1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
2 ; XUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
4 declare i32 @llvm.r600.read.tidig.x() nounwind readnone
5 declare float @llvm.fabs.f32(float) nounwind readnone
7 ; GCN-LABEL: {{^}}madmk_f32:
8 ; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
9 ; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
10 ; GCN: v_madmk_f32_e32 {{v[0-9]+}}, [[VA]], [[VB]], 0x41200000
11 define void @madmk_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
12 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
13 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
14 %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1
15 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
17 %a = load float, float addrspace(1)* %gep.0, align 4
18 %b = load float, float addrspace(1)* %gep.1, align 4
20 %mul = fmul float %a, 10.0
21 %madmk = fadd float %mul, %b
22 store float %madmk, float addrspace(1)* %out.gep, align 4
26 ; GCN-LABEL: {{^}}madmk_2_use_f32:
27 ; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
28 ; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
29 ; GCN-DAG: buffer_load_dword [[VC:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8
30 ; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x41200000
31 ; GCN-DAG: v_mad_f32 {{v[0-9]+}}, [[VA]], [[VK]], [[VB]]
32 ; GCN-DAG: v_mad_f32 {{v[0-9]+}}, [[VA]], [[VK]], [[VC]]
34 define void @madmk_2_use_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
35 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
37 %in.gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
38 %in.gep.1 = getelementptr float, float addrspace(1)* %in.gep.0, i32 1
39 %in.gep.2 = getelementptr float, float addrspace(1)* %in.gep.0, i32 2
41 %out.gep.0 = getelementptr float, float addrspace(1)* %out, i32 %tid
42 %out.gep.1 = getelementptr float, float addrspace(1)* %in.gep.0, i32 1
44 %a = load float, float addrspace(1)* %in.gep.0, align 4
45 %b = load float, float addrspace(1)* %in.gep.1, align 4
46 %c = load float, float addrspace(1)* %in.gep.2, align 4
48 %mul0 = fmul float %a, 10.0
49 %mul1 = fmul float %a, 10.0
50 %madmk0 = fadd float %mul0, %b
51 %madmk1 = fadd float %mul1, %c
53 store float %madmk0, float addrspace(1)* %out.gep.0, align 4
54 store float %madmk1, float addrspace(1)* %out.gep.1, align 4
58 ; We don't get any benefit if the constant is an inline immediate.
59 ; GCN-LABEL: {{^}}madmk_inline_imm_f32:
60 ; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
61 ; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
62 ; GCN: v_mad_f32 {{v[0-9]+}}, 4.0, [[VA]], [[VB]]
63 define void @madmk_inline_imm_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
64 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
65 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
66 %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1
67 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
69 %a = load float, float addrspace(1)* %gep.0, align 4
70 %b = load float, float addrspace(1)* %gep.1, align 4
72 %mul = fmul float %a, 4.0
73 %madmk = fadd float %mul, %b
74 store float %madmk, float addrspace(1)* %out.gep, align 4
78 ; GCN-LABEL: {{^}}s_s_madmk_f32:
79 ; GCN-NOT: v_madmk_f32
82 define void @s_s_madmk_f32(float addrspace(1)* noalias %out, float %a, float %b) nounwind {
83 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
84 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
86 %mul = fmul float %a, 10.0
87 %madmk = fadd float %mul, %b
88 store float %madmk, float addrspace(1)* %out.gep, align 4
92 ; GCN-LABEL: {{^}}v_s_madmk_f32:
93 ; GCN-NOT: v_madmk_f32
96 define void @v_s_madmk_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in, float %b) nounwind {
97 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
98 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
99 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
100 %a = load float, float addrspace(1)* %gep.0, align 4
102 %mul = fmul float %a, 10.0
103 %madmk = fadd float %mul, %b
104 store float %madmk, float addrspace(1)* %out.gep, align 4
108 ; GCN-LABEL: {{^}}scalar_vector_madmk_f32:
109 ; GCN-NOT: v_madmk_f32
112 define void @scalar_vector_madmk_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in, float %a) nounwind {
113 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
114 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
115 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
116 %b = load float, float addrspace(1)* %gep.0, align 4
118 %mul = fmul float %a, 10.0
119 %madmk = fadd float %mul, %b
120 store float %madmk, float addrspace(1)* %out.gep, align 4
124 ; GCN-LABEL: {{^}}no_madmk_src0_modifier_f32:
125 ; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
126 ; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
127 ; GCN: v_mad_f32 {{v[0-9]+}}, |{{v[0-9]+}}|, {{v[0-9]+}}, {{[sv][0-9]+}}
128 define void @no_madmk_src0_modifier_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
129 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
130 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
131 %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1
132 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
134 %a = load float, float addrspace(1)* %gep.0, align 4
135 %b = load float, float addrspace(1)* %gep.1, align 4
137 %a.fabs = call float @llvm.fabs.f32(float %a) nounwind readnone
139 %mul = fmul float %a.fabs, 10.0
140 %madmk = fadd float %mul, %b
141 store float %madmk, float addrspace(1)* %out.gep, align 4
145 ; GCN-LABEL: {{^}}no_madmk_src2_modifier_f32:
146 ; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
147 ; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
148 ; GCN: v_mad_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, |{{[sv][0-9]+}}|
149 define void @no_madmk_src2_modifier_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
150 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
151 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
152 %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1
153 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
155 %a = load float, float addrspace(1)* %gep.0, align 4
156 %b = load float, float addrspace(1)* %gep.1, align 4
158 %b.fabs = call float @llvm.fabs.f32(float %b) nounwind readnone
160 %mul = fmul float %a, 10.0
161 %madmk = fadd float %mul, %b.fabs
162 store float %madmk, float addrspace(1)* %out.gep, align 4
166 ; GCN-LABEL: {{^}}madmk_add_inline_imm_f32:
167 ; GCN: buffer_load_dword [[A:v[0-9]+]]
168 ; GCN: v_mov_b32_e32 [[VK:v[0-9]+]], 0x41200000
169 ; GCN: v_mad_f32 {{v[0-9]+}}, [[VK]], [[A]], 2.0
170 define void @madmk_add_inline_imm_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
171 %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
172 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
173 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
175 %a = load float, float addrspace(1)* %gep.0, align 4
177 %mul = fmul float %a, 10.0
178 %madmk = fadd float %mul, 2.0
179 store float %madmk, float addrspace(1)* %out.gep, align 4