105 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X863; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X644 5; When loading the shift amount from memory, avoid generating the splat.6 7define void @shift5a(<4 x i32> %val, ptr %dst, ptr %pamt) nounwind {8; X86-LABEL: shift5a:9; X86: # %bb.0: # %entry10; X86-NEXT: movl {{[0-9]+}}(%esp), %eax11; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx12; X86-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero13; X86-NEXT: pslld %xmm1, %xmm014; X86-NEXT: movdqa %xmm0, (%eax)15; X86-NEXT: retl16;17; X64-LABEL: shift5a:18; X64: # %bb.0: # %entry19; X64-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero20; X64-NEXT: pslld %xmm1, %xmm021; X64-NEXT: movdqa %xmm0, (%rdi)22; X64-NEXT: retq23entry:24 %amt = load i32, ptr %pamt25 %tmp0 = insertelement <4 x i32> undef, i32 %amt, i32 026 %shamt = shufflevector <4 x i32> %tmp0, <4 x i32> undef, <4 x i32> zeroinitializer27 %shl = shl <4 x i32> %val, %shamt28 store <4 x i32> %shl, ptr %dst29 ret void30}31 32 33define void @shift5b(<4 x i32> %val, ptr %dst, ptr %pamt) nounwind {34; X86-LABEL: shift5b:35; X86: # %bb.0: # %entry36; X86-NEXT: movl {{[0-9]+}}(%esp), %eax37; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx38; X86-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero39; X86-NEXT: psrad %xmm1, %xmm040; X86-NEXT: movdqa %xmm0, (%eax)41; X86-NEXT: retl42;43; X64-LABEL: shift5b:44; X64: # %bb.0: # %entry45; X64-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero46; X64-NEXT: psrad %xmm1, %xmm047; X64-NEXT: movdqa %xmm0, (%rdi)48; X64-NEXT: retq49entry:50 %amt = load i32, ptr %pamt51 %tmp0 = insertelement <4 x i32> undef, i32 %amt, i32 052 %shamt = shufflevector <4 x i32> %tmp0, <4 x i32> undef, <4 x i32> zeroinitializer53 %shr = ashr <4 x i32> %val, %shamt54 store <4 x i32> %shr, ptr %dst55 ret void56}57 58 59define void @shift5c(<4 x i32> %val, ptr %dst, i32 %amt) nounwind {60; X86-LABEL: shift5c:61; X86: # %bb.0: # %entry62; X86-NEXT: movl {{[0-9]+}}(%esp), %eax63; X86-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero64; X86-NEXT: pslld %xmm1, %xmm065; X86-NEXT: movdqa %xmm0, (%eax)66; X86-NEXT: retl67;68; X64-LABEL: shift5c:69; X64: # %bb.0: # %entry70; X64-NEXT: movd %esi, %xmm171; X64-NEXT: pslld %xmm1, %xmm072; X64-NEXT: movdqa %xmm0, (%rdi)73; X64-NEXT: retq74entry:75 %tmp0 = insertelement <4 x i32> undef, i32 %amt, i32 076 %shamt = shufflevector <4 x i32> %tmp0, <4 x i32> undef, <4 x i32> zeroinitializer77 %shl = shl <4 x i32> %val, %shamt78 store <4 x i32> %shl, ptr %dst79 ret void80}81 82 83define void @shift5d(<4 x i32> %val, ptr %dst, i32 %amt) nounwind {84; X86-LABEL: shift5d:85; X86: # %bb.0: # %entry86; X86-NEXT: movl {{[0-9]+}}(%esp), %eax87; X86-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero88; X86-NEXT: psrad %xmm1, %xmm089; X86-NEXT: movdqa %xmm0, (%eax)90; X86-NEXT: retl91;92; X64-LABEL: shift5d:93; X64: # %bb.0: # %entry94; X64-NEXT: movd %esi, %xmm195; X64-NEXT: psrad %xmm1, %xmm096; X64-NEXT: movdqa %xmm0, (%rdi)97; X64-NEXT: retq98entry:99 %tmp0 = insertelement <4 x i32> undef, i32 %amt, i32 0100 %shamt = shufflevector <4 x i32> %tmp0, <4 x i32> undef, <4 x i32> zeroinitializer101 %shr = ashr <4 x i32> %val, %shamt102 store <4 x i32> %shr, ptr %dst103 ret void104}105