| blob | 74e2b9ea71425822fd5893d59b48a1e918667a06 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=amdgcn -mcpu=hawaii -mattr=+flat-for-global < %s | FileCheck -enable-var-scope --check-prefixes=GCN,GCN-SAFE,SI,SI-SAFE %s
3 ; RUN: llc -enable-no-signed-zeros-fp-math -mtriple=amdgcn -mcpu=hawaii -mattr=+flat-for-global < %s | FileCheck -enable-var-scope --check-prefixes=GCN,GCN-NSZ,SI,SI-NSZ %s
5 ; RUN: llc -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -enable-var-scope --check-prefixes=GCN,GCN-SAFE,VI,VI-SAFE %s
6 ; RUN: llc -enable-no-signed-zeros-fp-math -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -enable-var-scope --check-prefixes=GCN,GCN-NSZ,VI,VI-NSZ %s
8 ; --------------------------------------------------------------------------------
9 ; fadd tests
10 ; --------------------------------------------------------------------------------
12 define float @v_fneg_add_f32(float %a, float %b) #0 {
13 ; GCN-SAFE-LABEL: v_fneg_add_f32:
14 ; GCN-SAFE: ; %bb.0:
15 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
16 ; GCN-SAFE-NEXT: v_add_f32_e32 v0, v0, v1
17 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
18 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
19 ;
20 ; GCN-NSZ-LABEL: v_fneg_add_f32:
21 ; GCN-NSZ: ; %bb.0:
22 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
23 ; GCN-NSZ-NEXT: v_sub_f32_e64 v0, -v0, v1
24 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
25 %add = fadd float %a, %b
26 %fneg = fneg float %add
27 ret float %fneg
28 }
30 define { float, float } @v_fneg_add_store_use_add_f32(float %a, float %b) #0 {
31 ; GCN-LABEL: v_fneg_add_store_use_add_f32:
32 ; GCN: ; %bb.0:
33 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
34 ; GCN-NEXT: v_add_f32_e32 v1, v0, v1
35 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
36 ; GCN-NEXT: s_setpc_b64 s[30:31]
37 %add = fadd float %a, %b
38 %fneg = fneg float %add
39 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
40 %insert.1 = insertvalue { float, float } %insert.0, float %add, 1
41 ret { float, float } %insert.1
42 }
44 define { float, float } @v_fneg_add_multi_use_add_f32(float %a, float %b) #0 {
45 ; GCN-SAFE-LABEL: v_fneg_add_multi_use_add_f32:
46 ; GCN-SAFE: ; %bb.0:
47 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
48 ; GCN-SAFE-NEXT: v_add_f32_e32 v1, v0, v1
49 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
50 ; GCN-SAFE-NEXT: v_mul_f32_e32 v1, 4.0, v1
51 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
52 ;
53 ; GCN-NSZ-LABEL: v_fneg_add_multi_use_add_f32:
54 ; GCN-NSZ: ; %bb.0:
55 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
56 ; GCN-NSZ-NEXT: v_sub_f32_e64 v0, -v0, v1
57 ; GCN-NSZ-NEXT: v_mul_f32_e32 v1, -4.0, v0
58 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
59 %add = fadd float %a, %b
60 %fneg = fneg float %add
61 %use1 = fmul float %add, 4.0
63 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
64 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
65 ret { float, float } %insert.1
66 }
68 define float @v_fneg_add_fneg_x_f32(float %a, float %b) #0 {
69 ; GCN-SAFE-LABEL: v_fneg_add_fneg_x_f32:
70 ; GCN-SAFE: ; %bb.0:
71 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
72 ; GCN-SAFE-NEXT: v_sub_f32_e32 v0, v1, v0
73 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
74 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
75 ;
76 ; GCN-NSZ-LABEL: v_fneg_add_fneg_x_f32:
77 ; GCN-NSZ: ; %bb.0:
78 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
79 ; GCN-NSZ-NEXT: v_sub_f32_e32 v0, v0, v1
80 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
81 %fneg.a = fneg float %a
82 %add = fadd float %fneg.a, %b
83 %fneg = fneg float %add
84 ret float %fneg
85 }
87 define float @v_fneg_add_x_fneg_f32(float %a, float %b) #0 {
88 ; GCN-SAFE-LABEL: v_fneg_add_x_fneg_f32:
89 ; GCN-SAFE: ; %bb.0:
90 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
91 ; GCN-SAFE-NEXT: v_sub_f32_e32 v0, v0, v1
92 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
93 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
94 ;
95 ; GCN-NSZ-LABEL: v_fneg_add_x_fneg_f32:
96 ; GCN-NSZ: ; %bb.0:
97 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
98 ; GCN-NSZ-NEXT: v_sub_f32_e32 v0, v1, v0
99 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
100 %fneg.b = fneg float %b
101 %add = fadd float %a, %fneg.b
102 %fneg = fneg float %add
103 ret float %fneg
104 }
106 define float @v_fneg_add_fneg_fneg_f32(float %a, float %b) #0 {
107 ; GCN-SAFE-LABEL: v_fneg_add_fneg_fneg_f32:
108 ; GCN-SAFE: ; %bb.0:
109 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
110 ; GCN-SAFE-NEXT: v_sub_f32_e64 v0, -v0, v1
111 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
112 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
113 ;
114 ; GCN-NSZ-LABEL: v_fneg_add_fneg_fneg_f32:
115 ; GCN-NSZ: ; %bb.0:
116 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
117 ; GCN-NSZ-NEXT: v_add_f32_e32 v0, v0, v1
118 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
119 %fneg.a = fneg float %a
120 %fneg.b = fneg float %b
121 %add = fadd float %fneg.a, %fneg.b
122 %fneg = fneg float %add
123 ret float %fneg
124 }
126 define { float, float } @v_fneg_add_store_use_fneg_x_f32(float %a, float %b) #0 {
127 ; GCN-SAFE-LABEL: v_fneg_add_store_use_fneg_x_f32:
128 ; GCN-SAFE: ; %bb.0:
129 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
130 ; GCN-SAFE-NEXT: v_xor_b32_e32 v2, 0x80000000, v0
131 ; GCN-SAFE-NEXT: v_sub_f32_e32 v0, v1, v0
132 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
133 ; GCN-SAFE-NEXT: v_mov_b32_e32 v1, v2
134 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
135 ;
136 ; GCN-NSZ-LABEL: v_fneg_add_store_use_fneg_x_f32:
137 ; GCN-NSZ: ; %bb.0:
138 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
139 ; GCN-NSZ-NEXT: v_xor_b32_e32 v2, 0x80000000, v0
140 ; GCN-NSZ-NEXT: v_sub_f32_e32 v0, v0, v1
141 ; GCN-NSZ-NEXT: v_mov_b32_e32 v1, v2
142 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
143 %fneg.a = fneg float %a
144 %add = fadd float %fneg.a, %b
145 %fneg = fneg float %add
146 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
147 %insert.1 = insertvalue { float, float } %insert.0, float %fneg.a, 1
148 ret { float, float } %insert.1
149 }
151 define { float, float } @v_fneg_add_multi_use_fneg_x_f32(float %a, float %b, float %c) #0 {
152 ; GCN-SAFE-LABEL: v_fneg_add_multi_use_fneg_x_f32:
153 ; GCN-SAFE: ; %bb.0:
154 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
155 ; GCN-SAFE-NEXT: v_sub_f32_e32 v1, v1, v0
156 ; GCN-SAFE-NEXT: v_xor_b32_e32 v3, 0x80000000, v1
157 ; GCN-SAFE-NEXT: v_mul_f32_e64 v1, -v0, v2
158 ; GCN-SAFE-NEXT: v_mov_b32_e32 v0, v3
159 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
160 ;
161 ; GCN-NSZ-LABEL: v_fneg_add_multi_use_fneg_x_f32:
162 ; GCN-NSZ: ; %bb.0:
163 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
164 ; GCN-NSZ-NEXT: v_sub_f32_e32 v3, v0, v1
165 ; GCN-NSZ-NEXT: v_mul_f32_e64 v1, -v0, v2
166 ; GCN-NSZ-NEXT: v_mov_b32_e32 v0, v3
167 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
168 %fneg.a = fneg float %a
169 %add = fadd float %fneg.a, %b
170 %fneg = fneg float %add
171 %use1 = fmul float %fneg.a, %c
173 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
174 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
175 ret { float, float } %insert.1
176 }
178 ; This one asserted with -enable-no-signed-zeros-fp-math
179 define amdgpu_ps float @fneg_fadd_0_f32(float inreg %tmp2, float inreg %tmp6, <4 x i32> %arg) #0 {
180 ; SI-SAFE-LABEL: fneg_fadd_0_f32:
181 ; SI-SAFE: ; %bb.0: ; %.entry
182 ; SI-SAFE-NEXT: v_div_scale_f32 v0, s[2:3], s1, s1, 1.0
183 ; SI-SAFE-NEXT: v_rcp_f32_e32 v1, v0
184 ; SI-SAFE-NEXT: v_div_scale_f32 v2, vcc, 1.0, s1, 1.0
185 ; SI-SAFE-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
186 ; SI-SAFE-NEXT: v_fma_f32 v3, -v0, v1, 1.0
187 ; SI-SAFE-NEXT: v_fma_f32 v1, v3, v1, v1
188 ; SI-SAFE-NEXT: v_mul_f32_e32 v3, v2, v1
189 ; SI-SAFE-NEXT: v_fma_f32 v4, -v0, v3, v2
190 ; SI-SAFE-NEXT: v_fma_f32 v3, v4, v1, v3
191 ; SI-SAFE-NEXT: v_fma_f32 v0, -v0, v3, v2
192 ; SI-SAFE-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
193 ; SI-SAFE-NEXT: v_div_fmas_f32 v0, v0, v1, v3
194 ; SI-SAFE-NEXT: v_div_fixup_f32 v0, v0, s1, 1.0
195 ; SI-SAFE-NEXT: v_mad_f32 v0, v0, 0, 0
196 ; SI-SAFE-NEXT: v_mov_b32_e32 v1, s0
197 ; SI-SAFE-NEXT: v_cmp_ngt_f32_e32 vcc, s0, v0
198 ; SI-SAFE-NEXT: v_cndmask_b32_e64 v0, -v0, v1, vcc
199 ; SI-SAFE-NEXT: v_mov_b32_e32 v1, 0x7fc00000
200 ; SI-SAFE-NEXT: v_cmp_nlt_f32_e32 vcc, 0, v0
201 ; SI-SAFE-NEXT: v_cndmask_b32_e64 v0, v1, 0, vcc
202 ; SI-SAFE-NEXT: ; return to shader part epilog
203 ;
204 ; SI-NSZ-LABEL: fneg_fadd_0_f32:
205 ; SI-NSZ: ; %bb.0: ; %.entry
206 ; SI-NSZ-NEXT: v_div_scale_f32 v0, s[2:3], s1, s1, 1.0
207 ; SI-NSZ-NEXT: v_rcp_f32_e32 v1, v0
208 ; SI-NSZ-NEXT: v_div_scale_f32 v2, vcc, 1.0, s1, 1.0
209 ; SI-NSZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
210 ; SI-NSZ-NEXT: v_fma_f32 v3, -v0, v1, 1.0
211 ; SI-NSZ-NEXT: v_fma_f32 v1, v3, v1, v1
212 ; SI-NSZ-NEXT: v_mul_f32_e32 v3, v2, v1
213 ; SI-NSZ-NEXT: v_fma_f32 v4, -v0, v3, v2
214 ; SI-NSZ-NEXT: v_fma_f32 v3, v4, v1, v3
215 ; SI-NSZ-NEXT: v_fma_f32 v0, -v0, v3, v2
216 ; SI-NSZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
217 ; SI-NSZ-NEXT: v_div_fmas_f32 v0, v0, v1, v3
218 ; SI-NSZ-NEXT: v_div_fixup_f32 v0, v0, s1, 1.0
219 ; SI-NSZ-NEXT: v_mul_f32_e32 v0, 0, v0
220 ; SI-NSZ-NEXT: v_mov_b32_e32 v1, s0
221 ; SI-NSZ-NEXT: v_cmp_ngt_f32_e32 vcc, s0, v0
222 ; SI-NSZ-NEXT: v_cndmask_b32_e64 v0, -v0, v1, vcc
223 ; SI-NSZ-NEXT: v_mov_b32_e32 v1, 0x7fc00000
224 ; SI-NSZ-NEXT: v_cmp_nlt_f32_e32 vcc, 0, v0
225 ; SI-NSZ-NEXT: v_cndmask_b32_e64 v0, v1, 0, vcc
226 ; SI-NSZ-NEXT: ; return to shader part epilog
227 ;
228 ; VI-SAFE-LABEL: fneg_fadd_0_f32:
229 ; VI-SAFE: ; %bb.0: ; %.entry
230 ; VI-SAFE-NEXT: v_div_scale_f32 v0, s[2:3], s1, s1, 1.0
231 ; VI-SAFE-NEXT: v_div_scale_f32 v1, vcc, 1.0, s1, 1.0
232 ; VI-SAFE-NEXT: v_rcp_f32_e32 v2, v0
233 ; VI-SAFE-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
234 ; VI-SAFE-NEXT: v_fma_f32 v3, -v0, v2, 1.0
235 ; VI-SAFE-NEXT: v_fma_f32 v2, v3, v2, v2
236 ; VI-SAFE-NEXT: v_mul_f32_e32 v3, v1, v2
237 ; VI-SAFE-NEXT: v_fma_f32 v4, -v0, v3, v1
238 ; VI-SAFE-NEXT: v_fma_f32 v3, v4, v2, v3
239 ; VI-SAFE-NEXT: v_fma_f32 v0, -v0, v3, v1
240 ; VI-SAFE-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
241 ; VI-SAFE-NEXT: v_div_fmas_f32 v0, v0, v2, v3
242 ; VI-SAFE-NEXT: v_mov_b32_e32 v2, s0
243 ; VI-SAFE-NEXT: v_mov_b32_e32 v1, 0x7fc00000
244 ; VI-SAFE-NEXT: v_div_fixup_f32 v0, v0, s1, 1.0
245 ; VI-SAFE-NEXT: v_mad_f32 v0, v0, 0, 0
246 ; VI-SAFE-NEXT: v_cmp_ngt_f32_e32 vcc, s0, v0
247 ; VI-SAFE-NEXT: v_cndmask_b32_e64 v0, -v0, v2, vcc
248 ; VI-SAFE-NEXT: v_cmp_nlt_f32_e32 vcc, 0, v0
249 ; VI-SAFE-NEXT: v_cndmask_b32_e64 v0, v1, 0, vcc
250 ; VI-SAFE-NEXT: ; return to shader part epilog
251 ;
252 ; VI-NSZ-LABEL: fneg_fadd_0_f32:
253 ; VI-NSZ: ; %bb.0: ; %.entry
254 ; VI-NSZ-NEXT: v_div_scale_f32 v0, s[2:3], s1, s1, 1.0
255 ; VI-NSZ-NEXT: v_div_scale_f32 v1, vcc, 1.0, s1, 1.0
256 ; VI-NSZ-NEXT: v_rcp_f32_e32 v2, v0
257 ; VI-NSZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
258 ; VI-NSZ-NEXT: v_fma_f32 v3, -v0, v2, 1.0
259 ; VI-NSZ-NEXT: v_fma_f32 v2, v3, v2, v2
260 ; VI-NSZ-NEXT: v_mul_f32_e32 v3, v1, v2
261 ; VI-NSZ-NEXT: v_fma_f32 v4, -v0, v3, v1
262 ; VI-NSZ-NEXT: v_fma_f32 v3, v4, v2, v3
263 ; VI-NSZ-NEXT: v_fma_f32 v0, -v0, v3, v1
264 ; VI-NSZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
265 ; VI-NSZ-NEXT: v_div_fmas_f32 v0, v0, v2, v3
266 ; VI-NSZ-NEXT: v_mov_b32_e32 v2, s0
267 ; VI-NSZ-NEXT: v_mov_b32_e32 v1, 0x7fc00000
268 ; VI-NSZ-NEXT: v_div_fixup_f32 v0, v0, s1, 1.0
269 ; VI-NSZ-NEXT: v_mul_f32_e32 v0, 0, v0
270 ; VI-NSZ-NEXT: v_cmp_ngt_f32_e32 vcc, s0, v0
271 ; VI-NSZ-NEXT: v_cndmask_b32_e64 v0, -v0, v2, vcc
272 ; VI-NSZ-NEXT: v_cmp_nlt_f32_e32 vcc, 0, v0
273 ; VI-NSZ-NEXT: v_cndmask_b32_e64 v0, v1, 0, vcc
274 ; VI-NSZ-NEXT: ; return to shader part epilog
275 .entry:
276 %tmp7 = fdiv float 1.000000e+00, %tmp6
277 %tmp8 = fmul float 0.000000e+00, %tmp7
278 %tmp9 = fmul reassoc nnan arcp contract float 0.000000e+00, %tmp8
279 %.i188 = fadd float %tmp9, 0.000000e+00
280 %tmp10 = fcmp uge float %.i188, %tmp2
281 %tmp11 = fneg float %.i188
282 %.i092 = select i1 %tmp10, float %tmp2, float %tmp11
283 %tmp12 = fcmp ule float %.i092, 0.000000e+00
284 %.i198 = select i1 %tmp12, float 0.000000e+00, float 0x7FF8000000000000
285 ret float %.i198
286 }
288 ; This is a workaround because -enable-no-signed-zeros-fp-math does not set up
289 ; function attribute unsafe-fp-math automatically. Combine with the previous test
290 ; when that is done.
291 define amdgpu_ps float @fneg_fadd_0_nsz_f32(float inreg %tmp2, float inreg %tmp6, <4 x i32> %arg) #2 {
292 ; SI-SAFE-LABEL: fneg_fadd_0_nsz_f32:
293 ; SI-SAFE: ; %bb.0: ; %.entry
294 ; SI-SAFE-NEXT: v_min_legacy_f32_e64 v0, 0, s0
295 ; SI-SAFE-NEXT: s_brev_b32 s0, 1
296 ; SI-SAFE-NEXT: v_mov_b32_e32 v1, 0x7fc00000
297 ; SI-SAFE-NEXT: v_cmp_ngt_f32_e32 vcc, s0, v0
298 ; SI-SAFE-NEXT: v_cndmask_b32_e64 v0, v1, 0, vcc
299 ; SI-SAFE-NEXT: ; return to shader part epilog
300 ;
301 ; GCN-NSZ-LABEL: fneg_fadd_0_nsz_f32:
302 ; GCN-NSZ: ; %bb.0: ; %.entry
303 ; GCN-NSZ-NEXT: v_rcp_f32_e32 v0, s1
304 ; GCN-NSZ-NEXT: v_mov_b32_e32 v1, s0
305 ; GCN-NSZ-NEXT: v_mul_f32_e32 v0, 0, v0
306 ; GCN-NSZ-NEXT: v_cmp_ngt_f32_e32 vcc, s0, v0
307 ; GCN-NSZ-NEXT: v_cndmask_b32_e64 v0, -v0, v1, vcc
308 ; GCN-NSZ-NEXT: v_mov_b32_e32 v1, 0x7fc00000
309 ; GCN-NSZ-NEXT: v_cmp_nlt_f32_e32 vcc, 0, v0
310 ; GCN-NSZ-NEXT: v_cndmask_b32_e64 v0, v1, 0, vcc
311 ; GCN-NSZ-NEXT: ; return to shader part epilog
312 ;
313 ; VI-SAFE-LABEL: fneg_fadd_0_nsz_f32:
314 ; VI-SAFE: ; %bb.0: ; %.entry
315 ; VI-SAFE-NEXT: v_rcp_f32_e32 v0, s1
316 ; VI-SAFE-NEXT: v_mov_b32_e32 v1, s0
317 ; VI-SAFE-NEXT: v_mul_f32_e32 v0, 0, v0
318 ; VI-SAFE-NEXT: v_add_f32_e32 v0, 0, v0
319 ; VI-SAFE-NEXT: v_cmp_ngt_f32_e32 vcc, s0, v0
320 ; VI-SAFE-NEXT: v_cndmask_b32_e64 v0, -v0, v1, vcc
321 ; VI-SAFE-NEXT: v_mov_b32_e32 v1, 0x7fc00000
322 ; VI-SAFE-NEXT: v_cmp_nlt_f32_e32 vcc, 0, v0
323 ; VI-SAFE-NEXT: v_cndmask_b32_e64 v0, v1, 0, vcc
324 ; VI-SAFE-NEXT: ; return to shader part epilog
325 .entry:
326 %tmp7 = fdiv afn float 1.000000e+00, %tmp6
327 %tmp8 = fmul float 0.000000e+00, %tmp7
328 %tmp9 = fmul reassoc nnan arcp contract float 0.000000e+00, %tmp8
329 %.i188 = fadd float %tmp9, 0.000000e+00
330 %tmp10 = fcmp uge float %.i188, %tmp2
331 %tmp11 = fneg float %.i188
332 %.i092 = select i1 %tmp10, float %tmp2, float %tmp11
333 %tmp12 = fcmp ule float %.i092, 0.000000e+00
334 %.i198 = select i1 %tmp12, float 0.000000e+00, float 0x7FF8000000000000
335 ret float %.i198
336 }
338 define double @v_fneg_add_f64(double %a, double %b) #0 {
339 ; GCN-SAFE-LABEL: v_fneg_add_f64:
340 ; GCN-SAFE: ; %bb.0:
341 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
342 ; GCN-SAFE-NEXT: v_add_f64 v[0:1], v[0:1], v[2:3]
343 ; GCN-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
344 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
345 ;
346 ; GCN-NSZ-LABEL: v_fneg_add_f64:
347 ; GCN-NSZ: ; %bb.0:
348 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
349 ; GCN-NSZ-NEXT: v_add_f64 v[0:1], -v[0:1], -v[2:3]
350 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
351 %add = fadd double %a, %b
352 %fneg = fneg double %add
353 ret double %fneg
354 }
356 define { double, double } @v_fneg_add_store_use_add_f64(double %a, double %b) #0 {
357 ; GCN-LABEL: v_fneg_add_store_use_add_f64:
358 ; GCN: ; %bb.0:
359 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
360 ; GCN-NEXT: v_add_f64 v[2:3], v[0:1], v[2:3]
361 ; GCN-NEXT: v_xor_b32_e32 v1, 0x80000000, v3
362 ; GCN-NEXT: v_mov_b32_e32 v0, v2
363 ; GCN-NEXT: s_setpc_b64 s[30:31]
364 %add = fadd double %a, %b
365 %fneg = fneg double %add
366 %insert.0 = insertvalue { double, double } poison, double %fneg, 0
367 %insert.1 = insertvalue { double, double } %insert.0, double %add, 1
368 ret { double, double } %insert.1
369 }
371 define { double, double } @v_fneg_add_multi_use_add_f64(double %a, double %b) #0 {
372 ; SI-SAFE-LABEL: v_fneg_add_multi_use_add_f64:
373 ; SI-SAFE: ; %bb.0:
374 ; SI-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
375 ; SI-SAFE-NEXT: v_add_f64 v[0:1], v[0:1], v[2:3]
376 ; SI-SAFE-NEXT: v_xor_b32_e32 v4, 0x80000000, v1
377 ; SI-SAFE-NEXT: v_mul_f64 v[2:3], v[0:1], 4.0
378 ; SI-SAFE-NEXT: v_mov_b32_e32 v1, v4
379 ; SI-SAFE-NEXT: s_setpc_b64 s[30:31]
380 ;
381 ; GCN-NSZ-LABEL: v_fneg_add_multi_use_add_f64:
382 ; GCN-NSZ: ; %bb.0:
383 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
384 ; GCN-NSZ-NEXT: v_add_f64 v[0:1], -v[0:1], -v[2:3]
385 ; GCN-NSZ-NEXT: v_mul_f64 v[2:3], v[0:1], -4.0
386 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
387 ;
388 ; VI-SAFE-LABEL: v_fneg_add_multi_use_add_f64:
389 ; VI-SAFE: ; %bb.0:
390 ; VI-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
391 ; VI-SAFE-NEXT: v_add_f64 v[0:1], v[0:1], v[2:3]
392 ; VI-SAFE-NEXT: v_mul_f64 v[2:3], v[0:1], 4.0
393 ; VI-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
394 ; VI-SAFE-NEXT: s_setpc_b64 s[30:31]
395 %add = fadd double %a, %b
396 %fneg = fneg double %add
397 %use1 = fmul double %add, 4.0
399 %insert.0 = insertvalue { double, double } poison, double %fneg, 0
400 %insert.1 = insertvalue { double, double } %insert.0, double %use1, 1
401 ret { double, double } %insert.1
402 }
404 define double @v_fneg_add_fneg_x_f64(double %a, double %b) #0 {
405 ; GCN-SAFE-LABEL: v_fneg_add_fneg_x_f64:
406 ; GCN-SAFE: ; %bb.0:
407 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
408 ; GCN-SAFE-NEXT: v_add_f64 v[0:1], v[2:3], -v[0:1]
409 ; GCN-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
410 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
411 ;
412 ; GCN-NSZ-LABEL: v_fneg_add_fneg_x_f64:
413 ; GCN-NSZ: ; %bb.0:
414 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
415 ; GCN-NSZ-NEXT: v_add_f64 v[0:1], v[0:1], -v[2:3]
416 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
417 %fneg.a = fneg double %a
418 %add = fadd double %fneg.a, %b
419 %fneg = fneg double %add
420 ret double %fneg
421 }
423 define double @v_fneg_add_x_fneg_f64(double %a, double %b) #0 {
424 ; GCN-SAFE-LABEL: v_fneg_add_x_fneg_f64:
425 ; GCN-SAFE: ; %bb.0:
426 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
427 ; GCN-SAFE-NEXT: v_add_f64 v[0:1], v[0:1], -v[2:3]
428 ; GCN-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
429 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
430 ;
431 ; GCN-NSZ-LABEL: v_fneg_add_x_fneg_f64:
432 ; GCN-NSZ: ; %bb.0:
433 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
434 ; GCN-NSZ-NEXT: v_add_f64 v[0:1], v[2:3], -v[0:1]
435 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
436 %fneg.b = fneg double %b
437 %add = fadd double %a, %fneg.b
438 %fneg = fneg double %add
439 ret double %fneg
440 }
442 define double @v_fneg_add_fneg_fneg_f64(double %a, double %b) #0 {
443 ; GCN-SAFE-LABEL: v_fneg_add_fneg_fneg_f64:
444 ; GCN-SAFE: ; %bb.0:
445 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
446 ; GCN-SAFE-NEXT: v_add_f64 v[0:1], -v[0:1], -v[2:3]
447 ; GCN-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
448 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
449 ;
450 ; GCN-NSZ-LABEL: v_fneg_add_fneg_fneg_f64:
451 ; GCN-NSZ: ; %bb.0:
452 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
453 ; GCN-NSZ-NEXT: v_add_f64 v[0:1], v[0:1], v[2:3]
454 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
455 %fneg.a = fneg double %a
456 %fneg.b = fneg double %b
457 %add = fadd double %fneg.a, %fneg.b
458 %fneg = fneg double %add
459 ret double %fneg
460 }
462 define { double, double } @v_fneg_add_store_use_fneg_x_f64(double %a, double %b) #0 {
463 ; SI-SAFE-LABEL: v_fneg_add_store_use_fneg_x_f64:
464 ; SI-SAFE: ; %bb.0:
465 ; SI-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
466 ; SI-SAFE-NEXT: v_mov_b32_e32 v5, v1
467 ; SI-SAFE-NEXT: v_mov_b32_e32 v4, v0
468 ; SI-SAFE-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
469 ; SI-SAFE-NEXT: v_xor_b32_e32 v3, 0x80000000, v5
470 ; SI-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
471 ; SI-SAFE-NEXT: v_mov_b32_e32 v2, v4
472 ; SI-SAFE-NEXT: s_setpc_b64 s[30:31]
473 ;
474 ; SI-NSZ-LABEL: v_fneg_add_store_use_fneg_x_f64:
475 ; SI-NSZ: ; %bb.0:
476 ; SI-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
477 ; SI-NSZ-NEXT: v_mov_b32_e32 v5, v1
478 ; SI-NSZ-NEXT: v_mov_b32_e32 v4, v0
479 ; SI-NSZ-NEXT: v_add_f64 v[0:1], v[4:5], -v[2:3]
480 ; SI-NSZ-NEXT: v_xor_b32_e32 v3, 0x80000000, v5
481 ; SI-NSZ-NEXT: v_mov_b32_e32 v2, v4
482 ; SI-NSZ-NEXT: s_setpc_b64 s[30:31]
483 ;
484 ; VI-SAFE-LABEL: v_fneg_add_store_use_fneg_x_f64:
485 ; VI-SAFE: ; %bb.0:
486 ; VI-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
487 ; VI-SAFE-NEXT: v_add_f64 v[4:5], v[2:3], -v[0:1]
488 ; VI-SAFE-NEXT: v_xor_b32_e32 v3, 0x80000000, v1
489 ; VI-SAFE-NEXT: v_mov_b32_e32 v2, v0
490 ; VI-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v5
491 ; VI-SAFE-NEXT: v_mov_b32_e32 v0, v4
492 ; VI-SAFE-NEXT: s_setpc_b64 s[30:31]
493 ;
494 ; VI-NSZ-LABEL: v_fneg_add_store_use_fneg_x_f64:
495 ; VI-NSZ: ; %bb.0:
496 ; VI-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
497 ; VI-NSZ-NEXT: v_add_f64 v[4:5], v[0:1], -v[2:3]
498 ; VI-NSZ-NEXT: v_xor_b32_e32 v3, 0x80000000, v1
499 ; VI-NSZ-NEXT: v_mov_b32_e32 v2, v0
500 ; VI-NSZ-NEXT: v_mov_b32_e32 v0, v4
501 ; VI-NSZ-NEXT: v_mov_b32_e32 v1, v5
502 ; VI-NSZ-NEXT: s_setpc_b64 s[30:31]
503 %fneg.a = fneg double %a
504 %add = fadd double %fneg.a, %b
505 %fneg = fneg double %add
506 %insert.0 = insertvalue { double, double } poison, double %fneg, 0
507 %insert.1 = insertvalue { double, double } %insert.0, double %fneg.a, 1
508 ret { double, double } %insert.1
509 }
511 define { double, double } @v_fneg_add_multi_use_fneg_x_f64(double %a, double %b, double %c) #0 {
512 ; SI-SAFE-LABEL: v_fneg_add_multi_use_fneg_x_f64:
513 ; SI-SAFE: ; %bb.0:
514 ; SI-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
515 ; SI-SAFE-NEXT: v_add_f64 v[6:7], v[2:3], -v[0:1]
516 ; SI-SAFE-NEXT: v_mul_f64 v[2:3], -v[0:1], v[4:5]
517 ; SI-SAFE-NEXT: v_xor_b32_e32 v7, 0x80000000, v7
518 ; SI-SAFE-NEXT: v_mov_b32_e32 v0, v6
519 ; SI-SAFE-NEXT: v_mov_b32_e32 v1, v7
520 ; SI-SAFE-NEXT: s_setpc_b64 s[30:31]
521 ;
522 ; GCN-NSZ-LABEL: v_fneg_add_multi_use_fneg_x_f64:
523 ; GCN-NSZ: ; %bb.0:
524 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
525 ; GCN-NSZ-NEXT: v_add_f64 v[6:7], v[0:1], -v[2:3]
526 ; GCN-NSZ-NEXT: v_mul_f64 v[2:3], -v[0:1], v[4:5]
527 ; GCN-NSZ-NEXT: v_mov_b32_e32 v0, v6
528 ; GCN-NSZ-NEXT: v_mov_b32_e32 v1, v7
529 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
530 ;
531 ; VI-SAFE-LABEL: v_fneg_add_multi_use_fneg_x_f64:
532 ; VI-SAFE: ; %bb.0:
533 ; VI-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
534 ; VI-SAFE-NEXT: v_add_f64 v[6:7], v[2:3], -v[0:1]
535 ; VI-SAFE-NEXT: v_mul_f64 v[2:3], -v[0:1], v[4:5]
536 ; VI-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v7
537 ; VI-SAFE-NEXT: v_mov_b32_e32 v0, v6
538 ; VI-SAFE-NEXT: s_setpc_b64 s[30:31]
539 %fneg.a = fneg double %a
540 %add = fadd double %fneg.a, %b
541 %fneg = fneg double %add
542 %use1 = fmul double %fneg.a, %c
544 %insert.0 = insertvalue { double, double } poison, double %fneg, 0
545 %insert.1 = insertvalue { double, double } %insert.0, double %use1, 1
546 ret { double, double } %insert.1
547 }
549 ; This one asserted with -enable-no-signed-zeros-fp-math
550 define amdgpu_ps double @fneg_fadd_0_f64(double inreg %tmp2, double inreg %tmp6, <4 x i32> %arg) #0 {
551 ; SI-SAFE-LABEL: fneg_fadd_0_f64:
552 ; SI-SAFE: ; %bb.0: ; %.entry
553 ; SI-SAFE-NEXT: v_div_scale_f64 v[0:1], s[4:5], s[2:3], s[2:3], 1.0
554 ; SI-SAFE-NEXT: v_rcp_f64_e32 v[2:3], v[0:1]
555 ; SI-SAFE-NEXT: v_fma_f64 v[4:5], -v[0:1], v[2:3], 1.0
556 ; SI-SAFE-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[2:3]
557 ; SI-SAFE-NEXT: v_div_scale_f64 v[4:5], vcc, 1.0, s[2:3], 1.0
558 ; SI-SAFE-NEXT: v_fma_f64 v[6:7], -v[0:1], v[2:3], 1.0
559 ; SI-SAFE-NEXT: v_fma_f64 v[2:3], v[2:3], v[6:7], v[2:3]
560 ; SI-SAFE-NEXT: v_mul_f64 v[6:7], v[4:5], v[2:3]
561 ; SI-SAFE-NEXT: v_fma_f64 v[0:1], -v[0:1], v[6:7], v[4:5]
562 ; SI-SAFE-NEXT: v_div_fmas_f64 v[0:1], v[0:1], v[2:3], v[6:7]
563 ; SI-SAFE-NEXT: v_mov_b32_e32 v2, s1
564 ; SI-SAFE-NEXT: v_mov_b32_e32 v3, s0
565 ; SI-SAFE-NEXT: v_div_fixup_f64 v[0:1], v[0:1], s[2:3], 1.0
566 ; SI-SAFE-NEXT: v_mul_f64 v[0:1], v[0:1], 0
567 ; SI-SAFE-NEXT: v_add_f64 v[0:1], v[0:1], 0
568 ; SI-SAFE-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[0:1]
569 ; SI-SAFE-NEXT: v_xor_b32_e32 v4, 0x80000000, v1
570 ; SI-SAFE-NEXT: v_cndmask_b32_e32 v1, v4, v2, vcc
571 ; SI-SAFE-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
572 ; SI-SAFE-NEXT: v_cmp_nlt_f64_e32 vcc, 0, v[0:1]
573 ; SI-SAFE-NEXT: s_and_b64 s[0:1], vcc, exec
574 ; SI-SAFE-NEXT: s_cselect_b32 s1, 0, 0x7ff80000
575 ; SI-SAFE-NEXT: s_mov_b32 s0, 0
576 ; SI-SAFE-NEXT: ; return to shader part epilog
577 ;
578 ; SI-NSZ-LABEL: fneg_fadd_0_f64:
579 ; SI-NSZ: ; %bb.0: ; %.entry
580 ; SI-NSZ-NEXT: v_div_scale_f64 v[0:1], s[4:5], s[2:3], s[2:3], 1.0
581 ; SI-NSZ-NEXT: s_mov_b32 s4, 0
582 ; SI-NSZ-NEXT: s_brev_b32 s5, 1
583 ; SI-NSZ-NEXT: v_rcp_f64_e32 v[2:3], v[0:1]
584 ; SI-NSZ-NEXT: v_fma_f64 v[4:5], -v[0:1], v[2:3], 1.0
585 ; SI-NSZ-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[2:3]
586 ; SI-NSZ-NEXT: v_div_scale_f64 v[4:5], vcc, 1.0, s[2:3], 1.0
587 ; SI-NSZ-NEXT: v_fma_f64 v[6:7], -v[0:1], v[2:3], 1.0
588 ; SI-NSZ-NEXT: v_fma_f64 v[2:3], v[2:3], v[6:7], v[2:3]
589 ; SI-NSZ-NEXT: v_mul_f64 v[6:7], v[4:5], v[2:3]
590 ; SI-NSZ-NEXT: v_fma_f64 v[0:1], -v[0:1], v[6:7], v[4:5]
591 ; SI-NSZ-NEXT: v_div_fmas_f64 v[0:1], v[0:1], v[2:3], v[6:7]
592 ; SI-NSZ-NEXT: v_mov_b32_e32 v2, s1
593 ; SI-NSZ-NEXT: v_mov_b32_e32 v3, s0
594 ; SI-NSZ-NEXT: v_div_fixup_f64 v[0:1], v[0:1], s[2:3], 1.0
595 ; SI-NSZ-NEXT: v_mul_f64 v[0:1], v[0:1], s[4:5]
596 ; SI-NSZ-NEXT: v_cmp_nlt_f64_e64 vcc, -v[0:1], s[0:1]
597 ; SI-NSZ-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
598 ; SI-NSZ-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
599 ; SI-NSZ-NEXT: v_cmp_nlt_f64_e32 vcc, 0, v[0:1]
600 ; SI-NSZ-NEXT: s_and_b64 s[0:1], vcc, exec
601 ; SI-NSZ-NEXT: s_cselect_b32 s1, 0, 0x7ff80000
602 ; SI-NSZ-NEXT: s_mov_b32 s0, 0
603 ; SI-NSZ-NEXT: ; return to shader part epilog
604 ;
605 ; VI-SAFE-LABEL: fneg_fadd_0_f64:
606 ; VI-SAFE: ; %bb.0: ; %.entry
607 ; VI-SAFE-NEXT: v_div_scale_f64 v[0:1], s[4:5], s[2:3], s[2:3], 1.0
608 ; VI-SAFE-NEXT: v_rcp_f64_e32 v[2:3], v[0:1]
609 ; VI-SAFE-NEXT: v_fma_f64 v[4:5], -v[0:1], v[2:3], 1.0
610 ; VI-SAFE-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[2:3]
611 ; VI-SAFE-NEXT: v_div_scale_f64 v[4:5], vcc, 1.0, s[2:3], 1.0
612 ; VI-SAFE-NEXT: v_fma_f64 v[6:7], -v[0:1], v[2:3], 1.0
613 ; VI-SAFE-NEXT: v_fma_f64 v[2:3], v[2:3], v[6:7], v[2:3]
614 ; VI-SAFE-NEXT: v_mul_f64 v[6:7], v[4:5], v[2:3]
615 ; VI-SAFE-NEXT: v_fma_f64 v[0:1], -v[0:1], v[6:7], v[4:5]
616 ; VI-SAFE-NEXT: v_mov_b32_e32 v4, s0
617 ; VI-SAFE-NEXT: v_div_fmas_f64 v[0:1], v[0:1], v[2:3], v[6:7]
618 ; VI-SAFE-NEXT: v_mov_b32_e32 v2, s1
619 ; VI-SAFE-NEXT: v_div_fixup_f64 v[0:1], v[0:1], s[2:3], 1.0
620 ; VI-SAFE-NEXT: v_mul_f64 v[0:1], v[0:1], 0
621 ; VI-SAFE-NEXT: v_add_f64 v[0:1], v[0:1], 0
622 ; VI-SAFE-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[0:1]
623 ; VI-SAFE-NEXT: v_xor_b32_e32 v3, 0x80000000, v1
624 ; VI-SAFE-NEXT: v_cndmask_b32_e32 v1, v3, v2, vcc
625 ; VI-SAFE-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
626 ; VI-SAFE-NEXT: v_cmp_nlt_f64_e32 vcc, 0, v[0:1]
627 ; VI-SAFE-NEXT: s_and_b64 s[0:1], vcc, exec
628 ; VI-SAFE-NEXT: s_cselect_b32 s1, 0, 0x7ff80000
629 ; VI-SAFE-NEXT: s_mov_b32 s0, 0
630 ; VI-SAFE-NEXT: ; return to shader part epilog
631 ;
632 ; VI-NSZ-LABEL: fneg_fadd_0_f64:
633 ; VI-NSZ: ; %bb.0: ; %.entry
634 ; VI-NSZ-NEXT: v_div_scale_f64 v[0:1], s[4:5], s[2:3], s[2:3], 1.0
635 ; VI-NSZ-NEXT: v_rcp_f64_e32 v[2:3], v[0:1]
636 ; VI-NSZ-NEXT: v_fma_f64 v[4:5], -v[0:1], v[2:3], 1.0
637 ; VI-NSZ-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[2:3]
638 ; VI-NSZ-NEXT: v_div_scale_f64 v[4:5], vcc, 1.0, s[2:3], 1.0
639 ; VI-NSZ-NEXT: v_fma_f64 v[6:7], -v[0:1], v[2:3], 1.0
640 ; VI-NSZ-NEXT: v_fma_f64 v[2:3], v[2:3], v[6:7], v[2:3]
641 ; VI-NSZ-NEXT: v_mul_f64 v[6:7], v[4:5], v[2:3]
642 ; VI-NSZ-NEXT: v_fma_f64 v[0:1], -v[0:1], v[6:7], v[4:5]
643 ; VI-NSZ-NEXT: v_div_fmas_f64 v[0:1], v[0:1], v[2:3], v[6:7]
644 ; VI-NSZ-NEXT: v_mov_b32_e32 v2, s1
645 ; VI-NSZ-NEXT: v_mov_b32_e32 v3, s0
646 ; VI-NSZ-NEXT: v_div_fixup_f64 v[0:1], v[0:1], s[2:3], 1.0
647 ; VI-NSZ-NEXT: s_mov_b32 s2, 0
648 ; VI-NSZ-NEXT: s_brev_b32 s3, 1
649 ; VI-NSZ-NEXT: v_mul_f64 v[0:1], v[0:1], s[2:3]
650 ; VI-NSZ-NEXT: v_cmp_nlt_f64_e64 vcc, -v[0:1], s[0:1]
651 ; VI-NSZ-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
652 ; VI-NSZ-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
653 ; VI-NSZ-NEXT: v_cmp_nlt_f64_e32 vcc, 0, v[0:1]
654 ; VI-NSZ-NEXT: s_and_b64 s[0:1], vcc, exec
655 ; VI-NSZ-NEXT: s_cselect_b32 s1, 0, 0x7ff80000
656 ; VI-NSZ-NEXT: s_mov_b32 s0, 0
657 ; VI-NSZ-NEXT: ; return to shader part epilog
658 .entry:
659 %tmp7 = fdiv double 1.000000e+00, %tmp6
660 %tmp8 = fmul double 0.000000e+00, %tmp7
661 %tmp9 = fmul reassoc nnan arcp contract double 0.000000e+00, %tmp8
662 %.i188 = fadd double %tmp9, 0.000000e+00
663 %tmp10 = fcmp uge double %.i188, %tmp2
664 %tmp11 = fneg double %.i188
665 %.i092 = select i1 %tmp10, double %tmp2, double %tmp11
666 %tmp12 = fcmp ule double %.i092, 0.000000e+00
667 %.i198 = select i1 %tmp12, double 0.000000e+00, double 0x7FF8000000000000
668 ret double %.i198
669 }
671 ; This is a workaround because -enable-no-signed-zeros-fp-math does not set up
672 ; function attribute unsafe-fp-math automatically. Combine with the previous test
673 ; when that is done.
674 define amdgpu_ps double @fneg_fadd_0_nsz_f64(double inreg %tmp2, double inreg %tmp6, <4 x i32> %arg) #2 {
675 ; GCN-SAFE-LABEL: fneg_fadd_0_nsz_f64:
676 ; GCN-SAFE: ; %bb.0: ; %.entry
677 ; GCN-SAFE-NEXT: v_cmp_ngt_f64_e64 s[2:3], s[0:1], 0
678 ; GCN-SAFE-NEXT: s_and_b64 s[2:3], s[2:3], exec
679 ; GCN-SAFE-NEXT: s_cselect_b32 s1, s1, 0x80000000
680 ; GCN-SAFE-NEXT: s_cselect_b32 s0, s0, 0
681 ; GCN-SAFE-NEXT: v_cmp_ngt_f64_e64 s[0:1], s[0:1], 0
682 ; GCN-SAFE-NEXT: s_and_b64 s[0:1], s[0:1], exec
683 ; GCN-SAFE-NEXT: s_cselect_b32 s1, 0, 0x7ff80000
684 ; GCN-SAFE-NEXT: s_mov_b32 s0, 0
685 ; GCN-SAFE-NEXT: ; return to shader part epilog
686 ;
687 ; SI-NSZ-LABEL: fneg_fadd_0_nsz_f64:
688 ; SI-NSZ: ; %bb.0: ; %.entry
689 ; SI-NSZ-NEXT: v_rcp_f64_e32 v[0:1], s[2:3]
690 ; SI-NSZ-NEXT: v_fma_f64 v[2:3], -s[2:3], v[0:1], 1.0
691 ; SI-NSZ-NEXT: v_fma_f64 v[0:1], v[2:3], v[0:1], v[0:1]
692 ; SI-NSZ-NEXT: v_fma_f64 v[2:3], -s[2:3], v[0:1], 1.0
693 ; SI-NSZ-NEXT: v_fma_f64 v[0:1], v[2:3], v[0:1], v[0:1]
694 ; SI-NSZ-NEXT: v_fma_f64 v[2:3], -s[2:3], v[0:1], 1.0
695 ; SI-NSZ-NEXT: s_mov_b32 s2, 0
696 ; SI-NSZ-NEXT: v_fma_f64 v[0:1], v[2:3], v[0:1], v[0:1]
697 ; SI-NSZ-NEXT: s_brev_b32 s3, 1
698 ; SI-NSZ-NEXT: v_mul_f64 v[0:1], v[0:1], s[2:3]
699 ; SI-NSZ-NEXT: v_mov_b32_e32 v2, s1
700 ; SI-NSZ-NEXT: v_cmp_nlt_f64_e64 vcc, -v[0:1], s[0:1]
701 ; SI-NSZ-NEXT: v_mov_b32_e32 v3, s0
702 ; SI-NSZ-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
703 ; SI-NSZ-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
704 ; SI-NSZ-NEXT: v_cmp_nlt_f64_e32 vcc, 0, v[0:1]
705 ; SI-NSZ-NEXT: s_and_b64 s[0:1], vcc, exec
706 ; SI-NSZ-NEXT: s_cselect_b32 s1, 0, 0x7ff80000
707 ; SI-NSZ-NEXT: s_mov_b32 s0, 0
708 ; SI-NSZ-NEXT: ; return to shader part epilog
709 ;
710 ; VI-NSZ-LABEL: fneg_fadd_0_nsz_f64:
711 ; VI-NSZ: ; %bb.0: ; %.entry
712 ; VI-NSZ-NEXT: v_rcp_f64_e32 v[0:1], s[2:3]
713 ; VI-NSZ-NEXT: v_fma_f64 v[2:3], -s[2:3], v[0:1], 1.0
714 ; VI-NSZ-NEXT: v_fma_f64 v[0:1], v[2:3], v[0:1], v[0:1]
715 ; VI-NSZ-NEXT: v_fma_f64 v[2:3], -s[2:3], v[0:1], 1.0
716 ; VI-NSZ-NEXT: v_fma_f64 v[0:1], v[2:3], v[0:1], v[0:1]
717 ; VI-NSZ-NEXT: v_fma_f64 v[2:3], -s[2:3], v[0:1], 1.0
718 ; VI-NSZ-NEXT: s_mov_b32 s2, 0
719 ; VI-NSZ-NEXT: s_brev_b32 s3, 1
720 ; VI-NSZ-NEXT: v_fma_f64 v[0:1], v[2:3], v[0:1], v[0:1]
721 ; VI-NSZ-NEXT: v_mov_b32_e32 v2, s1
722 ; VI-NSZ-NEXT: v_mov_b32_e32 v3, s0
723 ; VI-NSZ-NEXT: v_mul_f64 v[0:1], v[0:1], s[2:3]
724 ; VI-NSZ-NEXT: v_cmp_nlt_f64_e64 vcc, -v[0:1], s[0:1]
725 ; VI-NSZ-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
726 ; VI-NSZ-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
727 ; VI-NSZ-NEXT: v_cmp_nlt_f64_e32 vcc, 0, v[0:1]
728 ; VI-NSZ-NEXT: s_and_b64 s[0:1], vcc, exec
729 ; VI-NSZ-NEXT: s_cselect_b32 s1, 0, 0x7ff80000
730 ; VI-NSZ-NEXT: s_mov_b32 s0, 0
731 ; VI-NSZ-NEXT: ; return to shader part epilog
732 .entry:
733 %tmp7 = fdiv afn double 1.000000e+00, %tmp6
734 %tmp8 = fmul double 0.000000e+00, %tmp7
735 %tmp9 = fmul reassoc nnan arcp contract double 0.000000e+00, %tmp8
736 %.i188 = fadd double %tmp9, 0.000000e+00
737 %tmp10 = fcmp uge double %.i188, %tmp2
738 %tmp11 = fneg double %.i188
739 %.i092 = select i1 %tmp10, double %tmp2, double %tmp11
740 %tmp12 = fcmp ule double %.i092, 0.000000e+00
741 %.i198 = select i1 %tmp12, double 0.000000e+00, double 0x7FF8000000000000
742 ret double %.i198
743 }
745 ; --------------------------------------------------------------------------------
746 ; fmul tests
747 ; --------------------------------------------------------------------------------
749 define float @v_fneg_mul_f32(float %a, float %b) #0 {
750 ; GCN-LABEL: v_fneg_mul_f32:
751 ; GCN: ; %bb.0:
752 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
753 ; GCN-NEXT: v_mul_f32_e64 v0, v0, -v1
754 ; GCN-NEXT: s_setpc_b64 s[30:31]
755 %mul = fmul float %a, %b
756 %fneg = fneg float %mul
757 ret float %fneg
758 }
760 define { float, float } @v_fneg_mul_store_use_mul_f32(float %a, float %b) #0 {
761 ; GCN-LABEL: v_fneg_mul_store_use_mul_f32:
762 ; GCN: ; %bb.0:
763 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
764 ; GCN-NEXT: v_mul_f32_e32 v1, v0, v1
765 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
766 ; GCN-NEXT: s_setpc_b64 s[30:31]
767 %mul = fmul float %a, %b
768 %fneg = fneg float %mul
769 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
770 %insert.1 = insertvalue { float, float } %insert.0, float %mul, 1
771 ret { float, float } %insert.1
772 }
774 define { float, float } @v_fneg_mul_multi_use_mul_f32(float %a, float %b) #0 {
775 ; GCN-LABEL: v_fneg_mul_multi_use_mul_f32:
776 ; GCN: ; %bb.0:
777 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
778 ; GCN-NEXT: v_mul_f32_e64 v0, v0, -v1
779 ; GCN-NEXT: v_mul_f32_e32 v1, -4.0, v0
780 ; GCN-NEXT: s_setpc_b64 s[30:31]
781 %mul = fmul float %a, %b
782 %fneg = fneg float %mul
783 %use1 = fmul float %mul, 4.0
784 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
785 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
786 ret { float, float } %insert.1
787 }
789 define float @v_fneg_mul_fneg_x_f32(float %a, float %b) #0 {
790 ; GCN-LABEL: v_fneg_mul_fneg_x_f32:
791 ; GCN: ; %bb.0:
792 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
793 ; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
794 ; GCN-NEXT: s_setpc_b64 s[30:31]
795 %fneg.a = fneg float %a
796 %mul = fmul float %fneg.a, %b
797 %fneg = fneg float %mul
798 ret float %fneg
799 }
801 define float @v_fneg_mul_x_fneg_f32(float %a, float %b) #0 {
802 ; GCN-LABEL: v_fneg_mul_x_fneg_f32:
803 ; GCN: ; %bb.0:
804 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
805 ; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
806 ; GCN-NEXT: s_setpc_b64 s[30:31]
807 %fneg.b = fneg float %b
808 %mul = fmul float %a, %fneg.b
809 %fneg = fneg float %mul
810 ret float %fneg
811 }
813 define float @v_fneg_mul_fneg_fneg_f32(float %a, float %b) #0 {
814 ; GCN-LABEL: v_fneg_mul_fneg_fneg_f32:
815 ; GCN: ; %bb.0:
816 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
817 ; GCN-NEXT: v_mul_f32_e64 v0, v0, -v1
818 ; GCN-NEXT: s_setpc_b64 s[30:31]
819 %fneg.a = fneg float %a
820 %fneg.b = fneg float %b
821 %mul = fmul float %fneg.a, %fneg.b
822 %fneg = fneg float %mul
823 ret float %fneg
824 }
826 define { float, float } @v_fneg_mul_store_use_fneg_x_f32(float %a, float %b) #0 {
827 ; GCN-LABEL: v_fneg_mul_store_use_fneg_x_f32:
828 ; GCN: ; %bb.0:
829 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
830 ; GCN-NEXT: v_xor_b32_e32 v2, 0x80000000, v0
831 ; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
832 ; GCN-NEXT: v_mov_b32_e32 v1, v2
833 ; GCN-NEXT: s_setpc_b64 s[30:31]
834 %fneg.a = fneg float %a
835 %mul = fmul float %fneg.a, %b
836 %fneg = fneg float %mul
837 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
838 %insert.1 = insertvalue { float, float } %insert.0, float %fneg.a, 1
839 ret { float, float } %insert.1
840 }
842 define { float, float } @v_fneg_mul_multi_use_fneg_x_f32(float %a, float %b, float %c) #0 {
843 ; GCN-LABEL: v_fneg_mul_multi_use_fneg_x_f32:
844 ; GCN: ; %bb.0:
845 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
846 ; GCN-NEXT: v_mul_f32_e32 v3, v0, v1
847 ; GCN-NEXT: v_mul_f32_e64 v1, -v0, v2
848 ; GCN-NEXT: v_mov_b32_e32 v0, v3
849 ; GCN-NEXT: s_setpc_b64 s[30:31]
850 %fneg.a = fneg float %a
851 %mul = fmul float %fneg.a, %b
852 %fneg = fneg float %mul
853 %use1 = fmul float %fneg.a, %c
854 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
855 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
856 ret { float, float } %insert.1
857 }
859 ; --------------------------------------------------------------------------------
860 ; fminnum tests
861 ; --------------------------------------------------------------------------------
863 define float @v_fneg_minnum_f32_ieee(float %a, float %b) #0 {
864 ; GCN-LABEL: v_fneg_minnum_f32_ieee:
865 ; GCN: ; %bb.0:
866 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
867 ; GCN-NEXT: v_mul_f32_e32 v1, -1.0, v1
868 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
869 ; GCN-NEXT: v_max_f32_e32 v0, v0, v1
870 ; GCN-NEXT: s_setpc_b64 s[30:31]
871 %min = call float @llvm.minnum.f32(float %a, float %b)
872 %fneg = fneg float %min
873 ret float %fneg
874 }
876 define float @v_fneg_minnum_f32_no_ieee(float %a, float %b) #4 {
877 ; GCN-LABEL: v_fneg_minnum_f32_no_ieee:
878 ; GCN: ; %bb.0:
879 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
880 ; GCN-NEXT: v_max_f32_e64 v0, -v0, -v1
881 ; GCN-NEXT: s_setpc_b64 s[30:31]
882 %min = call float @llvm.minnum.f32(float %a, float %b)
883 %fneg = fneg float %min
884 ret float %fneg
885 }
887 define float @v_fneg_self_minnum_f32_ieee(float %a) #0 {
888 ; GCN-LABEL: v_fneg_self_minnum_f32_ieee:
889 ; GCN: ; %bb.0:
890 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
891 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
892 ; GCN-NEXT: s_setpc_b64 s[30:31]
893 %min = call float @llvm.minnum.f32(float %a, float %a)
894 %min.fneg = fneg float %min
895 ret float %min.fneg
896 }
898 define float @v_fneg_self_minnum_f32_no_ieee(float %a) #4 {
899 ; GCN-LABEL: v_fneg_self_minnum_f32_no_ieee:
900 ; GCN: ; %bb.0:
901 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
902 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
903 ; GCN-NEXT: s_setpc_b64 s[30:31]
904 %min = call float @llvm.minnum.f32(float %a, float %a)
905 %min.fneg = fneg float %min
906 ret float %min.fneg
907 }
909 define float @v_fneg_posk_minnum_f32_ieee(float %a) #0 {
910 ; GCN-LABEL: v_fneg_posk_minnum_f32_ieee:
911 ; GCN: ; %bb.0:
912 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
913 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
914 ; GCN-NEXT: v_max_f32_e32 v0, -4.0, v0
915 ; GCN-NEXT: s_setpc_b64 s[30:31]
916 %min = call float @llvm.minnum.f32(float 4.0, float %a)
917 %fneg = fneg float %min
918 ret float %fneg
919 }
921 define float @v_fneg_posk_minnum_f32_no_ieee(float %a) #4 {
922 ; GCN-LABEL: v_fneg_posk_minnum_f32_no_ieee:
923 ; GCN: ; %bb.0:
924 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
925 ; GCN-NEXT: v_max_f32_e64 v0, -v0, -4.0
926 ; GCN-NEXT: s_setpc_b64 s[30:31]
927 %min = call float @llvm.minnum.f32(float 4.0, float %a)
928 %fneg = fneg float %min
929 ret float %fneg
930 }
932 define float @v_fneg_negk_minnum_f32_ieee(float %a) #0 {
933 ; GCN-LABEL: v_fneg_negk_minnum_f32_ieee:
934 ; GCN: ; %bb.0:
935 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
936 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
937 ; GCN-NEXT: v_max_f32_e32 v0, 4.0, v0
938 ; GCN-NEXT: s_setpc_b64 s[30:31]
939 %min = call float @llvm.minnum.f32(float -4.0, float %a)
940 %fneg = fneg float %min
941 ret float %fneg
942 }
944 define float @v_fneg_negk_minnum_f32_no_ieee(float %a) #4 {
945 ; GCN-LABEL: v_fneg_negk_minnum_f32_no_ieee:
946 ; GCN: ; %bb.0:
947 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
948 ; GCN-NEXT: v_max_f32_e64 v0, -v0, 4.0
949 ; GCN-NEXT: s_setpc_b64 s[30:31]
950 %min = call float @llvm.minnum.f32(float -4.0, float %a)
951 %fneg = fneg float %min
952 ret float %fneg
953 }
955 define float @v_fneg_0_minnum_f32(float %a) #0 {
956 ; GCN-LABEL: v_fneg_0_minnum_f32:
957 ; GCN: ; %bb.0:
958 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
959 ; GCN-NEXT: v_min_f32_e32 v0, 0, v0
960 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
961 ; GCN-NEXT: s_setpc_b64 s[30:31]
962 %min = call nnan float @llvm.minnum.f32(float 0.0, float %a)
963 %fneg = fneg float %min
964 ret float %fneg
965 }
967 define float @v_fneg_neg0_minnum_f32_ieee(float %a) #0 {
968 ; GCN-LABEL: v_fneg_neg0_minnum_f32_ieee:
969 ; GCN: ; %bb.0:
970 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
971 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
972 ; GCN-NEXT: v_max_f32_e32 v0, 0, v0
973 ; GCN-NEXT: s_setpc_b64 s[30:31]
974 %min = call float @llvm.minnum.f32(float -0.0, float %a)
975 %fneg = fneg float %min
976 ret float %fneg
977 }
979 define float @v_fneg_inv2pi_minnum_f32(float %a) #0 {
980 ; SI-LABEL: v_fneg_inv2pi_minnum_f32:
981 ; SI: ; %bb.0:
982 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
983 ; SI-NEXT: v_mul_f32_e32 v0, -1.0, v0
984 ; SI-NEXT: v_max_f32_e32 v0, 0xbe22f983, v0
985 ; SI-NEXT: s_setpc_b64 s[30:31]
986 ;
987 ; VI-LABEL: v_fneg_inv2pi_minnum_f32:
988 ; VI: ; %bb.0:
989 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
990 ; VI-NEXT: v_mul_f32_e32 v0, 1.0, v0
991 ; VI-NEXT: v_min_f32_e32 v0, 0.15915494, v0
992 ; VI-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
993 ; VI-NEXT: s_setpc_b64 s[30:31]
994 %min = call float @llvm.minnum.f32(float 0x3FC45F3060000000, float %a)
995 %fneg = fneg float %min
996 ret float %fneg
997 }
999 define float @v_fneg_neg_inv2pi_minnum_f32(float %a) #0 {
1000 ; SI-LABEL: v_fneg_neg_inv2pi_minnum_f32:
1001 ; SI: ; %bb.0:
1002 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1003 ; SI-NEXT: v_mul_f32_e32 v0, -1.0, v0
1004 ; SI-NEXT: v_max_f32_e32 v0, 0x3e22f983, v0
1005 ; SI-NEXT: s_setpc_b64 s[30:31]
1006 ;
1007 ; VI-LABEL: v_fneg_neg_inv2pi_minnum_f32:
1008 ; VI: ; %bb.0:
1009 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1010 ; VI-NEXT: v_mul_f32_e32 v0, -1.0, v0
1011 ; VI-NEXT: v_max_f32_e32 v0, 0.15915494, v0
1012 ; VI-NEXT: s_setpc_b64 s[30:31]
1013 %min = call float @llvm.minnum.f32(float 0xBFC45F3060000000, float %a)
1014 %fneg = fneg float %min
1015 ret float %fneg
1016 }
1018 define half @v_fneg_inv2pi_minnum_f16(half %a) #0 {
1019 ; SI-LABEL: v_fneg_inv2pi_minnum_f16:
1020 ; SI: ; %bb.0:
1021 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1022 ; SI-NEXT: v_cvt_f16_f32_e64 v0, -v0
1023 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1024 ; SI-NEXT: v_max_f32_e32 v0, 0xbe230000, v0
1025 ; SI-NEXT: s_setpc_b64 s[30:31]
1026 ;
1027 ; VI-LABEL: v_fneg_inv2pi_minnum_f16:
1028 ; VI: ; %bb.0:
1029 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1030 ; VI-NEXT: v_max_f16_e32 v0, v0, v0
1031 ; VI-NEXT: v_min_f16_e32 v0, 0.15915494, v0
1032 ; VI-NEXT: v_xor_b32_e32 v0, 0x8000, v0
1033 ; VI-NEXT: s_setpc_b64 s[30:31]
1034 %min = call half @llvm.minnum.f16(half 0xH3118, half %a)
1035 %fneg = fneg half %min
1036 ret half %fneg
1037 }
1039 define half @v_fneg_neg_inv2pi_minnum_f16(half %a) #0 {
1040 ; SI-LABEL: v_fneg_neg_inv2pi_minnum_f16:
1041 ; SI: ; %bb.0:
1042 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1043 ; SI-NEXT: v_cvt_f16_f32_e64 v0, -v0
1044 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1045 ; SI-NEXT: v_max_f32_e32 v0, 0x3e230000, v0
1046 ; SI-NEXT: s_setpc_b64 s[30:31]
1047 ;
1048 ; VI-LABEL: v_fneg_neg_inv2pi_minnum_f16:
1049 ; VI: ; %bb.0:
1050 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1051 ; VI-NEXT: v_max_f16_e64 v0, -v0, -v0
1052 ; VI-NEXT: v_max_f16_e32 v0, 0.15915494, v0
1053 ; VI-NEXT: s_setpc_b64 s[30:31]
1054 %min = call half @llvm.minnum.f16(half 0xHB118, half %a)
1055 %fneg = fneg half %min
1056 ret half %fneg
1057 }
1059 define double @v_fneg_inv2pi_minnum_f64(double %a) #0 {
1060 ; SI-LABEL: v_fneg_inv2pi_minnum_f64:
1061 ; SI: ; %bb.0:
1062 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1063 ; SI-NEXT: v_max_f64 v[0:1], -v[0:1], -v[0:1]
1064 ; SI-NEXT: s_mov_b32 s4, 0x6dc9c882
1065 ; SI-NEXT: s_mov_b32 s5, 0xbfc45f30
1066 ; SI-NEXT: v_max_f64 v[0:1], v[0:1], s[4:5]
1067 ; SI-NEXT: s_setpc_b64 s[30:31]
1068 ;
1069 ; VI-LABEL: v_fneg_inv2pi_minnum_f64:
1070 ; VI: ; %bb.0:
1071 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1072 ; VI-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
1073 ; VI-NEXT: v_min_f64 v[0:1], v[0:1], 0.15915494309189532
1074 ; VI-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
1075 ; VI-NEXT: s_setpc_b64 s[30:31]
1076 %min = call double @llvm.minnum.f64(double 0x3fc45f306dc9c882, double %a)
1077 %fneg = fneg double %min
1078 ret double %fneg
1079 }
1081 define double @v_fneg_neg_inv2pi_minnum_f64(double %a) #0 {
1082 ; SI-LABEL: v_fneg_neg_inv2pi_minnum_f64:
1083 ; SI: ; %bb.0:
1084 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1085 ; SI-NEXT: v_max_f64 v[0:1], -v[0:1], -v[0:1]
1086 ; SI-NEXT: s_mov_b32 s4, 0x6dc9c882
1087 ; SI-NEXT: s_mov_b32 s5, 0x3fc45f30
1088 ; SI-NEXT: v_max_f64 v[0:1], v[0:1], s[4:5]
1089 ; SI-NEXT: s_setpc_b64 s[30:31]
1090 ;
1091 ; VI-LABEL: v_fneg_neg_inv2pi_minnum_f64:
1092 ; VI: ; %bb.0:
1093 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1094 ; VI-NEXT: v_max_f64 v[0:1], -v[0:1], -v[0:1]
1095 ; VI-NEXT: v_max_f64 v[0:1], v[0:1], 0.15915494309189532
1096 ; VI-NEXT: s_setpc_b64 s[30:31]
1097 %min = call double @llvm.minnum.f64(double 0xbfc45f306dc9c882, double %a)
1098 %fneg = fneg double %min
1099 ret double %fneg
1100 }
1102 define float @v_fneg_neg0_minnum_f32_no_ieee(float %a) #4 {
1103 ; GCN-LABEL: v_fneg_neg0_minnum_f32_no_ieee:
1104 ; GCN: ; %bb.0:
1105 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1106 ; GCN-NEXT: v_max_f32_e64 v0, -v0, 0
1107 ; GCN-NEXT: s_setpc_b64 s[30:31]
1108 %min = call float @llvm.minnum.f32(float -0.0, float %a)
1109 %fneg = fneg float %min
1110 ret float %fneg
1111 }
1113 define float @v_fneg_0_minnum_foldable_use_f32_ieee(float %a, float %b) #0 {
1114 ; GCN-LABEL: v_fneg_0_minnum_foldable_use_f32_ieee:
1115 ; GCN: ; %bb.0:
1116 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1117 ; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
1118 ; GCN-NEXT: v_min_f32_e32 v0, 0, v0
1119 ; GCN-NEXT: v_mul_f32_e64 v0, -v0, v1
1120 ; GCN-NEXT: s_setpc_b64 s[30:31]
1121 %min = call float @llvm.minnum.f32(float 0.0, float %a)
1122 %fneg = fneg float %min
1123 %mul = fmul float %fneg, %b
1124 ret float %mul
1125 }
1127 define float @v_fneg_inv2pi_minnum_foldable_use_f32(float %a, float %b) #0 {
1128 ; SI-LABEL: v_fneg_inv2pi_minnum_foldable_use_f32:
1129 ; SI: ; %bb.0:
1130 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1131 ; SI-NEXT: v_mul_f32_e32 v0, -1.0, v0
1132 ; SI-NEXT: v_max_f32_e32 v0, 0xbe22f983, v0
1133 ; SI-NEXT: v_mul_f32_e32 v0, v0, v1
1134 ; SI-NEXT: s_setpc_b64 s[30:31]
1135 ;
1136 ; VI-LABEL: v_fneg_inv2pi_minnum_foldable_use_f32:
1137 ; VI: ; %bb.0:
1138 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1139 ; VI-NEXT: v_mul_f32_e32 v0, 1.0, v0
1140 ; VI-NEXT: v_min_f32_e32 v0, 0.15915494, v0
1141 ; VI-NEXT: v_mul_f32_e64 v0, -v0, v1
1142 ; VI-NEXT: s_setpc_b64 s[30:31]
1143 %min = call float @llvm.minnum.f32(float 0x3FC45F3060000000, float %a)
1144 %fneg = fneg float %min
1145 %mul = fmul float %fneg, %b
1146 ret float %mul
1147 }
1149 define float @v_fneg_0_minnum_foldable_use_f32_no_ieee(float %a, float %b) #4 {
1150 ; GCN-LABEL: v_fneg_0_minnum_foldable_use_f32_no_ieee:
1151 ; GCN: ; %bb.0:
1152 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1153 ; GCN-NEXT: v_min_f32_e32 v0, 0, v0
1154 ; GCN-NEXT: v_mul_f32_e64 v0, -v0, v1
1155 ; GCN-NEXT: s_setpc_b64 s[30:31]
1156 %min = call float @llvm.minnum.f32(float 0.0, float %a)
1157 %fneg = fneg float %min
1158 %mul = fmul float %fneg, %b
1159 ret float %mul
1160 }
1162 define { float, float } @v_fneg_minnum_multi_use_minnum_f32_ieee(float %a, float %b) #0 {
1163 ; GCN-LABEL: v_fneg_minnum_multi_use_minnum_f32_ieee:
1164 ; GCN: ; %bb.0:
1165 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1166 ; GCN-NEXT: v_mul_f32_e32 v1, -1.0, v1
1167 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
1168 ; GCN-NEXT: v_max_f32_e32 v0, v0, v1
1169 ; GCN-NEXT: v_mul_f32_e32 v1, -4.0, v0
1170 ; GCN-NEXT: s_setpc_b64 s[30:31]
1171 %min = call float @llvm.minnum.f32(float %a, float %b)
1172 %fneg = fneg float %min
1173 %use1 = fmul float %min, 4.0
1174 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1175 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
1176 ret { float, float } %insert.1
1177 }
1179 define <2 x float> @v_fneg_minnum_multi_use_minnum_f32_no_ieee(float %a, float %b) #4 {
1180 ; GCN-LABEL: v_fneg_minnum_multi_use_minnum_f32_no_ieee:
1181 ; GCN: ; %bb.0:
1182 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1183 ; GCN-NEXT: v_max_f32_e64 v0, -v0, -v1
1184 ; GCN-NEXT: v_mul_f32_e32 v1, -4.0, v0
1185 ; GCN-NEXT: s_setpc_b64 s[30:31]
1186 %min = call float @llvm.minnum.f32(float %a, float %b)
1187 %fneg = fneg float %min
1188 %use1 = fmul float %min, 4.0
1189 %ins0 = insertelement <2 x float> undef, float %fneg, i32 0
1190 %ins1 = insertelement <2 x float> %ins0, float %use1, i32 1
1191 ret <2 x float> %ins1
1192 }
1194 ; --------------------------------------------------------------------------------
1195 ; fmaxnum tests
1196 ; --------------------------------------------------------------------------------
1198 define float @v_fneg_maxnum_f32_ieee(float %a, float %b) #0 {
1199 ; GCN-LABEL: v_fneg_maxnum_f32_ieee:
1200 ; GCN: ; %bb.0:
1201 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1202 ; GCN-NEXT: v_mul_f32_e32 v1, -1.0, v1
1203 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
1204 ; GCN-NEXT: v_min_f32_e32 v0, v0, v1
1205 ; GCN-NEXT: s_setpc_b64 s[30:31]
1206 %max = call float @llvm.maxnum.f32(float %a, float %b)
1207 %fneg = fneg float %max
1208 ret float %fneg
1209 }
1211 define float @v_fneg_maxnum_f32_no_ieee(float %a, float %b) #4 {
1212 ; GCN-LABEL: v_fneg_maxnum_f32_no_ieee:
1213 ; GCN: ; %bb.0:
1214 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1215 ; GCN-NEXT: v_min_f32_e64 v0, -v0, -v1
1216 ; GCN-NEXT: s_setpc_b64 s[30:31]
1217 %max = call float @llvm.maxnum.f32(float %a, float %b)
1218 %fneg = fneg float %max
1219 ret float %fneg
1220 }
1222 define float @v_fneg_self_maxnum_f32_ieee(float %a) #0 {
1223 ; GCN-LABEL: v_fneg_self_maxnum_f32_ieee:
1224 ; GCN: ; %bb.0:
1225 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1226 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1227 ; GCN-NEXT: s_setpc_b64 s[30:31]
1228 %max = call float @llvm.maxnum.f32(float %a, float %a)
1229 %max.fneg = fneg float %max
1230 ret float %max.fneg
1231 }
1233 define float @v_fneg_self_maxnum_f32_no_ieee(float %a) #4 {
1234 ; GCN-LABEL: v_fneg_self_maxnum_f32_no_ieee:
1235 ; GCN: ; %bb.0:
1236 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1237 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1238 ; GCN-NEXT: s_setpc_b64 s[30:31]
1239 %max = call float @llvm.maxnum.f32(float %a, float %a)
1240 %max.fneg = fneg float %max
1241 ret float %max.fneg
1242 }
1244 define float @v_fneg_posk_maxnum_f32_ieee(float %a) #0 {
1245 ; GCN-LABEL: v_fneg_posk_maxnum_f32_ieee:
1246 ; GCN: ; %bb.0:
1247 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1248 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
1249 ; GCN-NEXT: v_min_f32_e32 v0, -4.0, v0
1250 ; GCN-NEXT: s_setpc_b64 s[30:31]
1251 %max = call float @llvm.maxnum.f32(float 4.0, float %a)
1252 %fneg = fneg float %max
1253 ret float %fneg
1254 }
1256 define float @v_fneg_posk_maxnum_f32_no_ieee(float %a) #4 {
1257 ; GCN-LABEL: v_fneg_posk_maxnum_f32_no_ieee:
1258 ; GCN: ; %bb.0:
1259 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1260 ; GCN-NEXT: v_min_f32_e64 v0, -v0, -4.0
1261 ; GCN-NEXT: s_setpc_b64 s[30:31]
1262 %max = call float @llvm.maxnum.f32(float 4.0, float %a)
1263 %fneg = fneg float %max
1264 ret float %fneg
1265 }
1267 define float @v_fneg_negk_maxnum_f32_ieee(float %a) #0 {
1268 ; GCN-LABEL: v_fneg_negk_maxnum_f32_ieee:
1269 ; GCN: ; %bb.0:
1270 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1271 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
1272 ; GCN-NEXT: v_min_f32_e32 v0, 4.0, v0
1273 ; GCN-NEXT: s_setpc_b64 s[30:31]
1274 %max = call float @llvm.maxnum.f32(float -4.0, float %a)
1275 %fneg = fneg float %max
1276 ret float %fneg
1277 }
1279 define float @v_fneg_negk_maxnum_f32_no_ieee(float %a) #4 {
1280 ; GCN-LABEL: v_fneg_negk_maxnum_f32_no_ieee:
1281 ; GCN: ; %bb.0:
1282 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1283 ; GCN-NEXT: v_min_f32_e64 v0, -v0, 4.0
1284 ; GCN-NEXT: s_setpc_b64 s[30:31]
1285 %max = call float @llvm.maxnum.f32(float -4.0, float %a)
1286 %fneg = fneg float %max
1287 ret float %fneg
1288 }
1290 define float @v_fneg_0_maxnum_f32(float %a) #0 {
1291 ; GCN-LABEL: v_fneg_0_maxnum_f32:
1292 ; GCN: ; %bb.0:
1293 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1294 ; GCN-NEXT: v_max_f32_e32 v0, 0, v0
1295 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1296 ; GCN-NEXT: s_setpc_b64 s[30:31]
1297 %max = call nnan float @llvm.maxnum.f32(float 0.0, float %a)
1298 %fneg = fneg float %max
1299 ret float %fneg
1300 }
1302 define float @v_fneg_neg0_maxnum_f32_ieee(float %a) #0 {
1303 ; GCN-LABEL: v_fneg_neg0_maxnum_f32_ieee:
1304 ; GCN: ; %bb.0:
1305 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1306 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
1307 ; GCN-NEXT: v_min_f32_e32 v0, 0, v0
1308 ; GCN-NEXT: s_setpc_b64 s[30:31]
1309 %max = call float @llvm.maxnum.f32(float -0.0, float %a)
1310 %fneg = fneg float %max
1311 ret float %fneg
1312 }
1314 define float @v_fneg_neg0_maxnum_f32_no_ieee(float %a) #4 {
1315 ; GCN-LABEL: v_fneg_neg0_maxnum_f32_no_ieee:
1316 ; GCN: ; %bb.0:
1317 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1318 ; GCN-NEXT: v_min_f32_e64 v0, -v0, 0
1319 ; GCN-NEXT: s_setpc_b64 s[30:31]
1320 %max = call float @llvm.maxnum.f32(float -0.0, float %a)
1321 %fneg = fneg float %max
1322 ret float %fneg
1323 }
1325 define float @v_fneg_0_maxnum_foldable_use_f32_ieee(float %a, float %b) #0 {
1326 ; GCN-LABEL: v_fneg_0_maxnum_foldable_use_f32_ieee:
1327 ; GCN: ; %bb.0:
1328 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1329 ; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
1330 ; GCN-NEXT: v_max_f32_e32 v0, 0, v0
1331 ; GCN-NEXT: v_mul_f32_e64 v0, -v0, v1
1332 ; GCN-NEXT: s_setpc_b64 s[30:31]
1333 %max = call float @llvm.maxnum.f32(float 0.0, float %a)
1334 %fneg = fneg float %max
1335 %mul = fmul float %fneg, %b
1336 ret float %mul
1337 }
1339 define float @v_fneg_0_maxnum_foldable_use_f32_no_ieee(float %a, float %b) #4 {
1340 ; GCN-LABEL: v_fneg_0_maxnum_foldable_use_f32_no_ieee:
1341 ; GCN: ; %bb.0:
1342 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1343 ; GCN-NEXT: v_max_f32_e32 v0, 0, v0
1344 ; GCN-NEXT: v_mul_f32_e64 v0, -v0, v1
1345 ; GCN-NEXT: s_setpc_b64 s[30:31]
1346 %max = call float @llvm.maxnum.f32(float 0.0, float %a)
1347 %fneg = fneg float %max
1348 %mul = fmul float %fneg, %b
1349 ret float %mul
1350 }
1352 define { float, float } @v_fneg_maxnum_multi_use_maxnum_f32_ieee(float %a, float %b) #0 {
1353 ; GCN-LABEL: v_fneg_maxnum_multi_use_maxnum_f32_ieee:
1354 ; GCN: ; %bb.0:
1355 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1356 ; GCN-NEXT: v_mul_f32_e32 v1, -1.0, v1
1357 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
1358 ; GCN-NEXT: v_min_f32_e32 v0, v0, v1
1359 ; GCN-NEXT: v_mul_f32_e32 v1, -4.0, v0
1360 ; GCN-NEXT: s_setpc_b64 s[30:31]
1361 %max = call float @llvm.maxnum.f32(float %a, float %b)
1362 %fneg = fneg float %max
1363 %use1 = fmul float %max, 4.0
1364 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1365 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
1366 ret { float, float } %insert.1
1367 }
1369 define <2 x float> @v_fneg_maxnum_multi_use_maxnum_f32_no_ieee(float %a, float %b) #4 {
1370 ; GCN-LABEL: v_fneg_maxnum_multi_use_maxnum_f32_no_ieee:
1371 ; GCN: ; %bb.0:
1372 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1373 ; GCN-NEXT: v_min_f32_e64 v0, -v0, -v1
1374 ; GCN-NEXT: v_mul_f32_e32 v1, -4.0, v0
1375 ; GCN-NEXT: s_setpc_b64 s[30:31]
1376 %max = call float @llvm.maxnum.f32(float %a, float %b)
1377 %fneg = fneg float %max
1378 %use1 = fmul float %max, 4.0
1379 %ins0 = insertelement <2 x float> undef, float %fneg, i32 0
1380 %ins1 = insertelement <2 x float> %ins0, float %use1, i32 1
1381 ret <2 x float> %ins1
1382 }
1384 ; --------------------------------------------------------------------------------
1385 ; fma tests
1386 ; --------------------------------------------------------------------------------
1388 define float @v_fneg_fma_f32(float %a, float %b, float %c) #0 {
1389 ; GCN-SAFE-LABEL: v_fneg_fma_f32:
1390 ; GCN-SAFE: ; %bb.0:
1391 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1392 ; GCN-SAFE-NEXT: v_fma_f32 v0, v0, v1, v2
1393 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1394 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1395 ;
1396 ; GCN-NSZ-LABEL: v_fneg_fma_f32:
1397 ; GCN-NSZ: ; %bb.0:
1398 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1399 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, -v1, -v2
1400 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1401 %fma = call float @llvm.fma.f32(float %a, float %b, float %c)
1402 %fneg = fneg float %fma
1403 ret float %fneg
1404 }
1406 define { float, float } @v_fneg_fma_store_use_fma_f32(float %a, float %b, float %c) #0 {
1407 ; GCN-LABEL: v_fneg_fma_store_use_fma_f32:
1408 ; GCN: ; %bb.0:
1409 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1410 ; GCN-NEXT: v_fma_f32 v1, v0, v1, v2
1411 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
1412 ; GCN-NEXT: s_setpc_b64 s[30:31]
1413 %fma = call float @llvm.fma.f32(float %a, float %b, float %c)
1414 %fneg = fneg float %fma
1415 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1416 %insert.1 = insertvalue { float, float } %insert.0, float %fma, 1
1417 ret { float, float } %insert.1
1418 }
1420 define { float, float } @v_fneg_fma_multi_use_fma_f32(float %a, float %b, float %c) #0 {
1421 ; GCN-SAFE-LABEL: v_fneg_fma_multi_use_fma_f32:
1422 ; GCN-SAFE: ; %bb.0:
1423 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1424 ; GCN-SAFE-NEXT: v_fma_f32 v1, v0, v1, v2
1425 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
1426 ; GCN-SAFE-NEXT: v_mul_f32_e32 v1, 4.0, v1
1427 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1428 ;
1429 ; GCN-NSZ-LABEL: v_fneg_fma_multi_use_fma_f32:
1430 ; GCN-NSZ: ; %bb.0:
1431 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1432 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, -v1, -v2
1433 ; GCN-NSZ-NEXT: v_mul_f32_e32 v1, -4.0, v0
1434 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1435 %fma = call float @llvm.fma.f32(float %a, float %b, float %c)
1436 %fneg = fneg float %fma
1437 %use1 = fmul float %fma, 4.0
1438 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1439 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
1440 ret { float, float } %insert.1
1441 }
1443 define float @v_fneg_fma_fneg_x_y_f32(float %a, float %b, float %c) #0 {
1444 ; GCN-SAFE-LABEL: v_fneg_fma_fneg_x_y_f32:
1445 ; GCN-SAFE: ; %bb.0:
1446 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1447 ; GCN-SAFE-NEXT: v_fma_f32 v0, -v0, v1, v2
1448 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1449 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1450 ;
1451 ; GCN-NSZ-LABEL: v_fneg_fma_fneg_x_y_f32:
1452 ; GCN-NSZ: ; %bb.0:
1453 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1454 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, v1, -v2
1455 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1456 %fneg.a = fneg float %a
1457 %fma = call float @llvm.fma.f32(float %fneg.a, float %b, float %c)
1458 %fneg = fneg float %fma
1459 ret float %fneg
1460 }
1462 define float @v_fneg_fma_x_fneg_y_f32(float %a, float %b, float %c) #0 {
1463 ; GCN-SAFE-LABEL: v_fneg_fma_x_fneg_y_f32:
1464 ; GCN-SAFE: ; %bb.0:
1465 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1466 ; GCN-SAFE-NEXT: v_fma_f32 v0, v0, -v1, v2
1467 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1468 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1469 ;
1470 ; GCN-NSZ-LABEL: v_fneg_fma_x_fneg_y_f32:
1471 ; GCN-NSZ: ; %bb.0:
1472 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1473 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, v1, -v2
1474 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1475 %fneg.b = fneg float %b
1476 %fma = call float @llvm.fma.f32(float %a, float %fneg.b, float %c)
1477 %fneg = fneg float %fma
1478 ret float %fneg
1479 }
1481 define float @v_fneg_fma_fneg_fneg_y_f32(float %a, float %b, float %c) #0 {
1482 ; GCN-SAFE-LABEL: v_fneg_fma_fneg_fneg_y_f32:
1483 ; GCN-SAFE: ; %bb.0:
1484 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1485 ; GCN-SAFE-NEXT: v_fma_f32 v0, v0, v1, v2
1486 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1487 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1488 ;
1489 ; GCN-NSZ-LABEL: v_fneg_fma_fneg_fneg_y_f32:
1490 ; GCN-NSZ: ; %bb.0:
1491 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1492 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, -v1, -v2
1493 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1494 %fneg.a = fneg float %a
1495 %fneg.b = fneg float %b
1496 %fma = call float @llvm.fma.f32(float %fneg.a, float %fneg.b, float %c)
1497 %fneg = fneg float %fma
1498 ret float %fneg
1499 }
1501 define float @v_fneg_fma_fneg_x_fneg_f32(float %a, float %b, float %c) #0 {
1502 ; GCN-SAFE-LABEL: v_fneg_fma_fneg_x_fneg_f32:
1503 ; GCN-SAFE: ; %bb.0:
1504 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1505 ; GCN-SAFE-NEXT: v_fma_f32 v0, -v0, v1, -v2
1506 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1507 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1508 ;
1509 ; GCN-NSZ-LABEL: v_fneg_fma_fneg_x_fneg_f32:
1510 ; GCN-NSZ: ; %bb.0:
1511 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1512 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, v1, v2
1513 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1514 %fneg.a = fneg float %a
1515 %fneg.c = fneg float %c
1516 %fma = call float @llvm.fma.f32(float %fneg.a, float %b, float %fneg.c)
1517 %fneg = fneg float %fma
1518 ret float %fneg
1519 }
1521 define float @v_fneg_fma_x_y_fneg_f32(float %a, float %b, float %c) #0 {
1522 ; GCN-SAFE-LABEL: v_fneg_fma_x_y_fneg_f32:
1523 ; GCN-SAFE: ; %bb.0:
1524 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1525 ; GCN-SAFE-NEXT: v_fma_f32 v0, v0, v1, -v2
1526 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1527 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1528 ;
1529 ; GCN-NSZ-LABEL: v_fneg_fma_x_y_fneg_f32:
1530 ; GCN-NSZ: ; %bb.0:
1531 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1532 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, -v1, v2
1533 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1534 %fneg.c = fneg float %c
1535 %fma = call float @llvm.fma.f32(float %a, float %b, float %fneg.c)
1536 %fneg = fneg float %fma
1537 ret float %fneg
1538 }
1540 define { float, float } @v_fneg_fma_store_use_fneg_x_y_f32(float %a, float %b, float %c) #0 {
1541 ; GCN-SAFE-LABEL: v_fneg_fma_store_use_fneg_x_y_f32:
1542 ; GCN-SAFE: ; %bb.0:
1543 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1544 ; GCN-SAFE-NEXT: v_xor_b32_e32 v3, 0x80000000, v0
1545 ; GCN-SAFE-NEXT: v_fma_f32 v0, -v0, v1, v2
1546 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1547 ; GCN-SAFE-NEXT: v_mov_b32_e32 v1, v3
1548 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1549 ;
1550 ; GCN-NSZ-LABEL: v_fneg_fma_store_use_fneg_x_y_f32:
1551 ; GCN-NSZ: ; %bb.0:
1552 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1553 ; GCN-NSZ-NEXT: v_xor_b32_e32 v3, 0x80000000, v0
1554 ; GCN-NSZ-NEXT: v_fma_f32 v0, v0, v1, -v2
1555 ; GCN-NSZ-NEXT: v_mov_b32_e32 v1, v3
1556 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1557 %fneg.a = fneg float %a
1558 %fma = call float @llvm.fma.f32(float %fneg.a, float %b, float %c)
1559 %fneg = fneg float %fma
1560 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1561 %insert.1 = insertvalue { float, float } %insert.0, float %fneg.a, 1
1562 ret { float, float } %insert.1
1563 }
1565 define { float, float } @v_fneg_fma_multi_use_fneg_x_y_f32(float %a, float %b, float %c, float %d) #0 {
1566 ; GCN-SAFE-LABEL: v_fneg_fma_multi_use_fneg_x_y_f32:
1567 ; GCN-SAFE: ; %bb.0:
1568 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1569 ; GCN-SAFE-NEXT: v_fma_f32 v1, -v0, v1, v2
1570 ; GCN-SAFE-NEXT: v_xor_b32_e32 v2, 0x80000000, v1
1571 ; GCN-SAFE-NEXT: v_mul_f32_e64 v1, -v0, v3
1572 ; GCN-SAFE-NEXT: v_mov_b32_e32 v0, v2
1573 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1574 ;
1575 ; GCN-NSZ-LABEL: v_fneg_fma_multi_use_fneg_x_y_f32:
1576 ; GCN-NSZ: ; %bb.0:
1577 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1578 ; GCN-NSZ-NEXT: v_fma_f32 v2, v0, v1, -v2
1579 ; GCN-NSZ-NEXT: v_mul_f32_e64 v1, -v0, v3
1580 ; GCN-NSZ-NEXT: v_mov_b32_e32 v0, v2
1581 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1582 %fneg.a = fneg float %a
1583 %fma = call float @llvm.fma.f32(float %fneg.a, float %b, float %c)
1584 %fneg = fneg float %fma
1585 %use1 = fmul float %fneg.a, %d
1586 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1587 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
1588 ret { float, float } %insert.1
1589 }
1591 ; --------------------------------------------------------------------------------
1592 ; fmad tests
1593 ; --------------------------------------------------------------------------------
1595 define float @v_fneg_fmad_f32(float %a, float %b, float %c) #0 {
1596 ; GCN-SAFE-LABEL: v_fneg_fmad_f32:
1597 ; GCN-SAFE: ; %bb.0:
1598 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1599 ; GCN-SAFE-NEXT: v_mac_f32_e32 v2, v0, v1
1600 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v2
1601 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1602 ;
1603 ; GCN-NSZ-LABEL: v_fneg_fmad_f32:
1604 ; GCN-NSZ: ; %bb.0:
1605 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1606 ; GCN-NSZ-NEXT: v_mad_f32 v0, v0, -v1, -v2
1607 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1608 %fma = call float @llvm.fmuladd.f32(float %a, float %b, float %c)
1609 %fneg = fneg float %fma
1610 ret float %fneg
1611 }
1613 define <4 x float> @v_fneg_fmad_v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c) #0 {
1614 ; GCN-SAFE-LABEL: v_fneg_fmad_v4f32:
1615 ; GCN-SAFE: ; %bb.0:
1616 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1617 ; GCN-SAFE-NEXT: v_mac_f32_e32 v11, v3, v7
1618 ; GCN-SAFE-NEXT: v_mac_f32_e32 v10, v2, v6
1619 ; GCN-SAFE-NEXT: v_mac_f32_e32 v9, v1, v5
1620 ; GCN-SAFE-NEXT: v_mac_f32_e32 v8, v0, v4
1621 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v8
1622 ; GCN-SAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v9
1623 ; GCN-SAFE-NEXT: v_xor_b32_e32 v2, 0x80000000, v10
1624 ; GCN-SAFE-NEXT: v_xor_b32_e32 v3, 0x80000000, v11
1625 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1626 ;
1627 ; GCN-NSZ-LABEL: v_fneg_fmad_v4f32:
1628 ; GCN-NSZ: ; %bb.0:
1629 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1630 ; GCN-NSZ-NEXT: v_mad_f32 v0, v0, -v4, -v8
1631 ; GCN-NSZ-NEXT: v_mad_f32 v1, v1, -v5, -v9
1632 ; GCN-NSZ-NEXT: v_mad_f32 v2, v2, -v6, -v10
1633 ; GCN-NSZ-NEXT: v_mad_f32 v3, v3, -v7, -v11
1634 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1635 %fma = call <4 x float> @llvm.fmuladd.v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c)
1636 %fneg = fneg <4 x float> %fma
1637 ret <4 x float> %fneg
1638 }
1640 define { float, float } @v_fneg_fmad_multi_use_fmad_f32(float %a, float %b, float %c) #0 {
1641 ; GCN-SAFE-LABEL: v_fneg_fmad_multi_use_fmad_f32:
1642 ; GCN-SAFE: ; %bb.0:
1643 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1644 ; GCN-SAFE-NEXT: v_mac_f32_e32 v2, v0, v1
1645 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v2
1646 ; GCN-SAFE-NEXT: v_mul_f32_e32 v1, 4.0, v2
1647 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
1648 ;
1649 ; GCN-NSZ-LABEL: v_fneg_fmad_multi_use_fmad_f32:
1650 ; GCN-NSZ: ; %bb.0:
1651 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1652 ; GCN-NSZ-NEXT: v_mad_f32 v0, v0, -v1, -v2
1653 ; GCN-NSZ-NEXT: v_mul_f32_e32 v1, -4.0, v0
1654 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
1655 %fma = call float @llvm.fmuladd.f32(float %a, float %b, float %c)
1656 %fneg = fneg float %fma
1657 %use1 = fmul float %fma, 4.0
1658 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1659 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
1660 ret { float, float } %insert.1
1661 }
1663 ; --------------------------------------------------------------------------------
1664 ; fp_extend tests
1665 ; --------------------------------------------------------------------------------
1667 define double @v_fneg_fp_extend_f32_to_f64(float %a) #0 {
1668 ; GCN-LABEL: v_fneg_fp_extend_f32_to_f64:
1669 ; GCN: ; %bb.0:
1670 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1671 ; GCN-NEXT: v_cvt_f64_f32_e64 v[0:1], -v0
1672 ; GCN-NEXT: s_setpc_b64 s[30:31]
1673 %fpext = fpext float %a to double
1674 %fneg = fneg double %fpext
1675 ret double %fneg
1676 }
1678 define double @v_fneg_fp_extend_fneg_f32_to_f64(float %a) #0 {
1679 ; GCN-LABEL: v_fneg_fp_extend_fneg_f32_to_f64:
1680 ; GCN: ; %bb.0:
1681 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1682 ; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
1683 ; GCN-NEXT: s_setpc_b64 s[30:31]
1684 %fneg.a = fneg float %a
1685 %fpext = fpext float %fneg.a to double
1686 %fneg = fneg double %fpext
1687 ret double %fneg
1688 }
1690 define { double, float } @v_fneg_fp_extend_store_use_fneg_f32_to_f64(float %a) #0 {
1691 ; GCN-LABEL: v_fneg_fp_extend_store_use_fneg_f32_to_f64:
1692 ; GCN: ; %bb.0:
1693 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1694 ; GCN-NEXT: v_mov_b32_e32 v2, v0
1695 ; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
1696 ; GCN-NEXT: v_xor_b32_e32 v2, 0x80000000, v2
1697 ; GCN-NEXT: s_setpc_b64 s[30:31]
1698 %fneg.a = fneg float %a
1699 %fpext = fpext float %fneg.a to double
1700 %fneg = fneg double %fpext
1701 %insert.0 = insertvalue { double, float } poison, double %fneg, 0
1702 %insert.1 = insertvalue { double, float } %insert.0, float %fneg.a, 1
1703 ret { double, float } %insert.1
1704 }
1706 define { double, double } @v_fneg_multi_use_fp_extend_fneg_f32_to_f64(float %a) #0 {
1707 ; GCN-LABEL: v_fneg_multi_use_fp_extend_fneg_f32_to_f64:
1708 ; GCN: ; %bb.0:
1709 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1710 ; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v0
1711 ; GCN-NEXT: v_xor_b32_e32 v1, 0x80000000, v3
1712 ; GCN-NEXT: v_mov_b32_e32 v0, v2
1713 ; GCN-NEXT: s_setpc_b64 s[30:31]
1714 %fpext = fpext float %a to double
1715 %fneg = fneg double %fpext
1716 %insert.0 = insertvalue { double, double } poison, double %fneg, 0
1717 %insert.1 = insertvalue { double, double } %insert.0, double %fpext, 1
1718 ret { double, double } %insert.1
1719 }
1721 define { double, double } @v_fneg_multi_foldable_use_fp_extend_fneg_f32_to_f64(float %a) #0 {
1722 ; SI-LABEL: v_fneg_multi_foldable_use_fp_extend_fneg_f32_to_f64:
1723 ; SI: ; %bb.0:
1724 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1725 ; SI-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
1726 ; SI-NEXT: v_xor_b32_e32 v4, 0x80000000, v1
1727 ; SI-NEXT: v_mul_f64 v[2:3], v[0:1], 4.0
1728 ; SI-NEXT: v_mov_b32_e32 v1, v4
1729 ; SI-NEXT: s_setpc_b64 s[30:31]
1730 ;
1731 ; VI-LABEL: v_fneg_multi_foldable_use_fp_extend_fneg_f32_to_f64:
1732 ; VI: ; %bb.0:
1733 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1734 ; VI-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
1735 ; VI-NEXT: v_mul_f64 v[2:3], v[0:1], 4.0
1736 ; VI-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
1737 ; VI-NEXT: s_setpc_b64 s[30:31]
1738 %fpext = fpext float %a to double
1739 %fneg = fneg double %fpext
1740 %mul = fmul double %fpext, 4.0
1741 %insert.0 = insertvalue { double, double } poison, double %fneg, 0
1742 %insert.1 = insertvalue { double, double } %insert.0, double %mul, 1
1743 ret { double, double } %insert.1
1744 }
1746 ; FIXME: Source modifiers not folded for f16->f32
1747 define { float, float } @v_fneg_multi_use_fp_extend_fneg_f16_to_f32(half %a) #0 {
1748 ; SI-LABEL: v_fneg_multi_use_fp_extend_fneg_f16_to_f32:
1749 ; SI: ; %bb.0:
1750 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1751 ; SI-NEXT: v_mov_b32_e32 v1, v0
1752 ; SI-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
1753 ; SI-NEXT: s_setpc_b64 s[30:31]
1754 ;
1755 ; VI-LABEL: v_fneg_multi_use_fp_extend_fneg_f16_to_f32:
1756 ; VI: ; %bb.0:
1757 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1758 ; VI-NEXT: v_cvt_f32_f16_e32 v1, v0
1759 ; VI-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
1760 ; VI-NEXT: s_setpc_b64 s[30:31]
1761 %fpext = fpext half %a to float
1762 %fneg = fneg float %fpext
1763 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1764 %insert.1 = insertvalue { float, float } %insert.0, float %fpext, 1
1765 ret { float, float } %insert.1
1766 }
1768 define { float, float } @v_fneg_multi_foldable_use_fp_extend_fneg_f16_to_f32(half %a) #0 {
1769 ; SI-LABEL: v_fneg_multi_foldable_use_fp_extend_fneg_f16_to_f32:
1770 ; SI: ; %bb.0:
1771 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1772 ; SI-NEXT: v_xor_b32_e32 v2, 0x80000000, v0
1773 ; SI-NEXT: v_mul_f32_e32 v1, 4.0, v0
1774 ; SI-NEXT: v_mov_b32_e32 v0, v2
1775 ; SI-NEXT: s_setpc_b64 s[30:31]
1776 ;
1777 ; VI-LABEL: v_fneg_multi_foldable_use_fp_extend_fneg_f16_to_f32:
1778 ; VI: ; %bb.0:
1779 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1780 ; VI-NEXT: v_cvt_f32_f16_e32 v1, v0
1781 ; VI-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
1782 ; VI-NEXT: v_mul_f32_e32 v1, 4.0, v1
1783 ; VI-NEXT: s_setpc_b64 s[30:31]
1784 %fpext = fpext half %a to float
1785 %fneg = fneg float %fpext
1786 %mul = fmul float %fpext, 4.0
1787 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1788 %insert.1 = insertvalue { float, float } %insert.0, float %mul, 1
1789 ret { float, float } %insert.1
1790 }
1792 ; --------------------------------------------------------------------------------
1793 ; fp_round tests
1794 ; --------------------------------------------------------------------------------
1796 define float @v_fneg_fp_round_f64_to_f32(double %a) #0 {
1797 ; GCN-LABEL: v_fneg_fp_round_f64_to_f32:
1798 ; GCN: ; %bb.0:
1799 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1800 ; GCN-NEXT: v_cvt_f32_f64_e64 v0, -v[0:1]
1801 ; GCN-NEXT: s_setpc_b64 s[30:31]
1802 %fpround = fptrunc double %a to float
1803 %fneg = fneg float %fpround
1804 ret float %fneg
1805 }
1807 define float @v_fneg_fp_round_fneg_f64_to_f32(double %a) #0 {
1808 ; GCN-LABEL: v_fneg_fp_round_fneg_f64_to_f32:
1809 ; GCN: ; %bb.0:
1810 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1811 ; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
1812 ; GCN-NEXT: s_setpc_b64 s[30:31]
1813 %fneg.a = fneg double %a
1814 %fpround = fptrunc double %fneg.a to float
1815 %fneg = fneg float %fpround
1816 ret float %fneg
1817 }
1819 define { float, double } @v_fneg_fp_round_store_use_fneg_f64_to_f32(double %a) #0 {
1820 ; GCN-LABEL: v_fneg_fp_round_store_use_fneg_f64_to_f32:
1821 ; GCN: ; %bb.0:
1822 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1823 ; GCN-NEXT: v_cvt_f32_f64_e32 v3, v[0:1]
1824 ; GCN-NEXT: v_xor_b32_e32 v2, 0x80000000, v1
1825 ; GCN-NEXT: v_mov_b32_e32 v1, v0
1826 ; GCN-NEXT: v_mov_b32_e32 v0, v3
1827 ; GCN-NEXT: s_setpc_b64 s[30:31]
1828 %fneg.a = fneg double %a
1829 %fpround = fptrunc double %fneg.a to float
1830 %fneg = fneg float %fpround
1831 %insert.0 = insertvalue { float, double } poison, float %fneg, 0
1832 %insert.1 = insertvalue { float, double } %insert.0, double %fneg.a, 1
1833 ret { float, double } %insert.1
1834 }
1836 define { float, double } @v_fneg_fp_round_multi_use_fneg_f64_to_f32(double %a, double %c) #0 {
1837 ; SI-LABEL: v_fneg_fp_round_multi_use_fneg_f64_to_f32:
1838 ; SI: ; %bb.0:
1839 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1840 ; SI-NEXT: v_cvt_f32_f64_e32 v4, v[0:1]
1841 ; SI-NEXT: v_mul_f64 v[1:2], -v[0:1], v[2:3]
1842 ; SI-NEXT: v_mov_b32_e32 v0, v4
1843 ; SI-NEXT: s_setpc_b64 s[30:31]
1844 ;
1845 ; VI-LABEL: v_fneg_fp_round_multi_use_fneg_f64_to_f32:
1846 ; VI: ; %bb.0:
1847 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1848 ; VI-NEXT: v_mul_f64 v[2:3], -v[0:1], v[2:3]
1849 ; VI-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
1850 ; VI-NEXT: v_mov_b32_e32 v1, v2
1851 ; VI-NEXT: v_mov_b32_e32 v2, v3
1852 ; VI-NEXT: s_setpc_b64 s[30:31]
1853 %fneg.a = fneg double %a
1854 %fpround = fptrunc double %fneg.a to float
1855 %fneg = fneg float %fpround
1856 %use1 = fmul double %fneg.a, %c
1857 %insert.0 = insertvalue { float, double } poison, float %fneg, 0
1858 %insert.1 = insertvalue { float, double } %insert.0, double %use1, 1
1859 ret { float, double } %insert.1
1860 }
1862 define half @v_fneg_fp_round_f32_to_f16(float %a) #0 {
1863 ; SI-LABEL: v_fneg_fp_round_f32_to_f16:
1864 ; SI: ; %bb.0:
1865 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1866 ; SI-NEXT: v_cvt_f16_f32_e64 v0, -v0
1867 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1868 ; SI-NEXT: s_setpc_b64 s[30:31]
1869 ;
1870 ; VI-LABEL: v_fneg_fp_round_f32_to_f16:
1871 ; VI: ; %bb.0:
1872 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1873 ; VI-NEXT: v_cvt_f16_f32_e64 v0, -v0
1874 ; VI-NEXT: s_setpc_b64 s[30:31]
1875 %fpround = fptrunc float %a to half
1876 %fneg = fneg half %fpround
1877 ret half %fneg
1878 }
1880 define half @v_fneg_fp_round_fneg_f32_to_f16(float %a) #0 {
1881 ; SI-LABEL: v_fneg_fp_round_fneg_f32_to_f16:
1882 ; SI: ; %bb.0:
1883 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1884 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1885 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1886 ; SI-NEXT: s_setpc_b64 s[30:31]
1887 ;
1888 ; VI-LABEL: v_fneg_fp_round_fneg_f32_to_f16:
1889 ; VI: ; %bb.0:
1890 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1891 ; VI-NEXT: v_cvt_f16_f32_e32 v0, v0
1892 ; VI-NEXT: s_setpc_b64 s[30:31]
1893 %fneg.a = fneg float %a
1894 %fpround = fptrunc float %fneg.a to half
1895 %fneg = fneg half %fpround
1896 ret half %fneg
1897 }
1899 define { float, float } @v_fneg_multi_use_fp_round_fneg_f64_to_f32(double %a) #0 {
1900 ; GCN-LABEL: v_fneg_multi_use_fp_round_fneg_f64_to_f32:
1901 ; GCN: ; %bb.0:
1902 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1903 ; GCN-NEXT: v_cvt_f32_f64_e32 v1, v[0:1]
1904 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
1905 ; GCN-NEXT: s_setpc_b64 s[30:31]
1906 %fpround = fptrunc double %a to float
1907 %fneg = fneg float %fpround
1908 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
1909 %insert.1 = insertvalue { float, float } %insert.0, float %fpround, 1
1910 ret { float, float } %insert.1
1911 }
1913 define { half, float } @v_fneg_fp_round_store_use_fneg_f32_to_f16(float %a) #0 {
1914 ; SI-LABEL: v_fneg_fp_round_store_use_fneg_f32_to_f16:
1915 ; SI: ; %bb.0:
1916 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1917 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v0
1918 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v1
1919 ; SI-NEXT: v_xor_b32_e32 v1, 0x80000000, v0
1920 ; SI-NEXT: v_mov_b32_e32 v0, v2
1921 ; SI-NEXT: s_setpc_b64 s[30:31]
1922 ;
1923 ; VI-LABEL: v_fneg_fp_round_store_use_fneg_f32_to_f16:
1924 ; VI: ; %bb.0:
1925 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1926 ; VI-NEXT: v_cvt_f16_f32_e32 v2, v0
1927 ; VI-NEXT: v_xor_b32_e32 v1, 0x80000000, v0
1928 ; VI-NEXT: v_mov_b32_e32 v0, v2
1929 ; VI-NEXT: s_setpc_b64 s[30:31]
1930 %fneg.a = fneg float %a
1931 %fpround = fptrunc float %fneg.a to half
1932 %fneg = fneg half %fpround
1933 %insert.0 = insertvalue { half, float } poison, half %fneg, 0
1934 %insert.1 = insertvalue { half, float } %insert.0, float %fneg.a, 1
1935 ret { half, float } %insert.1
1936 }
1938 define { half, float } @v_fneg_fp_round_multi_use_fneg_f32_to_f16(float %a, float %c) #0 {
1939 ; SI-LABEL: v_fneg_fp_round_multi_use_fneg_f32_to_f16:
1940 ; SI: ; %bb.0:
1941 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1942 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v0
1943 ; SI-NEXT: v_mul_f32_e64 v1, -v0, v1
1944 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
1945 ; SI-NEXT: v_mov_b32_e32 v0, v2
1946 ; SI-NEXT: s_setpc_b64 s[30:31]
1947 ;
1948 ; VI-LABEL: v_fneg_fp_round_multi_use_fneg_f32_to_f16:
1949 ; VI: ; %bb.0:
1950 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1951 ; VI-NEXT: v_cvt_f16_f32_e32 v2, v0
1952 ; VI-NEXT: v_mul_f32_e64 v1, -v0, v1
1953 ; VI-NEXT: v_mov_b32_e32 v0, v2
1954 ; VI-NEXT: s_setpc_b64 s[30:31]
1955 %fneg.a = fneg float %a
1956 %fpround = fptrunc float %fneg.a to half
1957 %fneg = fneg half %fpround
1958 %use1 = fmul float %fneg.a, %c
1959 %insert.0 = insertvalue { half, float } poison, half %fneg, 0
1960 %insert.1 = insertvalue { half, float } %insert.0, float %use1, 1
1961 ret { half, float } %insert.1
1962 }
1964 ; --------------------------------------------------------------------------------
1965 ; rcp tests
1966 ; --------------------------------------------------------------------------------
1968 define float @v_fneg_rcp_f32(float %a) #0 {
1969 ; GCN-LABEL: v_fneg_rcp_f32:
1970 ; GCN: ; %bb.0:
1971 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1972 ; GCN-NEXT: v_rcp_f32_e64 v0, -v0
1973 ; GCN-NEXT: s_setpc_b64 s[30:31]
1974 %rcp = call float @llvm.amdgcn.rcp.f32(float %a)
1975 %fneg = fneg float %rcp
1976 ret float %fneg
1977 }
1979 define float @v_fneg_rcp_fneg_f32(float %a) #0 {
1980 ; GCN-LABEL: v_fneg_rcp_fneg_f32:
1981 ; GCN: ; %bb.0:
1982 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1983 ; GCN-NEXT: v_rcp_f32_e32 v0, v0
1984 ; GCN-NEXT: s_setpc_b64 s[30:31]
1985 %fneg.a = fneg float %a
1986 %rcp = call float @llvm.amdgcn.rcp.f32(float %fneg.a)
1987 %fneg = fneg float %rcp
1988 ret float %fneg
1989 }
1991 define void @v_fneg_rcp_store_use_fneg_f32(float %a, ptr addrspace(1) %ptr0, ptr addrspace(1) %ptr1) #0 {
1992 ; GCN-LABEL: v_fneg_rcp_store_use_fneg_f32:
1993 ; GCN: ; %bb.0:
1994 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1995 ; GCN-NEXT: v_rcp_f32_e32 v5, v0
1996 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
1997 ; GCN-NEXT: flat_store_dword v[1:2], v5
1998 ; GCN-NEXT: flat_store_dword v[3:4], v0
1999 ; GCN-NEXT: s_waitcnt vmcnt(0)
2000 ; GCN-NEXT: s_setpc_b64 s[30:31]
2001 %fneg.a = fneg float %a
2002 %rcp = call float @llvm.amdgcn.rcp.f32(float %fneg.a)
2003 %fneg = fneg float %rcp
2004 store float %fneg, ptr addrspace(1) %ptr0
2005 store float %fneg.a, ptr addrspace(1) %ptr1
2006 ret void
2007 }
2009 define { float, float } @v_fneg_rcp_multi_use_fneg_f32(float %a, float %c) #0 {
2010 ; GCN-LABEL: v_fneg_rcp_multi_use_fneg_f32:
2011 ; GCN: ; %bb.0:
2012 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2013 ; GCN-NEXT: v_rcp_f32_e32 v2, v0
2014 ; GCN-NEXT: v_mul_f32_e64 v1, -v0, v1
2015 ; GCN-NEXT: v_mov_b32_e32 v0, v2
2016 ; GCN-NEXT: s_setpc_b64 s[30:31]
2017 %fneg.a = fneg float %a
2018 %rcp = call float @llvm.amdgcn.rcp.f32(float %fneg.a)
2019 %fneg = fneg float %rcp
2020 %use1 = fmul float %fneg.a, %c
2021 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
2022 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
2023 ret { float, float } %insert.1
2024 }
2026 ; Test getNegatedExpression works for rcp nodes
2027 define float @v_negated_rcp_f32(float %arg0, float %arg1) #1 {
2028 ; GCN-LABEL: v_negated_rcp_f32:
2029 ; GCN: ; %bb.0:
2030 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2031 ; GCN-NEXT: v_fma_f32 v0, v0, v1, 2.0
2032 ; GCN-NEXT: v_rcp_f32_e32 v0, v0
2033 ; GCN-NEXT: v_add_f32_e32 v0, v1, v0
2034 ; GCN-NEXT: s_setpc_b64 s[30:31]
2035 %neg.arg0 = fneg float %arg0
2036 %fma = call nsz float @llvm.fma.f32(float %neg.arg0, float %arg1, float -2.0)
2037 %rcp0 = call float @llvm.amdgcn.rcp.f32(float %fma)
2038 %mul = fsub nsz float %arg1, %rcp0
2039 ret float %mul
2040 }
2042 ; --------------------------------------------------------------------------------
2043 ; fmul_legacy tests
2044 ; --------------------------------------------------------------------------------
2046 define float @v_fneg_mul_legacy_f32(float %a, float %b) #0 {
2047 ; GCN-LABEL: v_fneg_mul_legacy_f32:
2048 ; GCN: ; %bb.0:
2049 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2050 ; GCN-NEXT: v_mul_legacy_f32_e64 v0, v0, -v1
2051 ; GCN-NEXT: s_setpc_b64 s[30:31]
2052 %mul = call float @llvm.amdgcn.fmul.legacy(float %a, float %b)
2053 %fneg = fneg float %mul
2054 ret float %fneg
2055 }
2057 define { float, float } @v_fneg_mul_legacy_store_use_mul_legacy_f32(float %a, float %b) #0 {
2058 ; GCN-LABEL: v_fneg_mul_legacy_store_use_mul_legacy_f32:
2059 ; GCN: ; %bb.0:
2060 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2061 ; GCN-NEXT: v_mul_legacy_f32_e32 v1, v0, v1
2062 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v1
2063 ; GCN-NEXT: s_setpc_b64 s[30:31]
2064 %mul = call float @llvm.amdgcn.fmul.legacy(float %a, float %b)
2065 %fneg = fneg float %mul
2066 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
2067 %insert.1 = insertvalue { float, float } %insert.0, float %mul, 1
2068 ret { float, float } %insert.1
2069 }
2071 define { float, float } @v_fneg_mul_legacy_multi_use_mul_legacy_f32(float %a, float %b) #0 {
2072 ; GCN-LABEL: v_fneg_mul_legacy_multi_use_mul_legacy_f32:
2073 ; GCN: ; %bb.0:
2074 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2075 ; GCN-NEXT: v_mul_legacy_f32_e64 v0, v0, -v1
2076 ; GCN-NEXT: v_mul_legacy_f32_e64 v1, -v0, 4.0
2077 ; GCN-NEXT: s_setpc_b64 s[30:31]
2078 %mul = call float @llvm.amdgcn.fmul.legacy(float %a, float %b)
2079 %fneg = fneg float %mul
2080 %use1 = call float @llvm.amdgcn.fmul.legacy(float %mul, float 4.0)
2081 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
2082 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
2083 ret { float, float } %insert.1
2084 }
2086 define float @v_fneg_mul_legacy_fneg_x_f32(float %a, float %b) #0 {
2087 ; GCN-LABEL: v_fneg_mul_legacy_fneg_x_f32:
2088 ; GCN: ; %bb.0:
2089 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2090 ; GCN-NEXT: v_mul_legacy_f32_e32 v0, v0, v1
2091 ; GCN-NEXT: s_setpc_b64 s[30:31]
2092 %fneg.a = fneg float %a
2093 %mul = call float @llvm.amdgcn.fmul.legacy(float %fneg.a, float %b)
2094 %fneg = fneg float %mul
2095 ret float %fneg
2096 }
2098 define float @v_fneg_mul_legacy_x_fneg_f32(float %a, float %b) #0 {
2099 ; GCN-LABEL: v_fneg_mul_legacy_x_fneg_f32:
2100 ; GCN: ; %bb.0:
2101 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2102 ; GCN-NEXT: v_mul_legacy_f32_e32 v0, v0, v1
2103 ; GCN-NEXT: s_setpc_b64 s[30:31]
2104 %fneg.b = fneg float %b
2105 %mul = call float @llvm.amdgcn.fmul.legacy(float %a, float %fneg.b)
2106 %fneg = fneg float %mul
2107 ret float %fneg
2108 }
2110 define float @v_fneg_mul_legacy_fneg_fneg_f32(float %a, float %b) #0 {
2111 ; GCN-LABEL: v_fneg_mul_legacy_fneg_fneg_f32:
2112 ; GCN: ; %bb.0:
2113 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2114 ; GCN-NEXT: v_mul_legacy_f32_e64 v0, v0, -v1
2115 ; GCN-NEXT: s_setpc_b64 s[30:31]
2116 %fneg.a = fneg float %a
2117 %fneg.b = fneg float %b
2118 %mul = call float @llvm.amdgcn.fmul.legacy(float %fneg.a, float %fneg.b)
2119 %fneg = fneg float %mul
2120 ret float %fneg
2121 }
2123 define { float, float } @v_fneg_mul_legacy_store_use_fneg_x_f32(float %a, float %b) #0 {
2124 ; GCN-LABEL: v_fneg_mul_legacy_store_use_fneg_x_f32:
2125 ; GCN: ; %bb.0:
2126 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2127 ; GCN-NEXT: v_xor_b32_e32 v2, 0x80000000, v0
2128 ; GCN-NEXT: v_mul_legacy_f32_e32 v0, v0, v1
2129 ; GCN-NEXT: v_mov_b32_e32 v1, v2
2130 ; GCN-NEXT: s_setpc_b64 s[30:31]
2131 %fneg.a = fneg float %a
2132 %mul = call float @llvm.amdgcn.fmul.legacy(float %fneg.a, float %b)
2133 %fneg = fneg float %mul
2135 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
2136 %insert.1 = insertvalue { float, float } %insert.0, float %fneg.a, 1
2137 ret { float, float } %insert.1
2138 }
2140 define { float, float } @v_fneg_mul_legacy_multi_use_fneg_x_f32(float %a, float %b, float %c) #0 {
2141 ; GCN-LABEL: v_fneg_mul_legacy_multi_use_fneg_x_f32:
2142 ; GCN: ; %bb.0:
2143 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2144 ; GCN-NEXT: v_mul_legacy_f32_e32 v3, v0, v1
2145 ; GCN-NEXT: v_mul_legacy_f32_e64 v1, -v0, v2
2146 ; GCN-NEXT: v_mov_b32_e32 v0, v3
2147 ; GCN-NEXT: s_setpc_b64 s[30:31]
2148 %fneg.a = fneg float %a
2149 %mul = call float @llvm.amdgcn.fmul.legacy(float %fneg.a, float %b)
2150 %fneg = fneg float %mul
2151 %use1 = call float @llvm.amdgcn.fmul.legacy(float %fneg.a, float %c)
2152 %insert.0 = insertvalue { float, float } poison, float %fneg, 0
2153 %insert.1 = insertvalue { float, float } %insert.0, float %use1, 1
2154 ret { float, float } %insert.1
2155 }
2157 ; --------------------------------------------------------------------------------
2158 ; sin tests
2159 ; --------------------------------------------------------------------------------
2161 define float @v_fneg_sin_f32(float %a) #0 {
2162 ; GCN-LABEL: v_fneg_sin_f32:
2163 ; GCN: ; %bb.0:
2164 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2165 ; GCN-NEXT: v_mul_f32_e32 v0, 0xbe22f983, v0
2166 ; GCN-NEXT: v_fract_f32_e32 v0, v0
2167 ; GCN-NEXT: v_sin_f32_e32 v0, v0
2168 ; GCN-NEXT: s_setpc_b64 s[30:31]
2169 %sin = call float @llvm.sin.f32(float %a)
2170 %fneg = fneg float %sin
2171 ret float %fneg
2172 }
2174 define float @v_fneg_amdgcn_sin_f32(float %a) #0 {
2175 ; GCN-LABEL: v_fneg_amdgcn_sin_f32:
2176 ; GCN: ; %bb.0:
2177 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2178 ; GCN-NEXT: v_sin_f32_e64 v0, -v0
2179 ; GCN-NEXT: s_setpc_b64 s[30:31]
2180 %sin = call float @llvm.amdgcn.sin.f32(float %a)
2181 %fneg = fneg float %sin
2182 ret float %fneg
2183 }
2185 ; --------------------------------------------------------------------------------
2186 ; ftrunc tests
2187 ; --------------------------------------------------------------------------------
2189 define float @v_fneg_trunc_f32(float %a) #0 {
2190 ; GCN-LABEL: v_fneg_trunc_f32:
2191 ; GCN: ; %bb.0:
2192 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2193 ; GCN-NEXT: v_trunc_f32_e64 v0, -v0
2194 ; GCN-NEXT: s_setpc_b64 s[30:31]
2195 %trunc = call float @llvm.trunc.f32(float %a)
2196 %fneg = fneg float %trunc
2197 ret float %fneg
2198 }
2200 ; --------------------------------------------------------------------------------
2201 ; fround tests
2202 ; --------------------------------------------------------------------------------
2204 define float @v_fneg_round_f32(float %a) #0 {
2205 ; GCN-SAFE-LABEL: v_fneg_round_f32:
2206 ; GCN-SAFE: ; %bb.0:
2207 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2208 ; GCN-SAFE-NEXT: v_trunc_f32_e32 v1, v0
2209 ; GCN-SAFE-NEXT: v_sub_f32_e32 v2, v0, v1
2210 ; GCN-SAFE-NEXT: v_cmp_ge_f32_e64 s[4:5], |v2|, 0.5
2211 ; GCN-SAFE-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s[4:5]
2212 ; GCN-SAFE-NEXT: s_brev_b32 s4, -2
2213 ; GCN-SAFE-NEXT: v_bfi_b32 v0, s4, v2, v0
2214 ; GCN-SAFE-NEXT: v_add_f32_e32 v0, v1, v0
2215 ; GCN-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
2216 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
2217 ;
2218 ; GCN-NSZ-LABEL: v_fneg_round_f32:
2219 ; GCN-NSZ: ; %bb.0:
2220 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2221 ; GCN-NSZ-NEXT: v_trunc_f32_e32 v1, v0
2222 ; GCN-NSZ-NEXT: v_sub_f32_e32 v2, v0, v1
2223 ; GCN-NSZ-NEXT: v_cmp_ge_f32_e64 s[4:5], |v2|, 0.5
2224 ; GCN-NSZ-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s[4:5]
2225 ; GCN-NSZ-NEXT: s_brev_b32 s4, -2
2226 ; GCN-NSZ-NEXT: v_bfi_b32 v0, s4, v2, v0
2227 ; GCN-NSZ-NEXT: v_sub_f32_e64 v0, -v1, v0
2228 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
2229 %round = call float @llvm.round.f32(float %a)
2230 %fneg = fneg float %round
2231 ret float %fneg
2232 }
2234 ; --------------------------------------------------------------------------------
2235 ; rint tests
2236 ; --------------------------------------------------------------------------------
2238 define float @v_fneg_rint_f32(float %a) #0 {
2239 ; GCN-LABEL: v_fneg_rint_f32:
2240 ; GCN: ; %bb.0:
2241 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2242 ; GCN-NEXT: v_rndne_f32_e64 v0, -v0
2243 ; GCN-NEXT: s_setpc_b64 s[30:31]
2244 %rint = call float @llvm.rint.f32(float %a)
2245 %fneg = fneg float %rint
2246 ret float %fneg
2247 }
2249 ; --------------------------------------------------------------------------------
2250 ; nearbyint tests
2251 ; --------------------------------------------------------------------------------
2253 define float @v_fneg_nearbyint_f32(float %a) #0 {
2254 ; GCN-LABEL: v_fneg_nearbyint_f32:
2255 ; GCN: ; %bb.0:
2256 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2257 ; GCN-NEXT: v_rndne_f32_e64 v0, -v0
2258 ; GCN-NEXT: s_setpc_b64 s[30:31]
2259 %nearbyint = call float @llvm.nearbyint.f32(float %a)
2260 %fneg = fneg float %nearbyint
2261 ret float %fneg
2262 }
2264 ; --------------------------------------------------------------------------------
2265 ; fcanonicalize tests
2266 ; --------------------------------------------------------------------------------
2268 define float @v_fneg_canonicalize_f32(float %a) #0 {
2269 ; GCN-LABEL: v_fneg_canonicalize_f32:
2270 ; GCN: ; %bb.0:
2271 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2272 ; GCN-NEXT: v_mul_f32_e32 v0, -1.0, v0
2273 ; GCN-NEXT: s_setpc_b64 s[30:31]
2274 %trunc = call float @llvm.canonicalize.f32(float %a)
2275 %fneg = fneg float %trunc
2276 ret float %fneg
2277 }
2279 ; --------------------------------------------------------------------------------
2280 ; arithmetic.fence tests
2281 ; --------------------------------------------------------------------------------
2283 define float @v_fneg_arithmetic_fence_f32(float %a) #0 {
2284 ; GCN-LABEL: v_fneg_arithmetic_fence_f32:
2285 ; GCN: ; %bb.0:
2286 ; GCN-NEXT: ;ARITH_FENCE
2287 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2288 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
2289 ; GCN-NEXT: s_setpc_b64 s[30:31]
2290 %fence = call float @llvm.arithmetic.fence.f32(float %a)
2291 %fneg = fneg float %fence
2292 ret float %fneg
2293 }
2295 define float @v_fneg_arithmetic_fence_fmul_f32(float %a, float %b) #0 {
2296 ; GCN-LABEL: v_fneg_arithmetic_fence_fmul_f32:
2297 ; GCN: ; %bb.0:
2298 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2299 ; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
2300 ; GCN-NEXT: ;ARITH_FENCE
2301 ; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
2302 ; GCN-NEXT: s_setpc_b64 s[30:31]
2303 %mul = fmul float %a, %b
2304 %fence = call float @llvm.arithmetic.fence.f32(float %mul)
2305 %fneg = fneg float %fence
2306 ret float %fneg
2307 }
2309 ; --------------------------------------------------------------------------------
2310 ; vintrp tests
2311 ; --------------------------------------------------------------------------------
2313 define { float, float } @v_fneg_interp_p1_f32(float %a, float %b) #0 {
2314 ; SI-LABEL: v_fneg_interp_p1_f32:
2315 ; SI: ; %bb.0:
2316 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2317 ; SI-NEXT: v_mul_f32_e64 v1, v0, -v1
2318 ; SI-NEXT: s_mov_b32 m0, 0
2319 ; SI-NEXT: v_interp_p1_f32 v0, v1, attr0.x
2320 ; SI-NEXT: v_interp_p1_f32 v1, v1, attr0.y
2321 ; SI-NEXT: s_setpc_b64 s[30:31]
2322 ;
2323 ; VI-LABEL: v_fneg_interp_p1_f32:
2324 ; VI: ; %bb.0:
2325 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2326 ; VI-NEXT: v_mul_f32_e64 v1, v0, -v1
2327 ; VI-NEXT: s_mov_b32 m0, 0
2328 ; VI-NEXT: v_interp_p1_f32_e32 v0, v1, attr0.x
2329 ; VI-NEXT: v_interp_p1_f32_e32 v1, v1, attr0.y
2330 ; VI-NEXT: s_setpc_b64 s[30:31]
2331 %mul = fmul float %a, %b
2332 %fneg = fneg float %mul
2333 %intrp0 = call float @llvm.amdgcn.interp.p1(float %fneg, i32 0, i32 0, i32 0)
2334 %intrp1 = call float @llvm.amdgcn.interp.p1(float %fneg, i32 1, i32 0, i32 0)
2335 %insert.0 = insertvalue { float, float } poison, float %intrp0, 0
2336 %insert.1 = insertvalue { float, float } %insert.0, float %intrp1, 1
2337 ret { float, float } %insert.1
2338 }
2340 define { float, float } @v_fneg_interp_p2_f32(float %a, float %b) #0 {
2341 ; SI-LABEL: v_fneg_interp_p2_f32:
2342 ; SI: ; %bb.0:
2343 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2344 ; SI-NEXT: v_mul_f32_e64 v2, v0, -v1
2345 ; SI-NEXT: v_mov_b32_e32 v1, 4.0
2346 ; SI-NEXT: v_mov_b32_e32 v0, 4.0
2347 ; SI-NEXT: s_mov_b32 m0, 0
2348 ; SI-NEXT: v_interp_p2_f32 v0, v2, attr0.x
2349 ; SI-NEXT: v_interp_p2_f32 v1, v2, attr0.y
2350 ; SI-NEXT: s_setpc_b64 s[30:31]
2351 ;
2352 ; VI-LABEL: v_fneg_interp_p2_f32:
2353 ; VI: ; %bb.0:
2354 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2355 ; VI-NEXT: v_mul_f32_e64 v2, v0, -v1
2356 ; VI-NEXT: v_mov_b32_e32 v1, 4.0
2357 ; VI-NEXT: v_mov_b32_e32 v0, 4.0
2358 ; VI-NEXT: s_mov_b32 m0, 0
2359 ; VI-NEXT: v_interp_p2_f32_e32 v0, v2, attr0.x
2360 ; VI-NEXT: v_interp_p2_f32_e32 v1, v2, attr0.y
2361 ; VI-NEXT: s_setpc_b64 s[30:31]
2362 %mul = fmul float %a, %b
2363 %fneg = fneg float %mul
2364 %intrp0 = call float @llvm.amdgcn.interp.p2(float 4.0, float %fneg, i32 0, i32 0, i32 0)
2365 %intrp1 = call float @llvm.amdgcn.interp.p2(float 4.0, float %fneg, i32 1, i32 0, i32 0)
2366 %insert.0 = insertvalue { float, float } poison, float %intrp0, 0
2367 %insert.1 = insertvalue { float, float } %insert.0, float %intrp1, 1
2368 ret { float, float } %insert.1
2369 }
2371 ; --------------------------------------------------------------------------------
2372 ; CopyToReg tests
2373 ; --------------------------------------------------------------------------------
2375 define void @v_fneg_copytoreg_f32(ptr addrspace(1) %out, float %a, float %b, float %c, i32 %d) #0 {
2376 ; SI-LABEL: v_fneg_copytoreg_f32:
2377 ; SI: ; %bb.0:
2378 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2379 ; SI-NEXT: v_and_b32_e32 v6, 0x3ff, v31
2380 ; SI-NEXT: v_lshlrev_b32_e32 v6, 2, v6
2381 ; SI-NEXT: v_add_i32_e32 v0, vcc, v0, v6
2382 ; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
2383 ; SI-NEXT: v_mul_f32_e32 v2, v2, v3
2384 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
2385 ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
2386 ; SI-NEXT: s_cbranch_execz .LBB118_2
2387 ; SI-NEXT: ; %bb.1: ; %if
2388 ; SI-NEXT: v_mul_f32_e64 v3, -v2, v4
2389 ; SI-NEXT: flat_store_dword v[0:1], v3
2390 ; SI-NEXT: s_waitcnt vmcnt(0)
2391 ; SI-NEXT: .LBB118_2: ; %endif
2392 ; SI-NEXT: s_or_b64 exec, exec, s[4:5]
2393 ; SI-NEXT: flat_store_dword v[0:1], v2
2394 ; SI-NEXT: s_waitcnt vmcnt(0)
2395 ; SI-NEXT: s_setpc_b64 s[30:31]
2396 ;
2397 ; VI-LABEL: v_fneg_copytoreg_f32:
2398 ; VI: ; %bb.0:
2399 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2400 ; VI-NEXT: v_and_b32_e32 v6, 0x3ff, v31
2401 ; VI-NEXT: v_lshlrev_b32_e32 v6, 2, v6
2402 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v6
2403 ; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
2404 ; VI-NEXT: v_mul_f32_e32 v2, v2, v3
2405 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
2406 ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
2407 ; VI-NEXT: s_cbranch_execz .LBB118_2
2408 ; VI-NEXT: ; %bb.1: ; %if
2409 ; VI-NEXT: v_mul_f32_e64 v3, -v2, v4
2410 ; VI-NEXT: flat_store_dword v[0:1], v3
2411 ; VI-NEXT: s_waitcnt vmcnt(0)
2412 ; VI-NEXT: .LBB118_2: ; %endif
2413 ; VI-NEXT: s_or_b64 exec, exec, s[4:5]
2414 ; VI-NEXT: flat_store_dword v[0:1], v2
2415 ; VI-NEXT: s_waitcnt vmcnt(0)
2416 ; VI-NEXT: s_setpc_b64 s[30:31]
2417 %tid = call i32 @llvm.amdgcn.workitem.id.x()
2418 %tid.ext = sext i32 %tid to i64
2419 %out.gep = getelementptr inbounds float, ptr addrspace(1) %out, i64 %tid.ext
2420 %mul = fmul float %a, %b
2421 %fneg = fneg float %mul
2422 %cmp0 = icmp eq i32 %d, 0
2423 br i1 %cmp0, label %if, label %endif
2425 if:
2426 %mul1 = fmul float %fneg, %c
2427 store volatile float %mul1, ptr addrspace(1) %out.gep
2428 br label %endif
2430 endif:
2431 store volatile float %mul, ptr addrspace(1) %out.gep
2432 ret void
2433 }
2435 ; --------------------------------------------------------------------------------
2436 ; inlineasm tests
2437 ; --------------------------------------------------------------------------------
2439 ; Can't fold into use, so should fold into source
2440 define float @v_fneg_inlineasm_f32(float %a, float %b, float %c, i32 %d) #0 {
2441 ; GCN-LABEL: v_fneg_inlineasm_f32:
2442 ; GCN: ; %bb.0:
2443 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2444 ; GCN-NEXT: v_mul_f32_e64 v0, v0, -v1
2445 ; GCN-NEXT: ;;#ASMSTART
2446 ; GCN-NEXT: ; use v0
2447 ; GCN-NEXT: ;;#ASMEND
2448 ; GCN-NEXT: s_setpc_b64 s[30:31]
2449 %mul = fmul float %a, %b
2450 %fneg = fneg float %mul
2451 call void asm sideeffect "; use $0", "v"(float %fneg) #0
2452 ret float %fneg
2453 }
2455 ; --------------------------------------------------------------------------------
2456 ; inlineasm tests
2457 ; --------------------------------------------------------------------------------
2459 ; Can't fold into use, so should fold into source
2460 define float @v_fneg_inlineasm_multi_use_src_f32(ptr addrspace(1) %out, float %a, float %b, float %c, i32 %d) #0 {
2461 ; GCN-LABEL: v_fneg_inlineasm_multi_use_src_f32:
2462 ; GCN: ; %bb.0:
2463 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2464 ; GCN-NEXT: v_mul_f32_e32 v0, v2, v3
2465 ; GCN-NEXT: v_xor_b32_e32 v1, 0x80000000, v0
2466 ; GCN-NEXT: ;;#ASMSTART
2467 ; GCN-NEXT: ; use v1
2468 ; GCN-NEXT: ;;#ASMEND
2469 ; GCN-NEXT: s_setpc_b64 s[30:31]
2470 %mul = fmul float %a, %b
2471 %fneg = fneg float %mul
2472 call void asm sideeffect "; use $0", "v"(float %fneg) #0
2473 ret float %mul
2474 }
2476 ; --------------------------------------------------------------------------------
2477 ; code size regression tests
2478 ; --------------------------------------------------------------------------------
2480 ; There are multiple users of the fneg that must use a VOP3
2481 ; instruction, so there is no penalty
2482 define { float, float } @multiuse_fneg_2_vop3_users_f32(float %a, float %b, float %c) #0 {
2483 ; GCN-LABEL: multiuse_fneg_2_vop3_users_f32:
2484 ; GCN: ; %bb.0:
2485 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2486 ; GCN-NEXT: v_fma_f32 v3, -v0, v1, v2
2487 ; GCN-NEXT: v_fma_f32 v1, -v0, v2, 2.0
2488 ; GCN-NEXT: v_mov_b32_e32 v0, v3
2489 ; GCN-NEXT: s_setpc_b64 s[30:31]
2490 %fneg.a = fneg float %a
2491 %fma0 = call float @llvm.fma.f32(float %fneg.a, float %b, float %c)
2492 %fma1 = call float @llvm.fma.f32(float %fneg.a, float %c, float 2.0)
2493 %insert.0 = insertvalue { float, float } poison, float %fma0, 0
2494 %insert.1 = insertvalue { float, float } %insert.0, float %fma1, 1
2495 ret { float, float } %insert.1
2496 }
2498 ; There are multiple users, but both require using a larger encoding
2499 ; for the modifier.
2500 define { float, float } @multiuse_fneg_2_vop2_users_f32(float %a, float %b, float %c) #0 {
2501 ; GCN-LABEL: multiuse_fneg_2_vop2_users_f32:
2502 ; GCN: ; %bb.0:
2503 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2504 ; GCN-NEXT: v_mul_f32_e64 v3, -v0, v1
2505 ; GCN-NEXT: v_mul_f32_e64 v1, -v0, v2
2506 ; GCN-NEXT: v_mov_b32_e32 v0, v3
2507 ; GCN-NEXT: s_setpc_b64 s[30:31]
2508 %fneg.a = fneg float %a
2509 %mul0 = fmul float %fneg.a, %b
2510 %mul1 = fmul float %fneg.a, %c
2511 %insert.0 = insertvalue { float, float } poison, float %mul0, 0
2512 %insert.1 = insertvalue { float, float } %insert.0, float %mul1, 1
2513 ret { float, float } %insert.1
2514 }
2516 ; One user is VOP3 so has no cost to folding the modifier, the other does.
2517 define { float, float } @multiuse_fneg_vop2_vop3_users_f32(float %a, float %b, float %c) #0 {
2518 ; GCN-LABEL: multiuse_fneg_vop2_vop3_users_f32:
2519 ; GCN: ; %bb.0:
2520 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2521 ; GCN-NEXT: v_fma_f32 v3, -v0, v1, 2.0
2522 ; GCN-NEXT: v_mul_f32_e64 v1, -v0, v2
2523 ; GCN-NEXT: v_mov_b32_e32 v0, v3
2524 ; GCN-NEXT: s_setpc_b64 s[30:31]
2525 %fneg.a = fneg float %a
2526 %fma0 = call float @llvm.fma.f32(float %fneg.a, float %b, float 2.0)
2527 %mul1 = fmul float %fneg.a, %c
2529 %insert.0 = insertvalue { float, float } poison, float %fma0, 0
2530 %insert.1 = insertvalue { float, float } %insert.0, float %mul1, 1
2531 ret { float, float } %insert.1
2532 }
2534 ; The use of the fneg requires a code size increase, but folding into
2535 ; the source does not
2536 define { float, float } @free_fold_src_code_size_cost_use_f32(ptr addrspace(1) %out, float %a, float %b, float %c, float %d) #0 {
2537 ; GCN-SAFE-LABEL: free_fold_src_code_size_cost_use_f32:
2538 ; GCN-SAFE: ; %bb.0:
2539 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2540 ; GCN-SAFE-NEXT: v_fma_f32 v1, v2, v3, 2.0
2541 ; GCN-SAFE-NEXT: v_mul_f32_e64 v0, -v1, v4
2542 ; GCN-SAFE-NEXT: v_mul_f32_e64 v1, -v1, v5
2543 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
2544 ;
2545 ; GCN-NSZ-LABEL: free_fold_src_code_size_cost_use_f32:
2546 ; GCN-NSZ: ; %bb.0:
2547 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2548 ; GCN-NSZ-NEXT: v_fma_f32 v1, v2, -v3, -2.0
2549 ; GCN-NSZ-NEXT: v_mul_f32_e32 v0, v1, v4
2550 ; GCN-NSZ-NEXT: v_mul_f32_e32 v1, v1, v5
2551 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
2552 %fma0 = call float @llvm.fma.f32(float %a, float %b, float 2.0)
2553 %fneg.fma0 = fneg float %fma0
2554 %mul1 = fmul float %fneg.fma0, %c
2555 %mul2 = fmul float %fneg.fma0, %d
2557 %insert.0 = insertvalue { float, float } poison, float %mul1, 0
2558 %insert.1 = insertvalue { float, float } %insert.0, float %mul2, 1
2559 ret { float, float } %insert.1
2560 }
2562 define { double, double } @free_fold_src_code_size_cost_use_f64(double %a, double %b, double %c, double %d) #0 {
2563 ; GCN-LABEL: free_fold_src_code_size_cost_use_f64:
2564 ; GCN: ; %bb.0:
2565 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2566 ; GCN-NEXT: v_fma_f64 v[2:3], v[0:1], v[2:3], 2.0
2567 ; GCN-NEXT: v_mul_f64 v[0:1], -v[2:3], v[4:5]
2568 ; GCN-NEXT: v_mul_f64 v[2:3], -v[2:3], v[6:7]
2569 ; GCN-NEXT: s_setpc_b64 s[30:31]
2570 %fma0 = call double @llvm.fma.f64(double %a, double %b, double 2.0)
2571 %fneg.fma0 = fneg double %fma0
2572 %mul1 = fmul double %fneg.fma0, %c
2573 %mul2 = fmul double %fneg.fma0, %d
2575 %insert.0 = insertvalue { double, double } poison, double %mul1, 0
2576 %insert.1 = insertvalue { double, double } %insert.0, double %mul2, 1
2577 ret { double, double } %insert.1
2578 }
2580 ; %trunc.a has one fneg use, but it requires a code size increase and
2581 ; %the fneg can instead be folded for free into the fma.
2582 define float @one_use_cost_to_fold_into_src_f32(float %a, float %b, float %c, float %d) #0 {
2583 ; GCN-LABEL: one_use_cost_to_fold_into_src_f32:
2584 ; GCN: ; %bb.0:
2585 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2586 ; GCN-NEXT: v_trunc_f32_e32 v0, v0
2587 ; GCN-NEXT: v_fma_f32 v0, -v0, v1, v2
2588 ; GCN-NEXT: s_setpc_b64 s[30:31]
2589 %trunc.a = call float @llvm.trunc.f32(float %a)
2590 %trunc.fneg.a = fneg float %trunc.a
2591 %fma0 = call float @llvm.fma.f32(float %trunc.fneg.a, float %b, float %c)
2592 ret float %fma0
2593 }
2595 define { float, float } @multi_use_cost_to_fold_into_src(float %a, float %b, float %c, float %d) #0 {
2596 ; GCN-LABEL: multi_use_cost_to_fold_into_src:
2597 ; GCN: ; %bb.0:
2598 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2599 ; GCN-NEXT: v_trunc_f32_e32 v4, v0
2600 ; GCN-NEXT: v_fma_f32 v0, -v4, v1, v2
2601 ; GCN-NEXT: v_mul_f32_e32 v1, v4, v3
2602 ; GCN-NEXT: s_setpc_b64 s[30:31]
2603 %trunc.a = call float @llvm.trunc.f32(float %a)
2604 %trunc.fneg.a = fneg float %trunc.a
2605 %fma0 = call float @llvm.fma.f32(float %trunc.fneg.a, float %b, float %c)
2606 %mul1 = fmul float %trunc.a, %d
2607 %insert.0 = insertvalue { float, float } poison, float %fma0, 0
2608 %insert.1 = insertvalue { float, float } %insert.0, float %mul1, 1
2609 ret { float, float } %insert.1
2610 }
2612 ; The AMDGPU combine to pull fneg into the FMA operands was being
2613 ; undone by the generic combine to pull the fneg out of the fma if
2614 ; !isFNegFree. We were reporting false for v2f32 even though it will
2615 ; be split into f32 where it will be free.
2616 define <2 x float> @fneg_fma_fneg_dagcombine_loop(<2 x float> %arg, <2 x float> %arg1, <2 x float> %arg2) #0 {
2617 ; GCN-LABEL: fneg_fma_fneg_dagcombine_loop:
2618 ; GCN: ; %bb.0: ; %bb
2619 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2620 ; GCN-NEXT: s_brev_b32 s4, 1
2621 ; GCN-NEXT: v_fma_f32 v3, v3, -v5, s4
2622 ; GCN-NEXT: v_fma_f32 v2, v2, -v4, s4
2623 ; GCN-NEXT: v_sub_f32_e32 v1, v3, v1
2624 ; GCN-NEXT: v_sub_f32_e32 v0, v2, v0
2625 ; GCN-NEXT: v_mul_f32_e32 v0, v0, v4
2626 ; GCN-NEXT: v_mul_f32_e32 v1, v1, v5
2627 ; GCN-NEXT: s_setpc_b64 s[30:31]
2628 bb:
2629 %i3 = call fast <2 x float> @llvm.fma.v2f32(<2 x float> %arg1, <2 x float> %arg2, <2 x float> zeroinitializer)
2630 %i4 = fadd fast <2 x float> %i3, %arg
2631 %i5 = fneg <2 x float> %i4
2632 %i6 = fmul fast <2 x float> %i5, %arg2
2633 ret <2 x float> %i6
2634 }
2636 ; This expects denormal flushing, so can't turn this fmul into fneg
2637 ; TODO: Keeping this as fmul saves encoding size
2638 define float @nnan_fmul_neg1_to_fneg(float %x, float %y) #0 {
2639 ; GCN-LABEL: nnan_fmul_neg1_to_fneg:
2640 ; GCN: ; %bb.0:
2641 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2642 ; GCN-NEXT: v_mul_f32_e64 v0, -v0, v1
2643 ; GCN-NEXT: s_setpc_b64 s[30:31]
2644 %mul = fmul float %x, -1.0
2645 %add = fmul nnan float %mul, %y
2646 ret float %add
2647 }
2649 ; It's legal to turn this fmul into an fneg since denormals are
2650 ; preserved and we know an snan can't happen from the flag.
2651 define float @denormal_fmul_neg1_to_fneg(float %x, float %y) {
2652 ; GCN-LABEL: denormal_fmul_neg1_to_fneg:
2653 ; GCN: ; %bb.0:
2654 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2655 ; GCN-NEXT: v_mul_f32_e64 v0, -v0, v1
2656 ; GCN-NEXT: s_setpc_b64 s[30:31]
2657 %mul = fmul nnan float %x, -1.0
2658 %add = fmul float %mul, %y
2659 ret float %add
2660 }
2662 ; know the source can't be an snan
2663 define float @denorm_snan_fmul_neg1_to_fneg(float %x, float %y) {
2664 ; GCN-LABEL: denorm_snan_fmul_neg1_to_fneg:
2665 ; GCN: ; %bb.0:
2666 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2667 ; GCN-NEXT: v_mul_f32_e64 v0, v0, -v0
2668 ; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
2669 ; GCN-NEXT: s_setpc_b64 s[30:31]
2670 %canonical = fmul float %x, %x
2671 %mul = fmul float %canonical, -1.0
2672 %add = fmul float %mul, %y
2673 ret float %add
2674 }
2676 define float @flush_snan_fmul_neg1_to_fneg(float %x, float %y) #0 {
2677 ; GCN-LABEL: flush_snan_fmul_neg1_to_fneg:
2678 ; GCN: ; %bb.0:
2679 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2680 ; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
2681 ; GCN-NEXT: v_mul_f32_e64 v0, -v0, v1
2682 ; GCN-NEXT: s_setpc_b64 s[30:31]
2683 %quiet = call float @llvm.canonicalize.f32(float %x)
2684 %mul = fmul float %quiet, -1.0
2685 %add = fmul float %mul, %y
2686 ret float %add
2687 }
2689 define float @fadd_select_fneg_fneg_f32(i32 %arg0, float %x, float %y, float %z) {
2690 ; GCN-LABEL: fadd_select_fneg_fneg_f32:
2691 ; GCN: ; %bb.0:
2692 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2693 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2694 ; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
2695 ; GCN-NEXT: v_sub_f32_e32 v0, v3, v0
2696 ; GCN-NEXT: s_setpc_b64 s[30:31]
2697 %cmp = icmp eq i32 %arg0, 0
2698 %neg.x = fneg float %x
2699 %neg.y = fneg float %y
2700 %select = select i1 %cmp, float %neg.x, float %neg.y
2701 %add = fadd float %select, %z
2702 ret float %add
2703 }
2705 define double @fadd_select_fneg_fneg_f64(i32 %arg0, double %x, double %y, double %z) {
2706 ; GCN-LABEL: fadd_select_fneg_fneg_f64:
2707 ; GCN: ; %bb.0:
2708 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2709 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2710 ; GCN-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
2711 ; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc
2712 ; GCN-NEXT: v_add_f64 v[0:1], v[5:6], -v[1:2]
2713 ; GCN-NEXT: s_setpc_b64 s[30:31]
2714 %cmp = icmp eq i32 %arg0, 0
2715 %neg.x = fneg double %x
2716 %neg.y = fneg double %y
2717 %select = select i1 %cmp, double %neg.x, double %neg.y
2718 %add = fadd double %select, %z
2719 ret double %add
2720 }
2722 define half @fadd_select_fneg_fneg_f16(i32 %arg0, half %x, half %y, half %z) {
2723 ; SI-LABEL: fadd_select_fneg_fneg_f16:
2724 ; SI: ; %bb.0:
2725 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2726 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
2727 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
2728 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
2729 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2730 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
2731 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
2732 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
2733 ; SI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
2734 ; SI-NEXT: v_sub_f32_e32 v0, v3, v0
2735 ; SI-NEXT: s_setpc_b64 s[30:31]
2736 ;
2737 ; VI-LABEL: fadd_select_fneg_fneg_f16:
2738 ; VI: ; %bb.0:
2739 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2740 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2741 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
2742 ; VI-NEXT: v_sub_f16_e32 v0, v3, v0
2743 ; VI-NEXT: s_setpc_b64 s[30:31]
2744 %cmp = icmp eq i32 %arg0, 0
2745 %neg.x = fneg half %x
2746 %neg.y = fneg half %y
2747 %select = select i1 %cmp, half %neg.x, half %neg.y
2748 %add = fadd half %select, %z
2749 ret half %add
2750 }
2752 ; FIXME: Terrible code for SI
2753 define <2 x half> @fadd_select_fneg_fneg_v2f16(i32 %arg0, <2 x half> %x, <2 x half> %y, <2 x half> %z) {
2754 ; SI-LABEL: fadd_select_fneg_fneg_v2f16:
2755 ; SI: ; %bb.0:
2756 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2757 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
2758 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
2759 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
2760 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
2761 ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
2762 ; SI-NEXT: v_or_b32_e32 v1, v1, v2
2763 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v4
2764 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v6
2765 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2766 ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
2767 ; SI-NEXT: v_or_b32_e32 v2, v3, v2
2768 ; SI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
2769 ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v0
2770 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v4
2771 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v5
2772 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
2773 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
2774 ; SI-NEXT: v_sub_f32_e32 v0, v4, v0
2775 ; SI-NEXT: v_sub_f32_e32 v1, v3, v1
2776 ; SI-NEXT: s_setpc_b64 s[30:31]
2777 ;
2778 ; VI-LABEL: fadd_select_fneg_fneg_v2f16:
2779 ; VI: ; %bb.0:
2780 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2781 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2782 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
2783 ; VI-NEXT: v_sub_f16_sdwa v1, v3, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
2784 ; VI-NEXT: v_sub_f16_e32 v0, v3, v0
2785 ; VI-NEXT: v_or_b32_e32 v0, v0, v1
2786 ; VI-NEXT: s_setpc_b64 s[30:31]
2787 %cmp = icmp eq i32 %arg0, 0
2788 %neg.x = fneg <2 x half> %x
2789 %neg.y = fneg <2 x half> %y
2790 %select = select i1 %cmp, <2 x half> %neg.x, <2 x half> %neg.y
2791 %add = fadd <2 x half> %select, %z
2792 ret <2 x half> %add
2793 }
2795 ; --------------------------------------------------------------------------------
2796 ; select tests
2797 ; --------------------------------------------------------------------------------
2799 define amdgpu_kernel void @s_fneg_select_infloop_regression_f32(float %arg, i1 %arg1, ptr addrspace(1) %ptr) {
2800 ; SI-LABEL: s_fneg_select_infloop_regression_f32:
2801 ; SI: ; %bb.0:
2802 ; SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
2803 ; SI-NEXT: s_waitcnt lgkmcnt(0)
2804 ; SI-NEXT: s_bitcmp1_b32 s1, 0
2805 ; SI-NEXT: v_mov_b32_e32 v0, s0
2806 ; SI-NEXT: s_cselect_b64 s[0:1], -1, 0
2807 ; SI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1]
2808 ; SI-NEXT: v_cndmask_b32_e64 v2, -v0, 0, s[0:1]
2809 ; SI-NEXT: v_mov_b32_e32 v0, s2
2810 ; SI-NEXT: v_mov_b32_e32 v1, s3
2811 ; SI-NEXT: flat_store_dword v[0:1], v2
2812 ; SI-NEXT: s_endpgm
2813 ;
2814 ; VI-LABEL: s_fneg_select_infloop_regression_f32:
2815 ; VI: ; %bb.0:
2816 ; VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
2817 ; VI-NEXT: s_waitcnt lgkmcnt(0)
2818 ; VI-NEXT: s_bitcmp1_b32 s1, 0
2819 ; VI-NEXT: v_mov_b32_e32 v0, s0
2820 ; VI-NEXT: s_cselect_b64 s[0:1], -1, 0
2821 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1]
2822 ; VI-NEXT: v_cndmask_b32_e64 v2, -v0, 0, s[0:1]
2823 ; VI-NEXT: v_mov_b32_e32 v0, s2
2824 ; VI-NEXT: v_mov_b32_e32 v1, s3
2825 ; VI-NEXT: flat_store_dword v[0:1], v2
2826 ; VI-NEXT: s_endpgm
2827 %i = select i1 %arg1, float 0.0, float %arg
2828 %i2 = fneg float %i
2829 %i3 = select i1 %arg1, float 0.0, float %i2
2830 store float %i3, ptr addrspace(1) %ptr, align 4
2831 ret void
2832 }
2834 define float @v_fneg_select_infloop_regression_f32(float %arg, i1 %arg1) {
2835 ; GCN-LABEL: v_fneg_select_infloop_regression_f32:
2836 ; GCN: ; %bb.0:
2837 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2838 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2839 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2840 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
2841 ; GCN-NEXT: v_cndmask_b32_e64 v0, -v0, 0, vcc
2842 ; GCN-NEXT: s_setpc_b64 s[30:31]
2843 %i = select i1 %arg1, float 0.0, float %arg
2844 %i2 = fneg float %i
2845 %i3 = select i1 %arg1, float 0.0, float %i2
2846 ret float %i3
2847 }
2849 define float @v_fneg_select_infloop_regression_f32_commute0(float %arg, i1 %arg1) {
2850 ; GCN-LABEL: v_fneg_select_infloop_regression_f32_commute0:
2851 ; GCN: ; %bb.0:
2852 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2853 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2854 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2855 ; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
2856 ; GCN-NEXT: v_cndmask_b32_e64 v0, -v0, 0, vcc
2857 ; GCN-NEXT: s_setpc_b64 s[30:31]
2858 %i = select i1 %arg1, float %arg, float 0.0
2859 %i2 = fneg float %i
2860 %i3 = select i1 %arg1, float 0.0, float %i2
2861 ret float %i3
2862 }
2864 define float @v_fneg_select_infloop_regression_f32_commute1(float %arg, i1 %arg1) {
2865 ; GCN-LABEL: v_fneg_select_infloop_regression_f32_commute1:
2866 ; GCN: ; %bb.0:
2867 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2868 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2869 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2870 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
2871 ; GCN-NEXT: v_cndmask_b32_e64 v0, 0, -v0, vcc
2872 ; GCN-NEXT: s_setpc_b64 s[30:31]
2873 %i = select i1 %arg1, float 0.0, float %arg
2874 %i2 = fneg float %i
2875 %i3 = select i1 %arg1, float %i2, float 0.0
2876 ret float %i3
2877 }
2879 define float @v_fneg_select_infloop_regression_f32_commute2(float %arg, i1 %arg1) {
2880 ; GCN-LABEL: v_fneg_select_infloop_regression_f32_commute2:
2881 ; GCN: ; %bb.0:
2882 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2883 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2884 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2885 ; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
2886 ; GCN-NEXT: v_cndmask_b32_e64 v0, 0, -v0, vcc
2887 ; GCN-NEXT: s_setpc_b64 s[30:31]
2888 %i = select i1 %arg1, float %arg, float 0.0
2889 %i2 = fneg float %i
2890 %i3 = select i1 %arg1, float %i2, float 0.0
2891 ret float %i3
2892 }
2894 ; Check with an inline constant that's equally cheap to negate
2895 define float @v_fneg_select_infloop_regression_inline_imm_f32(float %arg, i1 %arg1) {
2896 ; GCN-LABEL: v_fneg_select_infloop_regression_inline_imm_f32:
2897 ; GCN: ; %bb.0:
2898 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2899 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2900 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2901 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 2.0, vcc
2902 ; GCN-NEXT: v_cndmask_b32_e64 v0, -v0, 2.0, vcc
2903 ; GCN-NEXT: s_setpc_b64 s[30:31]
2904 %i = select i1 %arg1, float 2.0, float %arg
2905 %i2 = fneg float %i
2906 %i3 = select i1 %arg1, float 2.0, float %i2
2907 ret float %i3
2908 }
2910 define float @v_fneg_select_infloop_regression_inline_imm_f32_commute0(float %arg, i1 %arg1) {
2911 ; GCN-LABEL: v_fneg_select_infloop_regression_inline_imm_f32_commute0:
2912 ; GCN: ; %bb.0:
2913 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2914 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2915 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2916 ; GCN-NEXT: v_cndmask_b32_e32 v0, 2.0, v0, vcc
2917 ; GCN-NEXT: v_cndmask_b32_e64 v0, -v0, 2.0, vcc
2918 ; GCN-NEXT: s_setpc_b64 s[30:31]
2919 %i = select i1 %arg1, float %arg, float 2.0
2920 %i2 = fneg float %i
2921 %i3 = select i1 %arg1, float 2.0, float %i2
2922 ret float %i3
2923 }
2925 define float @v_fneg_select_infloop_regression_inline_imm_f32_commute1(float %arg, i1 %arg1) {
2926 ; GCN-LABEL: v_fneg_select_infloop_regression_inline_imm_f32_commute1:
2927 ; GCN: ; %bb.0:
2928 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2929 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2930 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2931 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 2.0, vcc
2932 ; GCN-NEXT: v_cndmask_b32_e64 v0, 2.0, -v0, vcc
2933 ; GCN-NEXT: s_setpc_b64 s[30:31]
2934 %i = select i1 %arg1, float 2.0, float %arg
2935 %i2 = fneg float %i
2936 %i3 = select i1 %arg1, float %i2, float 2.0
2937 ret float %i3
2938 }
2940 define float @v_fneg_select_infloop_regression_inline_imm_f32_commute2(float %arg, i1 %arg1) {
2941 ; GCN-LABEL: v_fneg_select_infloop_regression_inline_imm_f32_commute2:
2942 ; GCN: ; %bb.0:
2943 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2944 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2945 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2946 ; GCN-NEXT: v_cndmask_b32_e32 v0, 2.0, v0, vcc
2947 ; GCN-NEXT: v_cndmask_b32_e64 v0, 2.0, -v0, vcc
2948 ; GCN-NEXT: s_setpc_b64 s[30:31]
2949 %i = select i1 %arg1, float %arg, float 2.0
2950 %i2 = fneg float %i
2951 %i3 = select i1 %arg1, float %i2, float 2.0
2952 ret float %i3
2953 }
2955 ; Check with an inline constant that's equally cheap to negate
2956 define float @v_fneg_select_infloop_regression_neg_inline_imm_f32(float %arg, i1 %arg1) {
2957 ; GCN-LABEL: v_fneg_select_infloop_regression_neg_inline_imm_f32:
2958 ; GCN: ; %bb.0:
2959 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2960 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2961 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2962 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, -2.0, vcc
2963 ; GCN-NEXT: v_cndmask_b32_e64 v0, -v0, -2.0, vcc
2964 ; GCN-NEXT: s_setpc_b64 s[30:31]
2965 %i = select i1 %arg1, float -2.0, float %arg
2966 %i2 = fneg float %i
2967 %i3 = select i1 %arg1, float -2.0, float %i2
2968 ret float %i3
2969 }
2971 define float @v_fneg_select_infloop_regression_neg_inline_imm_f32_commute0(float %arg, i1 %arg1) {
2972 ; GCN-LABEL: v_fneg_select_infloop_regression_neg_inline_imm_f32_commute0:
2973 ; GCN: ; %bb.0:
2974 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2975 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2976 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2977 ; GCN-NEXT: v_cndmask_b32_e32 v0, -2.0, v0, vcc
2978 ; GCN-NEXT: v_cndmask_b32_e64 v0, -v0, -2.0, vcc
2979 ; GCN-NEXT: s_setpc_b64 s[30:31]
2980 %i = select i1 %arg1, float %arg, float -2.0
2981 %i2 = fneg float %i
2982 %i3 = select i1 %arg1, float -2.0, float %i2
2983 ret float %i3
2984 }
2986 define float @v_fneg_select_infloop_regression_neg_inline_imm_f32_commute1(float %arg, i1 %arg1) {
2987 ; GCN-LABEL: v_fneg_select_infloop_regression_neg_inline_imm_f32_commute1:
2988 ; GCN: ; %bb.0:
2989 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2990 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
2991 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
2992 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, -2.0, vcc
2993 ; GCN-NEXT: v_cndmask_b32_e64 v0, -2.0, -v0, vcc
2994 ; GCN-NEXT: s_setpc_b64 s[30:31]
2995 %i = select i1 %arg1, float -2.0, float %arg
2996 %i2 = fneg float %i
2997 %i3 = select i1 %arg1, float %i2, float -2.0
2998 ret float %i3
2999 }
3001 define float @v_fneg_select_infloop_regression_neg_inline_imm_f32_commute2(float %arg, i1 %arg1) {
3002 ; GCN-LABEL: v_fneg_select_infloop_regression_neg_inline_imm_f32_commute2:
3003 ; GCN: ; %bb.0:
3004 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3005 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
3006 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
3007 ; GCN-NEXT: v_cndmask_b32_e32 v0, -2.0, v0, vcc
3008 ; GCN-NEXT: v_cndmask_b32_e64 v0, -2.0, -v0, vcc
3009 ; GCN-NEXT: s_setpc_b64 s[30:31]
3010 %i = select i1 %arg1, float %arg, float -2.0
3011 %i2 = fneg float %i
3012 %i3 = select i1 %arg1, float %i2, float -2.0
3013 ret float %i3
3014 }
3016 define amdgpu_kernel void @s_fneg_select_infloop_regression_f64(double %arg, i1 %arg1, ptr addrspace(1) %ptr) {
3017 ; SI-LABEL: s_fneg_select_infloop_regression_f64:
3018 ; SI: ; %bb.0:
3019 ; SI-NEXT: s_load_dword s4, s[2:3], 0xb
3020 ; SI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x9
3021 ; SI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0xd
3022 ; SI-NEXT: v_bfrev_b32_e32 v0, 1
3023 ; SI-NEXT: s_waitcnt lgkmcnt(0)
3024 ; SI-NEXT: s_bitcmp1_b32 s4, 0
3025 ; SI-NEXT: s_cselect_b64 s[4:5], -1, 0
3026 ; SI-NEXT: v_mov_b32_e32 v1, s1
3027 ; SI-NEXT: s_and_b64 s[6:7], s[4:5], exec
3028 ; SI-NEXT: v_cndmask_b32_e64 v0, -v1, v0, s[4:5]
3029 ; SI-NEXT: s_cselect_b32 s0, 0, s0
3030 ; SI-NEXT: v_mov_b32_e32 v2, s2
3031 ; SI-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[4:5]
3032 ; SI-NEXT: v_mov_b32_e32 v0, s0
3033 ; SI-NEXT: v_mov_b32_e32 v3, s3
3034 ; SI-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
3035 ; SI-NEXT: s_endpgm
3036 ;
3037 ; VI-LABEL: s_fneg_select_infloop_regression_f64:
3038 ; VI: ; %bb.0:
3039 ; VI-NEXT: s_load_dword s4, s[2:3], 0x2c
3040 ; VI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x24
3041 ; VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x34
3042 ; VI-NEXT: v_bfrev_b32_e32 v0, 1
3043 ; VI-NEXT: s_waitcnt lgkmcnt(0)
3044 ; VI-NEXT: s_bitcmp1_b32 s4, 0
3045 ; VI-NEXT: s_cselect_b64 s[4:5], -1, 0
3046 ; VI-NEXT: v_mov_b32_e32 v1, s1
3047 ; VI-NEXT: s_and_b64 s[6:7], s[4:5], exec
3048 ; VI-NEXT: v_cndmask_b32_e64 v0, -v1, v0, s[4:5]
3049 ; VI-NEXT: s_cselect_b32 s0, 0, s0
3050 ; VI-NEXT: v_mov_b32_e32 v2, s2
3051 ; VI-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[4:5]
3052 ; VI-NEXT: v_mov_b32_e32 v0, s0
3053 ; VI-NEXT: v_mov_b32_e32 v3, s3
3054 ; VI-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
3055 ; VI-NEXT: s_endpgm
3056 %i = select i1 %arg1, double 0.0, double %arg
3057 %i2 = fneg double %i
3058 %i3 = select i1 %arg1, double 0.0, double %i2
3059 store double %i3, ptr addrspace(1) %ptr, align 4
3060 ret void
3061 }
3063 define double @v_fneg_select_infloop_regression_f64(double %arg, i1 %arg1) {
3064 ; GCN-LABEL: v_fneg_select_infloop_regression_f64:
3065 ; GCN: ; %bb.0:
3066 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3067 ; GCN-NEXT: v_and_b32_e32 v2, 1, v2
3068 ; GCN-NEXT: v_bfrev_b32_e32 v3, 1
3069 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
3070 ; GCN-NEXT: v_cndmask_b32_e64 v1, -v1, v3, vcc
3071 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3072 ; GCN-NEXT: v_cndmask_b32_e64 v1, v1, 0, vcc
3073 ; GCN-NEXT: s_setpc_b64 s[30:31]
3074 %i = select i1 %arg1, double 0.0, double %arg
3075 %i2 = fneg double %i
3076 %i3 = select i1 %arg1, double 0.0, double %i2
3077 ret double %i3
3078 }
3080 define amdgpu_kernel void @s_fneg_select_infloop_regression_f16(half %arg, i1 %arg1, ptr addrspace(1) %ptr) {
3081 ; SI-LABEL: s_fneg_select_infloop_regression_f16:
3082 ; SI: ; %bb.0:
3083 ; SI-NEXT: s_load_dword s4, s[2:3], 0x9
3084 ; SI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0xb
3085 ; SI-NEXT: s_waitcnt lgkmcnt(0)
3086 ; SI-NEXT: v_cvt_f32_f16_e32 v0, s4
3087 ; SI-NEXT: s_bitcmp1_b32 s4, 16
3088 ; SI-NEXT: s_cselect_b64 s[2:3], -1, 0
3089 ; SI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[2:3]
3090 ; SI-NEXT: v_cndmask_b32_e64 v0, -v0, 0, s[2:3]
3091 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v0
3092 ; SI-NEXT: v_mov_b32_e32 v0, s0
3093 ; SI-NEXT: v_mov_b32_e32 v1, s1
3094 ; SI-NEXT: flat_store_short v[0:1], v2
3095 ; SI-NEXT: s_endpgm
3096 ;
3097 ; VI-LABEL: s_fneg_select_infloop_regression_f16:
3098 ; VI: ; %bb.0:
3099 ; VI-NEXT: s_load_dword s4, s[2:3], 0x24
3100 ; VI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2c
3101 ; VI-NEXT: s_waitcnt lgkmcnt(0)
3102 ; VI-NEXT: s_bitcmp1_b32 s4, 16
3103 ; VI-NEXT: v_mov_b32_e32 v0, s4
3104 ; VI-NEXT: s_cselect_b64 s[2:3], -1, 0
3105 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[2:3]
3106 ; VI-NEXT: v_xor_b32_e32 v0, 0x8000, v0
3107 ; VI-NEXT: v_cndmask_b32_e64 v2, v0, 0, s[2:3]
3108 ; VI-NEXT: v_mov_b32_e32 v0, s0
3109 ; VI-NEXT: v_mov_b32_e32 v1, s1
3110 ; VI-NEXT: flat_store_short v[0:1], v2
3111 ; VI-NEXT: s_endpgm
3112 %i = select i1 %arg1, half 0.0, half %arg
3113 %i2 = fneg half %i
3114 %i3 = select i1 %arg1, half 0.0, half %i2
3115 store half %i3, ptr addrspace(1) %ptr, align 4
3116 ret void
3117 }
3119 define half @v_fneg_select_infloop_regression_f16(half %arg, i1 %arg1) {
3120 ; SI-LABEL: v_fneg_select_infloop_regression_f16:
3121 ; SI: ; %bb.0:
3122 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3123 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
3124 ; SI-NEXT: v_and_b32_e32 v1, 1, v1
3125 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
3126 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
3127 ; SI-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3128 ; SI-NEXT: v_cndmask_b32_e64 v0, -v0, 0, vcc
3129 ; SI-NEXT: s_setpc_b64 s[30:31]
3130 ;
3131 ; VI-LABEL: v_fneg_select_infloop_regression_f16:
3132 ; VI: ; %bb.0:
3133 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3134 ; VI-NEXT: v_and_b32_e32 v1, 1, v1
3135 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
3136 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3137 ; VI-NEXT: v_xor_b32_e32 v0, 0x8000, v0
3138 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3139 ; VI-NEXT: s_setpc_b64 s[30:31]
3140 %i = select i1 %arg1, half 0.0, half %arg
3141 %i2 = fneg half %i
3142 %i3 = select i1 %arg1, half 0.0, half %i2
3143 ret half %i3
3144 }
3146 define amdgpu_kernel void @s_fneg_select_infloop_regression_v2f16(<2 x half> %arg, i1 %arg1, ptr addrspace(1) %ptr) {
3147 ; SI-LABEL: s_fneg_select_infloop_regression_v2f16:
3148 ; SI: ; %bb.0:
3149 ; SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
3150 ; SI-NEXT: s_waitcnt lgkmcnt(0)
3151 ; SI-NEXT: s_and_b32 s1, 1, s1
3152 ; SI-NEXT: s_cselect_b32 s0, 0, s0
3153 ; SI-NEXT: s_xor_b32 s0, s0, 0x80008000
3154 ; SI-NEXT: s_cmp_eq_u32 s1, 1
3155 ; SI-NEXT: s_cselect_b32 s0, 0, s0
3156 ; SI-NEXT: v_mov_b32_e32 v0, s2
3157 ; SI-NEXT: v_mov_b32_e32 v1, s3
3158 ; SI-NEXT: v_mov_b32_e32 v2, s0
3159 ; SI-NEXT: flat_store_dword v[0:1], v2
3160 ; SI-NEXT: s_endpgm
3161 ;
3162 ; VI-LABEL: s_fneg_select_infloop_regression_v2f16:
3163 ; VI: ; %bb.0:
3164 ; VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
3165 ; VI-NEXT: s_waitcnt lgkmcnt(0)
3166 ; VI-NEXT: s_and_b32 s1, 1, s1
3167 ; VI-NEXT: s_cselect_b32 s0, 0, s0
3168 ; VI-NEXT: s_xor_b32 s0, s0, 0x80008000
3169 ; VI-NEXT: s_cmp_eq_u32 s1, 1
3170 ; VI-NEXT: s_cselect_b32 s0, 0, s0
3171 ; VI-NEXT: v_mov_b32_e32 v0, s2
3172 ; VI-NEXT: v_mov_b32_e32 v1, s3
3173 ; VI-NEXT: v_mov_b32_e32 v2, s0
3174 ; VI-NEXT: flat_store_dword v[0:1], v2
3175 ; VI-NEXT: s_endpgm
3176 %i = select i1 %arg1, <2 x half> zeroinitializer, <2 x half> %arg
3177 %i2 = fneg <2 x half> %i
3178 %i3 = select i1 %arg1, <2 x half> zeroinitializer, <2 x half> %i2
3179 store <2 x half> %i3, ptr addrspace(1) %ptr, align 4
3180 ret void
3181 }
3183 define <2 x half> @v_fneg_select_infloop_regression_v2f16(<2 x half> %arg, i1 %arg1) {
3184 ; SI-LABEL: v_fneg_select_infloop_regression_v2f16:
3185 ; SI: ; %bb.0:
3186 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3187 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
3188 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
3189 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
3190 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
3191 ; SI-NEXT: v_and_b32_e32 v1, 1, v2
3192 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
3193 ; SI-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3194 ; SI-NEXT: v_xor_b32_e32 v0, 0x80008000, v0
3195 ; SI-NEXT: v_cndmask_b32_e64 v1, v0, 0, vcc
3196 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v1
3197 ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
3198 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
3199 ; SI-NEXT: s_setpc_b64 s[30:31]
3200 ;
3201 ; VI-LABEL: v_fneg_select_infloop_regression_v2f16:
3202 ; VI: ; %bb.0:
3203 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3204 ; VI-NEXT: v_and_b32_e32 v1, 1, v1
3205 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
3206 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3207 ; VI-NEXT: v_xor_b32_e32 v0, 0x80008000, v0
3208 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3209 ; VI-NEXT: s_setpc_b64 s[30:31]
3210 %i = select i1 %arg1, <2 x half> zeroinitializer, <2 x half> %arg
3211 %i2 = fneg <2 x half> %i
3212 %i3 = select i1 %arg1, <2 x half> zeroinitializer, <2 x half> %i2
3213 ret <2 x half> %i3
3214 }
3216 define amdgpu_kernel void @s_fneg_select_infloop_regression_v2f32(<2 x float> %arg, i1 %arg1, ptr addrspace(1) %ptr) {
3217 ; SI-LABEL: s_fneg_select_infloop_regression_v2f32:
3218 ; SI: ; %bb.0:
3219 ; SI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x9
3220 ; SI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0xd
3221 ; SI-NEXT: v_bfrev_b32_e32 v0, 1
3222 ; SI-NEXT: s_waitcnt lgkmcnt(0)
3223 ; SI-NEXT: s_bitcmp1_b32 s6, 0
3224 ; SI-NEXT: v_mov_b32_e32 v1, s4
3225 ; SI-NEXT: s_cselect_b64 s[2:3], -1, 0
3226 ; SI-NEXT: v_cndmask_b32_e64 v2, -v1, v0, s[2:3]
3227 ; SI-NEXT: v_mov_b32_e32 v1, s5
3228 ; SI-NEXT: v_cndmask_b32_e64 v0, -v1, v0, s[2:3]
3229 ; SI-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[2:3]
3230 ; SI-NEXT: v_cndmask_b32_e64 v0, v2, 0, s[2:3]
3231 ; SI-NEXT: v_mov_b32_e32 v3, s1
3232 ; SI-NEXT: v_mov_b32_e32 v2, s0
3233 ; SI-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
3234 ; SI-NEXT: s_endpgm
3235 ;
3236 ; VI-LABEL: s_fneg_select_infloop_regression_v2f32:
3237 ; VI: ; %bb.0:
3238 ; VI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x24
3239 ; VI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x34
3240 ; VI-NEXT: v_bfrev_b32_e32 v0, 1
3241 ; VI-NEXT: s_waitcnt lgkmcnt(0)
3242 ; VI-NEXT: s_bitcmp1_b32 s6, 0
3243 ; VI-NEXT: v_mov_b32_e32 v1, s4
3244 ; VI-NEXT: s_cselect_b64 s[2:3], -1, 0
3245 ; VI-NEXT: v_cndmask_b32_e64 v2, -v1, v0, s[2:3]
3246 ; VI-NEXT: v_mov_b32_e32 v1, s5
3247 ; VI-NEXT: v_cndmask_b32_e64 v0, -v1, v0, s[2:3]
3248 ; VI-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[2:3]
3249 ; VI-NEXT: v_cndmask_b32_e64 v0, v2, 0, s[2:3]
3250 ; VI-NEXT: v_mov_b32_e32 v3, s1
3251 ; VI-NEXT: v_mov_b32_e32 v2, s0
3252 ; VI-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
3253 ; VI-NEXT: s_endpgm
3254 %i = select i1 %arg1, <2 x float> zeroinitializer, <2 x float> %arg
3255 %i2 = fneg <2 x float> %i
3256 %i3 = select i1 %arg1, <2 x float> zeroinitializer, <2 x float> %i2
3257 store <2 x float> %i3, ptr addrspace(1) %ptr, align 4
3258 ret void
3259 }
3261 define <2 x float> @v_fneg_select_infloop_regression_v2f32(<2 x float> %arg, i1 %arg1) {
3262 ; GCN-LABEL: v_fneg_select_infloop_regression_v2f32:
3263 ; GCN: ; %bb.0:
3264 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3265 ; GCN-NEXT: v_and_b32_e32 v2, 1, v2
3266 ; GCN-NEXT: v_bfrev_b32_e32 v3, 1
3267 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
3268 ; GCN-NEXT: v_cndmask_b32_e64 v1, -v1, v3, vcc
3269 ; GCN-NEXT: v_cndmask_b32_e64 v0, -v0, v3, vcc
3270 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3271 ; GCN-NEXT: v_cndmask_b32_e64 v1, v1, 0, vcc
3272 ; GCN-NEXT: s_setpc_b64 s[30:31]
3273 %i = select i1 %arg1, <2 x float> zeroinitializer, <2 x float> %arg
3274 %i2 = fneg <2 x float> %i
3275 %i3 = select i1 %arg1, <2 x float> zeroinitializer, <2 x float> %i2
3276 ret <2 x float> %i3
3277 }
3279 define amdgpu_kernel void @s_fabs_select_infloop_regression_f32(float %arg, i1 %arg1, ptr addrspace(1) %ptr) {
3280 ; SI-LABEL: s_fabs_select_infloop_regression_f32:
3281 ; SI: ; %bb.0:
3282 ; SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
3283 ; SI-NEXT: s_waitcnt lgkmcnt(0)
3284 ; SI-NEXT: s_bitcmp1_b32 s1, 0
3285 ; SI-NEXT: v_mov_b32_e32 v0, s0
3286 ; SI-NEXT: s_cselect_b64 s[0:1], -1, 0
3287 ; SI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1]
3288 ; SI-NEXT: v_cndmask_b32_e64 v2, |v0|, 0, s[0:1]
3289 ; SI-NEXT: v_mov_b32_e32 v0, s2
3290 ; SI-NEXT: v_mov_b32_e32 v1, s3
3291 ; SI-NEXT: flat_store_dword v[0:1], v2
3292 ; SI-NEXT: s_endpgm
3293 ;
3294 ; VI-LABEL: s_fabs_select_infloop_regression_f32:
3295 ; VI: ; %bb.0:
3296 ; VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
3297 ; VI-NEXT: s_waitcnt lgkmcnt(0)
3298 ; VI-NEXT: s_bitcmp1_b32 s1, 0
3299 ; VI-NEXT: v_mov_b32_e32 v0, s0
3300 ; VI-NEXT: s_cselect_b64 s[0:1], -1, 0
3301 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1]
3302 ; VI-NEXT: v_cndmask_b32_e64 v2, |v0|, 0, s[0:1]
3303 ; VI-NEXT: v_mov_b32_e32 v0, s2
3304 ; VI-NEXT: v_mov_b32_e32 v1, s3
3305 ; VI-NEXT: flat_store_dword v[0:1], v2
3306 ; VI-NEXT: s_endpgm
3307 %i = select i1 %arg1, float 0.0, float %arg
3308 %i2 = call float @llvm.fabs.f32(float %i)
3309 %i3 = select i1 %arg1, float 0.0, float %i2
3310 store float %i3, ptr addrspace(1) %ptr, align 4
3311 ret void
3312 }
3314 define float @v_fabs_select_infloop_regression_f32(float %arg, i1 %arg1) {
3315 ; GCN-LABEL: v_fabs_select_infloop_regression_f32:
3316 ; GCN: ; %bb.0:
3317 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3318 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
3319 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
3320 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3321 ; GCN-NEXT: v_cndmask_b32_e64 v0, |v0|, 0, vcc
3322 ; GCN-NEXT: s_setpc_b64 s[30:31]
3323 %i = select i1 %arg1, float 0.0, float %arg
3324 %i2 = call float @llvm.fabs.f32(float %i)
3325 %i3 = select i1 %arg1, float 0.0, float %i2
3326 ret float %i3
3327 }
3329 define amdgpu_kernel void @s_fneg_fabs_select_infloop_regression(float %arg, i1 %arg1, ptr addrspace(1) %ptr) {
3330 ; SI-LABEL: s_fneg_fabs_select_infloop_regression:
3331 ; SI: ; %bb.0:
3332 ; SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
3333 ; SI-NEXT: s_waitcnt lgkmcnt(0)
3334 ; SI-NEXT: s_bitcmp1_b32 s1, 0
3335 ; SI-NEXT: v_mov_b32_e32 v0, s0
3336 ; SI-NEXT: s_cselect_b64 s[0:1], -1, 0
3337 ; SI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1]
3338 ; SI-NEXT: v_cndmask_b32_e64 v2, -|v0|, 0, s[0:1]
3339 ; SI-NEXT: v_mov_b32_e32 v0, s2
3340 ; SI-NEXT: v_mov_b32_e32 v1, s3
3341 ; SI-NEXT: flat_store_dword v[0:1], v2
3342 ; SI-NEXT: s_endpgm
3343 ;
3344 ; VI-LABEL: s_fneg_fabs_select_infloop_regression:
3345 ; VI: ; %bb.0:
3346 ; VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
3347 ; VI-NEXT: s_waitcnt lgkmcnt(0)
3348 ; VI-NEXT: s_bitcmp1_b32 s1, 0
3349 ; VI-NEXT: v_mov_b32_e32 v0, s0
3350 ; VI-NEXT: s_cselect_b64 s[0:1], -1, 0
3351 ; VI-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1]
3352 ; VI-NEXT: v_cndmask_b32_e64 v2, -|v0|, 0, s[0:1]
3353 ; VI-NEXT: v_mov_b32_e32 v0, s2
3354 ; VI-NEXT: v_mov_b32_e32 v1, s3
3355 ; VI-NEXT: flat_store_dword v[0:1], v2
3356 ; VI-NEXT: s_endpgm
3357 %i = select i1 %arg1, float 0.0, float %arg
3358 %i2 = call float @llvm.fabs.f32(float %i)
3359 %neg.i2 = fneg float %i2
3360 %i3 = select i1 %arg1, float 0.0, float %neg.i2
3361 store float %i3, ptr addrspace(1) %ptr, align 4
3362 ret void
3363 }
3365 define float @v_fneg_fabs_select_infloop_regression(float %arg, i1 %arg1) {
3366 ; GCN-LABEL: v_fneg_fabs_select_infloop_regression:
3367 ; GCN: ; %bb.0:
3368 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3369 ; GCN-NEXT: v_and_b32_e32 v1, 1, v1
3370 ; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
3371 ; GCN-NEXT: v_cndmask_b32_e64 v0, v0, 0, vcc
3372 ; GCN-NEXT: v_cndmask_b32_e64 v0, -|v0|, 0, vcc
3373 ; GCN-NEXT: s_setpc_b64 s[30:31]
3374 %i = select i1 %arg1, float 0.0, float %arg
3375 %i2 = call float @llvm.fabs.f32(float %i)
3376 %neg.i2 = fneg float %i2
3377 %i3 = select i1 %arg1, float 0.0, float %neg.i2
3378 ret float %i3
3379 }
3381 define float @v_fmul_0_fsub_0_infloop_regression(float %arg) {
3382 ; GCN-SAFE-LABEL: v_fmul_0_fsub_0_infloop_regression:
3383 ; GCN-SAFE: ; %bb.0: ; %bb
3384 ; GCN-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3385 ; GCN-SAFE-NEXT: v_mul_f32_e32 v0, 0, v0
3386 ; GCN-SAFE-NEXT: v_sub_f32_e32 v0, 0, v0
3387 ; GCN-SAFE-NEXT: s_setpc_b64 s[30:31]
3388 ;
3389 ; GCN-NSZ-LABEL: v_fmul_0_fsub_0_infloop_regression:
3390 ; GCN-NSZ: ; %bb.0: ; %bb
3391 ; GCN-NSZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3392 ; GCN-NSZ-NEXT: v_mul_f32_e32 v0, 0x80000000, v0
3393 ; GCN-NSZ-NEXT: s_setpc_b64 s[30:31]
3394 bb:
3395 %i = fmul float %arg, 0.0
3396 %i1 = fsub float 0.0, %i
3397 ret float %i1
3398 }
3400 declare i32 @llvm.amdgcn.workitem.id.x() #1
3401 declare float @llvm.fma.f32(float, float, float) #1
3402 declare <2 x float> @llvm.fma.v2f32(<2 x float>, <2 x float>, <2 x float>)
3403 declare float @llvm.fmuladd.f32(float, float, float) #1
3404 declare <4 x float> @llvm.fmuladd.v4f32(<4 x float>, <4 x float>, <4 x float>) #1
3405 declare float @llvm.fabs.f32(float) #1
3406 declare float @llvm.sin.f32(float) #1
3407 declare float @llvm.trunc.f32(float) #1
3408 declare float @llvm.round.f32(float) #1
3409 declare float @llvm.rint.f32(float) #1
3410 declare float @llvm.nearbyint.f32(float) #1
3411 declare float @llvm.roundeven.f32(float) #1
3412 declare float @llvm.canonicalize.f32(float) #1
3413 declare float @llvm.arithmetic.fence.f32(float) #1
3414 declare float @llvm.minnum.f32(float, float) #1
3415 declare float @llvm.maxnum.f32(float, float) #1
3416 declare double @llvm.minnum.f64(double, double) #1
3417 declare double @llvm.fma.f64(double, double, double) #1
3419 declare float @llvm.amdgcn.sin.f32(float) #1
3420 declare float @llvm.amdgcn.rcp.f32(float) #1
3421 declare float @llvm.amdgcn.rcp.legacy(float) #1
3422 declare float @llvm.amdgcn.fmul.legacy(float, float) #1
3423 declare float @llvm.amdgcn.interp.p1(float, i32, i32, i32) #0
3424 declare float @llvm.amdgcn.interp.p2(float, float, i32, i32, i32) #0
3426 declare half @llvm.fma.f16(half, half, half) #1
3427 declare <2 x half> @llvm.fma.v2f16(<2 x half>, <2 x half>, <2 x half>)
3428 declare half @llvm.fmuladd.f16(half, half, half) #1
3429 declare <4 x half> @llvm.fmuladd.v4f16(<4 x half>, <4 x half>, <4 x half>) #1
3430 declare half @llvm.sin.f16(half) #1
3431 declare half @llvm.trunc.f16(half) #1
3432 declare half @llvm.round.f16(half) #1
3433 declare half @llvm.rint.f16(half) #1
3434 declare half @llvm.nearbyint.f16(half) #1
3435 declare half @llvm.canonicalize.f16(half) #1
3436 declare half @llvm.minnum.f16(half, half) #1
3437 declare half @llvm.maxnum.f16(half, half) #1
3438 declare half @llvm.amdgcn.sin.f16(half) #1
3439 declare half @llvm.amdgcn.rcp.f16(half) #1
3441 attributes #0 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
3442 attributes #1 = { nounwind readnone }
3443 attributes #2 = { nounwind "unsafe-fp-math"="true" }
3444 attributes #3 = { nounwind "no-signed-zeros-fp-math"="true" }
3445 attributes #4 = { nounwind "amdgpu-ieee"="false" "denormal-fp-math-f32"="preserve-sign,preserve-sign" }