280 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mattr=+sse2 -mtriple=i686-apple-darwin -mcpu=core2 | FileCheck %s -check-prefix=SSE2-Darwin3; RUN: llc < %s -mattr=+sse2 -mtriple=i686-pc-mingw32 -mcpu=core2 | FileCheck %s -check-prefix=SSE2-Mingw324; RUN: llc < %s -mattr=+sse,-sse2 -mtriple=i686-apple-darwin -mcpu=core2 | FileCheck %s -check-prefix=SSE15; RUN: llc < %s -mattr=-sse -mtriple=i686-apple-darwin -mcpu=core2 | FileCheck %s -check-prefix=NOSSE6; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=core2 | FileCheck %s -check-prefix=X86-647; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=nehalem | FileCheck %s -check-prefix=NHM_648 9@.str = internal constant [25 x i8] c"image\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00"10@.str2 = internal constant [30 x i8] c"xxxxxxxxxxxxxxxxxxxxxxxxxxxxx\00", align 411 12define void @t1(i32 %argc, ptr %argv) nounwind {13; SSE2-Darwin-LABEL: t1:14; SSE2-Darwin: ## %bb.0: ## %entry15; SSE2-Darwin-NEXT: subl $28, %esp16; SSE2-Darwin-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero17; SSE2-Darwin-NEXT: movsd %xmm0, {{[0-9]+}}(%esp)18; SSE2-Darwin-NEXT: movaps _.str, %xmm019; SSE2-Darwin-NEXT: movaps %xmm0, (%esp)20; SSE2-Darwin-NEXT: movb $0, {{[0-9]+}}(%esp)21; SSE2-Darwin-NEXT: ud222;23; SSE2-Mingw32-LABEL: t1:24; SSE2-Mingw32: # %bb.0: # %entry25; SSE2-Mingw32-NEXT: subl $28, %esp26; SSE2-Mingw32-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero27; SSE2-Mingw32-NEXT: movsd %xmm0, {{[0-9]+}}(%esp)28; SSE2-Mingw32-NEXT: movaps _.str, %xmm029; SSE2-Mingw32-NEXT: movups %xmm0, (%esp)30; SSE2-Mingw32-NEXT: movb $0, {{[0-9]+}}(%esp)31;32; SSE1-LABEL: t1:33; SSE1: ## %bb.0: ## %entry34; SSE1-NEXT: subl $28, %esp35; SSE1-NEXT: movaps _.str, %xmm036; SSE1-NEXT: movaps %xmm0, (%esp)37; SSE1-NEXT: movb $0, {{[0-9]+}}(%esp)38; SSE1-NEXT: movl $0, {{[0-9]+}}(%esp)39; SSE1-NEXT: movl $0, {{[0-9]+}}(%esp)40; SSE1-NEXT: ud241;42; NOSSE-LABEL: t1:43; NOSSE: ## %bb.0: ## %entry44; NOSSE-NEXT: subl $28, %esp45; NOSSE-NEXT: movb $0, {{[0-9]+}}(%esp)46; NOSSE-NEXT: movl $0, {{[0-9]+}}(%esp)47; NOSSE-NEXT: movl $0, {{[0-9]+}}(%esp)48; NOSSE-NEXT: movl $0, {{[0-9]+}}(%esp)49; NOSSE-NEXT: movl $0, {{[0-9]+}}(%esp)50; NOSSE-NEXT: movl $101, {{[0-9]+}}(%esp)51; NOSSE-NEXT: movl $1734438249, (%esp) ## imm = 0x67616D6952; NOSSE-NEXT: ud253;54; X86-64-LABEL: t1:55; X86-64: ## %bb.0: ## %entry56; X86-64-NEXT: movaps _.str(%rip), %xmm057; X86-64-NEXT: movaps %xmm0, -{{[0-9]+}}(%rsp)58; X86-64-NEXT: movb $0, -{{[0-9]+}}(%rsp)59; X86-64-NEXT: movq $0, -{{[0-9]+}}(%rsp)60; X86-64-NEXT: ud261;62; NHM_64-LABEL: t1:63; NHM_64: ## %bb.0: ## %entry64; NHM_64-NEXT: movups _.str+9(%rip), %xmm065; NHM_64-NEXT: movups %xmm0, -{{[0-9]+}}(%rsp)66; NHM_64-NEXT: movaps _.str(%rip), %xmm067; NHM_64-NEXT: movaps %xmm0, -{{[0-9]+}}(%rsp)68; NHM_64-NEXT: ud269entry:70 %tmp1 = alloca [25 x i8]71 call void @llvm.memcpy.p0.p0.i32(ptr align 1 %tmp1, ptr align 1 @.str, i32 25, i1 false)72 unreachable73}74 75;rdar://777470476%struct.s0 = type { [2 x double] }77 78define void @t2(ptr nocapture %a, ptr nocapture %b) nounwind ssp {79; SSE2-Darwin-LABEL: t2:80; SSE2-Darwin: ## %bb.0: ## %entry81; SSE2-Darwin-NEXT: movl {{[0-9]+}}(%esp), %eax82; SSE2-Darwin-NEXT: movl {{[0-9]+}}(%esp), %ecx83; SSE2-Darwin-NEXT: movaps (%ecx), %xmm084; SSE2-Darwin-NEXT: movaps %xmm0, (%eax)85; SSE2-Darwin-NEXT: retl86;87; SSE2-Mingw32-LABEL: t2:88; SSE2-Mingw32: # %bb.0: # %entry89; SSE2-Mingw32-NEXT: movl {{[0-9]+}}(%esp), %eax90; SSE2-Mingw32-NEXT: movl {{[0-9]+}}(%esp), %ecx91; SSE2-Mingw32-NEXT: movaps (%ecx), %xmm092; SSE2-Mingw32-NEXT: movaps %xmm0, (%eax)93; SSE2-Mingw32-NEXT: retl94;95; SSE1-LABEL: t2:96; SSE1: ## %bb.0: ## %entry97; SSE1-NEXT: movl {{[0-9]+}}(%esp), %eax98; SSE1-NEXT: movl {{[0-9]+}}(%esp), %ecx99; SSE1-NEXT: movaps (%ecx), %xmm0100; SSE1-NEXT: movaps %xmm0, (%eax)101; SSE1-NEXT: retl102;103; NOSSE-LABEL: t2:104; NOSSE: ## %bb.0: ## %entry105; NOSSE-NEXT: movl {{[0-9]+}}(%esp), %eax106; NOSSE-NEXT: movl {{[0-9]+}}(%esp), %ecx107; NOSSE-NEXT: movl 12(%ecx), %edx108; NOSSE-NEXT: movl %edx, 12(%eax)109; NOSSE-NEXT: movl 8(%ecx), %edx110; NOSSE-NEXT: movl %edx, 8(%eax)111; NOSSE-NEXT: movl (%ecx), %edx112; NOSSE-NEXT: movl 4(%ecx), %ecx113; NOSSE-NEXT: movl %ecx, 4(%eax)114; NOSSE-NEXT: movl %edx, (%eax)115; NOSSE-NEXT: retl116;117; X86-64-LABEL: t2:118; X86-64: ## %bb.0: ## %entry119; X86-64-NEXT: movaps (%rsi), %xmm0120; X86-64-NEXT: movaps %xmm0, (%rdi)121; X86-64-NEXT: retq122;123; NHM_64-LABEL: t2:124; NHM_64: ## %bb.0: ## %entry125; NHM_64-NEXT: movaps (%rsi), %xmm0126; NHM_64-NEXT: movaps %xmm0, (%rdi)127; NHM_64-NEXT: retq128entry:129 tail call void @llvm.memcpy.p0.p0.i32(ptr align 16 %a, ptr align 16 %b, i32 16, i1 false)130 ret void131}132 133define void @t3(ptr nocapture %a, ptr nocapture %b) nounwind ssp {134; SSE2-Darwin-LABEL: t3:135; SSE2-Darwin: ## %bb.0: ## %entry136; SSE2-Darwin-NEXT: movl {{[0-9]+}}(%esp), %eax137; SSE2-Darwin-NEXT: movl {{[0-9]+}}(%esp), %ecx138; SSE2-Darwin-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero139; SSE2-Darwin-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero140; SSE2-Darwin-NEXT: movsd %xmm1, 8(%eax)141; SSE2-Darwin-NEXT: movsd %xmm0, (%eax)142; SSE2-Darwin-NEXT: retl143;144; SSE2-Mingw32-LABEL: t3:145; SSE2-Mingw32: # %bb.0: # %entry146; SSE2-Mingw32-NEXT: movl {{[0-9]+}}(%esp), %eax147; SSE2-Mingw32-NEXT: movl {{[0-9]+}}(%esp), %ecx148; SSE2-Mingw32-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero149; SSE2-Mingw32-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero150; SSE2-Mingw32-NEXT: movsd %xmm1, 8(%eax)151; SSE2-Mingw32-NEXT: movsd %xmm0, (%eax)152; SSE2-Mingw32-NEXT: retl153;154; SSE1-LABEL: t3:155; SSE1: ## %bb.0: ## %entry156; SSE1-NEXT: movl {{[0-9]+}}(%esp), %eax157; SSE1-NEXT: movl {{[0-9]+}}(%esp), %ecx158; SSE1-NEXT: movl 12(%ecx), %edx159; SSE1-NEXT: movl %edx, 12(%eax)160; SSE1-NEXT: movl 8(%ecx), %edx161; SSE1-NEXT: movl %edx, 8(%eax)162; SSE1-NEXT: movl (%ecx), %edx163; SSE1-NEXT: movl 4(%ecx), %ecx164; SSE1-NEXT: movl %ecx, 4(%eax)165; SSE1-NEXT: movl %edx, (%eax)166; SSE1-NEXT: retl167;168; NOSSE-LABEL: t3:169; NOSSE: ## %bb.0: ## %entry170; NOSSE-NEXT: movl {{[0-9]+}}(%esp), %eax171; NOSSE-NEXT: movl {{[0-9]+}}(%esp), %ecx172; NOSSE-NEXT: movl 12(%ecx), %edx173; NOSSE-NEXT: movl %edx, 12(%eax)174; NOSSE-NEXT: movl 8(%ecx), %edx175; NOSSE-NEXT: movl %edx, 8(%eax)176; NOSSE-NEXT: movl (%ecx), %edx177; NOSSE-NEXT: movl 4(%ecx), %ecx178; NOSSE-NEXT: movl %ecx, 4(%eax)179; NOSSE-NEXT: movl %edx, (%eax)180; NOSSE-NEXT: retl181;182; X86-64-LABEL: t3:183; X86-64: ## %bb.0: ## %entry184; X86-64-NEXT: movq (%rsi), %rax185; X86-64-NEXT: movq 8(%rsi), %rcx186; X86-64-NEXT: movq %rcx, 8(%rdi)187; X86-64-NEXT: movq %rax, (%rdi)188; X86-64-NEXT: retq189;190; NHM_64-LABEL: t3:191; NHM_64: ## %bb.0: ## %entry192; NHM_64-NEXT: movups (%rsi), %xmm0193; NHM_64-NEXT: movups %xmm0, (%rdi)194; NHM_64-NEXT: retq195entry:196 tail call void @llvm.memcpy.p0.p0.i32(ptr align 8 %a, ptr align 8 %b, i32 16, i1 false)197 ret void198}199 200;;; TODO: (1) Some of the loads and stores are certainly unaligned and (2) the first load and first201;;; store overlap with the second load and second store respectively.202;;;203;;; Are any of the sequences ideal?204define void @t4() nounwind {205; SSE2-Darwin-LABEL: t4:206; SSE2-Darwin: ## %bb.0: ## %entry207; SSE2-Darwin-NEXT: subl $32, %esp208; SSE2-Darwin-NEXT: movw $120, {{[0-9]+}}(%esp)209; SSE2-Darwin-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878210; SSE2-Darwin-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878211; SSE2-Darwin-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878212; SSE2-Darwin-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878213; SSE2-Darwin-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878214; SSE2-Darwin-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878215; SSE2-Darwin-NEXT: movl $2021161080, (%esp) ## imm = 0x78787878216; SSE2-Darwin-NEXT: ud2217;218; SSE2-Mingw32-LABEL: t4:219; SSE2-Mingw32: # %bb.0: # %entry220; SSE2-Mingw32-NEXT: subl $32, %esp221; SSE2-Mingw32-NEXT: movw $120, {{[0-9]+}}(%esp)222; SSE2-Mingw32-NEXT: movl $2021161080, {{[0-9]+}}(%esp) # imm = 0x78787878223; SSE2-Mingw32-NEXT: movl $2021161080, {{[0-9]+}}(%esp) # imm = 0x78787878224; SSE2-Mingw32-NEXT: movl $2021161080, {{[0-9]+}}(%esp) # imm = 0x78787878225; SSE2-Mingw32-NEXT: movl $2021161080, {{[0-9]+}}(%esp) # imm = 0x78787878226; SSE2-Mingw32-NEXT: movl $2021161080, {{[0-9]+}}(%esp) # imm = 0x78787878227; SSE2-Mingw32-NEXT: movl $2021161080, {{[0-9]+}}(%esp) # imm = 0x78787878228; SSE2-Mingw32-NEXT: movl $2021161080, (%esp) # imm = 0x78787878229;230; SSE1-LABEL: t4:231; SSE1: ## %bb.0: ## %entry232; SSE1-NEXT: subl $32, %esp233; SSE1-NEXT: movw $120, {{[0-9]+}}(%esp)234; SSE1-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878235; SSE1-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878236; SSE1-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878237; SSE1-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878238; SSE1-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878239; SSE1-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878240; SSE1-NEXT: movl $2021161080, (%esp) ## imm = 0x78787878241; SSE1-NEXT: ud2242;243; NOSSE-LABEL: t4:244; NOSSE: ## %bb.0: ## %entry245; NOSSE-NEXT: subl $32, %esp246; NOSSE-NEXT: movw $120, {{[0-9]+}}(%esp)247; NOSSE-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878248; NOSSE-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878249; NOSSE-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878250; NOSSE-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878251; NOSSE-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878252; NOSSE-NEXT: movl $2021161080, {{[0-9]+}}(%esp) ## imm = 0x78787878253; NOSSE-NEXT: movl $2021161080, (%esp) ## imm = 0x78787878254; NOSSE-NEXT: ud2255;256; X86-64-LABEL: t4:257; X86-64: ## %bb.0: ## %entry258; X86-64-NEXT: movabsq $33909456017848440, %rax ## imm = 0x78787878787878259; X86-64-NEXT: movq %rax, -{{[0-9]+}}(%rsp)260; X86-64-NEXT: movabsq $8680820740569200760, %rax ## imm = 0x7878787878787878261; X86-64-NEXT: movq %rax, -{{[0-9]+}}(%rsp)262; X86-64-NEXT: movq %rax, -{{[0-9]+}}(%rsp)263; X86-64-NEXT: movq %rax, -{{[0-9]+}}(%rsp)264; X86-64-NEXT: ud2265;266; NHM_64-LABEL: t4:267; NHM_64: ## %bb.0: ## %entry268; NHM_64-NEXT: movups _.str2+14(%rip), %xmm0269; NHM_64-NEXT: movups %xmm0, -{{[0-9]+}}(%rsp)270; NHM_64-NEXT: movups _.str2(%rip), %xmm0271; NHM_64-NEXT: movaps %xmm0, -{{[0-9]+}}(%rsp)272; NHM_64-NEXT: ud2273entry:274 %tmp1 = alloca [30 x i8]275 call void @llvm.memcpy.p0.p0.i32(ptr align 1 %tmp1, ptr align 1 @.str2, i32 30, i1 false)276 unreachable277}278 279declare void @llvm.memcpy.p0.p0.i32(ptr nocapture, ptr nocapture, i32, i1) nounwind280