| blob | 362e25fa932fc0ad633c230b68a741c63fe79ef1 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx900 -fp-contract=fast < %s | FileCheck -check-prefix=GFX9-CONTRACT %s
3 ; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx900 --denormal-fp-math=preserve-sign < %s | FileCheck -check-prefix=GFX9-DENORM %s
4 ; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1010 -fp-contract=fast < %s | FileCheck -check-prefix=GFX10-CONTRACT %s
5 ; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1010 --denormal-fp-math=preserve-sign < %s | FileCheck -check-prefix=GFX10-DENORM %s
6 ; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 -fp-contract=fast < %s | FileCheck -check-prefix=GFX11-CONTRACT %s
7 ; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 --denormal-fp-math=preserve-sign < %s | FileCheck -check-prefix=GFX11-DENORM %s
9 ; fadd (fma a, b, (fmul c, d)), e --> fma a, b, (fma c, d, e)
10 ; fadd e, (fma a, b, (fmul c, d)) --> fma a, b, (fma c, d, e)
12 define float @test_f32_add_mul(float %a, float %b, float %c, float %d, float %e) {
13 ; GFX9-CONTRACT-LABEL: test_f32_add_mul:
14 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
15 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
16 ; GFX9-CONTRACT-NEXT: v_fma_f32 v2, v2, v3, v4
17 ; GFX9-CONTRACT-NEXT: v_fma_f32 v0, v0, v1, v2
18 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
19 ;
20 ; GFX9-DENORM-LABEL: test_f32_add_mul:
21 ; GFX9-DENORM: ; %bb.0: ; %.entry
22 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
23 ; GFX9-DENORM-NEXT: v_mad_f32 v2, v2, v3, v4
24 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v2, v0, v1
25 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v0, v2
26 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
27 ;
28 ; GFX10-CONTRACT-LABEL: test_f32_add_mul:
29 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
30 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
31 ; GFX10-CONTRACT-NEXT: v_fma_f32 v2, v2, v3, v4
32 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v2, v0, v1
33 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
34 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
35 ;
36 ; GFX10-DENORM-LABEL: test_f32_add_mul:
37 ; GFX10-DENORM: ; %bb.0: ; %.entry
38 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
39 ; GFX10-DENORM-NEXT: v_fma_f32 v2, v2, v3, v4
40 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v2, v0, v1
41 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v0, v2
42 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
43 ;
44 ; GFX11-CONTRACT-LABEL: test_f32_add_mul:
45 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
46 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
47 ; GFX11-CONTRACT-NEXT: v_fma_f32 v2, v2, v3, v4
48 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
49 ; GFX11-CONTRACT-NEXT: v_fmac_f32_e32 v2, v0, v1
50 ; GFX11-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
51 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
52 ;
53 ; GFX11-DENORM-LABEL: test_f32_add_mul:
54 ; GFX11-DENORM: ; %bb.0: ; %.entry
55 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
56 ; GFX11-DENORM-NEXT: v_fma_f32 v2, v2, v3, v4
57 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
58 ; GFX11-DENORM-NEXT: v_fmac_f32_e32 v2, v0, v1
59 ; GFX11-DENORM-NEXT: v_mov_b32_e32 v0, v2
60 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
61 .entry:
62 %x = fmul fast float %c, %d
63 %y = call fast float @llvm.fmuladd.f32(float %a, float %b, float %x)
64 %z = fadd fast float %y, %e
65 ret float %z
66 }
68 define float @test_f32_add_mul_rhs(float %a, float %b, float %c, float %d, float %e) {
69 ; GFX9-CONTRACT-LABEL: test_f32_add_mul_rhs:
70 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
71 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
72 ; GFX9-CONTRACT-NEXT: v_fma_f32 v2, v2, v3, v4
73 ; GFX9-CONTRACT-NEXT: v_fma_f32 v0, v0, v1, v2
74 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
75 ;
76 ; GFX9-DENORM-LABEL: test_f32_add_mul_rhs:
77 ; GFX9-DENORM: ; %bb.0: ; %.entry
78 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
79 ; GFX9-DENORM-NEXT: v_mad_f32 v2, v2, v3, v4
80 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v2, v0, v1
81 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v0, v2
82 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
83 ;
84 ; GFX10-CONTRACT-LABEL: test_f32_add_mul_rhs:
85 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
86 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
87 ; GFX10-CONTRACT-NEXT: v_fma_f32 v2, v2, v3, v4
88 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v2, v0, v1
89 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
90 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
91 ;
92 ; GFX10-DENORM-LABEL: test_f32_add_mul_rhs:
93 ; GFX10-DENORM: ; %bb.0: ; %.entry
94 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
95 ; GFX10-DENORM-NEXT: v_fma_f32 v2, v2, v3, v4
96 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v2, v0, v1
97 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v0, v2
98 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
99 ;
100 ; GFX11-CONTRACT-LABEL: test_f32_add_mul_rhs:
101 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
102 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
103 ; GFX11-CONTRACT-NEXT: v_fma_f32 v2, v2, v3, v4
104 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
105 ; GFX11-CONTRACT-NEXT: v_fmac_f32_e32 v2, v0, v1
106 ; GFX11-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
107 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
108 ;
109 ; GFX11-DENORM-LABEL: test_f32_add_mul_rhs:
110 ; GFX11-DENORM: ; %bb.0: ; %.entry
111 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
112 ; GFX11-DENORM-NEXT: v_fma_f32 v2, v2, v3, v4
113 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
114 ; GFX11-DENORM-NEXT: v_fmac_f32_e32 v2, v0, v1
115 ; GFX11-DENORM-NEXT: v_mov_b32_e32 v0, v2
116 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
117 .entry:
118 %x = fmul fast float %c, %d
119 %y = call fast float @llvm.fmuladd.f32(float %a, float %b, float %x)
120 %z = fadd fast float %e, %y
121 ret float %z
122 }
124 define half @test_half_add_mul(half %a, half %b, half %c, half %d, half %e) {
125 ; GFX9-CONTRACT-LABEL: test_half_add_mul:
126 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
127 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
128 ; GFX9-CONTRACT-NEXT: v_fma_f16 v2, v2, v3, v4
129 ; GFX9-CONTRACT-NEXT: v_fma_f16 v0, v0, v1, v2
130 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
131 ;
132 ; GFX9-DENORM-LABEL: test_half_add_mul:
133 ; GFX9-DENORM: ; %bb.0: ; %.entry
134 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
135 ; GFX9-DENORM-NEXT: v_mad_legacy_f16 v2, v2, v3, v4
136 ; GFX9-DENORM-NEXT: v_mac_f16_e32 v2, v0, v1
137 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v0, v2
138 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
139 ;
140 ; GFX10-CONTRACT-LABEL: test_half_add_mul:
141 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
142 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
143 ; GFX10-CONTRACT-NEXT: v_fma_f16 v2, v2, v3, v4
144 ; GFX10-CONTRACT-NEXT: v_fmac_f16_e32 v2, v0, v1
145 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
146 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
147 ;
148 ; GFX10-DENORM-LABEL: test_half_add_mul:
149 ; GFX10-DENORM: ; %bb.0: ; %.entry
150 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
151 ; GFX10-DENORM-NEXT: v_mul_f16_e32 v2, v2, v3
152 ; GFX10-DENORM-NEXT: v_mul_f16_e32 v0, v0, v1
153 ; GFX10-DENORM-NEXT: v_add_f16_e32 v0, v0, v2
154 ; GFX10-DENORM-NEXT: v_add_f16_e32 v0, v0, v4
155 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
156 ;
157 ; GFX11-CONTRACT-LABEL: test_half_add_mul:
158 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
159 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
160 ; GFX11-CONTRACT-NEXT: v_fma_f16 v2, v2, v3, v4
161 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
162 ; GFX11-CONTRACT-NEXT: v_fmac_f16_e32 v2, v0, v1
163 ; GFX11-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
164 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
165 ;
166 ; GFX11-DENORM-LABEL: test_half_add_mul:
167 ; GFX11-DENORM: ; %bb.0: ; %.entry
168 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
169 ; GFX11-DENORM-NEXT: v_mul_f16_e32 v2, v2, v3
170 ; GFX11-DENORM-NEXT: v_mul_f16_e32 v0, v0, v1
171 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
172 ; GFX11-DENORM-NEXT: v_add_f16_e32 v0, v0, v2
173 ; GFX11-DENORM-NEXT: v_add_f16_e32 v0, v0, v4
174 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
175 .entry:
176 %x = fmul fast half %c, %d
177 %y = call fast half @llvm.fmuladd.f16(half %a, half %b, half %x)
178 %z = fadd fast half %y, %e
179 ret half %z
180 }
182 define half @test_half_add_mul_rhs(half %a, half %b, half %c, half %d, half %e) {
183 ; GFX9-CONTRACT-LABEL: test_half_add_mul_rhs:
184 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
185 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
186 ; GFX9-CONTRACT-NEXT: v_fma_f16 v2, v2, v3, v4
187 ; GFX9-CONTRACT-NEXT: v_fma_f16 v0, v0, v1, v2
188 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
189 ;
190 ; GFX9-DENORM-LABEL: test_half_add_mul_rhs:
191 ; GFX9-DENORM: ; %bb.0: ; %.entry
192 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
193 ; GFX9-DENORM-NEXT: v_mad_legacy_f16 v2, v2, v3, v4
194 ; GFX9-DENORM-NEXT: v_mac_f16_e32 v2, v0, v1
195 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v0, v2
196 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
197 ;
198 ; GFX10-CONTRACT-LABEL: test_half_add_mul_rhs:
199 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
200 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
201 ; GFX10-CONTRACT-NEXT: v_fma_f16 v2, v2, v3, v4
202 ; GFX10-CONTRACT-NEXT: v_fmac_f16_e32 v2, v0, v1
203 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
204 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
205 ;
206 ; GFX10-DENORM-LABEL: test_half_add_mul_rhs:
207 ; GFX10-DENORM: ; %bb.0: ; %.entry
208 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
209 ; GFX10-DENORM-NEXT: v_mul_f16_e32 v2, v2, v3
210 ; GFX10-DENORM-NEXT: v_mul_f16_e32 v0, v0, v1
211 ; GFX10-DENORM-NEXT: v_add_f16_e32 v0, v0, v2
212 ; GFX10-DENORM-NEXT: v_add_f16_e32 v0, v4, v0
213 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
214 ;
215 ; GFX11-CONTRACT-LABEL: test_half_add_mul_rhs:
216 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
217 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
218 ; GFX11-CONTRACT-NEXT: v_fma_f16 v2, v2, v3, v4
219 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
220 ; GFX11-CONTRACT-NEXT: v_fmac_f16_e32 v2, v0, v1
221 ; GFX11-CONTRACT-NEXT: v_mov_b32_e32 v0, v2
222 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
223 ;
224 ; GFX11-DENORM-LABEL: test_half_add_mul_rhs:
225 ; GFX11-DENORM: ; %bb.0: ; %.entry
226 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
227 ; GFX11-DENORM-NEXT: v_mul_f16_e32 v2, v2, v3
228 ; GFX11-DENORM-NEXT: v_mul_f16_e32 v0, v0, v1
229 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
230 ; GFX11-DENORM-NEXT: v_add_f16_e32 v0, v0, v2
231 ; GFX11-DENORM-NEXT: v_add_f16_e32 v0, v4, v0
232 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
233 .entry:
234 %x = fmul fast half %c, %d
235 %y = call fast half @llvm.fmuladd.f16(half %a, half %b, half %x)
236 %z = fadd fast half %e, %y
237 ret half %z
238 }
240 define double @test_double_add_mul(double %a, double %b, double %c, double %d, double %e) {
241 ; GFX9-CONTRACT-LABEL: test_double_add_mul:
242 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
243 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
244 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
245 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
246 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
247 ;
248 ; GFX9-DENORM-LABEL: test_double_add_mul:
249 ; GFX9-DENORM: ; %bb.0: ; %.entry
250 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
251 ; GFX9-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
252 ; GFX9-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
253 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
254 ;
255 ; GFX10-CONTRACT-LABEL: test_double_add_mul:
256 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
257 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
258 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
259 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
260 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
261 ;
262 ; GFX10-DENORM-LABEL: test_double_add_mul:
263 ; GFX10-DENORM: ; %bb.0: ; %.entry
264 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
265 ; GFX10-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
266 ; GFX10-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
267 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
268 ;
269 ; GFX11-CONTRACT-LABEL: test_double_add_mul:
270 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
271 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
272 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
273 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_1)
274 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
275 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
276 ;
277 ; GFX11-DENORM-LABEL: test_double_add_mul:
278 ; GFX11-DENORM: ; %bb.0: ; %.entry
279 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
280 ; GFX11-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
281 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_1)
282 ; GFX11-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
283 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
284 .entry:
285 %x = fmul fast double %c, %d
286 %y = call fast double @llvm.fmuladd.f64(double %a, double %b, double %x)
287 %z = fadd fast double %y, %e
288 ret double %z
289 }
291 define double @test_double_add_mul_rhs(double %a, double %b, double %c, double %d, double %e) {
292 ; GFX9-CONTRACT-LABEL: test_double_add_mul_rhs:
293 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
294 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
295 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
296 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
297 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
298 ;
299 ; GFX9-DENORM-LABEL: test_double_add_mul_rhs:
300 ; GFX9-DENORM: ; %bb.0: ; %.entry
301 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
302 ; GFX9-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
303 ; GFX9-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
304 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
305 ;
306 ; GFX10-CONTRACT-LABEL: test_double_add_mul_rhs:
307 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
308 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
309 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
310 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
311 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
312 ;
313 ; GFX10-DENORM-LABEL: test_double_add_mul_rhs:
314 ; GFX10-DENORM: ; %bb.0: ; %.entry
315 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
316 ; GFX10-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
317 ; GFX10-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
318 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
319 ;
320 ; GFX11-CONTRACT-LABEL: test_double_add_mul_rhs:
321 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
322 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
323 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
324 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_1)
325 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
326 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
327 ;
328 ; GFX11-DENORM-LABEL: test_double_add_mul_rhs:
329 ; GFX11-DENORM: ; %bb.0: ; %.entry
330 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
331 ; GFX11-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9]
332 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_1)
333 ; GFX11-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5]
334 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
335 .entry:
336 %x = fmul fast double %c, %d
337 %y = call fast double @llvm.fmuladd.f64(double %a, double %b, double %x)
338 %z = fadd fast double %e, %y
339 ret double %z
340 }
342 define <4 x float> @test_v4f32_add_mul(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> %e) {
343 ; GFX9-CONTRACT-LABEL: test_v4f32_add_mul:
344 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
345 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
346 ; GFX9-CONTRACT-NEXT: v_fma_f32 v8, v8, v12, v16
347 ; GFX9-CONTRACT-NEXT: v_fma_f32 v9, v9, v13, v17
348 ; GFX9-CONTRACT-NEXT: v_fma_f32 v10, v10, v14, v18
349 ; GFX9-CONTRACT-NEXT: v_fma_f32 v11, v11, v15, v19
350 ; GFX9-CONTRACT-NEXT: v_fma_f32 v0, v0, v4, v8
351 ; GFX9-CONTRACT-NEXT: v_fma_f32 v1, v1, v5, v9
352 ; GFX9-CONTRACT-NEXT: v_fma_f32 v2, v2, v6, v10
353 ; GFX9-CONTRACT-NEXT: v_fma_f32 v3, v3, v7, v11
354 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
355 ;
356 ; GFX9-DENORM-LABEL: test_v4f32_add_mul:
357 ; GFX9-DENORM: ; %bb.0: ; %.entry
358 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
359 ; GFX9-DENORM-NEXT: v_mad_f32 v8, v8, v12, v16
360 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v8, v0, v4
361 ; GFX9-DENORM-NEXT: v_mad_f32 v4, v9, v13, v17
362 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v4, v1, v5
363 ; GFX9-DENORM-NEXT: v_mad_f32 v5, v10, v14, v18
364 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v5, v2, v6
365 ; GFX9-DENORM-NEXT: v_mad_f32 v6, v11, v15, v19
366 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v6, v3, v7
367 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v0, v8
368 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v1, v4
369 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v2, v5
370 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v3, v6
371 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
372 ;
373 ; GFX10-CONTRACT-LABEL: test_v4f32_add_mul:
374 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
375 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
376 ; GFX10-CONTRACT-NEXT: v_fma_f32 v8, v8, v12, v16
377 ; GFX10-CONTRACT-NEXT: v_fma_f32 v9, v9, v13, v17
378 ; GFX10-CONTRACT-NEXT: v_fma_f32 v10, v10, v14, v18
379 ; GFX10-CONTRACT-NEXT: v_fma_f32 v11, v11, v15, v19
380 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v8, v0, v4
381 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v9, v1, v5
382 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v10, v2, v6
383 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v11, v3, v7
384 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v0, v8
385 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v1, v9
386 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v2, v10
387 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v3, v11
388 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
389 ;
390 ; GFX10-DENORM-LABEL: test_v4f32_add_mul:
391 ; GFX10-DENORM: ; %bb.0: ; %.entry
392 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
393 ; GFX10-DENORM-NEXT: v_fma_f32 v8, v8, v12, v16
394 ; GFX10-DENORM-NEXT: v_fma_f32 v9, v9, v13, v17
395 ; GFX10-DENORM-NEXT: v_fma_f32 v10, v10, v14, v18
396 ; GFX10-DENORM-NEXT: v_fma_f32 v11, v11, v15, v19
397 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v8, v0, v4
398 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v9, v1, v5
399 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v10, v2, v6
400 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v11, v3, v7
401 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v0, v8
402 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v1, v9
403 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v2, v10
404 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v3, v11
405 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
406 ;
407 ; GFX11-CONTRACT-LABEL: test_v4f32_add_mul:
408 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
409 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
410 ; GFX11-CONTRACT-NEXT: v_fma_f32 v8, v8, v12, v16
411 ; GFX11-CONTRACT-NEXT: v_fma_f32 v9, v9, v13, v17
412 ; GFX11-CONTRACT-NEXT: v_fma_f32 v10, v10, v14, v18
413 ; GFX11-CONTRACT-NEXT: v_fma_f32 v11, v11, v15, v19
414 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
415 ; GFX11-CONTRACT-NEXT: v_dual_fmac_f32 v8, v0, v4 :: v_dual_fmac_f32 v9, v1, v5
416 ; GFX11-CONTRACT-NEXT: v_dual_fmac_f32 v10, v2, v6 :: v_dual_fmac_f32 v11, v3, v7
417 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
418 ; GFX11-CONTRACT-NEXT: v_dual_mov_b32 v0, v8 :: v_dual_mov_b32 v1, v9
419 ; GFX11-CONTRACT-NEXT: v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v3, v11
420 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
421 ;
422 ; GFX11-DENORM-LABEL: test_v4f32_add_mul:
423 ; GFX11-DENORM: ; %bb.0: ; %.entry
424 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
425 ; GFX11-DENORM-NEXT: v_fma_f32 v8, v8, v12, v16
426 ; GFX11-DENORM-NEXT: v_fma_f32 v9, v9, v13, v17
427 ; GFX11-DENORM-NEXT: v_fma_f32 v10, v10, v14, v18
428 ; GFX11-DENORM-NEXT: v_fma_f32 v11, v11, v15, v19
429 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
430 ; GFX11-DENORM-NEXT: v_dual_fmac_f32 v8, v0, v4 :: v_dual_fmac_f32 v9, v1, v5
431 ; GFX11-DENORM-NEXT: v_dual_fmac_f32 v10, v2, v6 :: v_dual_fmac_f32 v11, v3, v7
432 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
433 ; GFX11-DENORM-NEXT: v_dual_mov_b32 v0, v8 :: v_dual_mov_b32 v1, v9
434 ; GFX11-DENORM-NEXT: v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v3, v11
435 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
436 .entry:
437 %x = fmul fast <4 x float> %c, %d
438 %y = call fast <4 x float> @llvm.fmuladd.v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %x)
439 %z = fadd fast <4 x float> %y, %e
440 ret <4 x float> %z
441 }
443 define <4 x float> @test_v4f32_add_mul_rhs(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> %e) {
444 ; GFX9-CONTRACT-LABEL: test_v4f32_add_mul_rhs:
445 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
446 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
447 ; GFX9-CONTRACT-NEXT: v_fma_f32 v8, v8, v12, v16
448 ; GFX9-CONTRACT-NEXT: v_fma_f32 v9, v9, v13, v17
449 ; GFX9-CONTRACT-NEXT: v_fma_f32 v10, v10, v14, v18
450 ; GFX9-CONTRACT-NEXT: v_fma_f32 v11, v11, v15, v19
451 ; GFX9-CONTRACT-NEXT: v_fma_f32 v0, v0, v4, v8
452 ; GFX9-CONTRACT-NEXT: v_fma_f32 v1, v1, v5, v9
453 ; GFX9-CONTRACT-NEXT: v_fma_f32 v2, v2, v6, v10
454 ; GFX9-CONTRACT-NEXT: v_fma_f32 v3, v3, v7, v11
455 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
456 ;
457 ; GFX9-DENORM-LABEL: test_v4f32_add_mul_rhs:
458 ; GFX9-DENORM: ; %bb.0: ; %.entry
459 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
460 ; GFX9-DENORM-NEXT: v_mad_f32 v8, v8, v12, v16
461 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v8, v0, v4
462 ; GFX9-DENORM-NEXT: v_mad_f32 v4, v9, v13, v17
463 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v4, v1, v5
464 ; GFX9-DENORM-NEXT: v_mad_f32 v5, v10, v14, v18
465 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v5, v2, v6
466 ; GFX9-DENORM-NEXT: v_mad_f32 v6, v11, v15, v19
467 ; GFX9-DENORM-NEXT: v_mac_f32_e32 v6, v3, v7
468 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v0, v8
469 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v1, v4
470 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v2, v5
471 ; GFX9-DENORM-NEXT: v_mov_b32_e32 v3, v6
472 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
473 ;
474 ; GFX10-CONTRACT-LABEL: test_v4f32_add_mul_rhs:
475 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
476 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
477 ; GFX10-CONTRACT-NEXT: v_fma_f32 v8, v8, v12, v16
478 ; GFX10-CONTRACT-NEXT: v_fma_f32 v9, v9, v13, v17
479 ; GFX10-CONTRACT-NEXT: v_fma_f32 v10, v10, v14, v18
480 ; GFX10-CONTRACT-NEXT: v_fma_f32 v11, v11, v15, v19
481 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v8, v0, v4
482 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v9, v1, v5
483 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v10, v2, v6
484 ; GFX10-CONTRACT-NEXT: v_fmac_f32_e32 v11, v3, v7
485 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v0, v8
486 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v1, v9
487 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v2, v10
488 ; GFX10-CONTRACT-NEXT: v_mov_b32_e32 v3, v11
489 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
490 ;
491 ; GFX10-DENORM-LABEL: test_v4f32_add_mul_rhs:
492 ; GFX10-DENORM: ; %bb.0: ; %.entry
493 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
494 ; GFX10-DENORM-NEXT: v_fma_f32 v8, v8, v12, v16
495 ; GFX10-DENORM-NEXT: v_fma_f32 v9, v9, v13, v17
496 ; GFX10-DENORM-NEXT: v_fma_f32 v10, v10, v14, v18
497 ; GFX10-DENORM-NEXT: v_fma_f32 v11, v11, v15, v19
498 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v8, v0, v4
499 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v9, v1, v5
500 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v10, v2, v6
501 ; GFX10-DENORM-NEXT: v_fmac_f32_e32 v11, v3, v7
502 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v0, v8
503 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v1, v9
504 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v2, v10
505 ; GFX10-DENORM-NEXT: v_mov_b32_e32 v3, v11
506 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
507 ;
508 ; GFX11-CONTRACT-LABEL: test_v4f32_add_mul_rhs:
509 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
510 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
511 ; GFX11-CONTRACT-NEXT: v_fma_f32 v8, v8, v12, v16
512 ; GFX11-CONTRACT-NEXT: v_fma_f32 v9, v9, v13, v17
513 ; GFX11-CONTRACT-NEXT: v_fma_f32 v10, v10, v14, v18
514 ; GFX11-CONTRACT-NEXT: v_fma_f32 v11, v11, v15, v19
515 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
516 ; GFX11-CONTRACT-NEXT: v_dual_fmac_f32 v8, v0, v4 :: v_dual_fmac_f32 v9, v1, v5
517 ; GFX11-CONTRACT-NEXT: v_dual_fmac_f32 v10, v2, v6 :: v_dual_fmac_f32 v11, v3, v7
518 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
519 ; GFX11-CONTRACT-NEXT: v_dual_mov_b32 v0, v8 :: v_dual_mov_b32 v1, v9
520 ; GFX11-CONTRACT-NEXT: v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v3, v11
521 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
522 ;
523 ; GFX11-DENORM-LABEL: test_v4f32_add_mul_rhs:
524 ; GFX11-DENORM: ; %bb.0: ; %.entry
525 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
526 ; GFX11-DENORM-NEXT: v_fma_f32 v8, v8, v12, v16
527 ; GFX11-DENORM-NEXT: v_fma_f32 v9, v9, v13, v17
528 ; GFX11-DENORM-NEXT: v_fma_f32 v10, v10, v14, v18
529 ; GFX11-DENORM-NEXT: v_fma_f32 v11, v11, v15, v19
530 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
531 ; GFX11-DENORM-NEXT: v_dual_fmac_f32 v8, v0, v4 :: v_dual_fmac_f32 v9, v1, v5
532 ; GFX11-DENORM-NEXT: v_dual_fmac_f32 v10, v2, v6 :: v_dual_fmac_f32 v11, v3, v7
533 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
534 ; GFX11-DENORM-NEXT: v_dual_mov_b32 v0, v8 :: v_dual_mov_b32 v1, v9
535 ; GFX11-DENORM-NEXT: v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v3, v11
536 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
537 .entry:
538 %x = fmul fast <4 x float> %c, %d
539 %y = call fast <4 x float> @llvm.fmuladd.v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %x)
540 %z = fadd fast <4 x float> %e, %y
541 ret <4 x float> %z
542 }
544 define <4 x half> @test_f16_add_mul(<4 x half> %a, <4 x half> %b, <4 x half> %c, <4 x half> %d, <4 x half> %e) {
545 ; GFX9-CONTRACT-LABEL: test_f16_add_mul:
546 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
547 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
548 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v4, v4, v6, v8
549 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v5, v5, v7, v9
550 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v0, v0, v2, v4
551 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v1, v1, v3, v5
552 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
553 ;
554 ; GFX9-DENORM-LABEL: test_f16_add_mul:
555 ; GFX9-DENORM: ; %bb.0: ; %.entry
556 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
557 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v4, v4, v6
558 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v5, v5, v7
559 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v0, v0, v2
560 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v1, v1, v3
561 ; GFX9-DENORM-NEXT: v_pk_add_f16 v0, v0, v4
562 ; GFX9-DENORM-NEXT: v_pk_add_f16 v1, v1, v5
563 ; GFX9-DENORM-NEXT: v_pk_add_f16 v0, v0, v8
564 ; GFX9-DENORM-NEXT: v_pk_add_f16 v1, v1, v9
565 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
566 ;
567 ; GFX10-CONTRACT-LABEL: test_f16_add_mul:
568 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
569 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
570 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v4, v4, v6, v8
571 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v5, v5, v7, v9
572 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v0, v0, v2, v4
573 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v1, v1, v3, v5
574 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
575 ;
576 ; GFX10-DENORM-LABEL: test_f16_add_mul:
577 ; GFX10-DENORM: ; %bb.0: ; %.entry
578 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
579 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v4, v4, v6
580 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v5, v5, v7
581 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v0, v0, v2
582 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v1, v1, v3
583 ; GFX10-DENORM-NEXT: v_pk_add_f16 v0, v0, v4
584 ; GFX10-DENORM-NEXT: v_pk_add_f16 v1, v1, v5
585 ; GFX10-DENORM-NEXT: v_pk_add_f16 v0, v0, v8
586 ; GFX10-DENORM-NEXT: v_pk_add_f16 v1, v1, v9
587 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
588 ;
589 ; GFX11-CONTRACT-LABEL: test_f16_add_mul:
590 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
591 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
592 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v4, v4, v6, v8
593 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v5, v5, v7, v9
594 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
595 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v0, v0, v2, v4
596 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v1, v1, v3, v5
597 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
598 ;
599 ; GFX11-DENORM-LABEL: test_f16_add_mul:
600 ; GFX11-DENORM: ; %bb.0: ; %.entry
601 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
602 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v4, v4, v6
603 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v5, v5, v7
604 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v0, v0, v2
605 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v1, v1, v3
606 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
607 ; GFX11-DENORM-NEXT: v_pk_add_f16 v0, v0, v4
608 ; GFX11-DENORM-NEXT: v_pk_add_f16 v1, v1, v5
609 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
610 ; GFX11-DENORM-NEXT: v_pk_add_f16 v0, v0, v8
611 ; GFX11-DENORM-NEXT: v_pk_add_f16 v1, v1, v9
612 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
613 .entry:
614 %x = fmul fast <4 x half> %c, %d
615 %y = call fast <4 x half> @llvm.fmuladd.v4f16(<4 x half> %a, <4 x half> %b, <4 x half> %x)
616 %z = fadd fast <4 x half> %y, %e
617 ret <4 x half> %z
618 }
620 define <4 x half> @test_f16_add_mul_rhs(<4 x half> %a, <4 x half> %b, <4 x half> %c, <4 x half> %d, <4 x half> %e) {
621 ; GFX9-CONTRACT-LABEL: test_f16_add_mul_rhs:
622 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
623 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
624 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v4, v4, v6, v8
625 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v5, v5, v7, v9
626 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v0, v0, v2, v4
627 ; GFX9-CONTRACT-NEXT: v_pk_fma_f16 v1, v1, v3, v5
628 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
629 ;
630 ; GFX9-DENORM-LABEL: test_f16_add_mul_rhs:
631 ; GFX9-DENORM: ; %bb.0: ; %.entry
632 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
633 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v4, v4, v6
634 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v5, v5, v7
635 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v0, v0, v2
636 ; GFX9-DENORM-NEXT: v_pk_mul_f16 v1, v1, v3
637 ; GFX9-DENORM-NEXT: v_pk_add_f16 v0, v0, v4
638 ; GFX9-DENORM-NEXT: v_pk_add_f16 v1, v1, v5
639 ; GFX9-DENORM-NEXT: v_pk_add_f16 v0, v8, v0
640 ; GFX9-DENORM-NEXT: v_pk_add_f16 v1, v9, v1
641 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
642 ;
643 ; GFX10-CONTRACT-LABEL: test_f16_add_mul_rhs:
644 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
645 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
646 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v4, v4, v6, v8
647 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v5, v5, v7, v9
648 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v0, v0, v2, v4
649 ; GFX10-CONTRACT-NEXT: v_pk_fma_f16 v1, v1, v3, v5
650 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
651 ;
652 ; GFX10-DENORM-LABEL: test_f16_add_mul_rhs:
653 ; GFX10-DENORM: ; %bb.0: ; %.entry
654 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
655 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v4, v4, v6
656 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v5, v5, v7
657 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v0, v0, v2
658 ; GFX10-DENORM-NEXT: v_pk_mul_f16 v1, v1, v3
659 ; GFX10-DENORM-NEXT: v_pk_add_f16 v0, v0, v4
660 ; GFX10-DENORM-NEXT: v_pk_add_f16 v1, v1, v5
661 ; GFX10-DENORM-NEXT: v_pk_add_f16 v0, v8, v0
662 ; GFX10-DENORM-NEXT: v_pk_add_f16 v1, v9, v1
663 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
664 ;
665 ; GFX11-CONTRACT-LABEL: test_f16_add_mul_rhs:
666 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
667 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
668 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v4, v4, v6, v8
669 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v5, v5, v7, v9
670 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
671 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v0, v0, v2, v4
672 ; GFX11-CONTRACT-NEXT: v_pk_fma_f16 v1, v1, v3, v5
673 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
674 ;
675 ; GFX11-DENORM-LABEL: test_f16_add_mul_rhs:
676 ; GFX11-DENORM: ; %bb.0: ; %.entry
677 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
678 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v4, v4, v6
679 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v5, v5, v7
680 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v0, v0, v2
681 ; GFX11-DENORM-NEXT: v_pk_mul_f16 v1, v1, v3
682 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
683 ; GFX11-DENORM-NEXT: v_pk_add_f16 v0, v0, v4
684 ; GFX11-DENORM-NEXT: v_pk_add_f16 v1, v1, v5
685 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
686 ; GFX11-DENORM-NEXT: v_pk_add_f16 v0, v8, v0
687 ; GFX11-DENORM-NEXT: v_pk_add_f16 v1, v9, v1
688 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
689 .entry:
690 %x = fmul fast <4 x half> %c, %d
691 %y = call fast <4 x half> @llvm.fmuladd.v4f16(<4 x half> %a, <4 x half> %b, <4 x half> %x)
692 %z = fadd fast <4 x half> %e, %y
693 ret <4 x half> %z
694 }
696 define <4 x double> @test_f64_add_mul(<4 x double> %a, <4 x double> %b, <4 x double> %c, <4 x double> %d, <4 x double> %e) {
697 ; GFX9-CONTRACT-LABEL: test_f64_add_mul:
698 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
699 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
700 ; GFX9-CONTRACT-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
701 ; GFX9-CONTRACT-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
702 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
703 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[31:32]
704 ; GFX9-CONTRACT-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:12
705 ; GFX9-CONTRACT-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:16
706 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
707 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
708 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[24:25]
709 ; GFX9-CONTRACT-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:20
710 ; GFX9-CONTRACT-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:24
711 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
712 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
713 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[24:25]
714 ; GFX9-CONTRACT-NEXT: buffer_load_dword v31, off, s[0:3], s32
715 ; GFX9-CONTRACT-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
716 ; GFX9-CONTRACT-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:32
717 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
718 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
719 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[24:25]
720 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
721 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
722 ;
723 ; GFX9-DENORM-LABEL: test_f64_add_mul:
724 ; GFX9-DENORM: ; %bb.0: ; %.entry
725 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
726 ; GFX9-DENORM-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
727 ; GFX9-DENORM-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
728 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
729 ; GFX9-DENORM-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[31:32]
730 ; GFX9-DENORM-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:12
731 ; GFX9-DENORM-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:16
732 ; GFX9-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
733 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
734 ; GFX9-DENORM-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[24:25]
735 ; GFX9-DENORM-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:20
736 ; GFX9-DENORM-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:24
737 ; GFX9-DENORM-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
738 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
739 ; GFX9-DENORM-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[24:25]
740 ; GFX9-DENORM-NEXT: buffer_load_dword v31, off, s[0:3], s32
741 ; GFX9-DENORM-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
742 ; GFX9-DENORM-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:32
743 ; GFX9-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
744 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
745 ; GFX9-DENORM-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[24:25]
746 ; GFX9-DENORM-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
747 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
748 ;
749 ; GFX10-CONTRACT-LABEL: test_f64_add_mul:
750 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
751 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
752 ; GFX10-CONTRACT-NEXT: s_clause 0x8
753 ; GFX10-CONTRACT-NEXT: buffer_load_dword v31, off, s[0:3], s32
754 ; GFX10-CONTRACT-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
755 ; GFX10-CONTRACT-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
756 ; GFX10-CONTRACT-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:12
757 ; GFX10-CONTRACT-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:16
758 ; GFX10-CONTRACT-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:20
759 ; GFX10-CONTRACT-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:24
760 ; GFX10-CONTRACT-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:28
761 ; GFX10-CONTRACT-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:32
762 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(6)
763 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
764 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(4)
765 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
766 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(2)
767 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
768 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0)
769 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
770 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
771 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
772 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
773 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
774 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
775 ;
776 ; GFX10-DENORM-LABEL: test_f64_add_mul:
777 ; GFX10-DENORM: ; %bb.0: ; %.entry
778 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
779 ; GFX10-DENORM-NEXT: s_clause 0x8
780 ; GFX10-DENORM-NEXT: buffer_load_dword v31, off, s[0:3], s32
781 ; GFX10-DENORM-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
782 ; GFX10-DENORM-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
783 ; GFX10-DENORM-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:12
784 ; GFX10-DENORM-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:16
785 ; GFX10-DENORM-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:20
786 ; GFX10-DENORM-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:24
787 ; GFX10-DENORM-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:28
788 ; GFX10-DENORM-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:32
789 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(6)
790 ; GFX10-DENORM-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
791 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(4)
792 ; GFX10-DENORM-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
793 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(2)
794 ; GFX10-DENORM-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
795 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0)
796 ; GFX10-DENORM-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
797 ; GFX10-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
798 ; GFX10-DENORM-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
799 ; GFX10-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
800 ; GFX10-DENORM-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
801 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
802 ;
803 ; GFX11-CONTRACT-LABEL: test_f64_add_mul:
804 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
805 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
806 ; GFX11-CONTRACT-NEXT: s_clause 0x8
807 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v31, off, s32
808 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v32, off, s32 offset:4
809 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v33, off, s32 offset:8
810 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v34, off, s32 offset:12
811 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v35, off, s32 offset:16
812 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v36, off, s32 offset:20
813 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v37, off, s32 offset:24
814 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v38, off, s32 offset:28
815 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v39, off, s32 offset:32
816 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(6)
817 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
818 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(4)
819 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
820 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(2)
821 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
822 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0)
823 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
824 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
825 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
826 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
827 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
828 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
829 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
830 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
831 ;
832 ; GFX11-DENORM-LABEL: test_f64_add_mul:
833 ; GFX11-DENORM: ; %bb.0: ; %.entry
834 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
835 ; GFX11-DENORM-NEXT: s_clause 0x8
836 ; GFX11-DENORM-NEXT: scratch_load_b32 v31, off, s32
837 ; GFX11-DENORM-NEXT: scratch_load_b32 v32, off, s32 offset:4
838 ; GFX11-DENORM-NEXT: scratch_load_b32 v33, off, s32 offset:8
839 ; GFX11-DENORM-NEXT: scratch_load_b32 v34, off, s32 offset:12
840 ; GFX11-DENORM-NEXT: scratch_load_b32 v35, off, s32 offset:16
841 ; GFX11-DENORM-NEXT: scratch_load_b32 v36, off, s32 offset:20
842 ; GFX11-DENORM-NEXT: scratch_load_b32 v37, off, s32 offset:24
843 ; GFX11-DENORM-NEXT: scratch_load_b32 v38, off, s32 offset:28
844 ; GFX11-DENORM-NEXT: scratch_load_b32 v39, off, s32 offset:32
845 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(6)
846 ; GFX11-DENORM-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
847 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(4)
848 ; GFX11-DENORM-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
849 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(2)
850 ; GFX11-DENORM-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
851 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0)
852 ; GFX11-DENORM-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
853 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
854 ; GFX11-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
855 ; GFX11-DENORM-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
856 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
857 ; GFX11-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
858 ; GFX11-DENORM-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
859 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
860 .entry:
861 %x = fmul fast <4 x double> %c, %d
862 %y = call fast <4 x double> @llvm.fmuladd.v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %x)
863 %z = fadd fast <4 x double> %y, %e
864 ret <4 x double> %z
865 }
867 define <4 x double> @test_f64_add_mul_rhs(<4 x double> %a, <4 x double> %b, <4 x double> %c, <4 x double> %d, <4 x double> %e) {
868 ; GFX9-CONTRACT-LABEL: test_f64_add_mul_rhs:
869 ; GFX9-CONTRACT: ; %bb.0: ; %.entry
870 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
871 ; GFX9-CONTRACT-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
872 ; GFX9-CONTRACT-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
873 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
874 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[31:32]
875 ; GFX9-CONTRACT-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:12
876 ; GFX9-CONTRACT-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:16
877 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
878 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
879 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[24:25]
880 ; GFX9-CONTRACT-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:20
881 ; GFX9-CONTRACT-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:24
882 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
883 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
884 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[24:25]
885 ; GFX9-CONTRACT-NEXT: buffer_load_dword v31, off, s[0:3], s32
886 ; GFX9-CONTRACT-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
887 ; GFX9-CONTRACT-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:32
888 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
889 ; GFX9-CONTRACT-NEXT: s_waitcnt vmcnt(0)
890 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[24:25]
891 ; GFX9-CONTRACT-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
892 ; GFX9-CONTRACT-NEXT: s_setpc_b64 s[30:31]
893 ;
894 ; GFX9-DENORM-LABEL: test_f64_add_mul_rhs:
895 ; GFX9-DENORM: ; %bb.0: ; %.entry
896 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
897 ; GFX9-DENORM-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
898 ; GFX9-DENORM-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
899 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
900 ; GFX9-DENORM-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[31:32]
901 ; GFX9-DENORM-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:12
902 ; GFX9-DENORM-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:16
903 ; GFX9-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
904 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
905 ; GFX9-DENORM-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[24:25]
906 ; GFX9-DENORM-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:20
907 ; GFX9-DENORM-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:24
908 ; GFX9-DENORM-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
909 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
910 ; GFX9-DENORM-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[24:25]
911 ; GFX9-DENORM-NEXT: buffer_load_dword v31, off, s[0:3], s32
912 ; GFX9-DENORM-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
913 ; GFX9-DENORM-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:32
914 ; GFX9-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
915 ; GFX9-DENORM-NEXT: s_waitcnt vmcnt(0)
916 ; GFX9-DENORM-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[24:25]
917 ; GFX9-DENORM-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
918 ; GFX9-DENORM-NEXT: s_setpc_b64 s[30:31]
919 ;
920 ; GFX10-CONTRACT-LABEL: test_f64_add_mul_rhs:
921 ; GFX10-CONTRACT: ; %bb.0: ; %.entry
922 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
923 ; GFX10-CONTRACT-NEXT: s_clause 0x8
924 ; GFX10-CONTRACT-NEXT: buffer_load_dword v31, off, s[0:3], s32
925 ; GFX10-CONTRACT-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
926 ; GFX10-CONTRACT-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
927 ; GFX10-CONTRACT-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:12
928 ; GFX10-CONTRACT-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:16
929 ; GFX10-CONTRACT-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:20
930 ; GFX10-CONTRACT-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:24
931 ; GFX10-CONTRACT-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:28
932 ; GFX10-CONTRACT-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:32
933 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(6)
934 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
935 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(4)
936 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
937 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(2)
938 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
939 ; GFX10-CONTRACT-NEXT: s_waitcnt vmcnt(0)
940 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
941 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
942 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
943 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
944 ; GFX10-CONTRACT-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
945 ; GFX10-CONTRACT-NEXT: s_setpc_b64 s[30:31]
946 ;
947 ; GFX10-DENORM-LABEL: test_f64_add_mul_rhs:
948 ; GFX10-DENORM: ; %bb.0: ; %.entry
949 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
950 ; GFX10-DENORM-NEXT: s_clause 0x8
951 ; GFX10-DENORM-NEXT: buffer_load_dword v31, off, s[0:3], s32
952 ; GFX10-DENORM-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
953 ; GFX10-DENORM-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
954 ; GFX10-DENORM-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:12
955 ; GFX10-DENORM-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:16
956 ; GFX10-DENORM-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:20
957 ; GFX10-DENORM-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:24
958 ; GFX10-DENORM-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:28
959 ; GFX10-DENORM-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:32
960 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(6)
961 ; GFX10-DENORM-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
962 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(4)
963 ; GFX10-DENORM-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
964 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(2)
965 ; GFX10-DENORM-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
966 ; GFX10-DENORM-NEXT: s_waitcnt vmcnt(0)
967 ; GFX10-DENORM-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
968 ; GFX10-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
969 ; GFX10-DENORM-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
970 ; GFX10-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
971 ; GFX10-DENORM-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
972 ; GFX10-DENORM-NEXT: s_setpc_b64 s[30:31]
973 ;
974 ; GFX11-CONTRACT-LABEL: test_f64_add_mul_rhs:
975 ; GFX11-CONTRACT: ; %bb.0: ; %.entry
976 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
977 ; GFX11-CONTRACT-NEXT: s_clause 0x8
978 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v31, off, s32
979 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v32, off, s32 offset:4
980 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v33, off, s32 offset:8
981 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v34, off, s32 offset:12
982 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v35, off, s32 offset:16
983 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v36, off, s32 offset:20
984 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v37, off, s32 offset:24
985 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v38, off, s32 offset:28
986 ; GFX11-CONTRACT-NEXT: scratch_load_b32 v39, off, s32 offset:32
987 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(6)
988 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
989 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(4)
990 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
991 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(2)
992 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
993 ; GFX11-CONTRACT-NEXT: s_waitcnt vmcnt(0)
994 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
995 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
996 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
997 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
998 ; GFX11-CONTRACT-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
999 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
1000 ; GFX11-CONTRACT-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
1001 ; GFX11-CONTRACT-NEXT: s_setpc_b64 s[30:31]
1002 ;
1003 ; GFX11-DENORM-LABEL: test_f64_add_mul_rhs:
1004 ; GFX11-DENORM: ; %bb.0: ; %.entry
1005 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1006 ; GFX11-DENORM-NEXT: s_clause 0x8
1007 ; GFX11-DENORM-NEXT: scratch_load_b32 v31, off, s32
1008 ; GFX11-DENORM-NEXT: scratch_load_b32 v32, off, s32 offset:4
1009 ; GFX11-DENORM-NEXT: scratch_load_b32 v33, off, s32 offset:8
1010 ; GFX11-DENORM-NEXT: scratch_load_b32 v34, off, s32 offset:12
1011 ; GFX11-DENORM-NEXT: scratch_load_b32 v35, off, s32 offset:16
1012 ; GFX11-DENORM-NEXT: scratch_load_b32 v36, off, s32 offset:20
1013 ; GFX11-DENORM-NEXT: scratch_load_b32 v37, off, s32 offset:24
1014 ; GFX11-DENORM-NEXT: scratch_load_b32 v38, off, s32 offset:28
1015 ; GFX11-DENORM-NEXT: scratch_load_b32 v39, off, s32 offset:32
1016 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(6)
1017 ; GFX11-DENORM-NEXT: v_fma_f64 v[16:17], v[16:17], v[24:25], v[32:33]
1018 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(4)
1019 ; GFX11-DENORM-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], v[34:35]
1020 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(2)
1021 ; GFX11-DENORM-NEXT: v_fma_f64 v[20:21], v[20:21], v[28:29], v[36:37]
1022 ; GFX11-DENORM-NEXT: s_waitcnt vmcnt(0)
1023 ; GFX11-DENORM-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], v[38:39]
1024 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
1025 ; GFX11-DENORM-NEXT: v_fma_f64 v[0:1], v[0:1], v[8:9], v[16:17]
1026 ; GFX11-DENORM-NEXT: v_fma_f64 v[2:3], v[2:3], v[10:11], v[18:19]
1027 ; GFX11-DENORM-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
1028 ; GFX11-DENORM-NEXT: v_fma_f64 v[4:5], v[4:5], v[12:13], v[20:21]
1029 ; GFX11-DENORM-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], v[22:23]
1030 ; GFX11-DENORM-NEXT: s_setpc_b64 s[30:31]
1031 .entry:
1032 %x = fmul fast <4 x double> %c, %d
1033 %y = call fast <4 x double> @llvm.fmuladd.v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %x)
1034 %z = fadd fast <4 x double> %e, %y
1035 ret <4 x double> %z
1036 }
1038 declare <4 x double> @llvm.fmuladd.v4f64(<4 x double>, <4 x double>, <4 x double>) #0
1039 declare <4 x float> @llvm.fmuladd.v4f32(<4 x float>, <4 x float>, <4 x float>) #0
1040 declare <4 x half> @llvm.fmuladd.v4f16(<4 x half>, <4 x half>, <4 x half>) #0
1041 declare double @llvm.fmuladd.f64(double, double, double) #0
1042 declare float @llvm.fmuladd.f32(float, float, float) #0
1043 declare half @llvm.fmuladd.f16(half, half, half) #0
1044 attributes #0 = { nounwind readnone }