; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X86
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X64

declare i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs, i32, <16 x i8>, i32, i8)
declare i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32, <16 x i8>, i32, i8)
declare <16 x i8> @llvm.x86.sse42.pcmpestrm128(<16 x i8> %lhs, i32, <16 x i8>, i32, i8)
declare i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs, <16 x i8>, i8)
declare i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8>, i8)
declare <16 x i8> @llvm.x86.sse42.pcmpistrm128(<16 x i8> %lhs, <16 x i8>, i8)

define i1 @pcmpestri_reg_eq_i8(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_reg_eq_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_reg_eq_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %c = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpestri_reg_idx_i8(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_reg_idx_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_reg_idx_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  ret i32 %idx
}

define i32 @pcmpestri_reg_diff_i8(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_reg_diff_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movl 8(%ebp), %eax
; X86-NEXT:    movl 12(%ebp), %edx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    cmpl $16, %ecx
; X86-NEXT:    jne .LBB2_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    jmp .LBB2_3
; X86-NEXT:  .LBB2_2: # %compare
; X86-NEXT:    movdqa %xmm0, (%esp)
; X86-NEXT:    andl $15, %ecx
; X86-NEXT:    movb (%esp,%ecx), %al
; X86-NEXT:    movdqa %xmm1, {{[0-9]+}}(%esp)
; X86-NEXT:    subb 16(%esp,%ecx), %al
; X86-NEXT:  .LBB2_3: # %exit
; X86-NEXT:    movzbl %al, %eax
; X86-NEXT:    movl %ebp, %esp
; X86-NEXT:    popl %ebp
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_reg_diff_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $16, %ecx
; X64-NEXT:    jne .LBB2_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB2_2: # %compare
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $15, %ecx
; X64-NEXT:    movb -24(%rsp,%rcx), %al
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subb -40(%rsp,%rcx), %al
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
entry:
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %eq = icmp eq i32 %idx, 16
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <16 x i8> %lhs, i32 %idx
  %rhs_c = extractelement <16 x i8> %rhs, i32 %idx
  %sub = sub i8 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i8 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i8 %result to i32
  ret i32 %result_ext
}

define i1 @pcmpestri_mem_eq_i8(i8* %lhs_ptr, i32 %lhs_len, i8* %rhs_ptr, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_mem_eq_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movdqu (%esi), %xmm0
; X86-NEXT:    pcmpestri $24, (%ecx), %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    popl %esi
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_mem_eq_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rdx, %r8
; X64-NEXT:    movl %esi, %eax
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    movl %ecx, %edx
; X64-NEXT:    pcmpestri $24, (%r8), %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i8* %lhs_ptr to <16 x i8>*
  %lhs = load <16 x i8>, <16 x i8>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i8* %rhs_ptr to <16 x i8>*
  %rhs = load <16 x i8>, <16 x i8>* %rhs_vptr, align 1
  %c = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpestri_mem_idx_i8(i8* %lhs_ptr, i32 %lhs_len, i8* %rhs_ptr, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_mem_idx_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movdqu (%esi), %xmm0
; X86-NEXT:    pcmpestri $24, (%ecx), %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    popl %esi
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_mem_idx_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rdx, %r8
; X64-NEXT:    movl %esi, %eax
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    movl %ecx, %edx
; X64-NEXT:    pcmpestri $24, (%r8), %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i8* %lhs_ptr to <16 x i8>*
  %lhs = load <16 x i8>, <16 x i8>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i8* %rhs_ptr to <16 x i8>*
  %rhs = load <16 x i8>, <16 x i8>* %rhs_vptr, align 1
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  ret i32 %idx
}

define i32 @pcmpestri_mem_diff_i8(i8* %lhs_ptr, i32 %lhs_len, i8* %rhs_ptr, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_mem_diff_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    pushl %esi
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movl 12(%ebp), %eax
; X86-NEXT:    movl 20(%ebp), %edx
; X86-NEXT:    movl 16(%ebp), %ecx
; X86-NEXT:    movl 8(%ebp), %esi
; X86-NEXT:    movdqu (%esi), %xmm1
; X86-NEXT:    movdqu (%ecx), %xmm0
; X86-NEXT:    pcmpestri $24, %xmm0, %xmm1
; X86-NEXT:    cmpl $16, %ecx
; X86-NEXT:    jne .LBB5_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    jmp .LBB5_3
; X86-NEXT:  .LBB5_2: # %compare
; X86-NEXT:    movdqa %xmm1, (%esp)
; X86-NEXT:    andl $15, %ecx
; X86-NEXT:    movb (%esp,%ecx), %al
; X86-NEXT:    movdqa %xmm0, {{[0-9]+}}(%esp)
; X86-NEXT:    subb 16(%esp,%ecx), %al
; X86-NEXT:  .LBB5_3: # %exit
; X86-NEXT:    movzbl %al, %eax
; X86-NEXT:    leal -4(%ebp), %esp
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %ebp
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_mem_diff_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %eax
; X64-NEXT:    movdqu (%rdi), %xmm1
; X64-NEXT:    movdqu (%rdx), %xmm0
; X64-NEXT:    movl %ecx, %edx
; X64-NEXT:    pcmpestri $24, %xmm0, %xmm1
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $16, %ecx
; X64-NEXT:    jne .LBB5_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB5_2: # %compare
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $15, %ecx
; X64-NEXT:    movb -24(%rsp,%rcx), %al
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subb -40(%rsp,%rcx), %al
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i8* %lhs_ptr to <16 x i8>*
  %lhs = load <16 x i8>, <16 x i8>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i8* %rhs_ptr to <16 x i8>*
  %rhs = load <16 x i8>, <16 x i8>* %rhs_vptr, align 1
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %eq = icmp eq i32 %idx, 16
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <16 x i8> %lhs, i32 %idx
  %rhs_c = extractelement <16 x i8> %rhs, i32 %idx
  %sub = sub i8 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i8 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i8 %result to i32
  ret i32 %result_ext
}

define i1 @pcmpestri_reg_eq_i16(<8 x i16> %lhs, i32 %lhs_len, <8 x i16> %rhs, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_reg_eq_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_reg_eq_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %c = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs_cast, i32 %lhs_len, <16 x i8> %rhs_cast, i32 %rhs_len, i8 24)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpestri_reg_idx_i16(<8 x i16> %lhs, i32 %lhs_len, <8 x i16> %rhs, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_reg_idx_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_reg_idx_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs_cast, i32 %lhs_len, <16 x i8> %rhs_cast, i32 %rhs_len, i8 24)
  ret i32 %idx
}

define i32 @pcmpestri_reg_diff_i16(<8 x i16> %lhs, i32 %lhs_len, <8 x i16> %rhs, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_reg_diff_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movl 8(%ebp), %eax
; X86-NEXT:    movl 12(%ebp), %edx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    cmpl $16, %ecx
; X86-NEXT:    jne .LBB8_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    jmp .LBB8_3
; X86-NEXT:  .LBB8_2: # %compare
; X86-NEXT:    movdqa %xmm0, (%esp)
; X86-NEXT:    addl %ecx, %ecx
; X86-NEXT:    andl $14, %ecx
; X86-NEXT:    movzwl (%esp,%ecx), %eax
; X86-NEXT:    movdqa %xmm1, {{[0-9]+}}(%esp)
; X86-NEXT:    subw 16(%esp,%ecx), %ax
; X86-NEXT:  .LBB8_3: # %exit
; X86-NEXT:    movzwl %ax, %eax
; X86-NEXT:    movl %ebp, %esp
; X86-NEXT:    popl %ebp
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_reg_diff_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $16, %ecx
; X64-NEXT:    jne .LBB8_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB8_2: # %compare
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $7, %ecx
; X64-NEXT:    movzwl -24(%rsp,%rcx,2), %eax
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subw -40(%rsp,%rcx,2), %ax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
entry:
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs_cast, i32 %lhs_len, <16 x i8> %rhs_cast, i32 %rhs_len, i8 24)
  %eq = icmp eq i32 %idx, 16
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <8 x i16> %lhs, i32 %idx
  %rhs_c = extractelement <8 x i16> %rhs, i32 %idx
  %sub = sub i16 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i16 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i16 %result to i32
  ret i32 %result_ext
}

define i1 @pcmpestri_mem_eq_i16(i16* %lhs_ptr, i32 %lhs_len, i16* %rhs_ptr, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_mem_eq_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movdqu (%esi), %xmm0
; X86-NEXT:    pcmpestri $25, (%ecx), %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    popl %esi
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_mem_eq_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rdx, %r8
; X64-NEXT:    movl %esi, %eax
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    movl %ecx, %edx
; X64-NEXT:    pcmpestri $25, (%r8), %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i16* %lhs_ptr to <8 x i16>*
  %lhs = load <8 x i16>, <8 x i16>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i16* %rhs_ptr to <8 x i16>*
  %rhs = load <8 x i16>, <8 x i16>* %rhs_vptr, align 1
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %c = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs_cast, i32 %lhs_len, <16 x i8> %rhs_cast, i32 %rhs_len, i8 25)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpestri_mem_idx_i16(i16* %lhs_ptr, i32 %lhs_len, i16* %rhs_ptr, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_mem_idx_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movdqu (%esi), %xmm0
; X86-NEXT:    pcmpestri $25, (%ecx), %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    popl %esi
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_mem_idx_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rdx, %r8
; X64-NEXT:    movl %esi, %eax
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    movl %ecx, %edx
; X64-NEXT:    pcmpestri $25, (%r8), %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i16* %lhs_ptr to <8 x i16>*
  %lhs = load <8 x i16>, <8 x i16>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i16* %rhs_ptr to <8 x i16>*
  %rhs = load <8 x i16>, <8 x i16>* %rhs_vptr, align 1
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs_cast, i32 %lhs_len, <16 x i8> %rhs_cast, i32 %rhs_len, i8 25)
  ret i32 %idx
}

define i32 @pcmpestri_mem_diff_i16(i16* %lhs_ptr, i32 %lhs_len, i16* %rhs_ptr, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_mem_diff_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    pushl %esi
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movl 12(%ebp), %eax
; X86-NEXT:    movl 20(%ebp), %edx
; X86-NEXT:    movl 16(%ebp), %ecx
; X86-NEXT:    movl 8(%ebp), %esi
; X86-NEXT:    movdqu (%esi), %xmm1
; X86-NEXT:    movdqu (%ecx), %xmm0
; X86-NEXT:    pcmpestri $25, %xmm0, %xmm1
; X86-NEXT:    cmpl $8, %ecx
; X86-NEXT:    jne .LBB11_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    jmp .LBB11_3
; X86-NEXT:  .LBB11_2: # %compare
; X86-NEXT:    movdqa %xmm1, (%esp)
; X86-NEXT:    addl %ecx, %ecx
; X86-NEXT:    andl $14, %ecx
; X86-NEXT:    movzwl (%esp,%ecx), %eax
; X86-NEXT:    movdqa %xmm0, {{[0-9]+}}(%esp)
; X86-NEXT:    subw 16(%esp,%ecx), %ax
; X86-NEXT:  .LBB11_3: # %exit
; X86-NEXT:    movzwl %ax, %eax
; X86-NEXT:    leal -4(%ebp), %esp
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %ebp
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_mem_diff_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %esi, %eax
; X64-NEXT:    movdqu (%rdi), %xmm1
; X64-NEXT:    movdqu (%rdx), %xmm0
; X64-NEXT:    movl %ecx, %edx
; X64-NEXT:    pcmpestri $25, %xmm0, %xmm1
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $8, %ecx
; X64-NEXT:    jne .LBB11_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB11_2: # %compare
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $7, %ecx
; X64-NEXT:    movzwl -24(%rsp,%rcx,2), %eax
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subw -40(%rsp,%rcx,2), %ax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i16* %lhs_ptr to <8 x i16>*
  %lhs = load <8 x i16>, <8 x i16>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i16* %rhs_ptr to <8 x i16>*
  %rhs = load <8 x i16>, <8 x i16>* %rhs_vptr, align 1
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs_cast, i32 %lhs_len, <16 x i8> %rhs_cast, i32 %rhs_len, i8 25)
  %eq = icmp eq i32 %idx, 8
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <8 x i16> %lhs, i32 %idx
  %rhs_c = extractelement <8 x i16> %rhs, i32 %idx
  %sub = sub i16 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i16 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i16 %result to i32
  ret i32 %result_ext
}

define i1 @pcmpistri_reg_eq_i8(<16 x i8> %lhs, <16 x i8> %rhs) nounwind {
; X86-LABEL: pcmpistri_reg_eq_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_reg_eq_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %c = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpistri_reg_idx_i8(<16 x i8> %lhs, <16 x i8> %rhs) nounwind {
; X86-LABEL: pcmpistri_reg_idx_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_reg_idx_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  ret i32 %idx
}

define i32 @pcmpistri_reg_diff_i8(<16 x i8> %lhs, <16 x i8> %rhs) nounwind {
; X86-LABEL: pcmpistri_reg_diff_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    cmpl $16, %ecx
; X86-NEXT:    jne .LBB14_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    movzbl %al, %eax
; X86-NEXT:    retl
; X86-NEXT:  .LBB14_2: # %compare
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movdqa %xmm0, (%esp)
; X86-NEXT:    andl $15, %ecx
; X86-NEXT:    movb (%esp,%ecx), %al
; X86-NEXT:    movdqa %xmm1, {{[0-9]+}}(%esp)
; X86-NEXT:    subb 16(%esp,%ecx), %al
; X86-NEXT:    movl %ebp, %esp
; X86-NEXT:    popl %ebp
; X86-NEXT:    movzbl %al, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_reg_diff_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $16, %ecx
; X64-NEXT:    jne .LBB14_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB14_2: # %compare
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $15, %ecx
; X64-NEXT:    movb -24(%rsp,%rcx), %al
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subb -40(%rsp,%rcx), %al
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
entry:
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %eq = icmp eq i32 %idx, 16
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <16 x i8> %lhs, i32 %idx
  %rhs_c = extractelement <16 x i8> %rhs, i32 %idx
  %sub = sub i8 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i8 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i8 %result to i32
  ret i32 %result_ext
}

define i1 @pcmpistri_mem_eq_i8(i8* %lhs_ptr, i8* %rhs_ptr) nounwind {
; X86-LABEL: pcmpistri_mem_eq_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movdqu (%ecx), %xmm0
; X86-NEXT:    pcmpistri $24, (%eax), %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_mem_eq_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    pcmpistri $24, (%rsi), %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i8* %lhs_ptr to <16 x i8>*
  %lhs = load <16 x i8>, <16 x i8>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i8* %rhs_ptr to <16 x i8>*
  %rhs = load <16 x i8>, <16 x i8>* %rhs_vptr, align 1
  %c = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpistri_mem_idx_i8(i8* %lhs_ptr, i8* %rhs_ptr) nounwind {
; X86-LABEL: pcmpistri_mem_idx_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movdqu (%ecx), %xmm0
; X86-NEXT:    pcmpistri $24, (%eax), %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_mem_idx_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    pcmpistri $24, (%rsi), %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i8* %lhs_ptr to <16 x i8>*
  %lhs = load <16 x i8>, <16 x i8>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i8* %rhs_ptr to <16 x i8>*
  %rhs = load <16 x i8>, <16 x i8>* %rhs_vptr, align 1
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  ret i32 %idx
}

define i32 @pcmpistri_mem_diff_i8(i8* %lhs_ptr, i8* %rhs_ptr) nounwind {
; X86-LABEL: pcmpistri_mem_diff_i8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movl 12(%ebp), %eax
; X86-NEXT:    movl 8(%ebp), %ecx
; X86-NEXT:    movdqu (%ecx), %xmm1
; X86-NEXT:    movdqu (%eax), %xmm0
; X86-NEXT:    pcmpistri $24, %xmm0, %xmm1
; X86-NEXT:    cmpl $16, %ecx
; X86-NEXT:    jne .LBB17_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    jmp .LBB17_3
; X86-NEXT:  .LBB17_2: # %compare
; X86-NEXT:    movdqa %xmm1, (%esp)
; X86-NEXT:    andl $15, %ecx
; X86-NEXT:    movb (%esp,%ecx), %al
; X86-NEXT:    movdqa %xmm0, {{[0-9]+}}(%esp)
; X86-NEXT:    subb 16(%esp,%ecx), %al
; X86-NEXT:  .LBB17_3: # %exit
; X86-NEXT:    movzbl %al, %eax
; X86-NEXT:    movl %ebp, %esp
; X86-NEXT:    popl %ebp
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_mem_diff_i8:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movdqu (%rdi), %xmm1
; X64-NEXT:    movdqu (%rsi), %xmm0
; X64-NEXT:    pcmpistri $24, %xmm0, %xmm1
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $16, %ecx
; X64-NEXT:    jne .LBB17_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB17_2: # %compare
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $15, %ecx
; X64-NEXT:    movb -24(%rsp,%rcx), %al
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subb -40(%rsp,%rcx), %al
; X64-NEXT:    movzbl %al, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i8* %lhs_ptr to <16 x i8>*
  %lhs = load <16 x i8>, <16 x i8>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i8* %rhs_ptr to <16 x i8>*
  %rhs = load <16 x i8>, <16 x i8>* %rhs_vptr, align 1
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %eq = icmp eq i32 %idx, 16
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <16 x i8> %lhs, i32 %idx
  %rhs_c = extractelement <16 x i8> %rhs, i32 %idx
  %sub = sub i8 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i8 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i8 %result to i32
  ret i32 %result_ext
}

define i1 @pcmpistri_reg_eq_i16(<8 x i16> %lhs, <8 x i16> %rhs) nounwind {
; X86-LABEL: pcmpistri_reg_eq_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_reg_eq_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %c = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs_cast, <16 x i8> %rhs_cast, i8 24)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpistri_reg_idx_i16(<8 x i16> %lhs, <8 x i16> %rhs) nounwind {
; X86-LABEL: pcmpistri_reg_idx_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_reg_idx_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs_cast, <16 x i8> %rhs_cast, i8 24)
  ret i32 %idx
}

define i32 @pcmpistri_reg_diff_i16(<8 x i16> %lhs, <8 x i16> %rhs) nounwind {
; X86-LABEL: pcmpistri_reg_diff_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    cmpl $16, %ecx
; X86-NEXT:    jne .LBB20_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    movzwl %ax, %eax
; X86-NEXT:    retl
; X86-NEXT:  .LBB20_2: # %compare
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movdqa %xmm0, (%esp)
; X86-NEXT:    addl %ecx, %ecx
; X86-NEXT:    andl $14, %ecx
; X86-NEXT:    movzwl (%esp,%ecx), %eax
; X86-NEXT:    movdqa %xmm1, {{[0-9]+}}(%esp)
; X86-NEXT:    subw 16(%esp,%ecx), %ax
; X86-NEXT:    movl %ebp, %esp
; X86-NEXT:    popl %ebp
; X86-NEXT:    movzwl %ax, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_reg_diff_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $16, %ecx
; X64-NEXT:    jne .LBB20_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB20_2: # %compare
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $7, %ecx
; X64-NEXT:    movzwl -24(%rsp,%rcx,2), %eax
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subw -40(%rsp,%rcx,2), %ax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
entry:
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs_cast, <16 x i8> %rhs_cast, i8 24)
  %eq = icmp eq i32 %idx, 16
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <8 x i16> %lhs, i32 %idx
  %rhs_c = extractelement <8 x i16> %rhs, i32 %idx
  %sub = sub i16 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i16 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i16 %result to i32
  ret i32 %result_ext
}

define i1 @pcmpistri_mem_eq_i16(i16* %lhs_ptr, i16* %rhs_ptr) nounwind {
; X86-LABEL: pcmpistri_mem_eq_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movdqu (%ecx), %xmm0
; X86-NEXT:    pcmpistri $25, (%eax), %xmm0
; X86-NEXT:    setae %al
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_mem_eq_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    pcmpistri $25, (%rsi), %xmm0
; X64-NEXT:    setae %al
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i16* %lhs_ptr to <8 x i16>*
  %lhs = load <8 x i16>, <8 x i16>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i16* %rhs_ptr to <8 x i16>*
  %rhs = load <8 x i16>, <8 x i16>* %rhs_vptr, align 1
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %c = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs_cast, <16 x i8> %rhs_cast, i8 25)
  %result = icmp eq i32 %c, 0
  ret i1 %result
}

define i32 @pcmpistri_mem_idx_i16(i16* %lhs_ptr, i16* %rhs_ptr) nounwind {
; X86-LABEL: pcmpistri_mem_idx_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movdqu (%ecx), %xmm0
; X86-NEXT:    pcmpistri $25, (%eax), %xmm0
; X86-NEXT:    movl %ecx, %eax
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_mem_idx_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movdqu (%rdi), %xmm0
; X64-NEXT:    pcmpistri $25, (%rsi), %xmm0
; X64-NEXT:    movl %ecx, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i16* %lhs_ptr to <8 x i16>*
  %lhs = load <8 x i16>, <8 x i16>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i16* %rhs_ptr to <8 x i16>*
  %rhs = load <8 x i16>, <8 x i16>* %rhs_vptr, align 1
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs_cast, <16 x i8> %rhs_cast, i8 25)
  ret i32 %idx
}

define i32 @pcmpistri_mem_diff_i16(i16* %lhs_ptr, i16* %rhs_ptr) nounwind {
; X86-LABEL: pcmpistri_mem_diff_i16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebp
; X86-NEXT:    movl %esp, %ebp
; X86-NEXT:    andl $-16, %esp
; X86-NEXT:    subl $48, %esp
; X86-NEXT:    movl 12(%ebp), %eax
; X86-NEXT:    movl 8(%ebp), %ecx
; X86-NEXT:    movdqu (%ecx), %xmm1
; X86-NEXT:    movdqu (%eax), %xmm0
; X86-NEXT:    pcmpistri $25, %xmm0, %xmm1
; X86-NEXT:    cmpl $8, %ecx
; X86-NEXT:    jne .LBB23_2
; X86-NEXT:  # %bb.1:
; X86-NEXT:    xorl %eax, %eax
; X86-NEXT:    jmp .LBB23_3
; X86-NEXT:  .LBB23_2: # %compare
; X86-NEXT:    movdqa %xmm1, (%esp)
; X86-NEXT:    addl %ecx, %ecx
; X86-NEXT:    andl $14, %ecx
; X86-NEXT:    movzwl (%esp,%ecx), %eax
; X86-NEXT:    movdqa %xmm0, {{[0-9]+}}(%esp)
; X86-NEXT:    subw 16(%esp,%ecx), %ax
; X86-NEXT:  .LBB23_3: # %exit
; X86-NEXT:    movzwl %ax, %eax
; X86-NEXT:    movl %ebp, %esp
; X86-NEXT:    popl %ebp
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistri_mem_diff_i16:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movdqu (%rdi), %xmm1
; X64-NEXT:    movdqu (%rsi), %xmm0
; X64-NEXT:    pcmpistri $25, %xmm0, %xmm1
; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
; X64-NEXT:    cmpl $8, %ecx
; X64-NEXT:    jne .LBB23_2
; X64-NEXT:  # %bb.1:
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
; X64-NEXT:  .LBB23_2: # %compare
; X64-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
; X64-NEXT:    andl $7, %ecx
; X64-NEXT:    movzwl -24(%rsp,%rcx,2), %eax
; X64-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
; X64-NEXT:    subw -40(%rsp,%rcx,2), %ax
; X64-NEXT:    movzwl %ax, %eax
; X64-NEXT:    retq
entry:
  %lhs_vptr = bitcast i16* %lhs_ptr to <8 x i16>*
  %lhs = load <8 x i16>, <8 x i16>* %lhs_vptr, align 1
  %rhs_vptr = bitcast i16* %rhs_ptr to <8 x i16>*
  %rhs = load <8 x i16>, <8 x i16>* %rhs_vptr, align 1
  %lhs_cast = bitcast <8 x i16> %lhs to <16 x i8>
  %rhs_cast = bitcast <8 x i16> %rhs to <16 x i8>
  %idx = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs_cast, <16 x i8> %rhs_cast, i8 25)
  %eq = icmp eq i32 %idx, 8
  br i1 %eq, label %exit, label %compare

compare:
  %lhs_c = extractelement <8 x i16> %lhs, i32 %idx
  %rhs_c = extractelement <8 x i16> %rhs, i32 %idx
  %sub = sub i16 %lhs_c, %rhs_c
  br label %exit

exit:
  %result = phi i16 [ 0, %entry ], [ %sub, %compare ]
  %result_ext = zext i16 %result to i32
  ret i32 %result_ext
}

define void @pcmpestr_index_flag(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i32* %iptr, i32* %fptr) nounwind {
; X86-LABEL: pcmpestr_index_flag:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebx
; X86-NEXT:    pushl %edi
; X86-NEXT:    pushl %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    xorl %ebx, %ebx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    setb %bl
; X86-NEXT:    movl %ecx, (%edi)
; X86-NEXT:    movl %ebx, (%esi)
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %edi
; X86-NEXT:    popl %ebx
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestr_index_flag:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rcx, %r8
; X64-NEXT:    movq %rdx, %r9
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    xorl %esi, %esi
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    setb %sil
; X64-NEXT:    movl %ecx, (%r9)
; X64-NEXT:    movl %esi, (%r8)
; X64-NEXT:    retq
entry:
  %flag = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %index = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  store i32 %index, i32* %iptr
  store i32 %flag, i32* %fptr
  ret void
}

define void @pcmpestr_mask_flag(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, <16 x i8>* %mptr, i32* %fptr) nounwind {
; X86-LABEL: pcmpestr_mask_flag:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebx
; X86-NEXT:    pushl %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    xorl %ebx, %ebx
; X86-NEXT:    pcmpestrm $24, %xmm1, %xmm0
; X86-NEXT:    setb %bl
; X86-NEXT:    movdqa %xmm0, (%esi)
; X86-NEXT:    movl %ebx, (%ecx)
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %ebx
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestr_mask_flag:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rdx, %r8
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    xorl %esi, %esi
; X64-NEXT:    pcmpestrm $24, %xmm1, %xmm0
; X64-NEXT:    setb %sil
; X64-NEXT:    movdqa %xmm0, (%r8)
; X64-NEXT:    movl %esi, (%rcx)
; X64-NEXT:    retq
entry:
  %flag = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %mask = call <16 x i8> @llvm.x86.sse42.pcmpestrm128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  store <16 x i8> %mask, <16 x i8>* %mptr
  store i32 %flag, i32* %fptr
  ret void
}

define void @pcmpestr_mask_index(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, <16 x i8>* %mptr, i32* %iptr) nounwind {
; X86-LABEL: pcmpestr_mask_index:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %edi
; X86-NEXT:    pushl %esi
; X86-NEXT:    movdqa %xmm0, %xmm2
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    pcmpestrm $24, %xmm1, %xmm0
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm2
; X86-NEXT:    movdqa %xmm0, (%edi)
; X86-NEXT:    movl %ecx, (%esi)
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %edi
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestr_mask_index:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rcx, %r8
; X64-NEXT:    movq %rdx, %r9
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    movdqa %xmm0, %xmm2
; X64-NEXT:    pcmpestrm $24, %xmm1, %xmm0
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm2
; X64-NEXT:    movdqa %xmm0, (%r9)
; X64-NEXT:    movl %ecx, (%r8)
; X64-NEXT:    retq
entry:
  %index = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %mask = call <16 x i8> @llvm.x86.sse42.pcmpestrm128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  store <16 x i8> %mask, <16 x i8>* %mptr
  store i32 %index, i32* %iptr
  ret void
}

define void @pcmpestr_mask_index_flag(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, <16 x i8>* %mptr, i32* %iptr, i32* %fptr) nounwind {
; X86-LABEL: pcmpestr_mask_index_flag:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebp
; X86-NEXT:    pushl %ebx
; X86-NEXT:    pushl %edi
; X86-NEXT:    pushl %esi
; X86-NEXT:    movdqa %xmm0, %xmm2
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    pcmpestrm $24, %xmm1, %xmm0
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebp
; X86-NEXT:    xorl %ebx, %ebx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm2
; X86-NEXT:    setb %bl
; X86-NEXT:    movdqa %xmm0, (%ebp)
; X86-NEXT:    movl %ecx, (%edi)
; X86-NEXT:    movl %ebx, (%esi)
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %edi
; X86-NEXT:    popl %ebx
; X86-NEXT:    popl %ebp
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestr_mask_index_flag:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rcx, %r9
; X64-NEXT:    movq %rdx, %r10
; X64-NEXT:    movl %esi, %edx
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    movdqa %xmm0, %xmm2
; X64-NEXT:    pcmpestrm $24, %xmm1, %xmm0
; X64-NEXT:    xorl %esi, %esi
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm2
; X64-NEXT:    setb %sil
; X64-NEXT:    movdqa %xmm0, (%r10)
; X64-NEXT:    movl %ecx, (%r9)
; X64-NEXT:    movl %esi, (%r8)
; X64-NEXT:    retq
entry:
  %index = call i32 @llvm.x86.sse42.pcmpestri128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %mask = call <16 x i8> @llvm.x86.sse42.pcmpestrm128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  %flag = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  store <16 x i8> %mask, <16 x i8>* %mptr
  store i32 %index, i32* %iptr
  store i32 %flag, i32* %fptr
  ret void
}

define void @pcmpistr_index_flag(<16 x i8> %lhs, <16 x i8> %rhs, i32* %iptr, i32* %fptr) nounwind {
; X86-LABEL: pcmpistr_index_flag:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    xorl %ebx, %ebx
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    setb %bl
; X86-NEXT:    movl %ecx, (%edx)
; X86-NEXT:    movl %ebx, (%eax)
; X86-NEXT:    popl %ebx
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistr_index_flag:
; X64:       # %bb.0: # %entry
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    setb %al
; X64-NEXT:    movl %ecx, (%rdi)
; X64-NEXT:    movl %eax, (%rsi)
; X64-NEXT:    retq
entry:
  %flag = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %index = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  store i32 %index, i32* %iptr
  store i32 %flag, i32* %fptr
  ret void
}

define void @pcmpistr_mask_flag(<16 x i8> %lhs, <16 x i8> %rhs, <16 x i8>* %mptr, i32* %fptr) nounwind {
; X86-LABEL: pcmpistr_mask_flag:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    xorl %edx, %edx
; X86-NEXT:    pcmpistrm $24, %xmm1, %xmm0
; X86-NEXT:    setb %dl
; X86-NEXT:    movdqa %xmm0, (%ecx)
; X86-NEXT:    movl %edx, (%eax)
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistr_mask_flag:
; X64:       # %bb.0: # %entry
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    pcmpistrm $24, %xmm1, %xmm0
; X64-NEXT:    setb %al
; X64-NEXT:    movdqa %xmm0, (%rdi)
; X64-NEXT:    movl %eax, (%rsi)
; X64-NEXT:    retq
entry:
  %flag = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %mask = call <16 x i8> @llvm.x86.sse42.pcmpistrm128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  store <16 x i8> %mask, <16 x i8>* %mptr
  store i32 %flag, i32* %fptr
  ret void
}

define void @pcmpistr_mask_index(<16 x i8> %lhs, <16 x i8> %rhs, <16 x i8>* %mptr, i32* %iptr) nounwind {
; X86-LABEL: pcmpistr_mask_index:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X86-NEXT:    pcmpistrm $24, %xmm1, %xmm0
; X86-NEXT:    movdqa %xmm0, (%edx)
; X86-NEXT:    movl %ecx, (%eax)
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistr_mask_index:
; X64:       # %bb.0: # %entry
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm0
; X64-NEXT:    pcmpistrm $24, %xmm1, %xmm0
; X64-NEXT:    movdqa %xmm0, (%rdi)
; X64-NEXT:    movl %ecx, (%rsi)
; X64-NEXT:    retq
entry:
  %index = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %mask = call <16 x i8> @llvm.x86.sse42.pcmpistrm128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  store <16 x i8> %mask, <16 x i8>* %mptr
  store i32 %index, i32* %iptr
  ret void
}

define void @pcmpistr_mask_index_flag(<16 x i8> %lhs, <16 x i8> %rhs, <16 x i8>* %mptr, i32* %iptr, i32* %fptr) nounwind {
; X86-LABEL: pcmpistr_mask_index_flag:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebx
; X86-NEXT:    pushl %esi
; X86-NEXT:    movdqa %xmm0, %xmm2
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    pcmpistrm $24, %xmm1, %xmm0
; X86-NEXT:    xorl %ebx, %ebx
; X86-NEXT:    pcmpistri $24, %xmm1, %xmm2
; X86-NEXT:    setb %bl
; X86-NEXT:    movdqa %xmm0, (%esi)
; X86-NEXT:    movl %ecx, (%edx)
; X86-NEXT:    movl %ebx, (%eax)
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %ebx
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistr_mask_index_flag:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movdqa %xmm0, %xmm2
; X64-NEXT:    pcmpistrm $24, %xmm1, %xmm0
; X64-NEXT:    xorl %eax, %eax
; X64-NEXT:    pcmpistri $24, %xmm1, %xmm2
; X64-NEXT:    setb %al
; X64-NEXT:    movdqa %xmm0, (%rdi)
; X64-NEXT:    movl %ecx, (%rsi)
; X64-NEXT:    movl %eax, (%rdx)
; X64-NEXT:    retq
entry:
  %index = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %mask = call <16 x i8> @llvm.x86.sse42.pcmpistrm128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %flag = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  store <16 x i8> %mask, <16 x i8>* %mptr
  store i32 %index, i32* %iptr
  store i32 %flag, i32* %fptr
  ret void
}

; Make sure we don't fold loads when we need to emit pcmpistrm and pcmpistri.
define void @pcmpistr_mask_index_flag_load(<16 x i8> %lhs, <16 x i8>* %rhsptr, <16 x i8>* %mptr, i32* %iptr, i32* %fptr) nounwind {
; X86-LABEL: pcmpistr_mask_index_flag_load:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebx
; X86-NEXT:    pushl %esi
; X86-NEXT:    movdqa %xmm0, %xmm1
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movdqu (%ecx), %xmm2
; X86-NEXT:    pcmpistrm $24, %xmm2, %xmm0
; X86-NEXT:    xorl %ebx, %ebx
; X86-NEXT:    pcmpistri $24, %xmm2, %xmm1
; X86-NEXT:    setb %bl
; X86-NEXT:    movdqa %xmm0, (%esi)
; X86-NEXT:    movl %ecx, (%edx)
; X86-NEXT:    movl %ebx, (%eax)
; X86-NEXT:    popl %esi
; X86-NEXT:    popl %ebx
; X86-NEXT:    retl
;
; X64-LABEL: pcmpistr_mask_index_flag_load:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movq %rcx, %rax
; X64-NEXT:    movdqa %xmm0, %xmm1
; X64-NEXT:    movdqu (%rdi), %xmm2
; X64-NEXT:    pcmpistrm $24, %xmm2, %xmm0
; X64-NEXT:    xorl %edi, %edi
; X64-NEXT:    pcmpistri $24, %xmm2, %xmm1
; X64-NEXT:    setb %dil
; X64-NEXT:    movdqa %xmm0, (%rsi)
; X64-NEXT:    movl %ecx, (%rdx)
; X64-NEXT:    movl %edi, (%rax)
; X64-NEXT:    retq
entry:
  %rhs = load <16 x i8>, <16 x i8>* %rhsptr, align 1
  %index = call i32 @llvm.x86.sse42.pcmpistri128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %mask = call <16 x i8> @llvm.x86.sse42.pcmpistrm128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  %flag = call i32 @llvm.x86.sse42.pcmpistric128(<16 x i8> %lhs, <16 x i8> %rhs, i8 24)
  store <16 x i8> %mask, <16 x i8>* %mptr
  store i32 %index, i32* %iptr
  store i32 %flag, i32* %fptr
  ret void
}

; Make sure we don't fold nontemporal loads.
define i32 @pcmpestri_nontemporal(<16 x i8> %lhs, i32 %lhs_len, <16 x i8>* %rhsptr, i32 %rhs_len) nounwind {
; X86-LABEL: pcmpestri_nontemporal:
; X86:       # %bb.0: # %entry
; X86-NEXT:    pushl %ebx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    movntdqa (%ecx), %xmm1
; X86-NEXT:    xorl %ebx, %ebx
; X86-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X86-NEXT:    setb %bl
; X86-NEXT:    movl %ebx, %eax
; X86-NEXT:    popl %ebx
; X86-NEXT:    retl
;
; X64-LABEL: pcmpestri_nontemporal:
; X64:       # %bb.0: # %entry
; X64-NEXT:    movl %edi, %eax
; X64-NEXT:    movntdqa (%rsi), %xmm1
; X64-NEXT:    xorl %esi, %esi
; X64-NEXT:    pcmpestri $24, %xmm1, %xmm0
; X64-NEXT:    setb %sil
; X64-NEXT:    movl %esi, %eax
; X64-NEXT:    retq
entry:
  %rhs = load <16 x i8>, <16 x i8>* %rhsptr, align 16, !nontemporal !0
  %flag = call i32 @llvm.x86.sse42.pcmpestric128(<16 x i8> %lhs, i32 %lhs_len, <16 x i8> %rhs, i32 %rhs_len, i8 24)
  ret i32 %flag
}

!0 = !{ i32 1 }