Home | History | Annotate | Download | only in X86
      1 ; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+sse2 -mcpu=nehalem | FileCheck %s
      2 
      3 ; rdar: 12558838
      4 ; PR14221
      5 ; There is a mismatch between the intrinsic and the actual instruction.
      6 ; The actual instruction has a partial update of dest, while the intrinsic
      7 ; passes through the upper FP values. Here, we make sure the source and
      8 ; destination of rsqrtss are the same.
      9 define void @t1(<4 x float> %a) nounwind uwtable ssp {
     10 entry:
     11 ; CHECK-LABEL: t1:
     12 ; CHECK: rsqrtss %xmm0, %xmm0
     13   %0 = tail call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %a) nounwind
     14   %a.addr.0.extract = extractelement <4 x float> %0, i32 0
     15   %conv = fpext float %a.addr.0.extract to double
     16   %a.addr.4.extract = extractelement <4 x float> %0, i32 1
     17   %conv3 = fpext float %a.addr.4.extract to double
     18   tail call void @callee(double %conv, double %conv3) nounwind
     19   ret void
     20 }
     21 declare void @callee(double, double)
     22 declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone
     23 
     24 define void @t2(<4 x float> %a) nounwind uwtable ssp {
     25 entry:
     26 ; CHECK-LABEL: t2:
     27 ; CHECK: rcpss %xmm0, %xmm0
     28   %0 = tail call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %a) nounwind
     29   %a.addr.0.extract = extractelement <4 x float> %0, i32 0
     30   %conv = fpext float %a.addr.0.extract to double
     31   %a.addr.4.extract = extractelement <4 x float> %0, i32 1
     32   %conv3 = fpext float %a.addr.4.extract to double
     33   tail call void @callee(double %conv, double %conv3) nounwind
     34   ret void
     35 }
     36 declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
     37