1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN %s
2 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=GCN %s
4 declare float @llvm.fma.f32(float, float, float) #1
5 declare double @llvm.fma.f64(double, double, double) #1
6 declare float @llvm.fmuladd.f32(float, float, float) #1
7 declare float @llvm.amdgcn.div.fixup.f32(float, float, float) #1
10 ; GCN-LABEL: {{^}}test_sgpr_use_twice_binop:
11 ; GCN: s_load_dword [[SGPR:s[0-9]+]],
12 ; GCN: v_add_f32_e64 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]]
13 ; GCN: buffer_store_dword [[RESULT]]
14 define amdgpu_kernel void @test_sgpr_use_twice_binop(ptr addrspace(1) %out, float %a) #0 {
15 %dbl = fadd float %a, %a
16 store float %dbl, ptr addrspace(1) %out, align 4
20 ; GCN-LABEL: {{^}}test_sgpr_use_three_ternary_op:
21 ; GCN: s_load_dword [[SGPR:s[0-9]+]],
22 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], [[SGPR]]
23 ; GCN: buffer_store_dword [[RESULT]]
24 define amdgpu_kernel void @test_sgpr_use_three_ternary_op(ptr addrspace(1) %out, float %a) #0 {
25 %fma = call float @llvm.fma.f32(float %a, float %a, float %a) #1
26 store float %fma, ptr addrspace(1) %out, align 4
30 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_b:
31 ; SI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x9
32 ; VI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x24
33 ; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[#LOAD + 3]]
34 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], s[[#LOAD + 2]], s[[#LOAD + 2]], [[VGPR1]]
35 ; GCN: buffer_store_dword [[RESULT]]
36 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_b(ptr addrspace(1) %out, float %a, float %b) #0 {
37 %fma = call float @llvm.fma.f32(float %a, float %a, float %b) #1
38 store float %fma, ptr addrspace(1) %out, align 4
42 ; GCN-LABEL: {{^}}test_use_s_v_s:
43 ; SI: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, {{0x9|0x24}}
44 ; SI: buffer_load_dword [[VA0:v[0-9]+]]
45 ; SI-NEXT: s_waitcnt vmcnt(0)
46 ; SI-NEXT: buffer_load_dword [[VA1:v[0-9]+]]
47 ; SI-NEXT: s_waitcnt vmcnt(0)
51 ; VI: buffer_load_dword [[VA0:v[0-9]+]]
52 ; VI-NEXT: s_waitcnt vmcnt(0)
53 ; VI-NEXT: buffer_load_dword [[VA1:v[0-9]+]]
54 ; VI-NEXT: s_waitcnt vmcnt(0)
55 ; VI: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, {{0x9|0x24}}
58 ; GCN: v_mov_b32_e32 [[VB:v[0-9]+]], s[[#LOAD + 3]]
61 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[#LOAD + 2]], [[VA0]], [[VB]]
62 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[#LOAD + 2]], [[VA1]], [[VB]]
63 ; GCN: buffer_store_dword [[RESULT0]]
64 ; GCN: buffer_store_dword [[RESULT1]]
65 define amdgpu_kernel void @test_use_s_v_s(ptr addrspace(1) %out, float %a, float %b, ptr addrspace(1) %in) #0 {
66 %va0 = load volatile float, ptr addrspace(1) %in
67 %va1 = load volatile float, ptr addrspace(1) %in
68 %fma0 = call float @llvm.fma.f32(float %a, float %va0, float %b) #1
69 %fma1 = call float @llvm.fma.f32(float %a, float %va1, float %b) #1
70 store volatile float %fma0, ptr addrspace(1) %out
71 store volatile float %fma1, ptr addrspace(1) %out
75 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_b_a:
76 ; SI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x9
77 ; VI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x24
78 ; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[#LOAD + 3]]
79 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], s[[#LOAD + 2]], [[VGPR1]], s[[#LOAD + 2]]
80 ; GCN: buffer_store_dword [[RESULT]]
81 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_b_a(ptr addrspace(1) %out, float %a, float %b) #0 {
82 %fma = call float @llvm.fma.f32(float %a, float %b, float %a) #1
83 store float %fma, ptr addrspace(1) %out, align 4
87 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_b_a_a:
88 ; SI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x9
89 ; VI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x24
90 ; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[#LOAD + 3]]
91 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[VGPR1]], s[[#LOAD + 2]], s[[#LOAD + 2]]
92 ; GCN: buffer_store_dword [[RESULT]]
93 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_b_a_a(ptr addrspace(1) %out, float %a, float %b) #0 {
94 %fma = call float @llvm.fma.f32(float %b, float %a, float %a) #1
95 store float %fma, ptr addrspace(1) %out, align 4
99 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_imm:
100 ; GCN: s_load_dword [[SGPR:s[0-9]+]]
101 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], 2.0
102 ; GCN: buffer_store_dword [[RESULT]]
103 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_imm(ptr addrspace(1) %out, float %a) #0 {
104 %fma = call float @llvm.fma.f32(float %a, float %a, float 2.0) #1
105 store float %fma, ptr addrspace(1) %out, align 4
109 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_imm_a:
110 ; GCN: s_load_dword [[SGPR:s[0-9]+]]
111 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], 2.0, [[SGPR]]
112 ; GCN: buffer_store_dword [[RESULT]]
113 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_imm_a(ptr addrspace(1) %out, float %a) #0 {
114 %fma = call float @llvm.fma.f32(float %a, float 2.0, float %a) #1
115 store float %fma, ptr addrspace(1) %out, align 4
119 ; Don't use fma since fma c, x, y is canonicalized to fma x, c, y
120 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_imm_a_a:
121 ; GCN: s_load_dword [[SGPR:s[0-9]+]]
122 ; GCN: v_div_fixup_f32 [[RESULT:v[0-9]+]], 2.0, [[SGPR]], [[SGPR]]
123 ; GCN: buffer_store_dword [[RESULT]]
124 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_imm_a_a(ptr addrspace(1) %out, float %a) #0 {
125 %val = call float @llvm.amdgcn.div.fixup.f32(float 2.0, float %a, float %a) #1
126 store float %val, ptr addrspace(1) %out, align 4
130 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_kimm:
131 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
132 ; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000
133 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], [[VK]]
134 ; GCN: buffer_store_dword [[RESULT]]
135 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_kimm(ptr addrspace(1) %out, float %a) #0 {
136 %fma = call float @llvm.fma.f32(float %a, float %a, float 1024.0) #1
137 store float %fma, ptr addrspace(1) %out, align 4
141 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_k_s:
142 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
143 ; GCN-DAG: v_mov_b32_e32 [[VGPR:v[0-9]+]], [[SGPR]]
144 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
145 ; GCN: v_fma_f32 [[RESULT0:v[0-9]+]], [[SK]], [[SK]], [[VGPR]]
146 ; GCN: buffer_store_dword [[RESULT0]]
147 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_k_s(ptr addrspace(1) %out, float %a) #0 {
148 %fma = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %a) #1
149 store float %fma, ptr addrspace(1) %out
153 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_k_s_x2:
154 ; GCN-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, {{0x9|0x24}}
155 ; GCN-DAG: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s[[#LOAD + 2]]
156 ; GCN-DAG: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[#LOAD + 3]]
157 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
158 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], [[SK]], [[SK]], [[VGPR0]]
159 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], [[SK]], [[SK]], [[VGPR1]]
160 ; GCN: buffer_store_dword [[RESULT0]]
161 ; GCN: buffer_store_dword [[RESULT1]]
163 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_k_s_x2(ptr addrspace(1) %out, float %a, float %b) #0 {
164 %fma0 = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %a) #1
165 %fma1 = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %b) #1
166 store volatile float %fma0, ptr addrspace(1) %out
167 store volatile float %fma1, ptr addrspace(1) %out
171 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_s_k:
172 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
173 ; GCN-DAG: v_mov_b32_e32 [[VGPR:v[0-9]+]], [[SGPR]]
174 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
175 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[VGPR]], [[SK]], [[SK]]
176 ; GCN: buffer_store_dword [[RESULT]]
177 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_s_k(ptr addrspace(1) %out, float %a) #0 {
178 %fma = call float @llvm.fma.f32(float 1024.0, float %a, float 1024.0) #1
179 store float %fma, ptr addrspace(1) %out
183 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_s_k_x2:
184 ; GCN-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, {{0x9|0x24}}
185 ; GCN-DAG: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s[[#LOAD + 2]]
186 ; GCN-DAG: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[#LOAD + 3]]
187 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
188 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], [[VGPR0]], [[SK]], [[SK]]
189 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], [[VGPR1]], [[SK]], [[SK]]
190 ; GCN: buffer_store_dword [[RESULT0]]
191 ; GCN: buffer_store_dword [[RESULT1]]
193 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_s_k_x2(ptr addrspace(1) %out, float %a, float %b) #0 {
194 %fma0 = call float @llvm.fma.f32(float 1024.0, float %a, float 1024.0) #1
195 %fma1 = call float @llvm.fma.f32(float 1024.0, float %b, float 1024.0) #1
196 store volatile float %fma0, ptr addrspace(1) %out
197 store volatile float %fma1, ptr addrspace(1) %out
201 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_s_k_k:
202 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
203 ; GCN-DAG: v_mov_b32_e32 [[VGPR:v[0-9]+]], [[SGPR]]
204 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
205 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[VGPR]], [[SK]], [[SK]]
206 ; GCN: buffer_store_dword [[RESULT]]
207 define amdgpu_kernel void @test_literal_use_twice_ternary_op_s_k_k(ptr addrspace(1) %out, float %a) #0 {
208 %fma = call float @llvm.fma.f32(float %a, float 1024.0, float 1024.0) #1
209 store float %fma, ptr addrspace(1) %out
213 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_s_k_k_x2:
214 ; GCN-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, {{0x9|0x24}}
215 ; GCN-DAG: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s[[#LOAD + 2]]
216 ; GCN-DAG: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[#LOAD + 3]]
217 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
218 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], [[VGPR0]], [[SK]], [[SK]]
219 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], [[VGPR1]], [[SK]], [[SK]]
220 ; GCN: buffer_store_dword [[RESULT0]]
221 ; GCN: buffer_store_dword [[RESULT1]]
223 define amdgpu_kernel void @test_literal_use_twice_ternary_op_s_k_k_x2(ptr addrspace(1) %out, float %a, float %b) #0 {
224 %fma0 = call float @llvm.fma.f32(float %a, float 1024.0, float 1024.0) #1
225 %fma1 = call float @llvm.fma.f32(float %b, float 1024.0, float 1024.0) #1
226 store volatile float %fma0, ptr addrspace(1) %out
227 store volatile float %fma1, ptr addrspace(1) %out
231 ; GCN-LABEL: {{^}}test_s0_s1_k_f32:
232 ; SI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x9
233 ; VI-DAG: s_load_dwordx4 s[[[#LOAD:]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 0x24
234 ; GCN-DAG: v_mov_b32_e32 [[VK0:v[0-9]+]], 0x44800000
235 ; GCN-DAG: v_mov_b32_e32 [[VS1:v[0-9]+]], s[[#LOAD + 3]]
237 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[#LOAD + 2]], [[VS1]], [[VK0]]
238 ; GCN-DAG: v_mov_b32_e32 [[VK1:v[0-9]+]], 0x45800000
239 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[#LOAD + 2]], [[VS1]], [[VK1]]
241 ; GCN: buffer_store_dword [[RESULT0]]
242 ; GCN: buffer_store_dword [[RESULT1]]
243 define amdgpu_kernel void @test_s0_s1_k_f32(ptr addrspace(1) %out, float %a, float %b) #0 {
244 %fma0 = call float @llvm.fma.f32(float %a, float %b, float 1024.0) #1
245 %fma1 = call float @llvm.fma.f32(float %a, float %b, float 4096.0) #1
246 store volatile float %fma0, ptr addrspace(1) %out
247 store volatile float %fma1, ptr addrspace(1) %out
251 ; FIXME: Immediate in SGPRs just copied to VGPRs
252 ; GCN-LABEL: {{^}}test_s0_s1_k_f64:
253 ; GCN-DAG: s_load_dwordx2 [[SGPR0:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, {{0x13|0x4c}}
254 ; GCN-DAG: s_load_dwordx2 s[[[SGPR1_SUB0:[0-9]+]]:[[SGPR1_SUB1:[0-9]+]]], s{{\[[0-9]+:[0-9]+\]}}, {{0x1d|0x74}}
255 ; GCN-DAG: v_mov_b32_e32 v[[VK0_SUB1:[0-9]+]], 0x40900000
256 ; GCN-DAG: v_mov_b32_e32 v[[VZERO:[0-9]+]], 0{{$}}
258 ; GCN-DAG: v_mov_b32_e32 v[[VS1_SUB0:[0-9]+]], s[[SGPR1_SUB0]]
259 ; GCN-DAG: v_mov_b32_e32 v[[VS1_SUB1:[0-9]+]], s[[SGPR1_SUB1]]
260 ; GCN-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], [[SGPR0]], v[[[VS1_SUB0]]:[[VS1_SUB1]]], v[[[VZERO]]:[[VK0_SUB1]]]
262 ; GCN-DAG: v_mov_b32_e32 v[[VK1_SUB1:[0-9]+]], 0x40b00000
263 ; GCN-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[SGPR0]], v[[[VS1_SUB0]]:[[VS1_SUB1]]], v[{{[0-9]+}}:[[VK1_SUB1]]]
265 ; GCN: buffer_store_dwordx2 [[RESULT0]]
266 ; GCN: buffer_store_dwordx2 [[RESULT1]]
267 define amdgpu_kernel void @test_s0_s1_k_f64(ptr addrspace(1) %out, [8 x i32], double %a, [8 x i32], double %b) #0 {
268 %fma0 = call double @llvm.fma.f64(double %a, double %b, double 1024.0) #1
269 %fma1 = call double @llvm.fma.f64(double %a, double %b, double 4096.0) #1
270 store volatile double %fma0, ptr addrspace(1) %out
271 store volatile double %fma1, ptr addrspace(1) %out
275 attributes #0 = { nounwind }
276 attributes #1 = { nounwind readnone }