1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
|
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -passes=instcombine -S < %s | FileCheck %s
declare half @llvm.amdgcn.sqrt.f16(half)
declare float @llvm.amdgcn.sqrt.f32(float)
declare double @llvm.amdgcn.sqrt.f64(double)
declare half @llvm.amdgcn.rcp.f16(half)
declare float @llvm.amdgcn.rcp.f32(float)
declare double @llvm.amdgcn.rcp.f64(double)
declare half @llvm.amdgcn.rsq.f16(half)
declare float @llvm.amdgcn.rsq.f32(float)
declare double @llvm.amdgcn.rsq.f64(double)
declare half @llvm.sqrt.f16(half)
declare float @llvm.sqrt.f32(float)
declare double @llvm.sqrt.f64(double)
; Should allow precision increasing contraction to rsq call
define float @amdgcn_rcp_amdgcn_sqrt_f32_contract(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_amdgcn_sqrt_f32_contract
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1:[0-9]+]] {
; CHECK-NEXT: [[RSQ:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[X]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call contract float @llvm.amdgcn.sqrt.f32(float %x)
%rsq = call contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
; contract required on both calls
define float @amdgcn_rcp_amdgcn_sqrt_f32_missing_contract0(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_amdgcn_sqrt_f32_missing_contract0
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[SQRT]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call float @llvm.amdgcn.sqrt.f32(float %x)
%rsq = call contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
; contract required on both calls
define float @amdgcn_rcp_amdgcn_sqrt_f32_missing_contract1(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_amdgcn_sqrt_f32_missing_contract1
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract float @llvm.amdgcn.sqrt.f32(float [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call float @llvm.amdgcn.rcp.f32(float [[SQRT]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call contract float @llvm.amdgcn.sqrt.f32(float %x)
%rsq = call float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
; Reject from multiple uses of sqrt
define float @amdgcn_rcp_amdgcn_sqrt_f32_contract_multi_use(float %x, ptr %ptr) {
; CHECK-LABEL: define float @amdgcn_rcp_amdgcn_sqrt_f32_contract_multi_use
; CHECK-SAME: (float [[X:%.*]], ptr [[PTR:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract float @llvm.amdgcn.sqrt.f32(float [[X]])
; CHECK-NEXT: store float [[SQRT]], ptr [[PTR]], align 4
; CHECK-NEXT: [[RSQ:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[SQRT]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call contract float @llvm.amdgcn.sqrt.f32(float %x)
store float %sqrt, ptr %ptr
%rsq = call contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
; Test flags are or'd together
define float @amdgcn_rcp_nnan_amdgcn_sqrt_ninf_f32_contract(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_nnan_amdgcn_sqrt_ninf_f32_contract
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call nnan ninf contract float @llvm.amdgcn.rsq.f32(float [[X]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call nnan contract float @llvm.amdgcn.sqrt.f32(float %x)
%rsq = call ninf contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
; Should allow precision increasing contraction to rsq call
define half @amdgcn_rcp_amdgcn_sqrt_f16_contract(half %x) {
; CHECK-LABEL: define half @amdgcn_rcp_amdgcn_sqrt_f16_contract
; CHECK-SAME: (half [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call contract half @llvm.amdgcn.rsq.f16(half [[X]])
; CHECK-NEXT: ret half [[RSQ]]
;
%sqrt = call contract half @llvm.amdgcn.sqrt.f16(half %x)
%rsq = call contract half @llvm.amdgcn.rcp.f16(half %sqrt)
ret half %rsq
}
; contract required on both calls
define half @amdgcn_rcp_amdgcn_sqrt_f16_missing_contract0(half %x) {
; CHECK-LABEL: define half @amdgcn_rcp_amdgcn_sqrt_f16_missing_contract0
; CHECK-SAME: (half [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call half @llvm.sqrt.f16(half [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call contract half @llvm.amdgcn.rcp.f16(half [[SQRT]])
; CHECK-NEXT: ret half [[RSQ]]
;
%sqrt = call half @llvm.amdgcn.sqrt.f16(half %x)
%rsq = call contract half @llvm.amdgcn.rcp.f16(half %sqrt)
ret half %rsq
}
; contract required on both calls
define half @amdgcn_rcp_amdgcn_sqrt_f16_missing_contract1(half %x) {
; CHECK-LABEL: define half @amdgcn_rcp_amdgcn_sqrt_f16_missing_contract1
; CHECK-SAME: (half [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract half @llvm.sqrt.f16(half [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call half @llvm.amdgcn.rcp.f16(half [[SQRT]])
; CHECK-NEXT: ret half [[RSQ]]
;
%sqrt = call contract half @llvm.amdgcn.sqrt.f16(half %x)
%rsq = call half @llvm.amdgcn.rcp.f16(half %sqrt)
ret half %rsq
}
; Reject from multiple uses of sqrt
define half @amdgcn_rcp_amdgcn_sqrt_f16_contract_multi_use(half %x, ptr %ptr) {
; CHECK-LABEL: define half @amdgcn_rcp_amdgcn_sqrt_f16_contract_multi_use
; CHECK-SAME: (half [[X:%.*]], ptr [[PTR:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract half @llvm.sqrt.f16(half [[X]])
; CHECK-NEXT: store half [[SQRT]], ptr [[PTR]], align 2
; CHECK-NEXT: [[RSQ:%.*]] = call contract half @llvm.amdgcn.rcp.f16(half [[SQRT]])
; CHECK-NEXT: ret half [[RSQ]]
;
%sqrt = call contract half @llvm.amdgcn.sqrt.f16(half %x)
store half %sqrt, ptr %ptr
%rsq = call contract half @llvm.amdgcn.rcp.f16(half %sqrt)
ret half %rsq
}
; Test flags are or'd together
define half @amdgcn_rcp_nnan_amdgcn_sqrt_ninf_f16_contract(half %x) {
; CHECK-LABEL: define half @amdgcn_rcp_nnan_amdgcn_sqrt_ninf_f16_contract
; CHECK-SAME: (half [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call nnan ninf contract half @llvm.amdgcn.rsq.f16(half [[X]])
; CHECK-NEXT: ret half [[RSQ]]
;
%sqrt = call nnan contract half @llvm.amdgcn.sqrt.f16(half %x)
%rsq = call ninf contract half @llvm.amdgcn.rcp.f16(half %sqrt)
ret half %rsq
}
; Should allow precision increasing contraction to rsq call
define double @amdgcn_rcp_amdgcn_sqrt_f64_contract(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_amdgcn_sqrt_f64_contract
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call contract double @llvm.amdgcn.rsq.f64(double [[X]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call contract double @llvm.amdgcn.sqrt.f64(double %x)
%rsq = call contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
; contract required on both calls
define double @amdgcn_rcp_amdgcn_sqrt_f64_missing_contract0(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_amdgcn_sqrt_f64_missing_contract0
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call double @llvm.amdgcn.sqrt.f64(double [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call contract double @llvm.amdgcn.rcp.f64(double [[SQRT]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call double @llvm.amdgcn.sqrt.f64(double %x)
%rsq = call contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
; contract required on both calls
define double @amdgcn_rcp_amdgcn_sqrt_f64_missing_contract1(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_amdgcn_sqrt_f64_missing_contract1
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract double @llvm.amdgcn.sqrt.f64(double [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call double @llvm.amdgcn.rcp.f64(double [[SQRT]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call contract double @llvm.amdgcn.sqrt.f64(double %x)
%rsq = call double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
; Reject from multiple uses of sqrt
define double @amdgcn_rcp_amdgcn_sqrt_f64_contract_multi_use(double %x, ptr %ptr) {
; CHECK-LABEL: define double @amdgcn_rcp_amdgcn_sqrt_f64_contract_multi_use
; CHECK-SAME: (double [[X:%.*]], ptr [[PTR:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract double @llvm.amdgcn.sqrt.f64(double [[X]])
; CHECK-NEXT: store double [[SQRT]], ptr [[PTR]], align 8
; CHECK-NEXT: [[RSQ:%.*]] = call contract double @llvm.amdgcn.rcp.f64(double [[SQRT]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call contract double @llvm.amdgcn.sqrt.f64(double %x)
store double %sqrt, ptr %ptr
%rsq = call contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
; Test flags are or'd together
define double @amdgcn_rcp_nnan_amdgcn_sqrt_ninf_f64_contract(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_nnan_amdgcn_sqrt_ninf_f64_contract
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call nnan ninf contract double @llvm.amdgcn.rsq.f64(double [[X]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call nnan contract double @llvm.amdgcn.sqrt.f64(double %x)
%rsq = call ninf contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
; Do not contract with regular sqrt
define float @amdgcn_rcp_sqrt_f32_contract(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_sqrt_f32_contract
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract float @llvm.sqrt.f32(float [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[SQRT]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call contract float @llvm.sqrt.f32(float %x)
%rsq = call contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
; Do contract with regular sqrt for f16
define half @amdgcn_rcp_sqrt_f16_contract(half %x) {
; CHECK-LABEL: define half @amdgcn_rcp_sqrt_f16_contract
; CHECK-SAME: (half [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call contract half @llvm.amdgcn.rsq.f16(half [[X]])
; CHECK-NEXT: ret half [[RSQ]]
;
%sqrt = call contract half @llvm.sqrt.f16(half %x)
%rsq = call contract half @llvm.amdgcn.rcp.f16(half %sqrt)
ret half %rsq
}
; Do not contract with regular sqrt
define double @amdgcn_rcp_sqrt_f64_contract(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_sqrt_f64_contract
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract double @llvm.sqrt.f64(double [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call contract double @llvm.amdgcn.rcp.f64(double [[SQRT]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call contract double @llvm.sqrt.f64(double %x)
%rsq = call contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
define float @amdgcn_rcp_afn_sqrt_f32_contract(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_afn_sqrt_f32_contract
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call contract afn float @llvm.amdgcn.rsq.f32(float [[X]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call afn contract float @llvm.sqrt.f32(float %x)
%rsq = call contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
define float @amdgcn_rcp_fpmath3_sqrt_f32_contract(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_fpmath3_sqrt_f32_contract
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[X]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call contract float @llvm.sqrt.f32(float %x), !fpmath !0
%rsq = call contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
define float @amdgcn_rcp_fpmath1_sqrt_f32_contract(float %x) {
; CHECK-LABEL: define float @amdgcn_rcp_fpmath1_sqrt_f32_contract
; CHECK-SAME: (float [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[RSQ:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[X]])
; CHECK-NEXT: ret float [[RSQ]]
;
%sqrt = call contract float @llvm.sqrt.f32(float %x), !fpmath !1
%rsq = call contract float @llvm.amdgcn.rcp.f32(float %sqrt)
ret float %rsq
}
; Ignore f64 sqrt
define double @amdgcn_rcp_afn_sqrt_f64_contract(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_afn_sqrt_f64_contract
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract afn double @llvm.sqrt.f64(double [[X]])
; CHECK-NEXT: [[RSQ:%.*]] = call contract double @llvm.amdgcn.rcp.f64(double [[SQRT]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call afn contract double @llvm.sqrt.f64(double %x)
%rsq = call contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
; Ignore f64 sqrt
define double @amdgcn_rcp_fpmath3_sqrt_f64_contract(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_fpmath3_sqrt_f64_contract
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract double @llvm.sqrt.f64(double [[X]]), !fpmath !0
; CHECK-NEXT: [[RSQ:%.*]] = call contract double @llvm.amdgcn.rcp.f64(double [[SQRT]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call contract double @llvm.sqrt.f64(double %x), !fpmath !0
%rsq = call contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
; Ignore f64 sqrt
define double @amdgcn_rcp_fpmath1_sqrt_f64_contract(double %x) {
; CHECK-LABEL: define double @amdgcn_rcp_fpmath1_sqrt_f64_contract
; CHECK-SAME: (double [[X:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[SQRT:%.*]] = call contract double @llvm.sqrt.f64(double [[X]]), !fpmath !1
; CHECK-NEXT: [[RSQ:%.*]] = call contract double @llvm.amdgcn.rcp.f64(double [[SQRT]])
; CHECK-NEXT: ret double [[RSQ]]
;
%sqrt = call contract double @llvm.sqrt.f64(double %x), !fpmath !1
%rsq = call contract double @llvm.amdgcn.rcp.f64(double %sqrt)
ret double %rsq
}
!0 = !{float 3.0}
!1 = !{float 1.0}
|
