; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k1
; AVX512F-NEXT: kshiftlw $15, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: ## implicit-def: %XMM0
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_2
; AVX512F-NEXT: ## BB#1: ## %cond.load
; AVX512F-NEXT: movzbl (%rdi), %eax
; AVX512F-NEXT: LBB50_2: ## %else
; AVX512F-NEXT: kshiftlw $14, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_4
; AVX512F-NEXT: ## BB#3: ## %cond.load1
; AVX512F-NEXT: vpinsrb $1, 1(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_4: ## %else2
; AVX512F-NEXT: kshiftlw $13, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_6
; AVX512F-NEXT: ## BB#5: ## %cond.load4
; AVX512F-NEXT: vpinsrb $2, 2(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_6: ## %else5
; AVX512F-NEXT: kshiftlw $12, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_8
; AVX512F-NEXT: ## BB#7: ## %cond.load7
; AVX512F-NEXT: vpinsrb $3, 3(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_8: ## %else8
; AVX512F-NEXT: kshiftlw $11, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_10
; AVX512F-NEXT: ## BB#9: ## %cond.load10
; AVX512F-NEXT: vpinsrb $4, 4(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_10: ## %else11
; AVX512F-NEXT: kshiftlw $10, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_12
; AVX512F-NEXT: ## BB#11: ## %cond.load13
; AVX512F-NEXT: vpinsrb $5, 5(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_12: ## %else14
; AVX512F-NEXT: kshiftlw $9, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_14
; AVX512F-NEXT: ## BB#13: ## %cond.load16
; AVX512F-NEXT: vpinsrb $6, 6(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_14: ## %else17
; AVX512F-NEXT: kshiftlw $8, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_16
; AVX512F-NEXT: ## BB#15: ## %cond.load19
; AVX512F-NEXT: vpinsrb $7, 7(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_16: ## %else20
; AVX512F-NEXT: kshiftlw $7, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_18
; AVX512F-NEXT: ## BB#17: ## %cond.load22
; AVX512F-NEXT: vpinsrb $8, 8(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_18: ## %else23
; AVX512F-NEXT: kshiftlw $6, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_20
; AVX512F-NEXT: ## BB#19: ## %cond.load25
; AVX512F-NEXT: vpinsrb $9, 9(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_20: ## %else26
; AVX512F-NEXT: kshiftlw $5, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_22
; AVX512F-NEXT: ## BB#21: ## %cond.load28
; AVX512F-NEXT: vpinsrb $10, 10(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_22: ## %else29
; AVX512F-NEXT: kshiftlw $4, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_24
; AVX512F-NEXT: ## BB#23: ## %cond.load31
; AVX512F-NEXT: vpinsrb $11, 11(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_24: ## %else32
; AVX512F-NEXT: kshiftlw $3, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_26
; AVX512F-NEXT: ## BB#25: ## %cond.load34
; AVX512F-NEXT: vpinsrb $12, 12(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_26: ## %else35
; AVX512F-NEXT: kshiftlw $2, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_28
; AVX512F-NEXT: ## BB#27: ## %cond.load37
; AVX512F-NEXT: vpinsrb $13, 13(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_28: ## %else38
; AVX512F-NEXT: kshiftlw $1, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_30
; AVX512F-NEXT: ## BB#29: ## %cond.load40
; AVX512F-NEXT: vpinsrb $14, 14(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB50_30: ## %else41
; AVX512F-NEXT: kshiftrw $15, %k1, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB50_32
; AVX512F-NEXT: ## BB#31: ## %cond.load43
; AVX512F-NEXT: vpinsrb $15, 15(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_2
; AVX512F-NEXT: ## BB#1: ## %cond.load
; AVX512F-NEXT: movzbl (%rdi), %eax
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_4
; AVX512F-NEXT: ## BB#3: ## %cond.load1
; AVX512F-NEXT: vpinsrb $1, 1(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_6
; AVX512F-NEXT: ## BB#5: ## %cond.load4
; AVX512F-NEXT: vpinsrb $2, 2(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_8
; AVX512F-NEXT: ## BB#7: ## %cond.load7
; AVX512F-NEXT: vpinsrb $3, 3(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_10
; AVX512F-NEXT: ## BB#9: ## %cond.load10
; AVX512F-NEXT: vpinsrb $4, 4(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_12
; AVX512F-NEXT: ## BB#11: ## %cond.load13
; AVX512F-NEXT: vpinsrb $5, 5(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_14
; AVX512F-NEXT: ## BB#13: ## %cond.load16
; AVX512F-NEXT: vpinsrb $6, 6(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_16
; AVX512F-NEXT: ## BB#15: ## %cond.load19
; AVX512F-NEXT: vpinsrb $7, 7(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_18
; AVX512F-NEXT: ## BB#17: ## %cond.load22
; AVX512F-NEXT: vpinsrb $8, 8(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, (%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_20
; AVX512F-NEXT: ## BB#19: ## %cond.load25
; AVX512F-NEXT: vpinsrb $9, 9(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_22
; AVX512F-NEXT: ## BB#21: ## %cond.load28
; AVX512F-NEXT: vpinsrb $10, 10(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_24
; AVX512F-NEXT: ## BB#23: ## %cond.load31
; AVX512F-NEXT: vpinsrb $11, 11(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_26
; AVX512F-NEXT: ## BB#25: ## %cond.load34
; AVX512F-NEXT: vpinsrb $12, 12(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_28
; AVX512F-NEXT: ## BB#27: ## %cond.load37
; AVX512F-NEXT: vpinsrb $13, 13(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_30
; AVX512F-NEXT: ## BB#29: ## %cond.load40
; AVX512F-NEXT: vpinsrb $14, 14(%rdi), %xmm0, %xmm6
; AVX512F-NEXT: vptestmd %zmm1, %zmm1, %k1
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_32
; AVX512F-NEXT: ## BB#31: ## %cond.load43
; AVX512F-NEXT: vpinsrb $15, 15(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: kshiftlw $15, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_34
; AVX512F-NEXT: ## BB#33: ## %cond.load46
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $14, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_36
; AVX512F-NEXT: ## BB#35: ## %cond.load49
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $13, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_38
; AVX512F-NEXT: ## BB#37: ## %cond.load52
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $12, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_40
; AVX512F-NEXT: ## BB#39: ## %cond.load55
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $11, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_42
; AVX512F-NEXT: ## BB#41: ## %cond.load58
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $10, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_44
; AVX512F-NEXT: ## BB#43: ## %cond.load61
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $9, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_46
; AVX512F-NEXT: ## BB#45: ## %cond.load64
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $8, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_48
; AVX512F-NEXT: ## BB#47: ## %cond.load67
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $7, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_50
; AVX512F-NEXT: ## BB#49: ## %cond.load70
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $6, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_52
; AVX512F-NEXT: ## BB#51: ## %cond.load73
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $5, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_54
; AVX512F-NEXT: ## BB#53: ## %cond.load76
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $4, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_56
; AVX512F-NEXT: ## BB#55: ## %cond.load79
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $3, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_58
; AVX512F-NEXT: ## BB#57: ## %cond.load82
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $2, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_60
; AVX512F-NEXT: ## BB#59: ## %cond.load85
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX512F-NEXT: kshiftlw $1, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_62
; AVX512F-NEXT: ## BB#61: ## %cond.load88
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX512F-NEXT: vptestmd %zmm1, %zmm1, %k0
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_64
; AVX512F-NEXT: ## BB#63: ## %cond.load91
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_66
; AVX512F-NEXT: ## BB#65: ## %cond.load94
; AVX512F-NEXT: vpinsrb $0, 32(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_68
; AVX512F-NEXT: ## BB#67: ## %cond.load97
; AVX512F-NEXT: vpinsrb $1, 33(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_70
; AVX512F-NEXT: ## BB#69: ## %cond.load100
; AVX512F-NEXT: vpinsrb $2, 34(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_72
; AVX512F-NEXT: ## BB#71: ## %cond.load103
; AVX512F-NEXT: vpinsrb $3, 35(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_74
; AVX512F-NEXT: ## BB#73: ## %cond.load106
; AVX512F-NEXT: vpinsrb $4, 36(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_76
; AVX512F-NEXT: ## BB#75: ## %cond.load109
; AVX512F-NEXT: vpinsrb $5, 37(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_78
; AVX512F-NEXT: ## BB#77: ## %cond.load112
; AVX512F-NEXT: vpinsrb $6, 38(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_80
; AVX512F-NEXT: ## BB#79: ## %cond.load115
; AVX512F-NEXT: vpinsrb $7, 39(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_82
; AVX512F-NEXT: ## BB#81: ## %cond.load118
; AVX512F-NEXT: vpinsrb $8, 40(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_84
; AVX512F-NEXT: ## BB#83: ## %cond.load121
; AVX512F-NEXT: vpinsrb $9, 41(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_86
; AVX512F-NEXT: ## BB#85: ## %cond.load124
; AVX512F-NEXT: vpinsrb $10, 42(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_88
; AVX512F-NEXT: ## BB#87: ## %cond.load127
; AVX512F-NEXT: vpinsrb $11, 43(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_90
; AVX512F-NEXT: ## BB#89: ## %cond.load130
; AVX512F-NEXT: vpinsrb $12, 44(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_92
; AVX512F-NEXT: ## BB#91: ## %cond.load133
; AVX512F-NEXT: vpinsrb $13, 45(%rdi), %xmm1, %xmm3
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_94
; AVX512F-NEXT: ## BB#93: ## %cond.load136
; AVX512F-NEXT: vpinsrb $14, 46(%rdi), %xmm1, %xmm3
; AVX512F-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-NEXT: LBB52_94: ## %else137
-; AVX512F-NEXT: vptestmd %zmm2, %zmm2, %k7
+; AVX512F-NEXT: vptestmd %zmm2, %zmm2, %k5
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_96
; AVX512F-NEXT: ## BB#95: ## %cond.load139
; AVX512F-NEXT: vpinsrb $15, 47(%rdi), %xmm1, %xmm2
; AVX512F-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-NEXT: LBB52_96: ## %else140
-; AVX512F-NEXT: kshiftlw $15, %k7, %k0
+; AVX512F-NEXT: kshiftlw $15, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_98
; AVX512F-NEXT: ## BB#97: ## %cond.load142
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $0, 48(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_98: ## %else143
-; AVX512F-NEXT: kshiftlw $14, %k7, %k0
+; AVX512F-NEXT: kshiftlw $14, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_100
; AVX512F-NEXT: ## BB#99: ## %cond.load145
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $1, 49(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_100: ## %else146
-; AVX512F-NEXT: kshiftlw $13, %k7, %k0
+; AVX512F-NEXT: kshiftlw $13, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_102
; AVX512F-NEXT: ## BB#101: ## %cond.load148
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $2, 50(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_102: ## %else149
-; AVX512F-NEXT: kshiftlw $12, %k7, %k0
+; AVX512F-NEXT: kshiftlw $12, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_104
; AVX512F-NEXT: ## BB#103: ## %cond.load151
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $3, 51(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_104: ## %else152
-; AVX512F-NEXT: kshiftlw $11, %k7, %k0
+; AVX512F-NEXT: kshiftlw $11, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_106
; AVX512F-NEXT: ## BB#105: ## %cond.load154
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $4, 52(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_106: ## %else155
-; AVX512F-NEXT: kshiftlw $10, %k7, %k0
+; AVX512F-NEXT: kshiftlw $10, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_108
; AVX512F-NEXT: ## BB#107: ## %cond.load157
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $5, 53(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_108: ## %else158
-; AVX512F-NEXT: kshiftlw $9, %k7, %k0
+; AVX512F-NEXT: kshiftlw $9, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_110
; AVX512F-NEXT: ## BB#109: ## %cond.load160
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $6, 54(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_110: ## %else161
-; AVX512F-NEXT: kshiftlw $8, %k7, %k0
+; AVX512F-NEXT: kshiftlw $8, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_112
; AVX512F-NEXT: ## BB#111: ## %cond.load163
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $7, 55(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_112: ## %else164
-; AVX512F-NEXT: kshiftlw $7, %k7, %k0
+; AVX512F-NEXT: kshiftlw $7, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, -{{[0-9]+}}(%rsp) ## 2-byte Spill
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_114
; AVX512F-NEXT: ## BB#113: ## %cond.load166
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $8, 56(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_114: ## %else167
-; AVX512F-NEXT: kshiftlw $6, %k7, %k0
+; AVX512F-NEXT: kshiftlw $6, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k2
-; AVX512F-NEXT: kortestw %k2, %k2
+; AVX512F-NEXT: kmovw %k2, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_116
; AVX512F-NEXT: ## BB#115: ## %cond.load169
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $9, 57(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_116: ## %else170
-; AVX512F-NEXT: kshiftlw $5, %k7, %k0
+; AVX512F-NEXT: kshiftlw $5, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k3
-; AVX512F-NEXT: kortestw %k3, %k3
+; AVX512F-NEXT: kmovw %k3, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_118
; AVX512F-NEXT: ## BB#117: ## %cond.load172
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $10, 58(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_118: ## %else173
-; AVX512F-NEXT: kshiftlw $4, %k7, %k0
+; AVX512F-NEXT: kshiftlw $4, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k4
-; AVX512F-NEXT: kortestw %k4, %k4
+; AVX512F-NEXT: kmovw %k4, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_120
; AVX512F-NEXT: ## BB#119: ## %cond.load175
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $11, 59(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_120: ## %else176
-; AVX512F-NEXT: kshiftlw $3, %k7, %k0
-; AVX512F-NEXT: kshiftrw $15, %k0, %k5
-; AVX512F-NEXT: kortestw %k5, %k5
+; AVX512F-NEXT: kshiftlw $3, %k5, %k0
+; AVX512F-NEXT: kshiftrw $15, %k0, %k6
+; AVX512F-NEXT: kmovw %k6, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_122
; AVX512F-NEXT: ## BB#121: ## %cond.load178
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $12, 60(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_122: ## %else179
-; AVX512F-NEXT: kshiftlw $2, %k7, %k0
-; AVX512F-NEXT: kshiftrw $15, %k0, %k6
-; AVX512F-NEXT: kortestw %k6, %k6
+; AVX512F-NEXT: kshiftlw $2, %k5, %k0
+; AVX512F-NEXT: kshiftrw $15, %k0, %k7
+; AVX512F-NEXT: kmovw %k7, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_124
; AVX512F-NEXT: ## BB#123: ## %cond.load181
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $13, 61(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_124: ## %else182
-; AVX512F-NEXT: kshiftlw $1, %k7, %k0
+; AVX512F-NEXT: kshiftlw $1, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_126
; AVX512F-NEXT: ## BB#125: ## %cond.load184
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: vpinsrb $14, 62(%rdi), %xmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: LBB52_126: ## %else185
-; AVX512F-NEXT: kshiftrw $15, %k7, %k7
-; AVX512F-NEXT: kortestw %k7, %k7
+; AVX512F-NEXT: kshiftrw $15, %k5, %k5
+; AVX512F-NEXT: kmovw %k5, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB52_128
; AVX512F-NEXT: ## BB#127: ## %cond.load187
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX512F-NEXT: movl %eax, {{[0-9]+}}(%rsp) ## 4-byte Spill
; AVX512F-NEXT: kmovw %k3, %r12d
; AVX512F-NEXT: kmovw %k4, %r15d
-; AVX512F-NEXT: kmovw %k5, %r14d
-; AVX512F-NEXT: kmovw %k6, %ebx
+; AVX512F-NEXT: kmovw %k6, %r14d
+; AVX512F-NEXT: kmovw %k7, %ebx
; AVX512F-NEXT: kmovw %k0, %r11d
-; AVX512F-NEXT: kmovw %k7, %r10d
+; AVX512F-NEXT: kmovw %k5, %r10d
; AVX512F-NEXT: kmovw -{{[0-9]+}}(%rsp), %k0 ## 2-byte Reload
; AVX512F-NEXT: kmovw %k0, %eax
; AVX512F-NEXT: movl %eax, -{{[0-9]+}}(%rsp) ## 4-byte Spill
; AVX512F-NEXT: vptestmq %zmm0, %zmm0, %k1
; AVX512F-NEXT: kshiftlw $15, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: ## implicit-def: %XMM0
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_2
; AVX512F-NEXT: ## BB#1: ## %cond.load
; AVX512F-NEXT: movzwl (%rdi), %eax
; AVX512F-NEXT: LBB53_2: ## %else
; AVX512F-NEXT: kshiftlw $14, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_4
; AVX512F-NEXT: ## BB#3: ## %cond.load1
; AVX512F-NEXT: vpinsrw $1, 2(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB53_4: ## %else2
; AVX512F-NEXT: kshiftlw $13, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_6
; AVX512F-NEXT: ## BB#5: ## %cond.load4
; AVX512F-NEXT: vpinsrw $2, 4(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB53_6: ## %else5
; AVX512F-NEXT: kshiftlw $12, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_8
; AVX512F-NEXT: ## BB#7: ## %cond.load7
; AVX512F-NEXT: vpinsrw $3, 6(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB53_8: ## %else8
; AVX512F-NEXT: kshiftlw $11, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_10
; AVX512F-NEXT: ## BB#9: ## %cond.load10
; AVX512F-NEXT: vpinsrw $4, 8(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB53_10: ## %else11
; AVX512F-NEXT: kshiftlw $10, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_12
; AVX512F-NEXT: ## BB#11: ## %cond.load13
; AVX512F-NEXT: vpinsrw $5, 10(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB53_12: ## %else14
; AVX512F-NEXT: kshiftlw $9, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_14
; AVX512F-NEXT: ## BB#13: ## %cond.load16
; AVX512F-NEXT: vpinsrw $6, 12(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: LBB53_14: ## %else17
; AVX512F-NEXT: kshiftlw $8, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB53_16
; AVX512F-NEXT: ## BB#15: ## %cond.load19
; AVX512F-NEXT: vpinsrw $7, 14(%rdi), %xmm0, %xmm0
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k1
; AVX512F-NEXT: kshiftlw $15, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: ## implicit-def: %YMM0
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_2
; AVX512F-NEXT: ## BB#1: ## %cond.load
; AVX512F-NEXT: movzwl (%rdi), %eax
; AVX512F-NEXT: LBB54_2: ## %else
; AVX512F-NEXT: kshiftlw $14, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_4
; AVX512F-NEXT: ## BB#3: ## %cond.load1
; AVX512F-NEXT: vpinsrw $1, 2(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: LBB54_4: ## %else2
; AVX512F-NEXT: kshiftlw $13, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_6
; AVX512F-NEXT: ## BB#5: ## %cond.load4
; AVX512F-NEXT: vpinsrw $2, 4(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: LBB54_6: ## %else5
; AVX512F-NEXT: kshiftlw $12, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_8
; AVX512F-NEXT: ## BB#7: ## %cond.load7
; AVX512F-NEXT: vpinsrw $3, 6(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: LBB54_8: ## %else8
; AVX512F-NEXT: kshiftlw $11, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_10
; AVX512F-NEXT: ## BB#9: ## %cond.load10
; AVX512F-NEXT: vpinsrw $4, 8(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: LBB54_10: ## %else11
; AVX512F-NEXT: kshiftlw $10, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_12
; AVX512F-NEXT: ## BB#11: ## %cond.load13
; AVX512F-NEXT: vpinsrw $5, 10(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: LBB54_12: ## %else14
; AVX512F-NEXT: kshiftlw $9, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_14
; AVX512F-NEXT: ## BB#13: ## %cond.load16
; AVX512F-NEXT: vpinsrw $6, 12(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: LBB54_14: ## %else17
; AVX512F-NEXT: kshiftlw $8, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_16
; AVX512F-NEXT: ## BB#15: ## %cond.load19
; AVX512F-NEXT: vpinsrw $7, 14(%rdi), %xmm0, %xmm1
; AVX512F-NEXT: LBB54_16: ## %else20
; AVX512F-NEXT: kshiftlw $7, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_18
; AVX512F-NEXT: ## BB#17: ## %cond.load22
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: LBB54_18: ## %else23
; AVX512F-NEXT: kshiftlw $6, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_20
; AVX512F-NEXT: ## BB#19: ## %cond.load25
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: LBB54_20: ## %else26
; AVX512F-NEXT: kshiftlw $5, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_22
; AVX512F-NEXT: ## BB#21: ## %cond.load28
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: LBB54_22: ## %else29
; AVX512F-NEXT: kshiftlw $4, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_24
; AVX512F-NEXT: ## BB#23: ## %cond.load31
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: LBB54_24: ## %else32
; AVX512F-NEXT: kshiftlw $3, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_26
; AVX512F-NEXT: ## BB#25: ## %cond.load34
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: LBB54_26: ## %else35
; AVX512F-NEXT: kshiftlw $2, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_28
; AVX512F-NEXT: ## BB#27: ## %cond.load37
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: LBB54_28: ## %else38
; AVX512F-NEXT: kshiftlw $1, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_30
; AVX512F-NEXT: ## BB#29: ## %cond.load40
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-NEXT: LBB54_30: ## %else41
; AVX512F-NEXT: kshiftrw $15, %k1, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB54_32
; AVX512F-NEXT: ## BB#31: ## %cond.load43
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_2
; AVX512F-NEXT: ## BB#1: ## %cond.store
; AVX512F-NEXT: vpextrb $0, %xmm1, (%rdi)
; AVX512F-NEXT: LBB56_2: ## %else
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_4
; AVX512F-NEXT: ## BB#3: ## %cond.store1
; AVX512F-NEXT: vpextrb $1, %xmm1, 1(%rdi)
; AVX512F-NEXT: LBB56_4: ## %else2
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_6
; AVX512F-NEXT: ## BB#5: ## %cond.store3
; AVX512F-NEXT: vpextrb $2, %xmm1, 2(%rdi)
; AVX512F-NEXT: LBB56_6: ## %else4
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_8
; AVX512F-NEXT: ## BB#7: ## %cond.store5
; AVX512F-NEXT: vpextrb $3, %xmm1, 3(%rdi)
; AVX512F-NEXT: LBB56_8: ## %else6
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_10
; AVX512F-NEXT: ## BB#9: ## %cond.store7
; AVX512F-NEXT: vpextrb $4, %xmm1, 4(%rdi)
; AVX512F-NEXT: LBB56_10: ## %else8
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_12
; AVX512F-NEXT: ## BB#11: ## %cond.store9
; AVX512F-NEXT: vpextrb $5, %xmm1, 5(%rdi)
; AVX512F-NEXT: LBB56_12: ## %else10
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_14
; AVX512F-NEXT: ## BB#13: ## %cond.store11
; AVX512F-NEXT: vpextrb $6, %xmm1, 6(%rdi)
; AVX512F-NEXT: LBB56_14: ## %else12
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_16
; AVX512F-NEXT: ## BB#15: ## %cond.store13
; AVX512F-NEXT: vpextrb $7, %xmm1, 7(%rdi)
; AVX512F-NEXT: LBB56_16: ## %else14
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_18
; AVX512F-NEXT: ## BB#17: ## %cond.store15
; AVX512F-NEXT: vpextrb $8, %xmm1, 8(%rdi)
; AVX512F-NEXT: LBB56_18: ## %else16
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_20
; AVX512F-NEXT: ## BB#19: ## %cond.store17
; AVX512F-NEXT: vpextrb $9, %xmm1, 9(%rdi)
; AVX512F-NEXT: LBB56_20: ## %else18
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_22
; AVX512F-NEXT: ## BB#21: ## %cond.store19
; AVX512F-NEXT: vpextrb $10, %xmm1, 10(%rdi)
; AVX512F-NEXT: LBB56_22: ## %else20
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_24
; AVX512F-NEXT: ## BB#23: ## %cond.store21
; AVX512F-NEXT: vpextrb $11, %xmm1, 11(%rdi)
; AVX512F-NEXT: LBB56_24: ## %else22
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_26
; AVX512F-NEXT: ## BB#25: ## %cond.store23
; AVX512F-NEXT: vpextrb $12, %xmm1, 12(%rdi)
; AVX512F-NEXT: LBB56_26: ## %else24
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_28
; AVX512F-NEXT: ## BB#27: ## %cond.store25
; AVX512F-NEXT: vpextrb $13, %xmm1, 13(%rdi)
; AVX512F-NEXT: LBB56_28: ## %else26
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_30
; AVX512F-NEXT: ## BB#29: ## %cond.store27
; AVX512F-NEXT: vpextrb $14, %xmm1, 14(%rdi)
; AVX512F-NEXT: LBB56_30: ## %else28
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB56_32
; AVX512F-NEXT: ## BB#31: ## %cond.store29
; AVX512F-NEXT: vpextrb $15, %xmm1, 15(%rdi)
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_2
; AVX512F-NEXT: ## BB#1: ## %cond.store
; AVX512F-NEXT: vpextrb $0, %xmm4, (%rdi)
; AVX512F-NEXT: LBB58_2: ## %else
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_4
; AVX512F-NEXT: ## BB#3: ## %cond.store1
; AVX512F-NEXT: vpextrb $1, %xmm4, 1(%rdi)
; AVX512F-NEXT: LBB58_4: ## %else2
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_6
; AVX512F-NEXT: ## BB#5: ## %cond.store3
; AVX512F-NEXT: vpextrb $2, %xmm4, 2(%rdi)
; AVX512F-NEXT: LBB58_6: ## %else4
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_8
; AVX512F-NEXT: ## BB#7: ## %cond.store5
; AVX512F-NEXT: vpextrb $3, %xmm4, 3(%rdi)
; AVX512F-NEXT: LBB58_8: ## %else6
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_10
; AVX512F-NEXT: ## BB#9: ## %cond.store7
; AVX512F-NEXT: vpextrb $4, %xmm4, 4(%rdi)
; AVX512F-NEXT: LBB58_10: ## %else8
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_12
; AVX512F-NEXT: ## BB#11: ## %cond.store9
; AVX512F-NEXT: vpextrb $5, %xmm4, 5(%rdi)
; AVX512F-NEXT: LBB58_12: ## %else10
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_14
; AVX512F-NEXT: ## BB#13: ## %cond.store11
; AVX512F-NEXT: vpextrb $6, %xmm4, 6(%rdi)
; AVX512F-NEXT: LBB58_14: ## %else12
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_16
; AVX512F-NEXT: ## BB#15: ## %cond.store13
; AVX512F-NEXT: vpextrb $7, %xmm4, 7(%rdi)
; AVX512F-NEXT: LBB58_16: ## %else14
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_18
; AVX512F-NEXT: ## BB#17: ## %cond.store15
; AVX512F-NEXT: vpextrb $8, %xmm4, 8(%rdi)
; AVX512F-NEXT: LBB58_18: ## %else16
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_20
; AVX512F-NEXT: ## BB#19: ## %cond.store17
; AVX512F-NEXT: vpextrb $9, %xmm4, 9(%rdi)
; AVX512F-NEXT: LBB58_20: ## %else18
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_22
; AVX512F-NEXT: ## BB#21: ## %cond.store19
; AVX512F-NEXT: vpextrb $10, %xmm4, 10(%rdi)
; AVX512F-NEXT: LBB58_22: ## %else20
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_24
; AVX512F-NEXT: ## BB#23: ## %cond.store21
; AVX512F-NEXT: vpextrb $11, %xmm4, 11(%rdi)
; AVX512F-NEXT: LBB58_24: ## %else22
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_26
; AVX512F-NEXT: ## BB#25: ## %cond.store23
; AVX512F-NEXT: vpextrb $12, %xmm4, 12(%rdi)
; AVX512F-NEXT: vpmovsxbd %xmm1, %zmm0
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_28
; AVX512F-NEXT: ## BB#27: ## %cond.store25
; AVX512F-NEXT: vpextrb $13, %xmm4, 13(%rdi)
; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_30
; AVX512F-NEXT: ## BB#29: ## %cond.store27
; AVX512F-NEXT: vpextrb $14, %xmm4, 14(%rdi)
; AVX512F-NEXT: LBB58_30: ## %else28
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k1
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_32
; AVX512F-NEXT: ## BB#31: ## %cond.store29
; AVX512F-NEXT: vpextrb $15, %xmm4, 15(%rdi)
; AVX512F-NEXT: LBB58_32: ## %else30
; AVX512F-NEXT: kshiftlw $15, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_34
; AVX512F-NEXT: ## BB#33: ## %cond.store31
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_34: ## %else32
; AVX512F-NEXT: kshiftlw $14, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_36
; AVX512F-NEXT: ## BB#35: ## %cond.store33
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_36: ## %else34
; AVX512F-NEXT: kshiftlw $13, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_38
; AVX512F-NEXT: ## BB#37: ## %cond.store35
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_38: ## %else36
; AVX512F-NEXT: kshiftlw $12, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_40
; AVX512F-NEXT: ## BB#39: ## %cond.store37
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_40: ## %else38
; AVX512F-NEXT: kshiftlw $11, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_42
; AVX512F-NEXT: ## BB#41: ## %cond.store39
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_42: ## %else40
; AVX512F-NEXT: kshiftlw $10, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_44
; AVX512F-NEXT: ## BB#43: ## %cond.store41
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_44: ## %else42
; AVX512F-NEXT: kshiftlw $9, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_46
; AVX512F-NEXT: ## BB#45: ## %cond.store43
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_46: ## %else44
; AVX512F-NEXT: kshiftlw $8, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_48
; AVX512F-NEXT: ## BB#47: ## %cond.store45
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_48: ## %else46
; AVX512F-NEXT: kshiftlw $7, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_50
; AVX512F-NEXT: ## BB#49: ## %cond.store47
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_50: ## %else48
; AVX512F-NEXT: kshiftlw $6, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_52
; AVX512F-NEXT: ## BB#51: ## %cond.store49
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_52: ## %else50
; AVX512F-NEXT: kshiftlw $5, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_54
; AVX512F-NEXT: ## BB#53: ## %cond.store51
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_54: ## %else52
; AVX512F-NEXT: kshiftlw $4, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_56
; AVX512F-NEXT: ## BB#55: ## %cond.store53
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_56: ## %else54
; AVX512F-NEXT: kshiftlw $3, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_58
; AVX512F-NEXT: ## BB#57: ## %cond.store55
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: vpmovsxbd %xmm2, %zmm0
; AVX512F-NEXT: kshiftlw $2, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_60
; AVX512F-NEXT: ## BB#59: ## %cond.store57
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm1
; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
; AVX512F-NEXT: kshiftlw $1, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_62
; AVX512F-NEXT: ## BB#61: ## %cond.store59
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm1
; AVX512F-NEXT: LBB58_62: ## %else60
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_64
; AVX512F-NEXT: ## BB#63: ## %cond.store61
; AVX512F-NEXT: vextracti128 $1, %ymm4, %xmm0
; AVX512F-NEXT: LBB58_64: ## %else62
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_66
; AVX512F-NEXT: ## BB#65: ## %cond.store63
; AVX512F-NEXT: vpextrb $0, %xmm5, 32(%rdi)
; AVX512F-NEXT: LBB58_66: ## %else64
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_68
; AVX512F-NEXT: ## BB#67: ## %cond.store65
; AVX512F-NEXT: vpextrb $1, %xmm5, 33(%rdi)
; AVX512F-NEXT: LBB58_68: ## %else66
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_70
; AVX512F-NEXT: ## BB#69: ## %cond.store67
; AVX512F-NEXT: vpextrb $2, %xmm5, 34(%rdi)
; AVX512F-NEXT: LBB58_70: ## %else68
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_72
; AVX512F-NEXT: ## BB#71: ## %cond.store69
; AVX512F-NEXT: vpextrb $3, %xmm5, 35(%rdi)
; AVX512F-NEXT: LBB58_72: ## %else70
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_74
; AVX512F-NEXT: ## BB#73: ## %cond.store71
; AVX512F-NEXT: vpextrb $4, %xmm5, 36(%rdi)
; AVX512F-NEXT: LBB58_74: ## %else72
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_76
; AVX512F-NEXT: ## BB#75: ## %cond.store73
; AVX512F-NEXT: vpextrb $5, %xmm5, 37(%rdi)
; AVX512F-NEXT: LBB58_76: ## %else74
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_78
; AVX512F-NEXT: ## BB#77: ## %cond.store75
; AVX512F-NEXT: vpextrb $6, %xmm5, 38(%rdi)
; AVX512F-NEXT: LBB58_78: ## %else76
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_80
; AVX512F-NEXT: ## BB#79: ## %cond.store77
; AVX512F-NEXT: vpextrb $7, %xmm5, 39(%rdi)
; AVX512F-NEXT: LBB58_80: ## %else78
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_82
; AVX512F-NEXT: ## BB#81: ## %cond.store79
; AVX512F-NEXT: vpextrb $8, %xmm5, 40(%rdi)
; AVX512F-NEXT: LBB58_82: ## %else80
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_84
; AVX512F-NEXT: ## BB#83: ## %cond.store81
; AVX512F-NEXT: vpextrb $9, %xmm5, 41(%rdi)
; AVX512F-NEXT: LBB58_84: ## %else82
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_86
; AVX512F-NEXT: ## BB#85: ## %cond.store83
; AVX512F-NEXT: vpextrb $10, %xmm5, 42(%rdi)
; AVX512F-NEXT: LBB58_86: ## %else84
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_88
; AVX512F-NEXT: ## BB#87: ## %cond.store85
; AVX512F-NEXT: vpextrb $11, %xmm5, 43(%rdi)
; AVX512F-NEXT: LBB58_88: ## %else86
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_90
; AVX512F-NEXT: ## BB#89: ## %cond.store87
; AVX512F-NEXT: vpextrb $12, %xmm5, 44(%rdi)
; AVX512F-NEXT: vpmovsxbd %xmm3, %zmm0
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_92
; AVX512F-NEXT: ## BB#91: ## %cond.store89
; AVX512F-NEXT: vpextrb $13, %xmm5, 45(%rdi)
; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_94
; AVX512F-NEXT: ## BB#93: ## %cond.store91
; AVX512F-NEXT: vpextrb $14, %xmm5, 46(%rdi)
; AVX512F-NEXT: LBB58_94: ## %else92
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k1
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_96
; AVX512F-NEXT: ## BB#95: ## %cond.store93
; AVX512F-NEXT: vpextrb $15, %xmm5, 47(%rdi)
; AVX512F-NEXT: LBB58_96: ## %else94
; AVX512F-NEXT: kshiftlw $15, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_98
; AVX512F-NEXT: ## BB#97: ## %cond.store95
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_98: ## %else96
; AVX512F-NEXT: kshiftlw $14, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_100
; AVX512F-NEXT: ## BB#99: ## %cond.store97
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_100: ## %else98
; AVX512F-NEXT: kshiftlw $13, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_102
; AVX512F-NEXT: ## BB#101: ## %cond.store99
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_102: ## %else100
; AVX512F-NEXT: kshiftlw $12, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_104
; AVX512F-NEXT: ## BB#103: ## %cond.store101
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_104: ## %else102
; AVX512F-NEXT: kshiftlw $11, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_106
; AVX512F-NEXT: ## BB#105: ## %cond.store103
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_106: ## %else104
; AVX512F-NEXT: kshiftlw $10, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_108
; AVX512F-NEXT: ## BB#107: ## %cond.store105
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_108: ## %else106
; AVX512F-NEXT: kshiftlw $9, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_110
; AVX512F-NEXT: ## BB#109: ## %cond.store107
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_110: ## %else108
; AVX512F-NEXT: kshiftlw $8, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_112
; AVX512F-NEXT: ## BB#111: ## %cond.store109
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_112: ## %else110
; AVX512F-NEXT: kshiftlw $7, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_114
; AVX512F-NEXT: ## BB#113: ## %cond.store111
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_114: ## %else112
; AVX512F-NEXT: kshiftlw $6, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_116
; AVX512F-NEXT: ## BB#115: ## %cond.store113
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_116: ## %else114
; AVX512F-NEXT: kshiftlw $5, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_118
; AVX512F-NEXT: ## BB#117: ## %cond.store115
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_118: ## %else116
; AVX512F-NEXT: kshiftlw $4, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_120
; AVX512F-NEXT: ## BB#119: ## %cond.store117
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_120: ## %else118
; AVX512F-NEXT: kshiftlw $3, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_122
; AVX512F-NEXT: ## BB#121: ## %cond.store119
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_122: ## %else120
; AVX512F-NEXT: kshiftlw $2, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_124
; AVX512F-NEXT: ## BB#123: ## %cond.store121
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: LBB58_124: ## %else122
; AVX512F-NEXT: kshiftlw $1, %k1, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_126
; AVX512F-NEXT: ## BB#125: ## %cond.store123
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: vpextrb $14, %xmm0, 62(%rdi)
; AVX512F-NEXT: LBB58_126: ## %else124
; AVX512F-NEXT: kshiftrw $15, %k1, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB58_128
; AVX512F-NEXT: ## BB#127: ## %cond.store125
; AVX512F-NEXT: vextracti128 $1, %ymm5, %xmm0
; AVX512F-NEXT: vptestmq %zmm0, %zmm0, %k0
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_2
; AVX512F-NEXT: ## BB#1: ## %cond.store
; AVX512F-NEXT: vpextrw $0, %xmm1, (%rdi)
; AVX512F-NEXT: LBB59_2: ## %else
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_4
; AVX512F-NEXT: ## BB#3: ## %cond.store1
; AVX512F-NEXT: vpextrw $1, %xmm1, 2(%rdi)
; AVX512F-NEXT: LBB59_4: ## %else2
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_6
; AVX512F-NEXT: ## BB#5: ## %cond.store3
; AVX512F-NEXT: vpextrw $2, %xmm1, 4(%rdi)
; AVX512F-NEXT: LBB59_6: ## %else4
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_8
; AVX512F-NEXT: ## BB#7: ## %cond.store5
; AVX512F-NEXT: vpextrw $3, %xmm1, 6(%rdi)
; AVX512F-NEXT: LBB59_8: ## %else6
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_10
; AVX512F-NEXT: ## BB#9: ## %cond.store7
; AVX512F-NEXT: vpextrw $4, %xmm1, 8(%rdi)
; AVX512F-NEXT: LBB59_10: ## %else8
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_12
; AVX512F-NEXT: ## BB#11: ## %cond.store9
; AVX512F-NEXT: vpextrw $5, %xmm1, 10(%rdi)
; AVX512F-NEXT: LBB59_12: ## %else10
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_14
; AVX512F-NEXT: ## BB#13: ## %cond.store11
; AVX512F-NEXT: vpextrw $6, %xmm1, 12(%rdi)
; AVX512F-NEXT: LBB59_14: ## %else12
; AVX512F-NEXT: kshiftlw $8, %k0, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB59_16
; AVX512F-NEXT: ## BB#15: ## %cond.store13
; AVX512F-NEXT: vpextrw $7, %xmm1, 14(%rdi)
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_2
; AVX512F-NEXT: ## BB#1: ## %cond.store
; AVX512F-NEXT: vpextrw $0, %xmm1, (%rdi)
; AVX512F-NEXT: LBB60_2: ## %else
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_4
; AVX512F-NEXT: ## BB#3: ## %cond.store1
; AVX512F-NEXT: vpextrw $1, %xmm1, 2(%rdi)
; AVX512F-NEXT: LBB60_4: ## %else2
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_6
; AVX512F-NEXT: ## BB#5: ## %cond.store3
; AVX512F-NEXT: vpextrw $2, %xmm1, 4(%rdi)
; AVX512F-NEXT: LBB60_6: ## %else4
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_8
; AVX512F-NEXT: ## BB#7: ## %cond.store5
; AVX512F-NEXT: vpextrw $3, %xmm1, 6(%rdi)
; AVX512F-NEXT: LBB60_8: ## %else6
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_10
; AVX512F-NEXT: ## BB#9: ## %cond.store7
; AVX512F-NEXT: vpextrw $4, %xmm1, 8(%rdi)
; AVX512F-NEXT: LBB60_10: ## %else8
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_12
; AVX512F-NEXT: ## BB#11: ## %cond.store9
; AVX512F-NEXT: vpextrw $5, %xmm1, 10(%rdi)
; AVX512F-NEXT: LBB60_12: ## %else10
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_14
; AVX512F-NEXT: ## BB#13: ## %cond.store11
; AVX512F-NEXT: vpextrw $6, %xmm1, 12(%rdi)
; AVX512F-NEXT: LBB60_14: ## %else12
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_16
; AVX512F-NEXT: ## BB#15: ## %cond.store13
; AVX512F-NEXT: vpextrw $7, %xmm1, 14(%rdi)
; AVX512F-NEXT: LBB60_16: ## %else14
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_18
; AVX512F-NEXT: ## BB#17: ## %cond.store15
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512F-NEXT: LBB60_18: ## %else16
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_20
; AVX512F-NEXT: ## BB#19: ## %cond.store17
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512F-NEXT: LBB60_20: ## %else18
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_22
; AVX512F-NEXT: ## BB#21: ## %cond.store19
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512F-NEXT: LBB60_22: ## %else20
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_24
; AVX512F-NEXT: ## BB#23: ## %cond.store21
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512F-NEXT: LBB60_24: ## %else22
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_26
; AVX512F-NEXT: ## BB#25: ## %cond.store23
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512F-NEXT: LBB60_26: ## %else24
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_28
; AVX512F-NEXT: ## BB#27: ## %cond.store25
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512F-NEXT: LBB60_28: ## %else26
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
-; AVX512F-NEXT: kortestw %k1, %k1
+; AVX512F-NEXT: kmovw %k1, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_30
; AVX512F-NEXT: ## BB#29: ## %cond.store27
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512F-NEXT: vpextrw $6, %xmm0, 28(%rdi)
; AVX512F-NEXT: LBB60_30: ## %else28
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
-; AVX512F-NEXT: kortestw %k0, %k0
+; AVX512F-NEXT: kmovw %k0, %eax
+; AVX512F-NEXT: andl $1, %eax
+; AVX512F-NEXT: testb %al, %al
; AVX512F-NEXT: je LBB60_32
; AVX512F-NEXT: ## BB#31: ## %cond.store29
; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm0