; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=i686-unknown-linux | FileCheck %s --check-prefix=X86
; RUN: llc < %s -mtriple=x86_64-unknown-linux -mcpu=core2 | FileCheck %s --check-prefix=CORE2
; RUN: llc < %s -mtriple=x86_64-unknown-linux -mcpu=nehalem | FileCheck %s --check-prefix=NEHALEM
; RUN: llc < %s -mtriple=x86_64-unknown-linux -mcpu=sandybridge | FileCheck %s --check-prefix=SANDYBRIDGE
; RUN: llc < %s -mtriple=x86_64-unknown-linux -mcpu=skylake | FileCheck %s --check-prefix=SKYLAKE
; RUN: llc < %s -mtriple=x86_64-unknown-linux -mcpu=knl | FileCheck %s --check-prefix=KNL


declare void @llvm.memset.p0.i64(ptr nocapture, i8, i64, i1) nounwind

define void @memset_0(ptr %a) nounwind  {
; X86-LABEL: memset_0:
; X86:       # %bb.0: # %entry
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_0:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_0:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_0:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_0:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_0:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 0, i1 false)
	ret void
}

define void @memset_4(ptr %a) nounwind  {
; X86-LABEL: memset_4:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_4:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movl $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_4:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    movl $0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_4:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    movl $0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_4:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    movl $0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_4:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    movl $0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 4, i1 false)
	ret void
}

define void @memset_5(ptr %a) nounwind  {
; X86-LABEL: memset_5:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movb $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_5:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movb $0, 4(%rdi)
; CORE2-NEXT:    movl $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_5:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    movb $0, 4(%rdi)
; NEHALEM-NEXT:    movl $0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_5:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    movb $0, 4(%rdi)
; SANDYBRIDGE-NEXT:    movl $0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_5:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    movb $0, 4(%rdi)
; SKYLAKE-NEXT:    movl $0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_5:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    movb $0, 4(%rdi)
; KNL-NEXT:    movl $0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 5, i1 false)
	ret void
}

define void @memset_7(ptr %a) nounwind  {
; X86-LABEL: memset_7:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 3(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_7:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movl $0, 3(%rdi)
; CORE2-NEXT:    movl $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_7:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    movl $0, 3(%rdi)
; NEHALEM-NEXT:    movl $0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_7:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    movl $0, 3(%rdi)
; SANDYBRIDGE-NEXT:    movl $0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_7:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    movl $0, 3(%rdi)
; SKYLAKE-NEXT:    movl $0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_7:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    movl $0, 3(%rdi)
; KNL-NEXT:    movl $0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 7, i1 false)
	ret void
}

define void @memset_8(ptr %a) nounwind  {
; X86-LABEL: memset_8:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_8:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_8:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    movq $0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_8:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    movq $0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_8:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    movq $0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_8:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    movq $0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 8, i1 false)
	ret void
}

define void @memset_11(ptr %a) nounwind  {
; X86-LABEL: memset_11:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 7(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_11:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movl $0, 7(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_11:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    movl $0, 7(%rdi)
; NEHALEM-NEXT:    movq $0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_11:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    movl $0, 7(%rdi)
; SANDYBRIDGE-NEXT:    movq $0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_11:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    movl $0, 7(%rdi)
; SKYLAKE-NEXT:    movq $0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_11:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    movl $0, 7(%rdi)
; KNL-NEXT:    movq $0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 11, i1 false)
	ret void
}

define void @memset_13(ptr %a) nounwind  {
; X86-LABEL: memset_13:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movb $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_13:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movq $0, 5(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_13:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    movq $0, 5(%rdi)
; NEHALEM-NEXT:    movq $0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_13:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    movq $0, 5(%rdi)
; SANDYBRIDGE-NEXT:    movq $0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_13:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    movq $0, 5(%rdi)
; SKYLAKE-NEXT:    movq $0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_13:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    movq $0, 5(%rdi)
; KNL-NEXT:    movq $0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 13, i1 false)
	ret void
}

define void @memset_15(ptr %a) nounwind  {
; X86-LABEL: memset_15:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 11(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_15:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movq $0, 7(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_15:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    movq $0, 7(%rdi)
; NEHALEM-NEXT:    movq $0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_15:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    movq $0, 7(%rdi)
; SANDYBRIDGE-NEXT:    movq $0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_15:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    movq $0, 7(%rdi)
; SKYLAKE-NEXT:    movq $0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_15:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    movq $0, 7(%rdi)
; KNL-NEXT:    movq $0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 15, i1 false)
	ret void
}

define void @memset_16(ptr %a) nounwind  {
; X86-LABEL: memset_16:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_16:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movq $0, 8(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_16:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movups %xmm0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_16:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovups %xmm0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_16:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovups %xmm0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_16:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovups %xmm0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 16, i1 false)
	ret void
}

define void @memset_17(ptr %a) nounwind  {
; X86-LABEL: memset_17:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movb $0, 16(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_17:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movb $0, 16(%rdi)
; CORE2-NEXT:    movq $0, 8(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_17:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movups %xmm0, (%rdi)
; NEHALEM-NEXT:    movb $0, 16(%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_17:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovups %xmm0, (%rdi)
; SANDYBRIDGE-NEXT:    movb $0, 16(%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_17:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovups %xmm0, (%rdi)
; SKYLAKE-NEXT:    movb $0, 16(%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_17:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovups %xmm0, (%rdi)
; KNL-NEXT:    movb $0, 16(%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 17, i1 false)
	ret void
}

define void @memset_19(ptr %a) nounwind  {
; X86-LABEL: memset_19:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 15(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_19:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movl $0, 15(%rdi)
; CORE2-NEXT:    movq $0, 8(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_19:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movups %xmm0, (%rdi)
; NEHALEM-NEXT:    movl $0, 15(%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_19:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovups %xmm0, (%rdi)
; SANDYBRIDGE-NEXT:    movl $0, 15(%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_19:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovups %xmm0, (%rdi)
; SKYLAKE-NEXT:    movl $0, 15(%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_19:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovups %xmm0, (%rdi)
; KNL-NEXT:    movl $0, 15(%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 19, i1 false)
	ret void
}

define void @memset_31(ptr %a) nounwind  {
; X86-LABEL: memset_31:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 27(%eax)
; X86-NEXT:    movl $0, 24(%eax)
; X86-NEXT:    movl $0, 20(%eax)
; X86-NEXT:    movl $0, 16(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_31:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movq $0, 23(%rdi)
; CORE2-NEXT:    movq $0, 16(%rdi)
; CORE2-NEXT:    movq $0, 8(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_31:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movups %xmm0, 15(%rdi)
; NEHALEM-NEXT:    movups %xmm0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_31:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovups %xmm0, 15(%rdi)
; SANDYBRIDGE-NEXT:    vmovups %xmm0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_31:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovups %xmm0, 15(%rdi)
; SKYLAKE-NEXT:    vmovups %xmm0, (%rdi)
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_31:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovups %xmm0, 15(%rdi)
; KNL-NEXT:    vmovups %xmm0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 31, i1 false)
	ret void
}

define void @memset_32(ptr %a) nounwind  {
; X86-LABEL: memset_32:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 28(%eax)
; X86-NEXT:    movl $0, 24(%eax)
; X86-NEXT:    movl $0, 20(%eax)
; X86-NEXT:    movl $0, 16(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_32:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movq $0, 24(%rdi)
; CORE2-NEXT:    movq $0, 16(%rdi)
; CORE2-NEXT:    movq $0, 8(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_32:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movups %xmm0, 16(%rdi)
; NEHALEM-NEXT:    movups %xmm0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_32:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovups %xmm0, 16(%rdi)
; SANDYBRIDGE-NEXT:    vmovups %xmm0, (%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_32:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovups %ymm0, (%rdi)
; SKYLAKE-NEXT:    vzeroupper
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_32:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovups %ymm0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 32, i1 false)
	ret void
}

define void @memset_32_align32(ptr %a) nounwind  {
; X86-LABEL: memset_32_align32:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 28(%eax)
; X86-NEXT:    movl $0, 24(%eax)
; X86-NEXT:    movl $0, 20(%eax)
; X86-NEXT:    movl $0, 16(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_32_align32:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    xorps %xmm0, %xmm0
; CORE2-NEXT:    movaps %xmm0, 16(%rdi)
; CORE2-NEXT:    movaps %xmm0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_32_align32:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movaps %xmm0, 16(%rdi)
; NEHALEM-NEXT:    movaps %xmm0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_32_align32:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovaps %ymm0, (%rdi)
; SANDYBRIDGE-NEXT:    vzeroupper
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_32_align32:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovaps %ymm0, (%rdi)
; SKYLAKE-NEXT:    vzeroupper
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_32_align32:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovaps %ymm0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr align 32 %a, i8 0, i64 32, i1 false)
	ret void
}


define void @memset_35(ptr %a) nounwind  {
; X86-LABEL: memset_35:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 31(%eax)
; X86-NEXT:    movl $0, 28(%eax)
; X86-NEXT:    movl $0, 24(%eax)
; X86-NEXT:    movl $0, 20(%eax)
; X86-NEXT:    movl $0, 16(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_35:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movl $0, 31(%rdi)
; CORE2-NEXT:    movq $0, 24(%rdi)
; CORE2-NEXT:    movq $0, 16(%rdi)
; CORE2-NEXT:    movq $0, 8(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_35:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movups %xmm0, 16(%rdi)
; NEHALEM-NEXT:    movups %xmm0, (%rdi)
; NEHALEM-NEXT:    movl $0, 31(%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_35:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovups %xmm0, 16(%rdi)
; SANDYBRIDGE-NEXT:    vmovups %xmm0, (%rdi)
; SANDYBRIDGE-NEXT:    movl $0, 31(%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_35:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovups %ymm0, (%rdi)
; SKYLAKE-NEXT:    movl $0, 31(%rdi)
; SKYLAKE-NEXT:    vzeroupper
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_35:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovups %ymm0, (%rdi)
; KNL-NEXT:    movl $0, 31(%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 35, i1 false)
	ret void
}

define void @memset_64(ptr %a) nounwind  {
; X86-LABEL: memset_64:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 60(%eax)
; X86-NEXT:    movl $0, 56(%eax)
; X86-NEXT:    movl $0, 52(%eax)
; X86-NEXT:    movl $0, 48(%eax)
; X86-NEXT:    movl $0, 44(%eax)
; X86-NEXT:    movl $0, 40(%eax)
; X86-NEXT:    movl $0, 36(%eax)
; X86-NEXT:    movl $0, 32(%eax)
; X86-NEXT:    movl $0, 28(%eax)
; X86-NEXT:    movl $0, 24(%eax)
; X86-NEXT:    movl $0, 20(%eax)
; X86-NEXT:    movl $0, 16(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_64:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    movq $0, 56(%rdi)
; CORE2-NEXT:    movq $0, 48(%rdi)
; CORE2-NEXT:    movq $0, 40(%rdi)
; CORE2-NEXT:    movq $0, 32(%rdi)
; CORE2-NEXT:    movq $0, 24(%rdi)
; CORE2-NEXT:    movq $0, 16(%rdi)
; CORE2-NEXT:    movq $0, 8(%rdi)
; CORE2-NEXT:    movq $0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_64:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movups %xmm0, 48(%rdi)
; NEHALEM-NEXT:    movups %xmm0, 32(%rdi)
; NEHALEM-NEXT:    movups %xmm0, 16(%rdi)
; NEHALEM-NEXT:    movups %xmm0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_64:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovups %xmm0, 16(%rdi)
; SANDYBRIDGE-NEXT:    vmovups %xmm0, (%rdi)
; SANDYBRIDGE-NEXT:    vmovups %xmm0, 48(%rdi)
; SANDYBRIDGE-NEXT:    vmovups %xmm0, 32(%rdi)
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_64:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovups %ymm0, 32(%rdi)
; SKYLAKE-NEXT:    vmovups %ymm0, (%rdi)
; SKYLAKE-NEXT:    vzeroupper
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_64:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovups %zmm0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr %a, i8 0, i64 64, i1 false)
	ret void
}

define void @memset_64_align64(ptr %a) nounwind  {
; X86-LABEL: memset_64_align64:
; X86:       # %bb.0: # %entry
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl $0, 60(%eax)
; X86-NEXT:    movl $0, 56(%eax)
; X86-NEXT:    movl $0, 52(%eax)
; X86-NEXT:    movl $0, 48(%eax)
; X86-NEXT:    movl $0, 44(%eax)
; X86-NEXT:    movl $0, 40(%eax)
; X86-NEXT:    movl $0, 36(%eax)
; X86-NEXT:    movl $0, 32(%eax)
; X86-NEXT:    movl $0, 28(%eax)
; X86-NEXT:    movl $0, 24(%eax)
; X86-NEXT:    movl $0, 20(%eax)
; X86-NEXT:    movl $0, 16(%eax)
; X86-NEXT:    movl $0, 12(%eax)
; X86-NEXT:    movl $0, 8(%eax)
; X86-NEXT:    movl $0, 4(%eax)
; X86-NEXT:    movl $0, (%eax)
; X86-NEXT:    retl
;
; CORE2-LABEL: memset_64_align64:
; CORE2:       # %bb.0: # %entry
; CORE2-NEXT:    xorps %xmm0, %xmm0
; CORE2-NEXT:    movaps %xmm0, 48(%rdi)
; CORE2-NEXT:    movaps %xmm0, 32(%rdi)
; CORE2-NEXT:    movaps %xmm0, 16(%rdi)
; CORE2-NEXT:    movaps %xmm0, (%rdi)
; CORE2-NEXT:    retq
;
; NEHALEM-LABEL: memset_64_align64:
; NEHALEM:       # %bb.0: # %entry
; NEHALEM-NEXT:    xorps %xmm0, %xmm0
; NEHALEM-NEXT:    movaps %xmm0, 48(%rdi)
; NEHALEM-NEXT:    movaps %xmm0, 32(%rdi)
; NEHALEM-NEXT:    movaps %xmm0, 16(%rdi)
; NEHALEM-NEXT:    movaps %xmm0, (%rdi)
; NEHALEM-NEXT:    retq
;
; SANDYBRIDGE-LABEL: memset_64_align64:
; SANDYBRIDGE:       # %bb.0: # %entry
; SANDYBRIDGE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SANDYBRIDGE-NEXT:    vmovaps %ymm0, 32(%rdi)
; SANDYBRIDGE-NEXT:    vmovaps %ymm0, (%rdi)
; SANDYBRIDGE-NEXT:    vzeroupper
; SANDYBRIDGE-NEXT:    retq
;
; SKYLAKE-LABEL: memset_64_align64:
; SKYLAKE:       # %bb.0: # %entry
; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; SKYLAKE-NEXT:    vmovaps %ymm0, 32(%rdi)
; SKYLAKE-NEXT:    vmovaps %ymm0, (%rdi)
; SKYLAKE-NEXT:    vzeroupper
; SKYLAKE-NEXT:    retq
;
; KNL-LABEL: memset_64_align64:
; KNL:       # %bb.0: # %entry
; KNL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; KNL-NEXT:    vmovaps %zmm0, (%rdi)
; KNL-NEXT:    retq
entry:
	call void @llvm.memset.p0.i64(ptr align 64 %a, i8 0, i64 64, i1 false)
	ret void
}