1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt < %s -instcombine -S | FileCheck %s
4 ; testing-case "float fold(float a) { return 1.2f * a * 2.3f; }"
5 ; 1.2f and 2.3f is supposed to be fold.
6 define float @fold(float %a) {
8 ; CHECK-NEXT: [[MUL1:%.*]] = fmul fast float [[A:%.*]], 0x4006147AE0000000
9 ; CHECK-NEXT: ret float [[MUL1]]
11 %mul = fmul fast float %a, 0x3FF3333340000000
12 %mul1 = fmul fast float %mul, 0x4002666660000000
16 ; Same testing-case as the one used in fold() except that the operators have
18 define float @notfold(float %a) {
19 ; CHECK-LABEL: @notfold(
20 ; CHECK-NEXT: [[MUL:%.*]] = fmul fast float [[A:%.*]], 0x3FF3333340000000
21 ; CHECK-NEXT: [[MUL1:%.*]] = fmul float [[MUL]], 0x4002666660000000
22 ; CHECK-NEXT: ret float [[MUL1]]
24 %mul = fmul fast float %a, 0x3FF3333340000000
25 %mul1 = fmul float %mul, 0x4002666660000000
29 define float @fold2(float %a) {
30 ; CHECK-LABEL: @fold2(
31 ; CHECK-NEXT: [[MUL1:%.*]] = fmul fast float [[A:%.*]], 0x4006147AE0000000
32 ; CHECK-NEXT: ret float [[MUL1]]
34 %mul = fmul float %a, 0x3FF3333340000000
35 %mul1 = fmul fast float %mul, 0x4002666660000000
39 ; C * f1 + f1 = (C+1) * f1
40 ; TODO: The particular case where C is 2 (so the folded result is 3.0*f1) is
41 ; always safe, and so doesn't need any FMF.
42 ; That is, (x + x + x) and (3*x) each have only a single rounding.
43 define double @fold3(double %f1) {
44 ; CHECK-LABEL: @fold3(
45 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[F1:%.*]], 6.000000e+00
46 ; CHECK-NEXT: ret double [[TMP1]]
48 %t1 = fmul fast double 5.000000e+00, %f1
49 %t2 = fadd fast double %f1, %t1
53 ; Check again with 'reassoc' and 'nsz' ('nsz' not technically required).
54 define double @fold3_reassoc_nsz(double %f1) {
55 ; CHECK-LABEL: @fold3_reassoc_nsz(
56 ; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz double [[F1:%.*]], 6.000000e+00
57 ; CHECK-NEXT: ret double [[TMP1]]
59 %t1 = fmul reassoc nsz double 5.000000e+00, %f1
60 %t2 = fadd reassoc nsz double %f1, %t1
64 ; TODO: This doesn't require 'nsz'. It should fold to f1 * 6.0.
65 define double @fold3_reassoc(double %f1) {
66 ; CHECK-LABEL: @fold3_reassoc(
67 ; CHECK-NEXT: [[T1:%.*]] = fmul reassoc double [[F1:%.*]], 5.000000e+00
68 ; CHECK-NEXT: [[T2:%.*]] = fadd reassoc double [[T1]], [[F1]]
69 ; CHECK-NEXT: ret double [[T2]]
71 %t1 = fmul reassoc double 5.000000e+00, %f1
72 %t2 = fadd reassoc double %f1, %t1
76 ; (C1 - X) + (C2 - Y) => (C1+C2) - (X + Y)
77 define float @fold4(float %f1, float %f2) {
78 ; CHECK-LABEL: @fold4(
79 ; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[F1:%.*]], [[F2:%.*]]
80 ; CHECK-NEXT: [[TMP2:%.*]] = fsub fast float 9.000000e+00, [[TMP1]]
81 ; CHECK-NEXT: ret float [[TMP2]]
83 %sub = fsub float 4.000000e+00, %f1
84 %sub1 = fsub float 5.000000e+00, %f2
85 %add = fadd fast float %sub, %sub1
89 ; Check again with 'reassoc' and 'nsz' ('nsz' not technically required).
90 define float @fold4_reassoc_nsz(float %f1, float %f2) {
91 ; CHECK-LABEL: @fold4_reassoc_nsz(
92 ; CHECK-NEXT: [[TMP1:%.*]] = fadd reassoc nsz float [[F1:%.*]], [[F2:%.*]]
93 ; CHECK-NEXT: [[TMP2:%.*]] = fsub reassoc nsz float 9.000000e+00, [[TMP1]]
94 ; CHECK-NEXT: ret float [[TMP2]]
96 %sub = fsub float 4.000000e+00, %f1
97 %sub1 = fsub float 5.000000e+00, %f2
98 %add = fadd reassoc nsz float %sub, %sub1
102 ; TODO: This doesn't require 'nsz'. It should fold to (9.0 - (f1 + f2)).
103 define float @fold4_reassoc(float %f1, float %f2) {
104 ; CHECK-LABEL: @fold4_reassoc(
105 ; CHECK-NEXT: [[SUB:%.*]] = fsub float 4.000000e+00, [[F1:%.*]]
106 ; CHECK-NEXT: [[SUB1:%.*]] = fsub float 5.000000e+00, [[F2:%.*]]
107 ; CHECK-NEXT: [[ADD:%.*]] = fadd reassoc float [[SUB]], [[SUB1]]
108 ; CHECK-NEXT: ret float [[ADD]]
110 %sub = fsub float 4.000000e+00, %f1
111 %sub1 = fsub float 5.000000e+00, %f2
112 %add = fadd reassoc float %sub, %sub1
116 ; (X + C1) + C2 => X + (C1 + C2)
117 define float @fold5(float %f1) {
118 ; CHECK-LABEL: @fold5(
119 ; CHECK-NEXT: [[ADD1:%.*]] = fadd fast float [[F1:%.*]], 9.000000e+00
120 ; CHECK-NEXT: ret float [[ADD1]]
122 %add = fadd float %f1, 4.000000e+00
123 %add1 = fadd fast float %add, 5.000000e+00
127 ; Check again with 'reassoc' and 'nsz' ('nsz' not technically required).
128 define float @fold5_reassoc_nsz(float %f1) {
129 ; CHECK-LABEL: @fold5_reassoc_nsz(
130 ; CHECK-NEXT: [[ADD1:%.*]] = fadd reassoc nsz float [[F1:%.*]], 9.000000e+00
131 ; CHECK-NEXT: ret float [[ADD1]]
133 %add = fadd float %f1, 4.000000e+00
134 %add1 = fadd reassoc nsz float %add, 5.000000e+00
138 ; TODO: This doesn't require 'nsz'. It should fold to f1 + 9.0
139 define float @fold5_reassoc(float %f1) {
140 ; CHECK-LABEL: @fold5_reassoc(
141 ; CHECK-NEXT: [[ADD:%.*]] = fadd float [[F1:%.*]], 4.000000e+00
142 ; CHECK-NEXT: [[ADD1:%.*]] = fadd reassoc float [[ADD]], 5.000000e+00
143 ; CHECK-NEXT: ret float [[ADD1]]
145 %add = fadd float %f1, 4.000000e+00
146 %add1 = fadd reassoc float %add, 5.000000e+00
150 ; (X + X) + X + X => 4.0 * X
151 define float @fold6(float %f1) {
152 ; CHECK-LABEL: @fold6(
153 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast float [[F1:%.*]], 4.000000e+00
154 ; CHECK-NEXT: ret float [[TMP1]]
156 %t1 = fadd fast float %f1, %f1
157 %t2 = fadd fast float %f1, %t1
158 %t3 = fadd fast float %t2, %f1
162 ; Check again with 'reassoc' and 'nsz' ('nsz' not technically required).
163 define float @fold6_reassoc_nsz(float %f1) {
164 ; CHECK-LABEL: @fold6_reassoc_nsz(
165 ; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz float [[F1:%.*]], 4.000000e+00
166 ; CHECK-NEXT: ret float [[TMP1]]
168 %t1 = fadd reassoc nsz float %f1, %f1
169 %t2 = fadd reassoc nsz float %f1, %t1
170 %t3 = fadd reassoc nsz float %t2, %f1
174 ; TODO: This doesn't require 'nsz'. It should fold to f1 * 4.0.
175 define float @fold6_reassoc(float %f1) {
176 ; CHECK-LABEL: @fold6_reassoc(
177 ; CHECK-NEXT: [[T1:%.*]] = fadd reassoc float [[F1:%.*]], [[F1]]
178 ; CHECK-NEXT: [[T2:%.*]] = fadd reassoc float [[T1]], [[F1]]
179 ; CHECK-NEXT: [[T3:%.*]] = fadd reassoc float [[T2]], [[F1]]
180 ; CHECK-NEXT: ret float [[T3]]
182 %t1 = fadd reassoc float %f1, %f1
183 %t2 = fadd reassoc float %f1, %t1
184 %t3 = fadd reassoc float %t2, %f1
188 ; C1 * X + (X + X) = (C1 + 2) * X
189 define float @fold7(float %f1) {
190 ; CHECK-LABEL: @fold7(
191 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast float [[F1:%.*]], 7.000000e+00
192 ; CHECK-NEXT: ret float [[TMP1]]
194 %t1 = fmul fast float %f1, 5.000000e+00
195 %t2 = fadd fast float %f1, %f1
196 %t3 = fadd fast float %t1, %t2
200 ; Check again with 'reassoc' and 'nsz' ('nsz' not technically required).
201 define float @fold7_reassoc_nsz(float %f1) {
202 ; CHECK-LABEL: @fold7_reassoc_nsz(
203 ; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz float [[F1:%.*]], 7.000000e+00
204 ; CHECK-NEXT: ret float [[TMP1]]
206 %t1 = fmul reassoc nsz float %f1, 5.000000e+00
207 %t2 = fadd reassoc nsz float %f1, %f1
208 %t3 = fadd reassoc nsz float %t1, %t2
212 ; TODO: This doesn't require 'nsz'. It should fold to f1 * 7.0.
213 define float @fold7_reassoc(float %f1) {
214 ; CHECK-LABEL: @fold7_reassoc(
215 ; CHECK-NEXT: [[T1:%.*]] = fmul reassoc float [[F1:%.*]], 5.000000e+00
216 ; CHECK-NEXT: [[T2:%.*]] = fadd reassoc float [[F1]], [[F1]]
217 ; CHECK-NEXT: [[T3:%.*]] = fadd reassoc float [[T1]], [[T2]]
218 ; CHECK-NEXT: ret float [[T3]]
220 %t1 = fmul reassoc float %f1, 5.000000e+00
221 %t2 = fadd reassoc float %f1, %f1
222 %t3 = fadd reassoc float %t1, %t2
226 ; (X + X) + (X + X) + X => 5.0 * X
227 define float @fold8(float %f1) {
228 ; CHECK-LABEL: @fold8(
229 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast float [[F1:%.*]], 5.000000e+00
230 ; CHECK-NEXT: ret float [[TMP1]]
232 %t1 = fadd fast float %f1, %f1
233 %t2 = fadd fast float %f1, %f1
234 %t3 = fadd fast float %t1, %t2
235 %t4 = fadd fast float %t3, %f1
239 ; Check again with 'reassoc' and 'nsz' ('nsz' not technically required).
240 define float @fold8_reassoc_nsz(float %f1) {
241 ; CHECK-LABEL: @fold8_reassoc_nsz(
242 ; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz float [[F1:%.*]], 5.000000e+00
243 ; CHECK-NEXT: ret float [[TMP1]]
245 %t1 = fadd reassoc nsz float %f1, %f1
246 %t2 = fadd reassoc nsz float %f1, %f1
247 %t3 = fadd reassoc nsz float %t1, %t2
248 %t4 = fadd reassoc nsz float %t3, %f1
252 ; TODO: This doesn't require 'nsz'. It should fold to f1 * 5.0.
253 define float @fold8_reassoc(float %f1) {
254 ; CHECK-LABEL: @fold8_reassoc(
255 ; CHECK-NEXT: [[T1:%.*]] = fadd reassoc float [[F1:%.*]], [[F1]]
256 ; CHECK-NEXT: [[T2:%.*]] = fadd reassoc float [[F1]], [[F1]]
257 ; CHECK-NEXT: [[T3:%.*]] = fadd reassoc float [[T1]], [[T2]]
258 ; CHECK-NEXT: [[T4:%.*]] = fadd reassoc float [[T3]], [[F1]]
259 ; CHECK-NEXT: ret float [[T4]]
261 %t1 = fadd reassoc float %f1, %f1
262 %t2 = fadd reassoc float %f1, %f1
263 %t3 = fadd reassoc float %t1, %t2
264 %t4 = fadd reassoc float %t3, %f1
270 define float @fsub_fadd_common_op_fneg(float %x, float %y) {
271 ; CHECK-LABEL: @fsub_fadd_common_op_fneg(
272 ; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float -0.000000e+00, [[X:%.*]]
273 ; CHECK-NEXT: ret float [[TMP1]]
275 %a = fadd float %x, %y
276 %r = fsub fast float %y, %a
281 ; Check again with 'reassoc' and 'nsz'.
282 ; nsz is required because: 0.0 - (0.0 + 0.0) -> 0.0, not -0.0
284 define float @fsub_fadd_common_op_fneg_reassoc_nsz(float %x, float %y) {
285 ; CHECK-LABEL: @fsub_fadd_common_op_fneg_reassoc_nsz(
286 ; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float -0.000000e+00, [[X:%.*]]
287 ; CHECK-NEXT: ret float [[TMP1]]
289 %a = fadd float %x, %y
290 %r = fsub reassoc nsz float %y, %a
296 define <2 x float> @fsub_fadd_common_op_fneg_vec(<2 x float> %x, <2 x float> %y) {
297 ; CHECK-LABEL: @fsub_fadd_common_op_fneg_vec(
298 ; CHECK-NEXT: [[A:%.*]] = fadd <2 x float> [[X:%.*]], [[Y:%.*]]
299 ; CHECK-NEXT: [[R:%.*]] = fsub reassoc nsz <2 x float> [[Y]], [[A]]
300 ; CHECK-NEXT: ret <2 x float> [[R]]
302 %a = fadd <2 x float> %x, %y
303 %r = fsub nsz reassoc <2 x float> %y, %a
308 ; Commute operands of the 'add'.
310 define float @fsub_fadd_common_op_fneg_commute(float %x, float %y) {
311 ; CHECK-LABEL: @fsub_fadd_common_op_fneg_commute(
312 ; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float -0.000000e+00, [[X:%.*]]
313 ; CHECK-NEXT: ret float [[TMP1]]
315 %a = fadd float %y, %x
316 %r = fsub reassoc nsz float %y, %a
322 define <2 x float> @fsub_fadd_common_op_fneg_commute_vec(<2 x float> %x, <2 x float> %y) {
323 ; CHECK-LABEL: @fsub_fadd_common_op_fneg_commute_vec(
324 ; CHECK-NEXT: [[A:%.*]] = fadd <2 x float> [[Y:%.*]], [[X:%.*]]
325 ; CHECK-NEXT: [[R:%.*]] = fsub reassoc nsz <2 x float> [[Y]], [[A]]
326 ; CHECK-NEXT: ret <2 x float> [[R]]
328 %a = fadd <2 x float> %y, %x
329 %r = fsub reassoc nsz <2 x float> %y, %a
334 ; nsz is required because: (0.0 - 0.0) - 0.0 -> 0.0, not -0.0
336 define float @fsub_fsub_common_op_fneg(float %x, float %y) {
337 ; CHECK-LABEL: @fsub_fsub_common_op_fneg(
338 ; CHECK-NEXT: [[R:%.*]] = fsub reassoc nsz float -0.000000e+00, [[X:%.*]]
339 ; CHECK-NEXT: ret float [[R]]
341 %s = fsub float %y, %x
342 %r = fsub reassoc nsz float %s, %y
348 define <2 x float> @fsub_fsub_common_op_fneg_vec(<2 x float> %x, <2 x float> %y) {
349 ; CHECK-LABEL: @fsub_fsub_common_op_fneg_vec(
350 ; CHECK-NEXT: [[R:%.*]] = fsub reassoc nsz <2 x float> <float -0.000000e+00, float -0.000000e+00>, [[X:%.*]]
351 ; CHECK-NEXT: ret <2 x float> [[R]]
353 %s = fsub <2 x float> %y, %x
354 %r = fsub reassoc nsz <2 x float> %s, %y
358 ; TODO: This doesn't require 'nsz'. It should fold to 0 - f2
359 define float @fold9_reassoc(float %f1, float %f2) {
360 ; CHECK-LABEL: @fold9_reassoc(
361 ; CHECK-NEXT: [[T1:%.*]] = fadd float [[F1:%.*]], [[F2:%.*]]
362 ; CHECK-NEXT: [[T3:%.*]] = fsub reassoc float [[F1]], [[T1]]
363 ; CHECK-NEXT: ret float [[T3]]
365 %t1 = fadd float %f1, %f2
366 %t3 = fsub reassoc float %f1, %t1
370 ; Let C3 = C1 + C2. (f1 + C1) + (f2 + C2) => (f1 + f2) + C3 instead of
371 ; "(f1 + C3) + f2" or "(f2 + C3) + f1". Placing constant-addend at the
372 ; top of resulting simplified expression tree may potentially reveal some
373 ; optimization opportunities in the super-expression trees.
375 define float @fold10(float %f1, float %f2) {
376 ; CHECK-LABEL: @fold10(
377 ; CHECK-NEXT: [[T2:%.*]] = fadd fast float [[F1:%.*]], [[F2:%.*]]
378 ; CHECK-NEXT: [[T3:%.*]] = fadd fast float [[T2]], -1.000000e+00
379 ; CHECK-NEXT: ret float [[T3]]
381 %t1 = fadd fast float 2.000000e+00, %f1
382 %t2 = fsub fast float %f2, 3.000000e+00
383 %t3 = fadd fast float %t1, %t2
387 ; Check again with 'reassoc' and 'nsz'.
388 ; TODO: We may be able to remove the 'nsz' requirement.
389 define float @fold10_reassoc_nsz(float %f1, float %f2) {
390 ; CHECK-LABEL: @fold10_reassoc_nsz(
391 ; CHECK-NEXT: [[T2:%.*]] = fadd reassoc nsz float [[F1:%.*]], [[F2:%.*]]
392 ; CHECK-NEXT: [[T3:%.*]] = fadd reassoc nsz float [[T2]], -1.000000e+00
393 ; CHECK-NEXT: ret float [[T3]]
395 %t1 = fadd reassoc nsz float 2.000000e+00, %f1
396 %t2 = fsub reassoc nsz float %f2, 3.000000e+00
397 %t3 = fadd reassoc nsz float %t1, %t2
401 ; Observe that the fold is not done with only reassoc (the instructions are
402 ; canonicalized, but not folded).
403 ; TODO: As noted above, 'nsz' may not be required for this to be fully folded.
404 define float @fold10_reassoc(float %f1, float %f2) {
405 ; CHECK-LABEL: @fold10_reassoc(
406 ; CHECK-NEXT: [[T1:%.*]] = fadd reassoc float [[F1:%.*]], 2.000000e+00
407 ; CHECK-NEXT: [[T2:%.*]] = fadd reassoc float [[F2:%.*]], -3.000000e+00
408 ; CHECK-NEXT: [[T3:%.*]] = fadd reassoc float [[T1]], [[T2]]
409 ; CHECK-NEXT: ret float [[T3]]
411 %t1 = fadd reassoc float 2.000000e+00, %f1
412 %t2 = fsub reassoc float %f2, 3.000000e+00
413 %t3 = fadd reassoc float %t1, %t2
417 ; This used to crash/miscompile.
419 define float @fail1(float %f1, float %f2) {
420 ; CHECK-LABEL: @fail1(
421 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast float [[F1:%.*]], 3.000000e+00
422 ; CHECK-NEXT: [[TMP2:%.*]] = fadd fast float [[TMP1]], -3.000000e+00
423 ; CHECK-NEXT: ret float [[TMP2]]
425 %conv3 = fadd fast float %f1, -1.000000e+00
426 %add = fadd fast float %conv3, %conv3
427 %add2 = fadd fast float %add, %conv3
431 define double @fail2(double %f1, double %f2) {
432 ; CHECK-LABEL: @fail2(
433 ; CHECK-NEXT: [[TMP1:%.*]] = fadd fast double [[F2:%.*]], [[F2]]
434 ; CHECK-NEXT: [[TMP2:%.*]] = fsub fast double -0.000000e+00, [[TMP1]]
435 ; CHECK-NEXT: ret double [[TMP2]]
437 %t1 = fsub fast double %f1, %f2
438 %t2 = fadd fast double %f1, %f2
439 %t3 = fsub fast double %t1, %t2
443 ; c1 * x - x => (c1 - 1.0) * x
444 define float @fold13(float %x) {
445 ; CHECK-LABEL: @fold13(
446 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast float [[X:%.*]], 6.000000e+00
447 ; CHECK-NEXT: ret float [[TMP1]]
449 %mul = fmul fast float %x, 7.000000e+00
450 %sub = fsub fast float %mul, %x
454 ; Check again using the minimal subset of FMF.
455 define float @fold13_reassoc_nsz(float %x) {
456 ; CHECK-LABEL: @fold13_reassoc_nsz(
457 ; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz float [[X:%.*]], 6.000000e+00
458 ; CHECK-NEXT: ret float [[TMP1]]
460 %mul = fmul reassoc nsz float %x, 7.000000e+00
461 %sub = fsub reassoc nsz float %mul, %x
465 ; Verify the fold is not done with only 'reassoc' ('nsz' is required).
466 define float @fold13_reassoc(float %x) {
467 ; CHECK-LABEL: @fold13_reassoc(
468 ; CHECK-NEXT: [[MUL:%.*]] = fmul reassoc float [[X:%.*]], 7.000000e+00
469 ; CHECK-NEXT: [[SUB:%.*]] = fsub reassoc float [[MUL]], [[X]]
470 ; CHECK-NEXT: ret float [[SUB]]
472 %mul = fmul reassoc float %x, 7.000000e+00
473 %sub = fsub reassoc float %mul, %x
477 ; (select X+Y, X-Y) => X + (select Y, -Y)
478 ; This is always safe. No FMF required.
479 define float @fold16(float %x, float %y) {
480 ; CHECK-LABEL: @fold16(
481 ; CHECK-NEXT: [[CMP:%.*]] = fcmp ogt float [[X:%.*]], [[Y:%.*]]
482 ; CHECK-NEXT: [[TMP1:%.*]] = fsub float -0.000000e+00, [[Y]]
483 ; CHECK-NEXT: [[R_P:%.*]] = select i1 [[CMP]], float [[Y]], float [[TMP1]]
484 ; CHECK-NEXT: [[R:%.*]] = fadd float [[R_P]], [[X]]
485 ; CHECK-NEXT: ret float [[R]]
487 %cmp = fcmp ogt float %x, %y
488 %plus = fadd float %x, %y
489 %minus = fsub float %x, %y
490 %r = select i1 %cmp, float %plus, float %minus
494 ; =========================================================================
496 ; Testing-cases about negation
498 ; =========================================================================
499 define float @fneg1(float %f1, float %f2) {
500 ; CHECK-LABEL: @fneg1(
501 ; CHECK-NEXT: [[MUL:%.*]] = fmul float [[F1:%.*]], [[F2:%.*]]
502 ; CHECK-NEXT: ret float [[MUL]]
504 %sub = fsub float -0.000000e+00, %f1
505 %sub1 = fsub nsz float 0.000000e+00, %f2
506 %mul = fmul float %sub, %sub1
510 define float @fneg2(float %x) {
511 ; CHECK-LABEL: @fneg2(
512 ; CHECK-NEXT: [[SUB:%.*]] = fsub nsz float -0.000000e+00, [[X:%.*]]
513 ; CHECK-NEXT: ret float [[SUB]]
515 %sub = fsub nsz float 0.0, %x
519 define <2 x float> @fneg2_vec_undef(<2 x float> %x) {
520 ; CHECK-LABEL: @fneg2_vec_undef(
521 ; CHECK-NEXT: [[SUB:%.*]] = fsub nsz <2 x float> <float -0.000000e+00, float -0.000000e+00>, [[X:%.*]]
522 ; CHECK-NEXT: ret <2 x float> [[SUB]]
524 %sub = fsub nsz <2 x float> <float undef, float 0.0>, %x
528 ; =========================================================================
530 ; Testing-cases about div
532 ; =========================================================================
534 ; X/C1 / C2 => X * (1/(C2*C1))
535 define float @fdiv1(float %x) {
536 ; CHECK-LABEL: @fdiv1(
537 ; CHECK-NEXT: [[DIV1:%.*]] = fmul fast float [[X:%.*]], 0x3FD7303B60000000
538 ; CHECK-NEXT: ret float [[DIV1]]
540 %div = fdiv float %x, 0x3FF3333340000000
541 %div1 = fdiv fast float %div, 0x4002666660000000
543 ; 0x3FF3333340000000 = 1.2f
544 ; 0x4002666660000000 = 2.3f
545 ; 0x3FD7303B60000000 = 0.36231884057971014492
548 ; X*C1 / C2 => X * (C1/C2)
549 define float @fdiv2(float %x) {
550 ; CHECK-LABEL: @fdiv2(
551 ; CHECK-NEXT: [[DIV1:%.*]] = fmul fast float [[X:%.*]], 0x3FE0B21660000000
552 ; CHECK-NEXT: ret float [[DIV1]]
554 %mul = fmul float %x, 0x3FF3333340000000
555 %div1 = fdiv fast float %mul, 0x4002666660000000
558 ; 0x3FF3333340000000 = 1.2f
559 ; 0x4002666660000000 = 2.3f
560 ; 0x3FE0B21660000000 = 0.52173918485641479492
563 define <2 x float> @fdiv2_vec(<2 x float> %x) {
564 ; CHECK-LABEL: @fdiv2_vec(
565 ; CHECK-NEXT: [[DIV1:%.*]] = fmul fast <2 x float> [[X:%.*]], <float 3.000000e+00, float 3.000000e+00>
566 ; CHECK-NEXT: ret <2 x float> [[DIV1]]
568 %mul = fmul <2 x float> %x, <float 6.0, float 9.0>
569 %div1 = fdiv fast <2 x float> %mul, <float 2.0, float 3.0>
570 ret <2 x float> %div1
573 ; "X/C1 / C2 => X * (1/(C2*C1))" is disabled (for now) is C2/C1 is a denormal
575 define float @fdiv3(float %x) {
576 ; CHECK-LABEL: @fdiv3(
577 ; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], 0x47EFFFFFE0000000
578 ; CHECK-NEXT: [[DIV1:%.*]] = fmul fast float [[DIV]], 0x3FDBD37A80000000
579 ; CHECK-NEXT: ret float [[DIV1]]
581 %div = fdiv float %x, 0x47EFFFFFE0000000
582 %div1 = fdiv fast float %div, 0x4002666660000000
586 ; "X*C1 / C2 => X * (C1/C2)" is disabled if C1/C2 is a denormal
587 define float @fdiv4(float %x) {
588 ; CHECK-LABEL: @fdiv4(
589 ; CHECK-NEXT: [[MUL:%.*]] = fmul float [[X:%.*]], 0x47EFFFFFE0000000
590 ; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[MUL]], 0x3FC99999A0000000
591 ; CHECK-NEXT: ret float [[DIV]]
593 %mul = fmul float %x, 0x47EFFFFFE0000000
594 %div = fdiv float %mul, 0x3FC99999A0000000
598 ; =========================================================================
600 ; Testing-cases about factorization
602 ; =========================================================================
603 ; x*z + y*z => (x+y) * z
604 define float @fact_mul1(float %x, float %y, float %z) {
605 ; CHECK-LABEL: @fact_mul1(
606 ; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[X:%.*]], [[Y:%.*]]
607 ; CHECK-NEXT: [[TMP2:%.*]] = fmul fast float [[TMP1]], [[Z:%.*]]
608 ; CHECK-NEXT: ret float [[TMP2]]
610 %t1 = fmul fast float %x, %z
611 %t2 = fmul fast float %y, %z
612 %t3 = fadd fast float %t1, %t2
616 ; Check again using the minimal subset of FMF.
617 define float @fact_mul1_reassoc_nsz(float %x, float %y, float %z) {
618 ; CHECK-LABEL: @fact_mul1_reassoc_nsz(
619 ; CHECK-NEXT: [[TMP1:%.*]] = fadd reassoc nsz float [[X:%.*]], [[Y:%.*]]
620 ; CHECK-NEXT: [[TMP2:%.*]] = fmul reassoc nsz float [[TMP1]], [[Z:%.*]]
621 ; CHECK-NEXT: ret float [[TMP2]]
623 %t1 = fmul reassoc nsz float %x, %z
624 %t2 = fmul reassoc nsz float %y, %z
625 %t3 = fadd reassoc nsz float %t1, %t2
629 ; Verify the fold is not done with only 'reassoc' ('nsz' is required).
630 define float @fact_mul1_reassoc(float %x, float %y, float %z) {
631 ; CHECK-LABEL: @fact_mul1_reassoc(
632 ; CHECK-NEXT: [[T1:%.*]] = fmul reassoc float [[X:%.*]], [[Z:%.*]]
633 ; CHECK-NEXT: [[T2:%.*]] = fmul reassoc float [[Y:%.*]], [[Z]]
634 ; CHECK-NEXT: [[T3:%.*]] = fadd reassoc float [[T1]], [[T2]]
635 ; CHECK-NEXT: ret float [[T3]]
637 %t1 = fmul reassoc float %x, %z
638 %t2 = fmul reassoc float %y, %z
639 %t3 = fadd reassoc float %t1, %t2
643 ; z*x + y*z => (x+y) * z
644 define float @fact_mul2(float %x, float %y, float %z) {
645 ; CHECK-LABEL: @fact_mul2(
646 ; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[X:%.*]], [[Y:%.*]]
647 ; CHECK-NEXT: [[TMP2:%.*]] = fmul fast float [[TMP1]], [[Z:%.*]]
648 ; CHECK-NEXT: ret float [[TMP2]]
650 %t1 = fmul fast float %z, %x
651 %t2 = fmul fast float %y, %z
652 %t3 = fsub fast float %t1, %t2
656 ; Check again using the minimal subset of FMF.
657 define float @fact_mul2_reassoc_nsz(float %x, float %y, float %z) {
658 ; CHECK-LABEL: @fact_mul2_reassoc_nsz(
659 ; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[X:%.*]], [[Y:%.*]]
660 ; CHECK-NEXT: [[TMP2:%.*]] = fmul reassoc nsz float [[TMP1]], [[Z:%.*]]
661 ; CHECK-NEXT: ret float [[TMP2]]
663 %t1 = fmul reassoc nsz float %z, %x
664 %t2 = fmul reassoc nsz float %y, %z
665 %t3 = fsub reassoc nsz float %t1, %t2
669 ; Verify the fold is not done with only 'reassoc' ('nsz' is required).
670 define float @fact_mul2_reassoc(float %x, float %y, float %z) {
671 ; CHECK-LABEL: @fact_mul2_reassoc(
672 ; CHECK-NEXT: [[T1:%.*]] = fmul reassoc float [[Z:%.*]], [[X:%.*]]
673 ; CHECK-NEXT: [[T2:%.*]] = fmul reassoc float [[Y:%.*]], [[Z]]
674 ; CHECK-NEXT: [[T3:%.*]] = fsub reassoc float [[T1]], [[T2]]
675 ; CHECK-NEXT: ret float [[T3]]
677 %t1 = fmul reassoc float %z, %x
678 %t2 = fmul reassoc float %y, %z
679 %t3 = fsub reassoc float %t1, %t2
683 ; z*x - z*y => (x-y) * z
684 define float @fact_mul3(float %x, float %y, float %z) {
685 ; CHECK-LABEL: @fact_mul3(
686 ; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[X:%.*]], [[Y:%.*]]
687 ; CHECK-NEXT: [[TMP2:%.*]] = fmul fast float [[TMP1]], [[Z:%.*]]
688 ; CHECK-NEXT: ret float [[TMP2]]
690 %t2 = fmul fast float %z, %y
691 %t1 = fmul fast float %z, %x
692 %t3 = fsub fast float %t1, %t2
696 ; Check again using the minimal subset of FMF.
697 define float @fact_mul3_reassoc_nsz(float %x, float %y, float %z) {
698 ; CHECK-LABEL: @fact_mul3_reassoc_nsz(
699 ; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[X:%.*]], [[Y:%.*]]
700 ; CHECK-NEXT: [[TMP2:%.*]] = fmul reassoc nsz float [[TMP1]], [[Z:%.*]]
701 ; CHECK-NEXT: ret float [[TMP2]]
703 %t2 = fmul reassoc nsz float %z, %y
704 %t1 = fmul reassoc nsz float %z, %x
705 %t3 = fsub reassoc nsz float %t1, %t2
709 ; Verify the fold is not done with only 'reassoc' ('nsz' is required).
710 define float @fact_mul3_reassoc(float %x, float %y, float %z) {
711 ; CHECK-LABEL: @fact_mul3_reassoc(
712 ; CHECK-NEXT: [[T2:%.*]] = fmul reassoc float [[Z:%.*]], [[Y:%.*]]
713 ; CHECK-NEXT: [[T1:%.*]] = fmul reassoc float [[Z]], [[X:%.*]]
714 ; CHECK-NEXT: [[T3:%.*]] = fsub reassoc float [[T1]], [[T2]]
715 ; CHECK-NEXT: ret float [[T3]]
717 %t2 = fmul reassoc float %z, %y
718 %t1 = fmul reassoc float %z, %x
719 %t3 = fsub reassoc float %t1, %t2
723 ; x*z - z*y => (x-y) * z
724 define float @fact_mul4(float %x, float %y, float %z) {
725 ; CHECK-LABEL: @fact_mul4(
726 ; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[X:%.*]], [[Y:%.*]]
727 ; CHECK-NEXT: [[TMP2:%.*]] = fmul fast float [[TMP1]], [[Z:%.*]]
728 ; CHECK-NEXT: ret float [[TMP2]]
730 %t1 = fmul fast float %x, %z
731 %t2 = fmul fast float %z, %y
732 %t3 = fsub fast float %t1, %t2
736 ; Check again using the minimal subset of FMF.
737 define float @fact_mul4_reassoc_nsz(float %x, float %y, float %z) {
738 ; CHECK-LABEL: @fact_mul4_reassoc_nsz(
739 ; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[X:%.*]], [[Y:%.*]]
740 ; CHECK-NEXT: [[TMP2:%.*]] = fmul reassoc nsz float [[TMP1]], [[Z:%.*]]
741 ; CHECK-NEXT: ret float [[TMP2]]
743 %t1 = fmul reassoc nsz float %x, %z
744 %t2 = fmul reassoc nsz float %z, %y
745 %t3 = fsub reassoc nsz float %t1, %t2
749 ; Verify the fold is not done with only 'reassoc' ('nsz' is required).
750 define float @fact_mul4_reassoc(float %x, float %y, float %z) {
751 ; CHECK-LABEL: @fact_mul4_reassoc(
752 ; CHECK-NEXT: [[T1:%.*]] = fmul reassoc float [[X:%.*]], [[Z:%.*]]
753 ; CHECK-NEXT: [[T2:%.*]] = fmul reassoc float [[Z]], [[Y:%.*]]
754 ; CHECK-NEXT: [[T3:%.*]] = fsub reassoc float [[T1]], [[T2]]
755 ; CHECK-NEXT: ret float [[T3]]
757 %t1 = fmul reassoc float %x, %z
758 %t2 = fmul reassoc float %z, %y
759 %t3 = fsub reassoc float %t1, %t2
763 ; x/y + x/z, no xform
764 define float @fact_div1(float %x, float %y, float %z) {
765 ; CHECK-LABEL: @fact_div1(
766 ; CHECK-NEXT: [[T1:%.*]] = fdiv fast float [[X:%.*]], [[Y:%.*]]
767 ; CHECK-NEXT: [[T2:%.*]] = fdiv fast float [[X]], [[Z:%.*]]
768 ; CHECK-NEXT: [[T3:%.*]] = fadd fast float [[T1]], [[T2]]
769 ; CHECK-NEXT: ret float [[T3]]
771 %t1 = fdiv fast float %x, %y
772 %t2 = fdiv fast float %x, %z
773 %t3 = fadd fast float %t1, %t2
777 ; x/y + z/x; no xform
778 define float @fact_div2(float %x, float %y, float %z) {
779 ; CHECK-LABEL: @fact_div2(
780 ; CHECK-NEXT: [[T1:%.*]] = fdiv fast float [[X:%.*]], [[Y:%.*]]
781 ; CHECK-NEXT: [[T2:%.*]] = fdiv fast float [[Z:%.*]], [[X]]
782 ; CHECK-NEXT: [[T3:%.*]] = fadd fast float [[T1]], [[T2]]
783 ; CHECK-NEXT: ret float [[T3]]
785 %t1 = fdiv fast float %x, %y
786 %t2 = fdiv fast float %z, %x
787 %t3 = fadd fast float %t1, %t2
791 ; y/x + z/x => (y+z)/x
792 define float @fact_div3(float %x, float %y, float %z) {
793 ; CHECK-LABEL: @fact_div3(
794 ; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[Y:%.*]], [[Z:%.*]]
795 ; CHECK-NEXT: [[TMP2:%.*]] = fdiv fast float [[TMP1]], [[X:%.*]]
796 ; CHECK-NEXT: ret float [[TMP2]]
798 %t1 = fdiv fast float %y, %x
799 %t2 = fdiv fast float %z, %x
800 %t3 = fadd fast float %t1, %t2
804 ; Check again using the minimal subset of FMF.
805 define float @fact_div3_reassoc_nsz(float %x, float %y, float %z) {
806 ; CHECK-LABEL: @fact_div3_reassoc_nsz(
807 ; CHECK-NEXT: [[TMP1:%.*]] = fadd reassoc nsz float [[Y:%.*]], [[Z:%.*]]
808 ; CHECK-NEXT: [[TMP2:%.*]] = fdiv reassoc nsz float [[TMP1]], [[X:%.*]]
809 ; CHECK-NEXT: ret float [[TMP2]]
811 %t1 = fdiv reassoc nsz float %y, %x
812 %t2 = fdiv reassoc nsz float %z, %x
813 %t3 = fadd reassoc nsz float %t1, %t2
817 ; Verify the fold is not done with only 'reassoc' ('nsz' is required).
818 define float @fact_div3_reassoc(float %x, float %y, float %z) {
819 ; CHECK-LABEL: @fact_div3_reassoc(
820 ; CHECK-NEXT: [[T1:%.*]] = fdiv reassoc float [[Y:%.*]], [[X:%.*]]
821 ; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc float [[Z:%.*]], [[X]]
822 ; CHECK-NEXT: [[T3:%.*]] = fadd reassoc float [[T1]], [[T2]]
823 ; CHECK-NEXT: ret float [[T3]]
825 %t1 = fdiv reassoc float %y, %x
826 %t2 = fdiv reassoc float %z, %x
827 %t3 = fadd reassoc float %t1, %t2
831 ; y/x - z/x => (y-z)/x
832 define float @fact_div4(float %x, float %y, float %z) {
833 ; CHECK-LABEL: @fact_div4(
834 ; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[Y:%.*]], [[Z:%.*]]
835 ; CHECK-NEXT: [[TMP2:%.*]] = fdiv fast float [[TMP1]], [[X:%.*]]
836 ; CHECK-NEXT: ret float [[TMP2]]
838 %t1 = fdiv fast float %y, %x
839 %t2 = fdiv fast float %z, %x
840 %t3 = fsub fast float %t1, %t2
844 ; Check again using the minimal subset of FMF.
845 define float @fact_div4_reassoc_nsz(float %x, float %y, float %z) {
846 ; CHECK-LABEL: @fact_div4_reassoc_nsz(
847 ; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[Y:%.*]], [[Z:%.*]]
848 ; CHECK-NEXT: [[TMP2:%.*]] = fdiv reassoc nsz float [[TMP1]], [[X:%.*]]
849 ; CHECK-NEXT: ret float [[TMP2]]
851 %t1 = fdiv reassoc nsz float %y, %x
852 %t2 = fdiv reassoc nsz float %z, %x
853 %t3 = fsub reassoc nsz float %t1, %t2
857 ; Verify the fold is not done with only 'reassoc' ('nsz' is required).
858 define float @fact_div4_reassoc(float %x, float %y, float %z) {
859 ; CHECK-LABEL: @fact_div4_reassoc(
860 ; CHECK-NEXT: [[T1:%.*]] = fdiv reassoc float [[Y:%.*]], [[X:%.*]]
861 ; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc float [[Z:%.*]], [[X]]
862 ; CHECK-NEXT: [[T3:%.*]] = fsub reassoc float [[T1]], [[T2]]
863 ; CHECK-NEXT: ret float [[T3]]
865 %t1 = fdiv reassoc float %y, %x
866 %t2 = fdiv reassoc float %z, %x
867 %t3 = fsub reassoc float %t1, %t2
871 ; y/x - z/x => (y-z)/x is disabled if y-z is denormal.
872 define float @fact_div5(float %x) {
873 ; CHECK-LABEL: @fact_div5(
874 ; CHECK-NEXT: [[TMP1:%.*]] = fdiv fast float 0x3818000000000000, [[X:%.*]]
875 ; CHECK-NEXT: ret float [[TMP1]]
877 %t1 = fdiv fast float 0x3810000000000000, %x
878 %t2 = fdiv fast float 0x3800000000000000, %x
879 %t3 = fadd fast float %t1, %t2
883 ; y/x - z/x => (y-z)/x is disabled if y-z is denormal.
884 define float @fact_div6(float %x) {
885 ; CHECK-LABEL: @fact_div6(
886 ; CHECK-NEXT: [[T1:%.*]] = fdiv fast float 0x3810000000000000, [[X:%.*]]
887 ; CHECK-NEXT: [[T2:%.*]] = fdiv fast float 0x3800000000000000, [[X]]
888 ; CHECK-NEXT: [[T3:%.*]] = fsub fast float [[T1]], [[T2]]
889 ; CHECK-NEXT: ret float [[T3]]
891 %t1 = fdiv fast float 0x3810000000000000, %x
892 %t2 = fdiv fast float 0x3800000000000000, %x
893 %t3 = fsub fast float %t1, %t2
897 ; =========================================================================
899 ; Test-cases for square root
901 ; =========================================================================
903 ; A squared factor fed into a square root intrinsic should be hoisted out
906 declare double @llvm.sqrt.f64(double)
908 define double @sqrt_intrinsic_arg_squared(double %x) {
909 ; CHECK-LABEL: @sqrt_intrinsic_arg_squared(
910 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
911 ; CHECK-NEXT: ret double [[FABS]]
913 %mul = fmul fast double %x, %x
914 %sqrt = call fast double @llvm.sqrt.f64(double %mul)
918 ; Check all 6 combinations of a 3-way multiplication tree where
919 ; one factor is repeated.
921 define double @sqrt_intrinsic_three_args1(double %x, double %y) {
922 ; CHECK-LABEL: @sqrt_intrinsic_three_args1(
923 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
924 ; CHECK-NEXT: [[SQRT1:%.*]] = call fast double @llvm.sqrt.f64(double [[Y:%.*]])
925 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[FABS]], [[SQRT1]]
926 ; CHECK-NEXT: ret double [[TMP1]]
928 %mul = fmul fast double %y, %x
929 %mul2 = fmul fast double %mul, %x
930 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
934 define double @sqrt_intrinsic_three_args2(double %x, double %y) {
935 ; CHECK-LABEL: @sqrt_intrinsic_three_args2(
936 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
937 ; CHECK-NEXT: [[SQRT1:%.*]] = call fast double @llvm.sqrt.f64(double [[Y:%.*]])
938 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[FABS]], [[SQRT1]]
939 ; CHECK-NEXT: ret double [[TMP1]]
941 %mul = fmul fast double %x, %y
942 %mul2 = fmul fast double %mul, %x
943 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
947 define double @sqrt_intrinsic_three_args3(double %x, double %y) {
948 ; CHECK-LABEL: @sqrt_intrinsic_three_args3(
949 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
950 ; CHECK-NEXT: [[SQRT1:%.*]] = call fast double @llvm.sqrt.f64(double [[Y:%.*]])
951 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[FABS]], [[SQRT1]]
952 ; CHECK-NEXT: ret double [[TMP1]]
954 %mul = fmul fast double %x, %x
955 %mul2 = fmul fast double %mul, %y
956 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
960 define double @sqrt_intrinsic_three_args4(double %x, double %y) {
961 ; CHECK-LABEL: @sqrt_intrinsic_three_args4(
962 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
963 ; CHECK-NEXT: [[SQRT1:%.*]] = call fast double @llvm.sqrt.f64(double [[Y:%.*]])
964 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[FABS]], [[SQRT1]]
965 ; CHECK-NEXT: ret double [[TMP1]]
967 %mul = fmul fast double %y, %x
968 %mul2 = fmul fast double %x, %mul
969 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
973 define double @sqrt_intrinsic_three_args5(double %x, double %y) {
974 ; CHECK-LABEL: @sqrt_intrinsic_three_args5(
975 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
976 ; CHECK-NEXT: [[SQRT1:%.*]] = call fast double @llvm.sqrt.f64(double [[Y:%.*]])
977 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[FABS]], [[SQRT1]]
978 ; CHECK-NEXT: ret double [[TMP1]]
980 %mul = fmul fast double %x, %y
981 %mul2 = fmul fast double %x, %mul
982 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
986 define double @sqrt_intrinsic_three_args6(double %x, double %y) {
987 ; CHECK-LABEL: @sqrt_intrinsic_three_args6(
988 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
989 ; CHECK-NEXT: [[SQRT1:%.*]] = call fast double @llvm.sqrt.f64(double [[Y:%.*]])
990 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[FABS]], [[SQRT1]]
991 ; CHECK-NEXT: ret double [[TMP1]]
993 %mul = fmul fast double %x, %x
994 %mul2 = fmul fast double %y, %mul
995 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
999 ; If any operation is not 'fast', we can't simplify.
1001 define double @sqrt_intrinsic_not_so_fast(double %x, double %y) {
1002 ; CHECK-LABEL: @sqrt_intrinsic_not_so_fast(
1003 ; CHECK-NEXT: [[MUL:%.*]] = fmul double [[X:%.*]], [[X]]
1004 ; CHECK-NEXT: [[MUL2:%.*]] = fmul fast double [[MUL]], [[Y:%.*]]
1005 ; CHECK-NEXT: [[SQRT:%.*]] = call fast double @llvm.sqrt.f64(double [[MUL2]])
1006 ; CHECK-NEXT: ret double [[SQRT]]
1008 %mul = fmul double %x, %x
1009 %mul2 = fmul fast double %mul, %y
1010 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
1014 define double @sqrt_intrinsic_arg_4th(double %x) {
1015 ; CHECK-LABEL: @sqrt_intrinsic_arg_4th(
1016 ; CHECK-NEXT: [[MUL:%.*]] = fmul fast double [[X:%.*]], [[X]]
1017 ; CHECK-NEXT: ret double [[MUL]]
1019 %mul = fmul fast double %x, %x
1020 %mul2 = fmul fast double %mul, %mul
1021 %sqrt = call fast double @llvm.sqrt.f64(double %mul2)
1025 define double @sqrt_intrinsic_arg_5th(double %x) {
1026 ; CHECK-LABEL: @sqrt_intrinsic_arg_5th(
1027 ; CHECK-NEXT: [[MUL:%.*]] = fmul fast double [[X:%.*]], [[X]]
1028 ; CHECK-NEXT: [[SQRT1:%.*]] = call fast double @llvm.sqrt.f64(double [[X]])
1029 ; CHECK-NEXT: [[TMP1:%.*]] = fmul fast double [[MUL]], [[SQRT1]]
1030 ; CHECK-NEXT: ret double [[TMP1]]
1032 %mul = fmul fast double %x, %x
1033 %mul2 = fmul fast double %mul, %x
1034 %mul3 = fmul fast double %mul2, %mul
1035 %sqrt = call fast double @llvm.sqrt.f64(double %mul3)
1039 ; Check that square root calls have the same behavior.
1041 declare float @sqrtf(float)
1042 declare double @sqrt(double)
1043 declare fp128 @sqrtl(fp128)
1045 define float @sqrt_call_squared_f32(float %x) {
1046 ; CHECK-LABEL: @sqrt_call_squared_f32(
1047 ; CHECK-NEXT: [[FABS:%.*]] = call fast float @llvm.fabs.f32(float [[X:%.*]])
1048 ; CHECK-NEXT: ret float [[FABS]]
1050 %mul = fmul fast float %x, %x
1051 %sqrt = call fast float @sqrtf(float %mul)
1055 define double @sqrt_call_squared_f64(double %x) {
1056 ; CHECK-LABEL: @sqrt_call_squared_f64(
1057 ; CHECK-NEXT: [[FABS:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]])
1058 ; CHECK-NEXT: ret double [[FABS]]
1060 %mul = fmul fast double %x, %x
1061 %sqrt = call fast double @sqrt(double %mul)
1065 define fp128 @sqrt_call_squared_f128(fp128 %x) {
1066 ; CHECK-LABEL: @sqrt_call_squared_f128(
1067 ; CHECK-NEXT: [[FABS:%.*]] = call fast fp128 @llvm.fabs.f128(fp128 [[X:%.*]])
1068 ; CHECK-NEXT: ret fp128 [[FABS]]
1070 %mul = fmul fast fp128 %x, %x
1071 %sqrt = call fast fp128 @sqrtl(fp128 %mul)
1075 ; =========================================================================
1077 ; Test-cases for fmin / fmax
1079 ; =========================================================================
1081 declare double @fmax(double, double)
1082 declare double @fmin(double, double)
1083 declare float @fmaxf(float, float)
1084 declare float @fminf(float, float)
1085 declare fp128 @fmaxl(fp128, fp128)
1086 declare fp128 @fminl(fp128, fp128)
1088 ; No NaNs is the minimum requirement to replace these calls.
1089 ; This should always be set when unsafe-fp-math is true, but
1090 ; alternate the attributes for additional test coverage.
1091 ; 'nsz' is implied by the definition of fmax or fmin itself.
1093 ; Shrink and remove the call.
1094 define float @max1(float %a, float %b) {
1095 ; CHECK-LABEL: @max1(
1096 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp fast ogt float [[A:%.*]], [[B:%.*]]
1097 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], float [[A]], float [[B]]
1098 ; CHECK-NEXT: ret float [[TMP2]]
1100 %c = fpext float %a to double
1101 %d = fpext float %b to double
1102 %e = call fast double @fmax(double %c, double %d)
1103 %f = fptrunc double %e to float
1107 define float @max2(float %a, float %b) {
1108 ; CHECK-LABEL: @max2(
1109 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp nnan nsz ogt float [[A:%.*]], [[B:%.*]]
1110 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], float [[A]], float [[B]]
1111 ; CHECK-NEXT: ret float [[TMP2]]
1113 %c = call nnan float @fmaxf(float %a, float %b)
1118 define double @max3(double %a, double %b) {
1119 ; CHECK-LABEL: @max3(
1120 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp fast ogt double [[A:%.*]], [[B:%.*]]
1121 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], double [[A]], double [[B]]
1122 ; CHECK-NEXT: ret double [[TMP2]]
1124 %c = call fast double @fmax(double %a, double %b)
1128 define fp128 @max4(fp128 %a, fp128 %b) {
1129 ; CHECK-LABEL: @max4(
1130 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp nnan nsz ogt fp128 [[A:%.*]], [[B:%.*]]
1131 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], fp128 [[A]], fp128 [[B]]
1132 ; CHECK-NEXT: ret fp128 [[TMP2]]
1134 %c = call nnan fp128 @fmaxl(fp128 %a, fp128 %b)
1138 ; Shrink and remove the call.
1139 define float @min1(float %a, float %b) {
1140 ; CHECK-LABEL: @min1(
1141 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp nnan nsz olt float [[A:%.*]], [[B:%.*]]
1142 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], float [[A]], float [[B]]
1143 ; CHECK-NEXT: ret float [[TMP2]]
1145 %c = fpext float %a to double
1146 %d = fpext float %b to double
1147 %e = call nnan double @fmin(double %c, double %d)
1148 %f = fptrunc double %e to float
1152 define float @min2(float %a, float %b) {
1153 ; CHECK-LABEL: @min2(
1154 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp fast olt float [[A:%.*]], [[B:%.*]]
1155 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], float [[A]], float [[B]]
1156 ; CHECK-NEXT: ret float [[TMP2]]
1158 %c = call fast float @fminf(float %a, float %b)
1162 define double @min3(double %a, double %b) {
1163 ; CHECK-LABEL: @min3(
1164 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp nnan nsz olt double [[A:%.*]], [[B:%.*]]
1165 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], double [[A]], double [[B]]
1166 ; CHECK-NEXT: ret double [[TMP2]]
1168 %c = call nnan double @fmin(double %a, double %b)
1172 define fp128 @min4(fp128 %a, fp128 %b) {
1173 ; CHECK-LABEL: @min4(
1174 ; CHECK-NEXT: [[TMP1:%.*]] = fcmp fast olt fp128 [[A:%.*]], [[B:%.*]]
1175 ; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], fp128 [[A]], fp128 [[B]]
1176 ; CHECK-NEXT: ret fp128 [[TMP2]]
1178 %c = call fast fp128 @fminl(fp128 %a, fp128 %b)
1182 ; ((which ? 2.0 : a) + 1.0) => (which ? 3.0 : (a + 1.0))
1183 ; This is always safe. No FMF required.
1184 define float @test55(i1 %which, float %a) {
1185 ; CHECK-LABEL: @test55(
1186 ; CHECK-NEXT: entry:
1187 ; CHECK-NEXT: br i1 [[WHICH:%.*]], label [[FINAL:%.*]], label [[DELAY:%.*]]
1189 ; CHECK-NEXT: [[PHITMP:%.*]] = fadd float [[A:%.*]], 1.000000e+00
1190 ; CHECK-NEXT: br label [[FINAL]]
1192 ; CHECK-NEXT: [[A:%.*]] = phi float [ 3.000000e+00, [[ENTRY:%.*]] ], [ [[PHITMP]], [[DELAY]] ]
1193 ; CHECK-NEXT: ret float [[A]]
1196 br i1 %which, label %final, label %delay
1202 %A = phi float [ 2.0, %entry ], [ %a, %delay ]
1203 %value = fadd float %A, 1.0