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(float addrspace(1)* %out, float %a) #0 {
15 %dbl = fadd float %a, %a
16 store float %dbl, float 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(float addrspace(1)* %out, float %a) #0 {
25 %fma = call float @llvm.fma.f32(float %a, float %a, float %a) #1
26 store float %fma, float addrspace(1)* %out, align 4
30 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_b:
31 ; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb
32 ; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c
33 ; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]]
34 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], s[[SGPR0]], s[[SGPR0]], [[VGPR1]]
35 ; GCN: buffer_store_dword [[RESULT]]
36 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_b(float 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, float addrspace(1)* %out, align 4
42 ; GCN-LABEL: {{^}}test_use_s_v_s:
43 ; GCN-DAG: s_load_dwordx2 s{{\[}}[[SA:[0-9]+]]:[[SB:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
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)
57 ; GCN: v_mov_b32_e32 [[VB:v[0-9]+]], s[[SB]]
60 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[SA]], [[VA0]], [[VB]]
61 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[SA]], [[VA1]], [[VB]]
62 ; GCN: buffer_store_dword [[RESULT0]]
63 ; GCN: buffer_store_dword [[RESULT1]]
64 define amdgpu_kernel void @test_use_s_v_s(float addrspace(1)* %out, float %a, float %b, float addrspace(1)* %in) #0 {
65 %va0 = load volatile float, float addrspace(1)* %in
66 %va1 = load volatile float, float addrspace(1)* %in
67 %fma0 = call float @llvm.fma.f32(float %a, float %va0, float %b) #1
68 %fma1 = call float @llvm.fma.f32(float %a, float %va1, float %b) #1
69 store volatile float %fma0, float addrspace(1)* %out
70 store volatile float %fma1, float addrspace(1)* %out
74 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_b_a:
75 ; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb
76 ; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c
77 ; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]]
78 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], s[[SGPR0]], [[VGPR1]], s[[SGPR0]]
79 ; GCN: buffer_store_dword [[RESULT]]
80 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_b_a(float addrspace(1)* %out, float %a, float %b) #0 {
81 %fma = call float @llvm.fma.f32(float %a, float %b, float %a) #1
82 store float %fma, float addrspace(1)* %out, align 4
86 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_b_a_a:
87 ; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb
88 ; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c
89 ; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]]
90 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[VGPR1]], s[[SGPR0]], s[[SGPR0]]
91 ; GCN: buffer_store_dword [[RESULT]]
92 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_b_a_a(float addrspace(1)* %out, float %a, float %b) #0 {
93 %fma = call float @llvm.fma.f32(float %b, float %a, float %a) #1
94 store float %fma, float addrspace(1)* %out, align 4
98 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_imm:
99 ; GCN: s_load_dword [[SGPR:s[0-9]+]]
100 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], 2.0
101 ; GCN: buffer_store_dword [[RESULT]]
102 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_imm(float addrspace(1)* %out, float %a) #0 {
103 %fma = call float @llvm.fma.f32(float %a, float %a, float 2.0) #1
104 store float %fma, float addrspace(1)* %out, align 4
108 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_imm_a:
109 ; GCN: s_load_dword [[SGPR:s[0-9]+]]
110 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], 2.0, [[SGPR]]
111 ; GCN: buffer_store_dword [[RESULT]]
112 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_imm_a(float addrspace(1)* %out, float %a) #0 {
113 %fma = call float @llvm.fma.f32(float %a, float 2.0, float %a) #1
114 store float %fma, float addrspace(1)* %out, align 4
118 ; Don't use fma since fma c, x, y is canonicalized to fma x, c, y
119 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_imm_a_a:
120 ; GCN: s_load_dword [[SGPR:s[0-9]+]]
121 ; GCN: v_div_fixup_f32 [[RESULT:v[0-9]+]], 2.0, [[SGPR]], [[SGPR]]
122 ; GCN: buffer_store_dword [[RESULT]]
123 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_imm_a_a(float addrspace(1)* %out, float %a) #0 {
124 %val = call float @llvm.amdgcn.div.fixup.f32(float 2.0, float %a, float %a) #1
125 store float %val, float addrspace(1)* %out, align 4
129 ; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_kimm:
130 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
131 ; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000
132 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], [[VK]]
133 ; GCN: buffer_store_dword [[RESULT]]
134 define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_kimm(float addrspace(1)* %out, float %a) #0 {
135 %fma = call float @llvm.fma.f32(float %a, float %a, float 1024.0) #1
136 store float %fma, float addrspace(1)* %out, align 4
140 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_k_s:
141 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
142 ; GCN-DAG: v_mov_b32_e32 [[VGPR:v[0-9]+]], [[SGPR]]
143 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
144 ; GCN: v_fma_f32 [[RESULT0:v[0-9]+]], [[SK]], [[SK]], [[VGPR]]
145 ; GCN: buffer_store_dword [[RESULT0]]
146 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_k_s(float addrspace(1)* %out, float %a) #0 {
147 %fma = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %a) #1
148 store float %fma, float addrspace(1)* %out
152 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_k_s_x2:
153 ; GCN-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]{{\:}}[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
154 ; GCN-DAG: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s[[SGPR0]]
155 ; GCN-DAG: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]]
156 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
157 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], [[SK]], [[SK]], [[VGPR0]]
158 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], [[SK]], [[SK]], [[VGPR1]]
159 ; GCN: buffer_store_dword [[RESULT0]]
160 ; GCN: buffer_store_dword [[RESULT1]]
162 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_k_s_x2(float addrspace(1)* %out, float %a, float %b) #0 {
163 %fma0 = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %a) #1
164 %fma1 = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %b) #1
165 store volatile float %fma0, float addrspace(1)* %out
166 store volatile float %fma1, float addrspace(1)* %out
170 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_s_k:
171 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
172 ; GCN-DAG: v_mov_b32_e32 [[VGPR:v[0-9]+]], [[SGPR]]
173 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
174 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[VGPR]], [[SK]], [[SK]]
175 ; GCN: buffer_store_dword [[RESULT]]
176 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_s_k(float addrspace(1)* %out, float %a) #0 {
177 %fma = call float @llvm.fma.f32(float 1024.0, float %a, float 1024.0) #1
178 store float %fma, float addrspace(1)* %out
182 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_s_k_x2:
183 ; GCN-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]{{\:}}[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
184 ; GCN-DAG: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s[[SGPR0]]
185 ; GCN-DAG: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]]
186 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
187 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], [[VGPR0]], [[SK]], [[SK]]
188 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], [[VGPR1]], [[SK]], [[SK]]
189 ; GCN: buffer_store_dword [[RESULT0]]
190 ; GCN: buffer_store_dword [[RESULT1]]
192 define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_s_k_x2(float addrspace(1)* %out, float %a, float %b) #0 {
193 %fma0 = call float @llvm.fma.f32(float 1024.0, float %a, float 1024.0) #1
194 %fma1 = call float @llvm.fma.f32(float 1024.0, float %b, float 1024.0) #1
195 store volatile float %fma0, float addrspace(1)* %out
196 store volatile float %fma1, float addrspace(1)* %out
200 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_s_k_k:
201 ; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]]
202 ; GCN-DAG: v_mov_b32_e32 [[VGPR:v[0-9]+]], [[SGPR]]
203 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
204 ; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[VGPR]], [[SK]], [[SK]]
205 ; GCN: buffer_store_dword [[RESULT]]
206 define amdgpu_kernel void @test_literal_use_twice_ternary_op_s_k_k(float addrspace(1)* %out, float %a) #0 {
207 %fma = call float @llvm.fma.f32(float %a, float 1024.0, float 1024.0) #1
208 store float %fma, float addrspace(1)* %out
212 ; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_s_k_k_x2:
213 ; GCN-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]{{\:}}[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
214 ; GCN-DAG: v_mov_b32_e32 [[VGPR0:v[0-9]+]], s[[SGPR0]]
215 ; GCN-DAG: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]]
216 ; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x44800000
217 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], [[VGPR0]], [[SK]], [[SK]]
218 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], [[VGPR1]], [[SK]], [[SK]]
219 ; GCN: buffer_store_dword [[RESULT0]]
220 ; GCN: buffer_store_dword [[RESULT1]]
222 define amdgpu_kernel void @test_literal_use_twice_ternary_op_s_k_k_x2(float addrspace(1)* %out, float %a, float %b) #0 {
223 %fma0 = call float @llvm.fma.f32(float %a, float 1024.0, float 1024.0) #1
224 %fma1 = call float @llvm.fma.f32(float %b, float 1024.0, float 1024.0) #1
225 store volatile float %fma0, float addrspace(1)* %out
226 store volatile float %fma1, float addrspace(1)* %out
230 ; GCN-LABEL: {{^}}test_s0_s1_k_f32:
231 ; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb
232 ; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c
233 ; GCN-DAG: v_mov_b32_e32 [[VK0:v[0-9]+]], 0x44800000
234 ; GCN-DAG: v_mov_b32_e32 [[VS1:v[0-9]+]], s[[SGPR1]]
236 ; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[SGPR0]], [[VS1]], [[VK0]]
237 ; GCN-DAG: v_mov_b32_e32 [[VK1:v[0-9]+]], 0x45800000
238 ; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[SGPR0]], [[VS1]], [[VK1]]
240 ; GCN: buffer_store_dword [[RESULT0]]
241 ; GCN: buffer_store_dword [[RESULT1]]
242 define amdgpu_kernel void @test_s0_s1_k_f32(float addrspace(1)* %out, float %a, float %b) #0 {
243 %fma0 = call float @llvm.fma.f32(float %a, float %b, float 1024.0) #1
244 %fma1 = call float @llvm.fma.f32(float %a, float %b, float 4096.0) #1
245 store volatile float %fma0, float addrspace(1)* %out
246 store volatile float %fma1, float addrspace(1)* %out
250 ; FIXME: Immediate in SGPRs just copied to VGPRs
251 ; GCN-LABEL: {{^}}test_s0_s1_k_f64:
252 ; GCN-DAG: s_load_dwordx2 [[SGPR0:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, {{0x13|0x4c}}
253 ; GCN-DAG: s_load_dwordx2 s{{\[}}[[SGPR1_SUB0:[0-9]+]]:[[SGPR1_SUB1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0x1d|0x74}}
254 ; GCN-DAG: v_mov_b32_e32 v[[VK0_SUB1:[0-9]+]], 0x40900000
255 ; GCN-DAG: v_mov_b32_e32 v[[VZERO:[0-9]+]], 0{{$}}
257 ; GCN-DAG: v_mov_b32_e32 v[[VS1_SUB0:[0-9]+]], s[[SGPR1_SUB0]]
258 ; GCN-DAG: v_mov_b32_e32 v[[VS1_SUB1:[0-9]+]], s[[SGPR1_SUB1]]
259 ; GCN: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], [[SGPR0]], v{{\[}}[[VS1_SUB0]]:[[VS1_SUB1]]{{\]}}, v{{\[}}[[VZERO]]:[[VK0_SUB1]]{{\]}}
261 ; Same zero component is re-used for half of each immediate.
262 ; GCN: v_mov_b32_e32 v[[VK1_SUB1:[0-9]+]], 0x40b00000
263 ; GCN: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[SGPR0]], v{{\[}}[[VS1_SUB0]]:[[VS1_SUB1]]{{\]}}, v{{\[}}[[VZERO]]:[[VK1_SUB1]]{{\]}}
265 ; GCN: buffer_store_dwordx2 [[RESULT0]]
266 ; GCN: buffer_store_dwordx2 [[RESULT1]]
267 define amdgpu_kernel void @test_s0_s1_k_f64(double 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, double addrspace(1)* %out
271 store volatile double %fma1, double addrspace(1)* %out
275 attributes #0 = { nounwind }
276 attributes #1 = { nounwind readnone }