95 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s3 4; In the following 4 tests, the existing call to VZU/VZA ensures clean state before5; the call to the unknown, so we don't need to insert a second VZU at that point.6 7define <4 x float> @zeroupper_v4f32(ptr%x, <8 x float> %y) nounwind {8; CHECK-LABEL: zeroupper_v4f32:9; CHECK: # %bb.0:10; CHECK-NEXT: pushq %rbx11; CHECK-NEXT: subq $32, %rsp12; CHECK-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill13; CHECK-NEXT: movq %rdi, %rbx14; CHECK-NEXT: vzeroupper15; CHECK-NEXT: callq the_unknown@PLT16; CHECK-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload17; CHECK-NEXT: vaddps (%rbx), %ymm0, %ymm018; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm119; CHECK-NEXT: vaddps %xmm1, %xmm0, %xmm020; CHECK-NEXT: addq $32, %rsp21; CHECK-NEXT: popq %rbx22; CHECK-NEXT: vzeroupper23; CHECK-NEXT: retq24 call void @llvm.x86.avx.vzeroupper()25 call void @the_unknown()26 %loadx = load <8 x float>, ptr%x, align 3227 %sum = fadd <8 x float> %loadx, %y28 %lo = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>29 %hi = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>30 %res = fadd <4 x float> %lo, %hi31 ret <4 x float> %res32}33 34define <8 x float> @zeroupper_v8f32(<8 x float> %x) nounwind {35; CHECK-LABEL: zeroupper_v8f32:36; CHECK: # %bb.0:37; CHECK-NEXT: subq $40, %rsp38; CHECK-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill39; CHECK-NEXT: vzeroupper40; CHECK-NEXT: callq the_unknown@PLT41; CHECK-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload42; CHECK-NEXT: addq $40, %rsp43; CHECK-NEXT: retq44 call void @llvm.x86.avx.vzeroupper()45 call void @the_unknown()46 ret <8 x float> %x47}48 49define <4 x float> @zeroall_v4f32(ptr%x, <8 x float> %y) nounwind {50; CHECK-LABEL: zeroall_v4f32:51; CHECK: # %bb.0:52; CHECK-NEXT: pushq %rbx53; CHECK-NEXT: subq $32, %rsp54; CHECK-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill55; CHECK-NEXT: movq %rdi, %rbx56; CHECK-NEXT: vzeroall57; CHECK-NEXT: callq the_unknown@PLT58; CHECK-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload59; CHECK-NEXT: vaddps (%rbx), %ymm0, %ymm060; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm161; CHECK-NEXT: vaddps %xmm1, %xmm0, %xmm062; CHECK-NEXT: addq $32, %rsp63; CHECK-NEXT: popq %rbx64; CHECK-NEXT: vzeroupper65; CHECK-NEXT: retq66 call void @llvm.x86.avx.vzeroall()67 call void @the_unknown()68 %loadx = load <8 x float>, ptr%x, align 3269 %sum = fadd <8 x float> %loadx, %y70 %lo = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>71 %hi = shufflevector <8 x float> %sum, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>72 %res = fadd <4 x float> %lo, %hi73 ret <4 x float> %res74}75 76define <8 x float> @zeroall_v8f32(<8 x float> %x) nounwind {77; CHECK-LABEL: zeroall_v8f32:78; CHECK: # %bb.0:79; CHECK-NEXT: subq $40, %rsp80; CHECK-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill81; CHECK-NEXT: vzeroall82; CHECK-NEXT: callq the_unknown@PLT83; CHECK-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload84; CHECK-NEXT: addq $40, %rsp85; CHECK-NEXT: retq86 call void @llvm.x86.avx.vzeroall()87 call void @the_unknown()88 ret <8 x float> %x89}90 91declare void @llvm.x86.avx.vzeroupper() nounwind readnone92declare void @llvm.x86.avx.vzeroall() nounwind readnone93declare void @the_unknown() nounwind94 95