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 each scalar unary op are the same.
9
10define void @rsqrtss(<4 x float> %a) nounwind uwtable ssp {
11entry:
12; CHECK-LABEL: rsqrtss:
13; CHECK: rsqrtss %xmm0, %xmm0
14; CHECK-NEXT: cvtss2sd %xmm0
15; CHECK-NEXT: movshdup
16; CHECK-NEXT: cvtss2sd %xmm0
17; CHECK-NEXT: movap
18; CHECK-NEXT: jmp
19
20  %0 = tail call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %a) nounwind
21  %a.addr.0.extract = extractelement <4 x float> %0, i32 0
22  %conv = fpext float %a.addr.0.extract to double
23  %a.addr.4.extract = extractelement <4 x float> %0, i32 1
24  %conv3 = fpext float %a.addr.4.extract to double
25  tail call void @callee(double %conv, double %conv3) nounwind
26  ret void
27}
28declare void @callee(double, double)
29declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone
30
31define void @rcpss(<4 x float> %a) nounwind uwtable ssp {
32entry:
33; CHECK-LABEL: rcpss:
34; CHECK: rcpss %xmm0, %xmm0
35; CHECK-NEXT: cvtss2sd %xmm0
36; CHECK-NEXT: movshdup
37; CHECK-NEXT: cvtss2sd %xmm0
38; CHECK-NEXT: movap
39; CHECK-NEXT: jmp
40
41  %0 = tail call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %a) nounwind
42  %a.addr.0.extract = extractelement <4 x float> %0, i32 0
43  %conv = fpext float %a.addr.0.extract to double
44  %a.addr.4.extract = extractelement <4 x float> %0, i32 1
45  %conv3 = fpext float %a.addr.4.extract to double
46  tail call void @callee(double %conv, double %conv3) nounwind
47  ret void
48}
49declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
50
51define void @sqrtss(<4 x float> %a) nounwind uwtable ssp {
52entry:
53; CHECK-LABEL: sqrtss:
54; CHECK: sqrtss %xmm0, %xmm0
55; CHECK-NEXT: cvtss2sd %xmm0
56; CHECK-NEXT: movshdup
57; CHECK-NEXT: cvtss2sd %xmm0
58; CHECK-NEXT: movap
59; CHECK-NEXT: jmp
60
61  %0 = tail call <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float> %a) nounwind
62  %a.addr.0.extract = extractelement <4 x float> %0, i32 0
63  %conv = fpext float %a.addr.0.extract to double
64  %a.addr.4.extract = extractelement <4 x float> %0, i32 1
65  %conv3 = fpext float %a.addr.4.extract to double
66  tail call void @callee(double %conv, double %conv3) nounwind
67  ret void
68}
69declare <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float>) nounwind readnone
70
71define void @sqrtsd(<2 x double> %a) nounwind uwtable ssp {
72entry:
73; CHECK-LABEL: sqrtsd:
74; CHECK: sqrtsd %xmm0, %xmm0
75; CHECK-NEXT: cvtsd2ss %xmm0
76; CHECK-NEXT: shufpd
77; CHECK-NEXT: cvtsd2ss %xmm0
78; CHECK-NEXT: movap
79; CHECK-NEXT: jmp
80
81 %0 = tail call <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double> %a) nounwind
82 %a0 = extractelement <2 x double> %0, i32 0
83 %conv = fptrunc double %a0 to float
84 %a1 = extractelement <2 x double> %0, i32 1
85 %conv3 = fptrunc double %a1 to float
86 tail call void @callee2(float %conv, float %conv3) nounwind
87 ret void
88}
89
90declare void @callee2(float, float)
91declare <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double>) nounwind readnone
92
93