@ -40012,8 +40038,9 @@ define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, double* %
 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    vcmpeqpd (%rdi){1to2}, %xmm0, %k0
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
@@ -40044,8 +40071,9 @@ define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask(i2 zeroext %__u, <2 x i6
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask:
@@ -40078,8 +40106,9 @@ define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem(i2 zeroext %__u, <2
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem:
@@ -40113,8 +40142,9 @@ define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b(i2 zeroext %__u, <
 ; VLX:       # %bb.0: # %entry
 ; VLX-NEXT:    kmovd %edi, %k1
 ; VLX-NEXT:    vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovb %k0, -{{[0-9]+}}(%rsp)
-; VLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    andb $3, %al
+; VLX-NEXT:    movzbl %al, %eax
 ; VLX-NEXT:    retq
 ;
 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
diff --git a/test/CodeGen/X86/bitcast-and-setcc-128.ll b/test/CodeGen/X86/bitcast-and-setcc-128.ll
index 78c44e4dca3..ef847a8f99d 100644
--- a/test/CodeGen/X86/bitcast-and-setcc-128.ll
+++ b/test/CodeGen/X86/bitcast-and-setcc-128.ll
@@ -80,8 +80,7 @@ define i4 @v4i32(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c, <4 x i32> %d) {
 ; AVX512F-NEXT:    vpcmpgtd %xmm1, %xmm0, %k1
 ; AVX512F-NEXT:    vpcmpgtd %xmm3, %xmm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4i32:
@@ -89,8 +88,7 @@ define i4 @v4i32(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c, <4 x i32> %d) {
 ; AVX512BW-NEXT:    vpcmpgtd %xmm1, %xmm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtd %xmm3, %xmm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <4 x i32> %a, %b
   %x1 = icmp sgt <4 x i32> %c, %d
@@ -123,8 +121,7 @@ define i4 @v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d)
 ; AVX512F-NEXT:    vcmpltps %xmm0, %xmm1, %k1
 ; AVX512F-NEXT:    vcmpltps %xmm2, %xmm3, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4f32:
@@ -132,8 +129,7 @@ define i4 @v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d)
 ; AVX512BW-NEXT:    vcmpltps %xmm0, %xmm1, %k1
 ; AVX512BW-NEXT:    vcmpltps %xmm2, %xmm3, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = fcmp ogt <4 x float> %a, %b
   %x1 = fcmp ogt <4 x float> %c, %d
@@ -318,8 +314,7 @@ define i2 @v2i8(<2 x i8> %a, <2 x i8> %b, <2 x i8> %c, <2 x i8> %d) {
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512F-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i8:
@@ -335,8 +330,7 @@ define i2 @v2i8(<2 x i8> %a, <2 x i8> %b, <2 x i8> %c, <2 x i8> %d) {
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <2 x i8> %a, %b
   %x1 = icmp sgt <2 x i8> %c, %d
@@ -473,8 +467,7 @@ define i2 @v2i16(<2 x i16> %a, <2 x i16> %b, <2 x i16> %c, <2 x i16> %d) {
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512F-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i16:
@@ -490,8 +483,7 @@ define i2 @v2i16(<2 x i16> %a, <2 x i16> %b, <2 x i16> %c, <2 x i16> %d) {
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <2 x i16> %a, %b
   %x1 = icmp sgt <2 x i16> %c, %d
@@ -612,8 +604,7 @@ define i2 @v2i32(<2 x i32> %a, <2 x i32> %b, <2 x i32> %c, <2 x i32> %d) {
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512F-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i32:
@@ -629,8 +620,7 @@ define i2 @v2i32(<2 x i32> %a, <2 x i32> %b, <2 x i32> %c, <2 x i32> %d) {
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <2 x i32> %a, %b
   %x1 = icmp sgt <2 x i32> %c, %d
@@ -682,8 +672,7 @@ define i2 @v2i64(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c, <2 x i64> %d) {
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512F-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i64:
@@ -691,8 +680,7 @@ define i2 @v2i64(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c, <2 x i64> %d) {
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtq %xmm3, %xmm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <2 x i64> %a, %b
   %x1 = icmp sgt <2 x i64> %c, %d
@@ -725,8 +713,7 @@ define i2 @v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double>
 ; AVX512F-NEXT:    vcmpltpd %xmm0, %xmm1, %k1
 ; AVX512F-NEXT:    vcmpltpd %xmm2, %xmm3, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2f64:
@@ -734,8 +721,7 @@ define i2 @v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double>
 ; AVX512BW-NEXT:    vcmpltpd %xmm0, %xmm1, %k1
 ; AVX512BW-NEXT:    vcmpltpd %xmm2, %xmm3, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = fcmp ogt <2 x double> %a, %b
   %x1 = fcmp ogt <2 x double> %c, %d
@@ -792,8 +778,7 @@ define i4 @v4i8(<4 x i8> %a, <4 x i8> %b, <4 x i8> %c, <4 x i8> %d) {
 ; AVX512F-NEXT:    vpcmpgtd %xmm1, %xmm0, %k1
 ; AVX512F-NEXT:    vpcmpgtd %xmm3, %xmm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4i8:
@@ -809,8 +794,7 @@ define i4 @v4i8(<4 x i8> %a, <4 x i8> %b, <4 x i8> %c, <4 x i8> %d) {
 ; AVX512BW-NEXT:    vpcmpgtd %xmm1, %xmm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtd %xmm3, %xmm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <4 x i8> %a, %b
   %x1 = icmp sgt <4 x i8> %c, %d
@@ -867,8 +851,7 @@ define i4 @v4i16(<4 x i16> %a, <4 x i16> %b, <4 x i16> %c, <4 x i16> %d) {
 ; AVX512F-NEXT:    vpcmpgtd %xmm1, %xmm0, %k1
 ; AVX512F-NEXT:    vpcmpgtd %xmm3, %xmm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4i16:
@@ -884,8 +867,7 @@ define i4 @v4i16(<4 x i16> %a, <4 x i16> %b, <4 x i16> %c, <4 x i16> %d) {
 ; AVX512BW-NEXT:    vpcmpgtd %xmm1, %xmm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtd %xmm3, %xmm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <4 x i16> %a, %b
   %x1 = icmp sgt <4 x i16> %c, %d
diff --git a/test/CodeGen/X86/bitcast-and-setcc-256.ll b/test/CodeGen/X86/bitcast-and-setcc-256.ll
index fdce65516e3..455a77aaef1 100644
--- a/test/CodeGen/X86/bitcast-and-setcc-256.ll
+++ b/test/CodeGen/X86/bitcast-and-setcc-256.ll
@@ -94,8 +94,7 @@ define i4 @v4i64(<4 x i64> %a, <4 x i64> %b, <4 x i64> %c, <4 x i64> %d) {
 ; AVX512F-NEXT:    vpcmpgtq %ymm1, %ymm0, %k1
 ; AVX512F-NEXT:    vpcmpgtq %ymm3, %ymm2, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -104,8 +103,7 @@ define i4 @v4i64(<4 x i64> %a, <4 x i64> %b, <4 x i64> %c, <4 x i64> %d) {
 ; AVX512BW-NEXT:    vpcmpgtq %ymm1, %ymm0, %k1
 ; AVX512BW-NEXT:    vpcmpgtq %ymm3, %ymm2, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
   %x0 = icmp sgt <4 x i64> %a, %b
@@ -148,8 +146,7 @@ define i4 @v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c, <4 x double>
 ; AVX512F-NEXT:    vcmpltpd %ymm0, %ymm1, %k1
 ; AVX512F-NEXT:    vcmpltpd %ymm2, %ymm3, %k0 {%k1}
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -158,8 +155,7 @@ define i4 @v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c, <4 x double>
 ; AVX512BW-NEXT:    vcmpltpd %ymm0, %ymm1, %k1
 ; AVX512BW-NEXT:    vcmpltpd %ymm2, %ymm3, %k0 {%k1}
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
   %x0 = fcmp ogt <4 x double> %a, %b
diff --git a/test/CodeGen/X86/bitcast-setcc-128.ll b/test/CodeGen/X86/bitcast-setcc-128.ll
index 8fdacb7b79d..ab55f78422e 100644
--- a/test/CodeGen/X86/bitcast-setcc-128.ll
+++ b/test/CodeGen/X86/bitcast-setcc-128.ll
@@ -64,16 +64,14 @@ define i4 @v4i32(<4 x i32> %a, <4 x i32> %b) {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4i32:
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <4 x i32> %a, %b
   %res = bitcast <4 x i1> %x to i4
@@ -99,16 +97,14 @@ define i4 @v4f32(<4 x float> %a, <4 x float> %b) {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vcmpltps %xmm0, %xmm1, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4f32:
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    vcmpltps %xmm0, %xmm1, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = fcmp ogt <4 x float> %a, %b
   %res = bitcast <4 x i1> %x to i4
@@ -226,8 +222,7 @@ define i2 @v2i8(<2 x i8> %a, <2 x i8> %b) {
 ; AVX512F-NEXT:    vpsraq $56, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i8:
@@ -238,8 +233,7 @@ define i2 @v2i8(<2 x i8> %a, <2 x i8> %b) {
 ; AVX512BW-NEXT:    vpsraq $56, %xmm0, %xmm0
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <2 x i8> %a, %b
   %res = bitcast <2 x i1> %x to i2
@@ -320,8 +314,7 @@ define i2 @v2i16(<2 x i16> %a, <2 x i16> %b) {
 ; AVX512F-NEXT:    vpsraq $48, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i16:
@@ -332,8 +325,7 @@ define i2 @v2i16(<2 x i16> %a, <2 x i16> %b) {
 ; AVX512BW-NEXT:    vpsraq $48, %xmm0, %xmm0
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <2 x i16> %a, %b
   %res = bitcast <2 x i1> %x to i2
@@ -406,8 +398,7 @@ define i2 @v2i32(<2 x i32> %a, <2 x i32> %b) {
 ; AVX512F-NEXT:    vpsraq $32, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i32:
@@ -418,8 +409,7 @@ define i2 @v2i32(<2 x i32> %a, <2 x i32> %b) {
 ; AVX512BW-NEXT:    vpsraq $32, %xmm0, %xmm0
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <2 x i32> %a, %b
   %res = bitcast <2 x i1> %x to i2
@@ -455,16 +445,14 @@ define i2 @v2i64(<2 x i64> %a, <2 x i64> %b) {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2i64:
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <2 x i64> %a, %b
   %res = bitcast <2 x i1> %x to i2
@@ -490,16 +478,14 @@ define i2 @v2f64(<2 x double> %a, <2 x double> %b) {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vcmpltpd %xmm0, %xmm1, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v2f64:
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    vcmpltpd %xmm0, %xmm1, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = fcmp ogt <2 x double> %a, %b
   %res = bitcast <2 x i1> %x to i2
@@ -537,8 +523,7 @@ define i4 @v4i8(<4 x i8> %a, <4 x i8> %b) {
 ; AVX512F-NEXT:    vpsrad $24, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4i8:
@@ -549,8 +534,7 @@ define i4 @v4i8(<4 x i8> %a, <4 x i8> %b) {
 ; AVX512BW-NEXT:    vpsrad $24, %xmm0, %xmm0
 ; AVX512BW-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <4 x i8> %a, %b
   %res = bitcast <4 x i1> %x to i4
@@ -588,8 +572,7 @@ define i4 @v4i16(<4 x i16> %a, <4 x i16> %b) {
 ; AVX512F-NEXT:    vpsrad $16, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: v4i16:
@@ -600,8 +583,7 @@ define i4 @v4i16(<4 x i16> %a, <4 x i16> %b) {
 ; AVX512BW-NEXT:    vpsrad $16, %xmm0, %xmm0
 ; AVX512BW-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <4 x i16> %a, %b
   %res = bitcast <4 x i1> %x to i4
diff --git a/test/CodeGen/X86/bitcast-setcc-256.ll b/test/CodeGen/X86/bitcast-setcc-256.ll
index 48e28c9d26c..7875318c109 100644
--- a/test/CodeGen/X86/bitcast-setcc-256.ll
+++ b/test/CodeGen/X86/bitcast-setcc-256.ll
@@ -271,8 +271,7 @@ define i4 @v4i64(<4 x i64> %a, <4 x i64> %b) {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -280,8 +279,7 @@ define i4 @v4i64(<4 x i64> %a, <4 x i64> %b) {
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
   %x = icmp sgt <4 x i64> %a, %b
@@ -311,8 +309,7 @@ define i4 @v4f64(<4 x double> %a, <4 x double> %b) {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vcmpltpd %ymm0, %ymm1, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
-; AVX512F-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512F-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512F-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -320,8 +317,7 @@ define i4 @v4f64(<4 x double> %a, <4 x double> %b) {
 ; AVX512BW:       # %bb.0:
 ; AVX512BW-NEXT:    vcmpltpd %ymm0, %ymm1, %k0
 ; AVX512BW-NEXT:    kmovd %k0, %eax
-; AVX512BW-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
-; AVX512BW-NEXT:    movb -{{[0-9]+}}(%rsp), %al
+; AVX512BW-NEXT:    # kill: def %al killed %al killed %eax
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
   %x = fcmp ogt <4 x double> %a, %b