1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+sse2 -mcpu=nehalem | FileCheck %s
6 ; There is a mismatch between the intrinsic and the actual instruction.
7 ; The actual instruction has a partial update of dest, while the intrinsic
8 ; passes through the upper FP values. Here, we make sure the source and
9 ; destination of each scalar unary op are the same.
11 define void @rsqrtss(<4 x float> %a) nounwind uwtable ssp {
12 ; CHECK-LABEL: rsqrtss:
14 ; CHECK-NEXT: rsqrtss %xmm0, %xmm0
15 ; CHECK-NEXT: cvtss2sd %xmm0, %xmm2
16 ; CHECK-NEXT: movshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
17 ; CHECK-NEXT: cvtss2sd %xmm0, %xmm1
18 ; CHECK-NEXT: movaps %xmm2, %xmm0
19 ; CHECK-NEXT: jmp _callee ## TAILCALL
20 %t0 = tail call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %a) nounwind
21 %a.addr.0.extract = extractelement <4 x float> %t0, i32 0
22 %conv = fpext float %a.addr.0.extract to double
23 %a.addr.4.extract = extractelement <4 x float> %t0, i32 1
24 %conv3 = fpext float %a.addr.4.extract to double
25 tail call void @callee(double %conv, double %conv3) nounwind
28 declare void @callee(double, double)
29 declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone
31 define void @rcpss(<4 x float> %a) nounwind uwtable ssp {
34 ; CHECK-NEXT: rcpss %xmm0, %xmm0
35 ; CHECK-NEXT: cvtss2sd %xmm0, %xmm2
36 ; CHECK-NEXT: movshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
37 ; CHECK-NEXT: cvtss2sd %xmm0, %xmm1
38 ; CHECK-NEXT: movaps %xmm2, %xmm0
39 ; CHECK-NEXT: jmp _callee ## TAILCALL
40 %t0 = tail call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %a) nounwind
41 %a.addr.0.extract = extractelement <4 x float> %t0, i32 0
42 %conv = fpext float %a.addr.0.extract to double
43 %a.addr.4.extract = extractelement <4 x float> %t0, i32 1
44 %conv3 = fpext float %a.addr.4.extract to double
45 tail call void @callee(double %conv, double %conv3) nounwind
48 declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
50 define void @sqrtss(<4 x float> %a) nounwind uwtable ssp {
51 ; CHECK-LABEL: sqrtss:
53 ; CHECK-NEXT: sqrtss %xmm0, %xmm1
54 ; CHECK-NEXT: cvtss2sd %xmm1, %xmm2
55 ; CHECK-NEXT: movshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
56 ; CHECK-NEXT: xorps %xmm1, %xmm1
57 ; CHECK-NEXT: cvtss2sd %xmm0, %xmm1
58 ; CHECK-NEXT: movaps %xmm2, %xmm0
59 ; CHECK-NEXT: jmp _callee ## TAILCALL
60 %t0 = tail call <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float> %a) nounwind
61 %a.addr.0.extract = extractelement <4 x float> %t0, i32 0
62 %conv = fpext float %a.addr.0.extract to double
63 %a.addr.4.extract = extractelement <4 x float> %t0, i32 1
64 %conv3 = fpext float %a.addr.4.extract to double
65 tail call void @callee(double %conv, double %conv3) nounwind
68 declare <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float>) nounwind readnone
70 define void @sqrtsd(<2 x double> %a) nounwind uwtable ssp {
71 ; CHECK-LABEL: sqrtsd:
73 ; CHECK-NEXT: sqrtsd %xmm0, %xmm1
74 ; CHECK-NEXT: cvtsd2ss %xmm1, %xmm2
75 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]
76 ; CHECK-NEXT: xorps %xmm1, %xmm1
77 ; CHECK-NEXT: cvtsd2ss %xmm0, %xmm1
78 ; CHECK-NEXT: movaps %xmm2, %xmm0
79 ; CHECK-NEXT: jmp _callee2 ## TAILCALL
80 %t0 = tail call <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double> %a) nounwind
81 %a0 = extractelement <2 x double> %t0, i32 0
82 %conv = fptrunc double %a0 to float
83 %a1 = extractelement <2 x double> %t0, i32 1
84 %conv3 = fptrunc double %a1 to float
85 tail call void @callee2(float %conv, float %conv3) nounwind
89 declare void @callee2(float, float)
90 declare <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double>) nounwind readnone
92 define <2 x double> @load_fold_cvtss2sd_int(ptr%a) {
93 ; CHECK-LABEL: load_fold_cvtss2sd_int:
95 ; CHECK-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
96 ; CHECK-NEXT: cvtss2sd %xmm0, %xmm0
97 ; CHECK-NEXT: movq {{.*#+}} xmm0 = xmm0[0],zero
99 %ld = load <4 x float>, ptr%a
100 %x = call <2 x double> @llvm.x86.sse2.cvtss2sd(<2 x double> <double 0x0, double 0x0>, <4 x float> %ld)
104 define <2 x double> @load_fold_cvtss2sd_int_optsize(ptr%a) optsize {
105 ; CHECK-LABEL: load_fold_cvtss2sd_int_optsize:
107 ; CHECK-NEXT: cvtss2sd (%rdi), %xmm0
108 ; CHECK-NEXT: movq {{.*#+}} xmm0 = xmm0[0],zero
110 %ld = load <4 x float>, ptr%a
111 %x = call <2 x double> @llvm.x86.sse2.cvtss2sd(<2 x double> <double 0x0, double 0x0>, <4 x float> %ld)
115 define <2 x double> @load_fold_cvtss2sd_int_minsize(ptr%a) minsize {
116 ; CHECK-LABEL: load_fold_cvtss2sd_int_minsize:
118 ; CHECK-NEXT: cvtss2sd (%rdi), %xmm0
119 ; CHECK-NEXT: movq {{.*#+}} xmm0 = xmm0[0],zero
121 %ld = load <4 x float>, ptr%a
122 %x = call <2 x double> @llvm.x86.sse2.cvtss2sd(<2 x double> <double 0x0, double 0x0>, <4 x float> %ld)
126 declare <2 x double> @llvm.x86.sse2.cvtss2sd(<2 x double>, <4 x float>) nounwind readnone
128 define float @PR22206(<4 x float> %a) {
129 ; CHECK-LABEL: PR22206:
131 ; CHECK-NEXT: sqrtss %xmm0, %xmm1
132 ; CHECK-NEXT: addss %xmm1, %xmm0
134 %res = tail call <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float> %a) nounwind
135 %new = extractelement <4 x float> %res, i32 0
136 %orig = extractelement <4 x float> %a, i32 0
137 %add = fadd float %new, %orig