76 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+mmx,+sse2 | FileCheck %s --check-prefix=X86-SSE3; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+mmx,+avx2 | FileCheck %s --check-prefix=X86-AVX4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+mmx,+sse2 | FileCheck %s --check-prefix=X64-SSE5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+mmx,+avx2 | FileCheck %s --check-prefix=X64-AVX6 7define i32 @PR29222(i32) nounwind {8; X86-SSE-LABEL: PR29222:9; X86-SSE: # %bb.0:10; X86-SSE-NEXT: pushl %ebp11; X86-SSE-NEXT: movl %esp, %ebp12; X86-SSE-NEXT: andl $-8, %esp13; X86-SSE-NEXT: subl $8, %esp14; X86-SSE-NEXT: movd 8(%ebp), %mm015; X86-SSE-NEXT: pshufw $68, %mm0, %mm0 # mm0 = mm0[0,1,0,1]16; X86-SSE-NEXT: packsswb %mm0, %mm017; X86-SSE-NEXT: movq %mm0, (%esp)18; X86-SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero19; X86-SSE-NEXT: packsswb %xmm0, %xmm020; X86-SSE-NEXT: movd %xmm0, %eax21; X86-SSE-NEXT: movl %ebp, %esp22; X86-SSE-NEXT: popl %ebp23; X86-SSE-NEXT: retl24;25; X86-AVX-LABEL: PR29222:26; X86-AVX: # %bb.0:27; X86-AVX-NEXT: pushl %ebp28; X86-AVX-NEXT: movl %esp, %ebp29; X86-AVX-NEXT: andl $-8, %esp30; X86-AVX-NEXT: subl $8, %esp31; X86-AVX-NEXT: movd 8(%ebp), %mm032; X86-AVX-NEXT: pshufw $68, %mm0, %mm0 # mm0 = mm0[0,1,0,1]33; X86-AVX-NEXT: packsswb %mm0, %mm034; X86-AVX-NEXT: movq %mm0, (%esp)35; X86-AVX-NEXT: vpbroadcastq (%esp), %xmm036; X86-AVX-NEXT: vpacksswb %xmm0, %xmm0, %xmm037; X86-AVX-NEXT: vmovd %xmm0, %eax38; X86-AVX-NEXT: movl %ebp, %esp39; X86-AVX-NEXT: popl %ebp40; X86-AVX-NEXT: retl41;42; X64-SSE-LABEL: PR29222:43; X64-SSE: # %bb.0:44; X64-SSE-NEXT: movd %edi, %mm045; X64-SSE-NEXT: pshufw $68, %mm0, %mm0 # mm0 = mm0[0,1,0,1]46; X64-SSE-NEXT: packsswb %mm0, %mm047; X64-SSE-NEXT: movq2dq %mm0, %xmm048; X64-SSE-NEXT: packsswb %xmm0, %xmm049; X64-SSE-NEXT: movd %xmm0, %eax50; X64-SSE-NEXT: retq51;52; X64-AVX-LABEL: PR29222:53; X64-AVX: # %bb.0:54; X64-AVX-NEXT: movd %edi, %mm055; X64-AVX-NEXT: pshufw $68, %mm0, %mm0 # mm0 = mm0[0,1,0,1]56; X64-AVX-NEXT: packsswb %mm0, %mm057; X64-AVX-NEXT: movq2dq %mm0, %xmm058; X64-AVX-NEXT: vpacksswb %xmm0, %xmm0, %xmm059; X64-AVX-NEXT: vmovd %xmm0, %eax60; X64-AVX-NEXT: retq61 %2 = insertelement <2 x i32> undef, i32 %0, i32 062 %3 = shufflevector <2 x i32> %2, <2 x i32> undef, <2 x i32> zeroinitializer63 %4 = bitcast <2 x i32> %3 to <1 x i64>64 %5 = tail call <1 x i64> @llvm.x86.mmx.packsswb(<1 x i64> %4, <1 x i64> %4)65 %6 = bitcast <1 x i64> %5 to i6466 %7 = insertelement <2 x i64> undef, i64 %6, i32 067 %8 = bitcast <2 x i64> %7 to <8 x i16>68 %9 = tail call <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16> %8, <8 x i16> undef)69 %10 = bitcast <16 x i8> %9 to <4 x i32>70 %11 = extractelement <4 x i32> %10, i32 071 ret i32 %1172}73 74declare <1 x i64> @llvm.x86.mmx.packsswb(<1 x i64>, <1 x i64>)75declare <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16>, <8 x i16>)76