1 ; RUN: llc < %s -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr8 -mattr=+vsx | FileCheck %s
2 ; RUN: llc < %s -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr8 -mattr=+vsx -fast-isel -O0 | FileCheck -check-prefix=CHECK-FISL %s
3 define void @test1sp(float %a, float %b, float %c, float %e, float* nocapture %d) #0 {
5 %0 = tail call float @llvm.fma.f32(float %b, float %c, float %a)
6 store float %0, float* %d, align 4
7 %1 = tail call float @llvm.fma.f32(float %b, float %e, float %a)
8 %arrayidx1 = getelementptr inbounds float, float* %d, i64 1
9 store float %1, float* %arrayidx1, align 4
12 ; CHECK-LABEL: @test1sp
13 ; CHECK-DAG: li [[C1:[0-9]+]], 4
14 ; CHECK-DAG: xsmaddmsp 3, 2, 1
15 ; CHECK-DAG: xsmaddasp 1, 2, 4
16 ; CHECK-DAG: stxsspx 3, 0, 7
17 ; CHECK-DAG: stxsspx 1, 7, [[C1]]
20 ; CHECK-FISL-LABEL: @test1sp
21 ; CHECK-FISL-DAG: fmr 0, 1
22 ; CHECK-FISL-DAG: xsmaddasp 0, 2, 3
23 ; CHECK-FISL-DAG: stxsspx 0, 0, 7
24 ; CHECK-FISL-DAG: xsmaddasp 1, 2, 4
25 ; CHECK-FISL-DAG: li [[C1:[0-9]+]], 4
26 ; CHECK-FISL-DAG: stxsspx 1, 7, [[C1]]
30 define void @test2sp(float %a, float %b, float %c, float %e, float %f, float* nocapture %d) #0 {
32 %0 = tail call float @llvm.fma.f32(float %b, float %c, float %a)
33 store float %0, float* %d, align 4
34 %1 = tail call float @llvm.fma.f32(float %b, float %e, float %a)
35 %arrayidx1 = getelementptr inbounds float, float* %d, i64 1
36 store float %1, float* %arrayidx1, align 4
37 %2 = tail call float @llvm.fma.f32(float %b, float %f, float %a)
38 %arrayidx2 = getelementptr inbounds float, float* %d, i64 2
39 store float %2, float* %arrayidx2, align 4
42 ; CHECK-LABEL: @test2sp
43 ; CHECK-DAG: li [[C1:[0-9]+]], 4
44 ; CHECK-DAG: li [[C2:[0-9]+]], 8
45 ; FIXME: We now miss this because of copy ordering at the MI level.
46 ; CHECX-DAG: xsmaddmsp 3, 2, 1
47 ; CHECX-DAG: xsmaddmsp 4, 2, 1
48 ; CHECX-DAG: xsmaddasp 1, 2, 5
49 ; CHECX-DAG: stxsspx 3, 0, 8
50 ; CHECX-DAG: stxsspx 4, 8, [[C1]]
51 ; CHECX-DAG: stxsspx 1, 8, [[C2]]
54 ; CHECK-FISL-LABEL: @test2sp
55 ; CHECK-FISL-DAG: fmr 0, 1
56 ; CHECK-FISL-DAG: xsmaddasp 0, 2, 3
57 ; CHECK-FISL-DAG: stxsspx 0, 0, 8
58 ; CHECK-FISL-DAG: fmr 0, 1
59 ; CHECK-FISL-DAG: xsmaddasp 0, 2, 4
60 ; CHECK-FISL-DAG: li [[C1:[0-9]+]], 4
61 ; CHECK-FISL-DAG: stxsspx 0, 8, [[C1]]
62 ; CHECK-FISL-DAG: xsmaddasp 1, 2, 5
63 ; CHECK-FISL-DAG: li [[C2:[0-9]+]], 8
64 ; CHECK-FISL-DAG: stxsspx 1, 8, [[C2]]
68 define void @test3sp(float %a, float %b, float %c, float %e, float %f, float* nocapture %d) #0 {
70 %0 = tail call float @llvm.fma.f32(float %b, float %c, float %a)
71 store float %0, float* %d, align 4
72 %1 = tail call float @llvm.fma.f32(float %b, float %e, float %a)
73 %2 = tail call float @llvm.fma.f32(float %b, float %c, float %1)
74 %arrayidx1 = getelementptr inbounds float, float* %d, i64 3
75 store float %2, float* %arrayidx1, align 4
76 %3 = tail call float @llvm.fma.f32(float %b, float %f, float %a)
77 %arrayidx2 = getelementptr inbounds float, float* %d, i64 2
78 store float %3, float* %arrayidx2, align 4
79 %arrayidx3 = getelementptr inbounds float, float* %d, i64 1
80 store float %1, float* %arrayidx3, align 4
83 ; CHECK-LABEL: @test3sp
84 ; CHECK-DAG: fmr [[F1:[0-9]+]], 1
85 ; CHECK-DAG: li [[C1:[0-9]+]], 12
86 ; CHECK-DAG: li [[C2:[0-9]+]], 8
87 ; CHECK-DAG: li [[C3:[0-9]+]], 4
88 ; CHECK-DAG: xsmaddmsp 4, 2, 1
89 ; CHECK-DAG: xsmaddasp 1, 2, 5
91 ; Note: We could convert this next FMA to M-type as well, but it would require
92 ; re-ordering the instructions.
93 ; CHECK-DAG: xsmaddasp [[F1]], 2, 3
95 ; CHECK-DAG: xsmaddmsp 3, 2, 4
96 ; CHECK-DAG: stxsspx [[F1]], 0, 8
97 ; CHECK-DAG: stxsspx 3, 8, [[C1]]
98 ; CHECK-DAG: stxsspx 1, 8, [[C2]]
99 ; CHECK-DAG: stxsspx 4, 8, [[C3]]
102 ; CHECK-FISL-LABEL: @test3sp
103 ; CHECK-FISL-DAG: fmr [[F1:[0-9]+]], 1
104 ; CHECK-FISL-DAG: xsmaddasp [[F1]], 2, 4
105 ; CHECK-FISL-DAG: fmr 4, [[F1]]
106 ; CHECK-FISL-DAG: xsmaddasp 4, 2, 3
107 ; CHECK-FISL-DAG: li [[C1:[0-9]+]], 12
108 ; CHECK-FISL-DAG: stxsspx 4, 8, [[C1]]
109 ; CHECK-FISL-DAG: xsmaddasp 1, 2, 5
110 ; CHECK-FISL-DAG: li [[C2:[0-9]+]], 8
111 ; CHECK-FISL-DAG: stxsspx 1, 8, [[C2]]
112 ; CHECK-FISL-DAG: li [[C3:[0-9]+]], 4
113 ; CHECK-FISL-DAG: stxsspx 0, 8, [[C3]]
117 define void @test4sp(float %a, float %b, float %c, float %e, float %f, float* nocapture %d) #0 {
119 %0 = tail call float @llvm.fma.f32(float %b, float %c, float %a)
120 store float %0, float* %d, align 4
121 %1 = tail call float @llvm.fma.f32(float %b, float %e, float %a)
122 %arrayidx1 = getelementptr inbounds float, float* %d, i64 1
123 store float %1, float* %arrayidx1, align 4
124 %2 = tail call float @llvm.fma.f32(float %b, float %c, float %1)
125 %arrayidx3 = getelementptr inbounds float, float* %d, i64 3
126 store float %2, float* %arrayidx3, align 4
127 %3 = tail call float @llvm.fma.f32(float %b, float %f, float %a)
128 %arrayidx4 = getelementptr inbounds float, float* %d, i64 2
129 store float %3, float* %arrayidx4, align 4
132 ; CHECK-LABEL: @test4sp
133 ; CHECK-DAG: fmr [[F1:[0-9]+]], 1
134 ; CHECK-DAG: li [[C1:[0-9]+]], 4
135 ; CHECK-DAG: li [[C2:[0-9]+]], 8
136 ; CHECK-DAG: xsmaddmsp 4, 2, 1
138 ; Note: We could convert this next FMA to M-type as well, but it would require
139 ; re-ordering the instructions.
140 ; CHECK-DAG: xsmaddasp 1, 2, 5
142 ; CHECK-DAG: xsmaddasp [[F1]], 2, 3
143 ; CHECK-DAG: stxsspx [[F1]], 0, 8
144 ; CHECK-DAG: stxsspx 4, 8, [[C1]]
145 ; CHECK-DAG: li [[C3:[0-9]+]], 12
146 ; CHECK-DAG: xsmaddasp 4, 2, 3
147 ; CHECK-DAG: stxsspx 4, 8, [[C3]]
148 ; CHECK-DAG: stxsspx 1, 8, [[C2]]
151 ; CHECK-FISL-LABEL: @test4sp
152 ; CHECK-FISL-DAG: fmr [[F1:[0-9]+]], 1
153 ; CHECK-FISL-DAG: xsmaddasp [[F1]], 2, 3
154 ; CHECK-FISL-DAG: stxsspx 0, 0, 8
155 ; CHECK-FISL-DAG: fmr [[F1]], 1
156 ; CHECK-FISL-DAG: xsmaddasp [[F1]], 2, 4
157 ; CHECK-FISL-DAG: li [[C3:[0-9]+]], 4
158 ; CHECK-FISL-DAG: stxsspx 0, 8, [[C3]]
159 ; CHECK-FISL-DAG: xsmaddasp 0, 2, 3
160 ; CHECK-FISL-DAG: li [[C1:[0-9]+]], 12
161 ; CHECK-FISL-DAG: stxsspx 0, 8, [[C1]]
162 ; CHECK-FISL-DAG: xsmaddasp 1, 2, 5
163 ; CHECK-FISL-DAG: li [[C2:[0-9]+]], 8
164 ; CHECK-FISL-DAG: stxsspx 1, 8, [[C2]]
168 declare float @llvm.fma.f32(float, float, float) #0