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
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
|
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -S -instcombine < %s | FileCheck %s
define float @exact_inverse(float %x) {
; CHECK-LABEL: @exact_inverse(
; CHECK-NEXT: [[DIV:%.*]] = fmul float [[X:%.*]], 1.250000e-01
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 8.0
ret float %div
}
; Min normal float = 1.17549435E-38
define float @exact_inverse2(float %x) {
; CHECK-LABEL: @exact_inverse2(
; CHECK-NEXT: [[DIV:%.*]] = fmul float [[X:%.*]], 0x47D0000000000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 0x3810000000000000
ret float %div
}
; Max exponent = 1.70141183E+38; don't transform to multiply with denormal.
define float @exact_inverse_but_denorm(float %x) {
; CHECK-LABEL: @exact_inverse_but_denorm(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], 0x47E0000000000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 0x47E0000000000000
ret float %div
}
; Denormal = float 1.40129846E-45; inverse can't be represented.
define float @not_exact_inverse2(float %x) {
; CHECK-LABEL: @not_exact_inverse2(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], 0x36A0000000000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 0x36A0000000000000
ret float %div
}
; Fast math allows us to replace this fdiv.
define float @not_exact_but_allow_recip(float %x) {
; CHECK-LABEL: @not_exact_but_allow_recip(
; CHECK-NEXT: [[DIV:%.*]] = fmul arcp float [[X:%.*]], 0x3FD5555560000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv arcp float %x, 3.0
ret float %div
}
; Fast math allows us to replace this fdiv, but we don't to avoid a denormal.
; TODO: What if the function attributes tell us that denormals are flushed?
define float @not_exact_but_allow_recip_but_denorm(float %x) {
; CHECK-LABEL: @not_exact_but_allow_recip_but_denorm(
; CHECK-NEXT: [[DIV:%.*]] = fdiv arcp float [[X:%.*]], 0x47E0000100000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv arcp float %x, 0x47E0000100000000
ret float %div
}
define <2 x float> @exact_inverse_splat(<2 x float> %x) {
; CHECK-LABEL: @exact_inverse_splat(
; CHECK-NEXT: [[DIV:%.*]] = fmul <2 x float> [[X:%.*]], <float 2.500000e-01, float 2.500000e-01>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 4.0, float 4.0>
ret <2 x float> %div
}
; Fast math allows us to replace this fdiv.
define <2 x float> @not_exact_but_allow_recip_splat(<2 x float> %x) {
; CHECK-LABEL: @not_exact_but_allow_recip_splat(
; CHECK-NEXT: [[DIV:%.*]] = fmul arcp <2 x float> [[X:%.*]], <float 0x3FD5555560000000, float 0x3FD5555560000000>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv arcp <2 x float> %x, <float 3.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @exact_inverse_vec(<2 x float> %x) {
; CHECK-LABEL: @exact_inverse_vec(
; CHECK-NEXT: [[DIV:%.*]] = fmul <2 x float> [[X:%.*]], <float 2.500000e-01, float 1.250000e-01>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 4.0, float 8.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_splat(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_splat(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], <float 3.000000e+00, float 3.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 3.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_vec(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], <float 4.000000e+00, float 3.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 4.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_vec_arcp(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_vec_arcp(
; CHECK-NEXT: [[DIV:%.*]] = fmul arcp <2 x float> [[X:%.*]], <float 2.500000e-01, float 0x3FD5555560000000>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv arcp <2 x float> %x, <float 4.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_vec_arcp_with_undef_elt(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_vec_arcp_with_undef_elt(
; CHECK-NEXT: [[DIV:%.*]] = fdiv arcp <2 x float> [[X:%.*]], <float undef, float 3.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv arcp <2 x float> %x, <float undef, float 3.0>
ret <2 x float> %div
}
; (X / Y) / Z --> X / (Y * Z)
define float @div_with_div_numerator(float %x, float %y, float %z) {
; CHECK-LABEL: @div_with_div_numerator(
; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc arcp float [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv reassoc arcp float [[X:%.*]], [[TMP1]]
; CHECK-NEXT: ret float [[DIV2]]
;
%div1 = fdiv ninf float %x, %y
%div2 = fdiv arcp reassoc float %div1, %z
ret float %div2
}
; Z / (X / Y) --> (Z * Y) / X
define <2 x float> @div_with_div_denominator(<2 x float> %x, <2 x float> %y, <2 x float> %z) {
; CHECK-LABEL: @div_with_div_denominator(
; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc arcp <2 x float> [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv reassoc arcp <2 x float> [[TMP1]], [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV2]]
;
%div1 = fdiv nnan <2 x float> %x, %y
%div2 = fdiv arcp reassoc <2 x float> %z, %div1
ret <2 x float> %div2
}
; Don't create an extra multiply if we can't eliminate the first div.
declare void @use_f32(float)
define float @div_with_div_numerator_extra_use(float %x, float %y, float %z) {
; CHECK-LABEL: @div_with_div_numerator_extra_use(
; CHECK-NEXT: [[DIV1:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv fast float [[DIV1]], [[Z:%.*]]
; CHECK-NEXT: call void @use_f32(float [[DIV1]])
; CHECK-NEXT: ret float [[DIV2]]
;
%div1 = fdiv float %x, %y
%div2 = fdiv fast float %div1, %z
call void @use_f32(float %div1)
ret float %div2
}
define float @div_with_div_denominator_extra_use(float %x, float %y, float %z) {
; CHECK-LABEL: @div_with_div_denominator_extra_use(
; CHECK-NEXT: [[DIV1:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv fast float [[Z:%.*]], [[DIV1]]
; CHECK-NEXT: call void @use_f32(float [[DIV1]])
; CHECK-NEXT: ret float [[DIV2]]
;
%div1 = fdiv float %x, %y
%div2 = fdiv fast float %z, %div1
call void @use_f32(float %div1)
ret float %div2
}
define float @fneg_fneg(float %x, float %y) {
; CHECK-LABEL: @fneg_fneg(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fsub float -0.0, %x
%y.fneg = fsub float -0.0, %y
%div = fdiv float %x.fneg, %y.fneg
ret float %div
}
; The test above shows that no FMF are needed, but show that we are not dropping FMF.
define float @fneg_fneg_fast(float %x, float %y) {
; CHECK-LABEL: @fneg_fneg_fast(
; CHECK-NEXT: [[DIV:%.*]] = fdiv fast float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fsub float -0.0, %x
%y.fneg = fsub float -0.0, %y
%div = fdiv fast float %x.fneg, %y.fneg
ret float %div
}
define <2 x float> @fneg_fneg_vec(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @fneg_fneg_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%xneg = fsub <2 x float> <float -0.0, float -0.0>, %x
%yneg = fsub <2 x float> <float -0.0, float -0.0>, %y
%div = fdiv <2 x float> %xneg, %yneg
ret <2 x float> %div
}
define <2 x float> @fneg_fneg_vec_undef_elts(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @fneg_fneg_vec_undef_elts(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%xneg = fsub <2 x float> <float undef, float -0.0>, %x
%yneg = fsub <2 x float> <float -0.0, float undef>, %y
%div = fdiv <2 x float> %xneg, %yneg
ret <2 x float> %div
}
define float @fneg_dividend_constant_divisor(float %x) {
; CHECK-LABEL: @fneg_dividend_constant_divisor(
; CHECK-NEXT: [[DIV:%.*]] = fdiv nsz float [[X:%.*]], -3.000000e+00
; CHECK-NEXT: ret float [[DIV]]
;
%neg = fsub float -0.0, %x
%div = fdiv nsz float %neg, 3.0
ret float %div
}
define float @fneg_divisor_constant_dividend(float %x) {
; CHECK-LABEL: @fneg_divisor_constant_dividend(
; CHECK-NEXT: [[DIV:%.*]] = fdiv nnan float 3.000000e+00, [[X:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%neg = fsub float -0.0, %x
%div = fdiv nnan float -3.0, %neg
ret float %div
}
define <2 x float> @fneg_dividend_constant_divisor_vec(<2 x float> %x) {
; CHECK-LABEL: @fneg_dividend_constant_divisor_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv ninf <2 x float> [[X:%.*]], <float -3.000000e+00, float 8.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fsub <2 x float> <float -0.0, float -0.0>, %x
%div = fdiv ninf <2 x float> %neg, <float 3.0, float -8.0>
ret <2 x float> %div
}
define <2 x float> @fneg_dividend_constant_divisor_vec_undef_elt(<2 x float> %x) {
; CHECK-LABEL: @fneg_dividend_constant_divisor_vec_undef_elt(
; CHECK-NEXT: [[DIV:%.*]] = fdiv ninf <2 x float> [[X:%.*]], <float -3.000000e+00, float 8.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fsub <2 x float> <float undef, float -0.0>, %x
%div = fdiv ninf <2 x float> %neg, <float 3.0, float -8.0>
ret <2 x float> %div
}
define <2 x float> @fneg_divisor_constant_dividend_vec(<2 x float> %x) {
; CHECK-LABEL: @fneg_divisor_constant_dividend_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv afn <2 x float> <float 3.000000e+00, float -5.000000e+00>, [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fsub <2 x float> <float -0.0, float -0.0>, %x
%div = fdiv afn <2 x float> <float -3.0, float 5.0>, %neg
ret <2 x float> %div
}
; X / (X * Y) --> 1.0 / Y
define float @div_factor(float %x, float %y) {
; CHECK-LABEL: @div_factor(
; CHECK-NEXT: [[D:%.*]] = fdiv reassoc nnan float 1.000000e+00, [[Y:%.*]]
; CHECK-NEXT: ret float [[D]]
;
%m = fmul float %x, %y
%d = fdiv nnan reassoc float %x, %m
ret float %d;
}
; We can't do the transform without 'nnan' because if x is NAN and y is a number, this should return NAN.
define float @div_factor_too_strict(float %x, float %y) {
; CHECK-LABEL: @div_factor_too_strict(
; CHECK-NEXT: [[M:%.*]] = fmul float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[D:%.*]] = fdiv reassoc float [[X]], [[M]]
; CHECK-NEXT: ret float [[D]]
;
%m = fmul float %x, %y
%d = fdiv reassoc float %x, %m
ret float %d
}
; Commute, verify vector types, and show that we are not dropping extra FMF.
; X / (Y * X) --> 1.0 / Y
define <2 x float> @div_factor_commute(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @div_factor_commute(
; CHECK-NEXT: [[D:%.*]] = fdiv reassoc nnan ninf nsz <2 x float> <float 1.000000e+00, float 1.000000e+00>, [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[D]]
;
%m = fmul <2 x float> %y, %x
%d = fdiv nnan ninf nsz reassoc <2 x float> %x, %m
ret <2 x float> %d
}
; C1/(X*C2) => (C1/C2) / X
define <2 x float> @div_constant_dividend1(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend1(
; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc arcp <2 x float> <float 5.000000e+00, float 1.000000e+00>, [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fmul <2 x float> %x, <float 3.0e0, float 7.0e0>
%t2 = fdiv arcp reassoc <2 x float> <float 15.0e0, float 7.0e0>, %t1
ret <2 x float> %t2
}
define <2 x float> @div_constant_dividend1_arcp_only(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend1_arcp_only(
; CHECK-NEXT: [[T1:%.*]] = fmul <2 x float> [[X:%.*]], <float 3.000000e+00, float 7.000000e+00>
; CHECK-NEXT: [[T2:%.*]] = fdiv arcp <2 x float> <float 1.500000e+01, float 7.000000e+00>, [[T1]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fmul <2 x float> %x, <float 3.0e0, float 7.0e0>
%t2 = fdiv arcp <2 x float> <float 15.0e0, float 7.0e0>, %t1
ret <2 x float> %t2
}
; C1/(X/C2) => (C1*C2) / X
define <2 x float> @div_constant_dividend2(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend2(
; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc arcp <2 x float> <float 4.500000e+01, float 4.900000e+01>, [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fdiv <2 x float> %x, <float 3.0e0, float -7.0e0>
%t2 = fdiv arcp reassoc <2 x float> <float 15.0e0, float -7.0e0>, %t1
ret <2 x float> %t2
}
define <2 x float> @div_constant_dividend2_reassoc_only(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend2_reassoc_only(
; CHECK-NEXT: [[T1:%.*]] = fdiv <2 x float> [[X:%.*]], <float 3.000000e+00, float -7.000000e+00>
; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc <2 x float> <float 1.500000e+01, float -7.000000e+00>, [[T1]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fdiv <2 x float> %x, <float 3.0e0, float -7.0e0>
%t2 = fdiv reassoc <2 x float> <float 15.0e0, float -7.0e0>, %t1
ret <2 x float> %t2
}
; C1/(C2/X) => (C1/C2) * X
; This tests the combination of 2 folds: (C1 * X) / C2 --> (C1 / C2) * X
define <2 x float> @div_constant_dividend3(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend3(
; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc arcp <2 x float> [[X:%.*]], <float 1.500000e+01, float -7.000000e+00>
; CHECK-NEXT: [[T2:%.*]] = fmul reassoc arcp <2 x float> [[TMP1]], <float 0x3FD5555560000000, float 0x3FC24924A0000000>
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fdiv <2 x float> <float 3.0e0, float 7.0e0>, %x
%t2 = fdiv arcp reassoc <2 x float> <float 15.0e0, float -7.0e0>, %t1
ret <2 x float> %t2
}
|
