LLVM 22.0.0git
TargetLoweringBase.cpp
Go to the documentation of this file.
1//===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This implements the TargetLoweringBase class.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/ADT/BitVector.h"
14#include "llvm/ADT/DenseMap.h"
15#include "llvm/ADT/STLExtras.h"
18#include "llvm/ADT/StringRef.h"
19#include "llvm/ADT/Twine.h"
20#include "llvm/Analysis/Loads.h"
39#include "llvm/IR/Attributes.h"
40#include "llvm/IR/CallingConv.h"
41#include "llvm/IR/DataLayout.h"
43#include "llvm/IR/Function.h"
44#include "llvm/IR/GlobalValue.h"
46#include "llvm/IR/IRBuilder.h"
47#include "llvm/IR/Module.h"
48#include "llvm/IR/Type.h"
58#include <algorithm>
59#include <cassert>
60#include <cstdint>
61#include <cstring>
62#include <string>
63#include <tuple>
64#include <utility>
65
66using namespace llvm;
67
69 "jump-is-expensive", cl::init(false),
70 cl::desc("Do not create extra branches to split comparison logic."),
72
74 ("min-jump-table-entries", cl::init(4), cl::Hidden,
75 cl::desc("Set minimum number of entries to use a jump table."));
76
78 ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden,
79 cl::desc("Set maximum size of jump tables."));
80
81/// Minimum jump table density for normal functions.
83 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden,
84 cl::desc("Minimum density for building a jump table in "
85 "a normal function"));
86
87/// Minimum jump table density for -Os or -Oz functions.
89 "optsize-jump-table-density", cl::init(40), cl::Hidden,
90 cl::desc("Minimum density for building a jump table in "
91 "an optsize function"));
92
94 "min-bit-test-cmps", cl::init(2), cl::Hidden,
95 cl::desc("Set minimum of largest number of comparisons "
96 "to use bit test for switch."));
97
98// FIXME: This option is only to test if the strict fp operation processed
99// correctly by preventing mutating strict fp operation to normal fp operation
100// during development. When the backend supports strict float operation, this
101// option will be meaningless.
102static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation",
103 cl::desc("Don't mutate strict-float node to a legalize node"),
104 cl::init(false), cl::Hidden);
105
106LLVM_ABI RTLIB::Libcall RTLIB::getSHL(EVT VT) {
107 if (VT == MVT::i16)
108 return RTLIB::SHL_I16;
109 if (VT == MVT::i32)
110 return RTLIB::SHL_I32;
111 if (VT == MVT::i64)
112 return RTLIB::SHL_I64;
113 if (VT == MVT::i128)
114 return RTLIB::SHL_I128;
115
116 return RTLIB::UNKNOWN_LIBCALL;
117}
118
119LLVM_ABI RTLIB::Libcall RTLIB::getSRL(EVT VT) {
120 if (VT == MVT::i16)
121 return RTLIB::SRL_I16;
122 if (VT == MVT::i32)
123 return RTLIB::SRL_I32;
124 if (VT == MVT::i64)
125 return RTLIB::SRL_I64;
126 if (VT == MVT::i128)
127 return RTLIB::SRL_I128;
128
129 return RTLIB::UNKNOWN_LIBCALL;
130}
131
132LLVM_ABI RTLIB::Libcall RTLIB::getSRA(EVT VT) {
133 if (VT == MVT::i16)
134 return RTLIB::SRA_I16;
135 if (VT == MVT::i32)
136 return RTLIB::SRA_I32;
137 if (VT == MVT::i64)
138 return RTLIB::SRA_I64;
139 if (VT == MVT::i128)
140 return RTLIB::SRA_I128;
141
142 return RTLIB::UNKNOWN_LIBCALL;
143}
144
145LLVM_ABI RTLIB::Libcall RTLIB::getMUL(EVT VT) {
146 if (VT == MVT::i16)
147 return RTLIB::MUL_I16;
148 if (VT == MVT::i32)
149 return RTLIB::MUL_I32;
150 if (VT == MVT::i64)
151 return RTLIB::MUL_I64;
152 if (VT == MVT::i128)
153 return RTLIB::MUL_I128;
154 return RTLIB::UNKNOWN_LIBCALL;
155}
156
157LLVM_ABI RTLIB::Libcall RTLIB::getMULO(EVT VT) {
158 if (VT == MVT::i32)
159 return RTLIB::MULO_I32;
160 if (VT == MVT::i64)
161 return RTLIB::MULO_I64;
162 if (VT == MVT::i128)
163 return RTLIB::MULO_I128;
164 return RTLIB::UNKNOWN_LIBCALL;
165}
166
167LLVM_ABI RTLIB::Libcall RTLIB::getSDIV(EVT VT) {
168 if (VT == MVT::i16)
169 return RTLIB::SDIV_I16;
170 if (VT == MVT::i32)
171 return RTLIB::SDIV_I32;
172 if (VT == MVT::i64)
173 return RTLIB::SDIV_I64;
174 if (VT == MVT::i128)
175 return RTLIB::SDIV_I128;
176 return RTLIB::UNKNOWN_LIBCALL;
177}
178
179LLVM_ABI RTLIB::Libcall RTLIB::getUDIV(EVT VT) {
180 if (VT == MVT::i16)
181 return RTLIB::UDIV_I16;
182 if (VT == MVT::i32)
183 return RTLIB::UDIV_I32;
184 if (VT == MVT::i64)
185 return RTLIB::UDIV_I64;
186 if (VT == MVT::i128)
187 return RTLIB::UDIV_I128;
188 return RTLIB::UNKNOWN_LIBCALL;
189}
190
191LLVM_ABI RTLIB::Libcall RTLIB::getSREM(EVT VT) {
192 if (VT == MVT::i16)
193 return RTLIB::SREM_I16;
194 if (VT == MVT::i32)
195 return RTLIB::SREM_I32;
196 if (VT == MVT::i64)
197 return RTLIB::SREM_I64;
198 if (VT == MVT::i128)
199 return RTLIB::SREM_I128;
200 return RTLIB::UNKNOWN_LIBCALL;
201}
202
203LLVM_ABI RTLIB::Libcall RTLIB::getUREM(EVT VT) {
204 if (VT == MVT::i16)
205 return RTLIB::UREM_I16;
206 if (VT == MVT::i32)
207 return RTLIB::UREM_I32;
208 if (VT == MVT::i64)
209 return RTLIB::UREM_I64;
210 if (VT == MVT::i128)
211 return RTLIB::UREM_I128;
212 return RTLIB::UNKNOWN_LIBCALL;
213}
214
215LLVM_ABI RTLIB::Libcall RTLIB::getCTPOP(EVT VT) {
216 if (VT == MVT::i32)
217 return RTLIB::CTPOP_I32;
218 if (VT == MVT::i64)
219 return RTLIB::CTPOP_I64;
220 if (VT == MVT::i128)
221 return RTLIB::CTPOP_I128;
222 return RTLIB::UNKNOWN_LIBCALL;
223}
224
225/// GetFPLibCall - Helper to return the right libcall for the given floating
226/// point type, or UNKNOWN_LIBCALL if there is none.
227RTLIB::Libcall RTLIB::getFPLibCall(EVT VT,
228 RTLIB::Libcall Call_F32,
229 RTLIB::Libcall Call_F64,
230 RTLIB::Libcall Call_F80,
231 RTLIB::Libcall Call_F128,
232 RTLIB::Libcall Call_PPCF128) {
233 return
234 VT == MVT::f32 ? Call_F32 :
235 VT == MVT::f64 ? Call_F64 :
236 VT == MVT::f80 ? Call_F80 :
237 VT == MVT::f128 ? Call_F128 :
238 VT == MVT::ppcf128 ? Call_PPCF128 :
239 RTLIB::UNKNOWN_LIBCALL;
240}
241
242/// getFPEXT - Return the FPEXT_*_* value for the given types, or
243/// UNKNOWN_LIBCALL if there is none.
244RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
245 if (OpVT == MVT::f16) {
246 if (RetVT == MVT::f32)
247 return FPEXT_F16_F32;
248 if (RetVT == MVT::f64)
249 return FPEXT_F16_F64;
250 if (RetVT == MVT::f80)
251 return FPEXT_F16_F80;
252 if (RetVT == MVT::f128)
253 return FPEXT_F16_F128;
254 } else if (OpVT == MVT::f32) {
255 if (RetVT == MVT::f64)
256 return FPEXT_F32_F64;
257 if (RetVT == MVT::f128)
258 return FPEXT_F32_F128;
259 if (RetVT == MVT::ppcf128)
260 return FPEXT_F32_PPCF128;
261 } else if (OpVT == MVT::f64) {
262 if (RetVT == MVT::f128)
263 return FPEXT_F64_F128;
264 else if (RetVT == MVT::ppcf128)
265 return FPEXT_F64_PPCF128;
266 } else if (OpVT == MVT::f80) {
267 if (RetVT == MVT::f128)
268 return FPEXT_F80_F128;
269 } else if (OpVT == MVT::bf16) {
270 if (RetVT == MVT::f32)
271 return FPEXT_BF16_F32;
272 }
273
274 return UNKNOWN_LIBCALL;
275}
276
277/// getFPROUND - Return the FPROUND_*_* value for the given types, or
278/// UNKNOWN_LIBCALL if there is none.
279RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
280 if (RetVT == MVT::f16) {
281 if (OpVT == MVT::f32)
282 return FPROUND_F32_F16;
283 if (OpVT == MVT::f64)
284 return FPROUND_F64_F16;
285 if (OpVT == MVT::f80)
286 return FPROUND_F80_F16;
287 if (OpVT == MVT::f128)
288 return FPROUND_F128_F16;
289 if (OpVT == MVT::ppcf128)
290 return FPROUND_PPCF128_F16;
291 } else if (RetVT == MVT::bf16) {
292 if (OpVT == MVT::f32)
293 return FPROUND_F32_BF16;
294 if (OpVT == MVT::f64)
295 return FPROUND_F64_BF16;
296 if (OpVT == MVT::f80)
297 return FPROUND_F80_BF16;
298 if (OpVT == MVT::f128)
299 return FPROUND_F128_BF16;
300 } else if (RetVT == MVT::f32) {
301 if (OpVT == MVT::f64)
302 return FPROUND_F64_F32;
303 if (OpVT == MVT::f80)
304 return FPROUND_F80_F32;
305 if (OpVT == MVT::f128)
306 return FPROUND_F128_F32;
307 if (OpVT == MVT::ppcf128)
308 return FPROUND_PPCF128_F32;
309 } else if (RetVT == MVT::f64) {
310 if (OpVT == MVT::f80)
311 return FPROUND_F80_F64;
312 if (OpVT == MVT::f128)
313 return FPROUND_F128_F64;
314 if (OpVT == MVT::ppcf128)
315 return FPROUND_PPCF128_F64;
316 } else if (RetVT == MVT::f80) {
317 if (OpVT == MVT::f128)
318 return FPROUND_F128_F80;
319 }
320
321 return UNKNOWN_LIBCALL;
322}
323
324/// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
325/// UNKNOWN_LIBCALL if there is none.
326RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
327 if (OpVT == MVT::f16) {
328 if (RetVT == MVT::i32)
329 return FPTOSINT_F16_I32;
330 if (RetVT == MVT::i64)
331 return FPTOSINT_F16_I64;
332 if (RetVT == MVT::i128)
333 return FPTOSINT_F16_I128;
334 } else if (OpVT == MVT::f32) {
335 if (RetVT == MVT::i32)
336 return FPTOSINT_F32_I32;
337 if (RetVT == MVT::i64)
338 return FPTOSINT_F32_I64;
339 if (RetVT == MVT::i128)
340 return FPTOSINT_F32_I128;
341 } else if (OpVT == MVT::f64) {
342 if (RetVT == MVT::i32)
343 return FPTOSINT_F64_I32;
344 if (RetVT == MVT::i64)
345 return FPTOSINT_F64_I64;
346 if (RetVT == MVT::i128)
347 return FPTOSINT_F64_I128;
348 } else if (OpVT == MVT::f80) {
349 if (RetVT == MVT::i32)
350 return FPTOSINT_F80_I32;
351 if (RetVT == MVT::i64)
352 return FPTOSINT_F80_I64;
353 if (RetVT == MVT::i128)
354 return FPTOSINT_F80_I128;
355 } else if (OpVT == MVT::f128) {
356 if (RetVT == MVT::i32)
357 return FPTOSINT_F128_I32;
358 if (RetVT == MVT::i64)
359 return FPTOSINT_F128_I64;
360 if (RetVT == MVT::i128)
361 return FPTOSINT_F128_I128;
362 } else if (OpVT == MVT::ppcf128) {
363 if (RetVT == MVT::i32)
364 return FPTOSINT_PPCF128_I32;
365 if (RetVT == MVT::i64)
366 return FPTOSINT_PPCF128_I64;
367 if (RetVT == MVT::i128)
368 return FPTOSINT_PPCF128_I128;
369 }
370 return UNKNOWN_LIBCALL;
371}
372
373/// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
374/// UNKNOWN_LIBCALL if there is none.
375RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
376 if (OpVT == MVT::f16) {
377 if (RetVT == MVT::i32)
378 return FPTOUINT_F16_I32;
379 if (RetVT == MVT::i64)
380 return FPTOUINT_F16_I64;
381 if (RetVT == MVT::i128)
382 return FPTOUINT_F16_I128;
383 } else if (OpVT == MVT::f32) {
384 if (RetVT == MVT::i32)
385 return FPTOUINT_F32_I32;
386 if (RetVT == MVT::i64)
387 return FPTOUINT_F32_I64;
388 if (RetVT == MVT::i128)
389 return FPTOUINT_F32_I128;
390 } else if (OpVT == MVT::f64) {
391 if (RetVT == MVT::i32)
392 return FPTOUINT_F64_I32;
393 if (RetVT == MVT::i64)
394 return FPTOUINT_F64_I64;
395 if (RetVT == MVT::i128)
396 return FPTOUINT_F64_I128;
397 } else if (OpVT == MVT::f80) {
398 if (RetVT == MVT::i32)
399 return FPTOUINT_F80_I32;
400 if (RetVT == MVT::i64)
401 return FPTOUINT_F80_I64;
402 if (RetVT == MVT::i128)
403 return FPTOUINT_F80_I128;
404 } else if (OpVT == MVT::f128) {
405 if (RetVT == MVT::i32)
406 return FPTOUINT_F128_I32;
407 if (RetVT == MVT::i64)
408 return FPTOUINT_F128_I64;
409 if (RetVT == MVT::i128)
410 return FPTOUINT_F128_I128;
411 } else if (OpVT == MVT::ppcf128) {
412 if (RetVT == MVT::i32)
413 return FPTOUINT_PPCF128_I32;
414 if (RetVT == MVT::i64)
415 return FPTOUINT_PPCF128_I64;
416 if (RetVT == MVT::i128)
417 return FPTOUINT_PPCF128_I128;
418 }
419 return UNKNOWN_LIBCALL;
420}
421
422/// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
423/// UNKNOWN_LIBCALL if there is none.
424RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
425 if (OpVT == MVT::i32) {
426 if (RetVT == MVT::f16)
427 return SINTTOFP_I32_F16;
428 if (RetVT == MVT::f32)
429 return SINTTOFP_I32_F32;
430 if (RetVT == MVT::f64)
431 return SINTTOFP_I32_F64;
432 if (RetVT == MVT::f80)
433 return SINTTOFP_I32_F80;
434 if (RetVT == MVT::f128)
435 return SINTTOFP_I32_F128;
436 if (RetVT == MVT::ppcf128)
437 return SINTTOFP_I32_PPCF128;
438 } else if (OpVT == MVT::i64) {
439 if (RetVT == MVT::bf16)
440 return SINTTOFP_I64_BF16;
441 if (RetVT == MVT::f16)
442 return SINTTOFP_I64_F16;
443 if (RetVT == MVT::f32)
444 return SINTTOFP_I64_F32;
445 if (RetVT == MVT::f64)
446 return SINTTOFP_I64_F64;
447 if (RetVT == MVT::f80)
448 return SINTTOFP_I64_F80;
449 if (RetVT == MVT::f128)
450 return SINTTOFP_I64_F128;
451 if (RetVT == MVT::ppcf128)
452 return SINTTOFP_I64_PPCF128;
453 } else if (OpVT == MVT::i128) {
454 if (RetVT == MVT::f16)
455 return SINTTOFP_I128_F16;
456 if (RetVT == MVT::f32)
457 return SINTTOFP_I128_F32;
458 if (RetVT == MVT::f64)
459 return SINTTOFP_I128_F64;
460 if (RetVT == MVT::f80)
461 return SINTTOFP_I128_F80;
462 if (RetVT == MVT::f128)
463 return SINTTOFP_I128_F128;
464 if (RetVT == MVT::ppcf128)
465 return SINTTOFP_I128_PPCF128;
466 }
467 return UNKNOWN_LIBCALL;
468}
469
470/// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
471/// UNKNOWN_LIBCALL if there is none.
472RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
473 if (OpVT == MVT::i32) {
474 if (RetVT == MVT::f16)
475 return UINTTOFP_I32_F16;
476 if (RetVT == MVT::f32)
477 return UINTTOFP_I32_F32;
478 if (RetVT == MVT::f64)
479 return UINTTOFP_I32_F64;
480 if (RetVT == MVT::f80)
481 return UINTTOFP_I32_F80;
482 if (RetVT == MVT::f128)
483 return UINTTOFP_I32_F128;
484 if (RetVT == MVT::ppcf128)
485 return UINTTOFP_I32_PPCF128;
486 } else if (OpVT == MVT::i64) {
487 if (RetVT == MVT::bf16)
488 return UINTTOFP_I64_BF16;
489 if (RetVT == MVT::f16)
490 return UINTTOFP_I64_F16;
491 if (RetVT == MVT::f32)
492 return UINTTOFP_I64_F32;
493 if (RetVT == MVT::f64)
494 return UINTTOFP_I64_F64;
495 if (RetVT == MVT::f80)
496 return UINTTOFP_I64_F80;
497 if (RetVT == MVT::f128)
498 return UINTTOFP_I64_F128;
499 if (RetVT == MVT::ppcf128)
500 return UINTTOFP_I64_PPCF128;
501 } else if (OpVT == MVT::i128) {
502 if (RetVT == MVT::f16)
503 return UINTTOFP_I128_F16;
504 if (RetVT == MVT::f32)
505 return UINTTOFP_I128_F32;
506 if (RetVT == MVT::f64)
507 return UINTTOFP_I128_F64;
508 if (RetVT == MVT::f80)
509 return UINTTOFP_I128_F80;
510 if (RetVT == MVT::f128)
511 return UINTTOFP_I128_F128;
512 if (RetVT == MVT::ppcf128)
513 return UINTTOFP_I128_PPCF128;
514 }
515 return UNKNOWN_LIBCALL;
516}
517
518RTLIB::Libcall RTLIB::getPOWI(EVT RetVT) {
519 return getFPLibCall(RetVT, POWI_F32, POWI_F64, POWI_F80, POWI_F128,
520 POWI_PPCF128);
521}
522
523RTLIB::Libcall RTLIB::getPOW(EVT RetVT) {
524 return getFPLibCall(RetVT, POW_F32, POW_F64, POW_F80, POW_F128, POW_PPCF128);
525}
526
527RTLIB::Libcall RTLIB::getLDEXP(EVT RetVT) {
528 return getFPLibCall(RetVT, LDEXP_F32, LDEXP_F64, LDEXP_F80, LDEXP_F128,
529 LDEXP_PPCF128);
530}
531
532RTLIB::Libcall RTLIB::getFREXP(EVT RetVT) {
533 return getFPLibCall(RetVT, FREXP_F32, FREXP_F64, FREXP_F80, FREXP_F128,
534 FREXP_PPCF128);
535}
536
537RTLIB::Libcall RTLIB::getSIN(EVT RetVT) {
538 return getFPLibCall(RetVT, SIN_F32, SIN_F64, SIN_F80, SIN_F128, SIN_PPCF128);
539}
540
541RTLIB::Libcall RTLIB::getCOS(EVT RetVT) {
542 return getFPLibCall(RetVT, COS_F32, COS_F64, COS_F80, COS_F128, COS_PPCF128);
543}
544
545RTLIB::Libcall RTLIB::getSINCOS(EVT RetVT) {
546 // TODO: Tablegen should generate this function
547 if (RetVT.isVector()) {
548 if (!RetVT.isSimple())
549 return RTLIB::UNKNOWN_LIBCALL;
550 switch (RetVT.getSimpleVT().SimpleTy) {
551 case MVT::v4f32:
552 return RTLIB::SINCOS_V4F32;
553 case MVT::v2f64:
554 return RTLIB::SINCOS_V2F64;
555 case MVT::nxv4f32:
556 return RTLIB::SINCOS_NXV4F32;
557 case MVT::nxv2f64:
558 return RTLIB::SINCOS_NXV2F64;
559 default:
560 return RTLIB::UNKNOWN_LIBCALL;
561 }
562 }
563
564 return getFPLibCall(RetVT, SINCOS_F32, SINCOS_F64, SINCOS_F80, SINCOS_F128,
565 SINCOS_PPCF128);
566}
567
568RTLIB::Libcall RTLIB::getSINCOSPI(EVT RetVT) {
569 // TODO: Tablegen should generate this function
570 if (RetVT.isVector()) {
571 if (!RetVT.isSimple())
572 return RTLIB::UNKNOWN_LIBCALL;
573 switch (RetVT.getSimpleVT().SimpleTy) {
574 case MVT::v4f32:
575 return RTLIB::SINCOSPI_V4F32;
576 case MVT::v2f64:
577 return RTLIB::SINCOSPI_V2F64;
578 case MVT::nxv4f32:
579 return RTLIB::SINCOSPI_NXV4F32;
580 case MVT::nxv2f64:
581 return RTLIB::SINCOSPI_NXV2F64;
582 default:
583 return RTLIB::UNKNOWN_LIBCALL;
584 }
585 }
586
587 return getFPLibCall(RetVT, SINCOSPI_F32, SINCOSPI_F64, SINCOSPI_F80,
588 SINCOSPI_F128, SINCOSPI_PPCF128);
589}
590
591RTLIB::Libcall RTLIB::getSINCOS_STRET(EVT RetVT) {
592 return getFPLibCall(RetVT, SINCOS_STRET_F32, SINCOS_STRET_F64,
593 UNKNOWN_LIBCALL, UNKNOWN_LIBCALL, UNKNOWN_LIBCALL);
594}
595
596RTLIB::Libcall RTLIB::getMODF(EVT RetVT) {
597 // TODO: Tablegen should generate this function
598 if (RetVT.isVector()) {
599 if (!RetVT.isSimple())
600 return RTLIB::UNKNOWN_LIBCALL;
601 switch (RetVT.getSimpleVT().SimpleTy) {
602 case MVT::v4f32:
603 return RTLIB::MODF_V4F32;
604 case MVT::v2f64:
605 return RTLIB::MODF_V2F64;
606 case MVT::nxv4f32:
607 return RTLIB::MODF_NXV4F32;
608 case MVT::nxv2f64:
609 return RTLIB::MODF_NXV2F64;
610 default:
611 return RTLIB::UNKNOWN_LIBCALL;
612 }
613 }
614
615 return getFPLibCall(RetVT, MODF_F32, MODF_F64, MODF_F80, MODF_F128,
616 MODF_PPCF128);
617}
618
619RTLIB::Libcall RTLIB::getLROUND(EVT VT) {
620 if (VT == MVT::f32)
621 return RTLIB::LROUND_F32;
622 if (VT == MVT::f64)
623 return RTLIB::LROUND_F64;
624 if (VT == MVT::f80)
625 return RTLIB::LROUND_F80;
626 if (VT == MVT::f128)
627 return RTLIB::LROUND_F128;
628 if (VT == MVT::ppcf128)
629 return RTLIB::LROUND_PPCF128;
630
631 return RTLIB::UNKNOWN_LIBCALL;
632}
633
634RTLIB::Libcall RTLIB::getLLROUND(EVT VT) {
635 if (VT == MVT::f32)
636 return RTLIB::LLROUND_F32;
637 if (VT == MVT::f64)
638 return RTLIB::LLROUND_F64;
639 if (VT == MVT::f80)
640 return RTLIB::LLROUND_F80;
641 if (VT == MVT::f128)
642 return RTLIB::LLROUND_F128;
643 if (VT == MVT::ppcf128)
644 return RTLIB::LLROUND_PPCF128;
645
646 return RTLIB::UNKNOWN_LIBCALL;
647}
648
649RTLIB::Libcall RTLIB::getLRINT(EVT VT) {
650 if (VT == MVT::f32)
651 return RTLIB::LRINT_F32;
652 if (VT == MVT::f64)
653 return RTLIB::LRINT_F64;
654 if (VT == MVT::f80)
655 return RTLIB::LRINT_F80;
656 if (VT == MVT::f128)
657 return RTLIB::LRINT_F128;
658 if (VT == MVT::ppcf128)
659 return RTLIB::LRINT_PPCF128;
660 return RTLIB::UNKNOWN_LIBCALL;
661}
662
663RTLIB::Libcall RTLIB::getLLRINT(EVT VT) {
664 if (VT == MVT::f32)
665 return RTLIB::LLRINT_F32;
666 if (VT == MVT::f64)
667 return RTLIB::LLRINT_F64;
668 if (VT == MVT::f80)
669 return RTLIB::LLRINT_F80;
670 if (VT == MVT::f128)
671 return RTLIB::LLRINT_F128;
672 if (VT == MVT::ppcf128)
673 return RTLIB::LLRINT_PPCF128;
674 return RTLIB::UNKNOWN_LIBCALL;
675}
676
677RTLIB::Libcall RTLIB::getOutlineAtomicHelper(const Libcall (&LC)[5][4],
678 AtomicOrdering Order,
679 uint64_t MemSize) {
680 unsigned ModeN, ModelN;
681 switch (MemSize) {
682 case 1:
683 ModeN = 0;
684 break;
685 case 2:
686 ModeN = 1;
687 break;
688 case 4:
689 ModeN = 2;
690 break;
691 case 8:
692 ModeN = 3;
693 break;
694 case 16:
695 ModeN = 4;
696 break;
697 default:
698 return RTLIB::UNKNOWN_LIBCALL;
699 }
700
701 switch (Order) {
703 ModelN = 0;
704 break;
706 ModelN = 1;
707 break;
709 ModelN = 2;
710 break;
713 ModelN = 3;
714 break;
715 default:
716 return UNKNOWN_LIBCALL;
717 }
718
719 return LC[ModeN][ModelN];
720}
721
722RTLIB::Libcall RTLIB::getOUTLINE_ATOMIC(unsigned Opc, AtomicOrdering Order,
723 MVT VT) {
724 if (!VT.isScalarInteger())
725 return UNKNOWN_LIBCALL;
726 uint64_t MemSize = VT.getScalarSizeInBits() / 8;
727
728#define LCALLS(A, B) \
729 { A##B##_RELAX, A##B##_ACQ, A##B##_REL, A##B##_ACQ_REL }
730#define LCALL5(A) \
731 LCALLS(A, 1), LCALLS(A, 2), LCALLS(A, 4), LCALLS(A, 8), LCALLS(A, 16)
732 switch (Opc) {
733 case ISD::ATOMIC_CMP_SWAP: {
734 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_CAS)};
735 return getOutlineAtomicHelper(LC, Order, MemSize);
736 }
737 case ISD::ATOMIC_SWAP: {
738 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_SWP)};
739 return getOutlineAtomicHelper(LC, Order, MemSize);
740 }
741 case ISD::ATOMIC_LOAD_ADD: {
742 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDADD)};
743 return getOutlineAtomicHelper(LC, Order, MemSize);
744 }
745 case ISD::ATOMIC_LOAD_OR: {
746 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDSET)};
747 return getOutlineAtomicHelper(LC, Order, MemSize);
748 }
749 case ISD::ATOMIC_LOAD_CLR: {
750 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDCLR)};
751 return getOutlineAtomicHelper(LC, Order, MemSize);
752 }
753 case ISD::ATOMIC_LOAD_XOR: {
754 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDEOR)};
755 return getOutlineAtomicHelper(LC, Order, MemSize);
756 }
757 default:
758 return UNKNOWN_LIBCALL;
759 }
760#undef LCALLS
761#undef LCALL5
762}
763
764RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
765#define OP_TO_LIBCALL(Name, Enum) \
766 case Name: \
767 switch (VT.SimpleTy) { \
768 default: \
769 return UNKNOWN_LIBCALL; \
770 case MVT::i8: \
771 return Enum##_1; \
772 case MVT::i16: \
773 return Enum##_2; \
774 case MVT::i32: \
775 return Enum##_4; \
776 case MVT::i64: \
777 return Enum##_8; \
778 case MVT::i128: \
779 return Enum##_16; \
780 }
781
782 switch (Opc) {
783 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
784 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
785 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
786 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
787 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
788 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
789 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
790 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
791 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
792 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
793 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
794 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
795 }
796
797#undef OP_TO_LIBCALL
798
799 return UNKNOWN_LIBCALL;
800}
801
803 switch (ElementSize) {
804 case 1:
805 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
806 case 2:
807 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
808 case 4:
809 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
810 case 8:
811 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
812 case 16:
813 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
814 default:
815 return UNKNOWN_LIBCALL;
816 }
817}
818
820 switch (ElementSize) {
821 case 1:
822 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
823 case 2:
824 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
825 case 4:
826 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
827 case 8:
828 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
829 case 16:
830 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
831 default:
832 return UNKNOWN_LIBCALL;
833 }
834}
835
837 switch (ElementSize) {
838 case 1:
839 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
840 case 2:
841 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
842 case 4:
843 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
844 case 8:
845 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
846 case 16:
847 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
848 default:
849 return UNKNOWN_LIBCALL;
850 }
851}
852
854 RTLIB::LibcallImpl Impl) const {
855 switch (Impl) {
856 case RTLIB::impl___aeabi_dcmpeq__une:
857 case RTLIB::impl___aeabi_fcmpeq__une:
858 // Usage in the eq case, so we have to invert the comparison.
859 return ISD::SETEQ;
860 case RTLIB::impl___aeabi_dcmpeq__oeq:
861 case RTLIB::impl___aeabi_fcmpeq__oeq:
862 // Normal comparison to boolean value.
863 return ISD::SETNE;
864 case RTLIB::impl___aeabi_dcmplt:
865 case RTLIB::impl___aeabi_dcmple:
866 case RTLIB::impl___aeabi_dcmpge:
867 case RTLIB::impl___aeabi_dcmpgt:
868 case RTLIB::impl___aeabi_dcmpun:
869 case RTLIB::impl___aeabi_fcmplt:
870 case RTLIB::impl___aeabi_fcmple:
871 case RTLIB::impl___aeabi_fcmpge:
872 case RTLIB::impl___aeabi_fcmpgt:
873 /// The AEABI versions return a typical boolean value, so we can compare
874 /// against the integer result as simply != 0.
875 return ISD::SETNE;
876 default:
877 break;
878 }
879
880 // Assume libgcc/compiler-rt behavior. Most of the cases are really aliases of
881 // each other, and return a 3-way comparison style result of -1, 0, or 1
882 // depending on lt/eq/gt.
883 //
884 // FIXME: It would be cleaner to directly express this as a 3-way comparison
885 // soft FP libcall instead of individual compares.
886 RTLIB::Libcall LC = RTLIB::RuntimeLibcallsInfo::getLibcallFromImpl(Impl);
887 switch (LC) {
888 case RTLIB::OEQ_F32:
889 case RTLIB::OEQ_F64:
890 case RTLIB::OEQ_F128:
891 case RTLIB::OEQ_PPCF128:
892 return ISD::SETEQ;
893 case RTLIB::UNE_F32:
894 case RTLIB::UNE_F64:
895 case RTLIB::UNE_F128:
896 case RTLIB::UNE_PPCF128:
897 return ISD::SETNE;
898 case RTLIB::OGE_F32:
899 case RTLIB::OGE_F64:
900 case RTLIB::OGE_F128:
901 case RTLIB::OGE_PPCF128:
902 return ISD::SETGE;
903 case RTLIB::OLT_F32:
904 case RTLIB::OLT_F64:
905 case RTLIB::OLT_F128:
906 case RTLIB::OLT_PPCF128:
907 return ISD::SETLT;
908 case RTLIB::OLE_F32:
909 case RTLIB::OLE_F64:
910 case RTLIB::OLE_F128:
911 case RTLIB::OLE_PPCF128:
912 return ISD::SETLE;
913 case RTLIB::OGT_F32:
914 case RTLIB::OGT_F64:
915 case RTLIB::OGT_F128:
916 case RTLIB::OGT_PPCF128:
917 return ISD::SETGT;
918 case RTLIB::UO_F32:
919 case RTLIB::UO_F64:
920 case RTLIB::UO_F128:
921 case RTLIB::UO_PPCF128:
922 return ISD::SETNE;
923 default:
924 llvm_unreachable("not a compare libcall");
925 }
926}
927
928/// NOTE: The TargetMachine owns TLOF.
930 const TargetSubtargetInfo &STI)
931 : TM(tm),
932 RuntimeLibcallInfo(TM.getTargetTriple(), TM.Options.ExceptionModel,
933 TM.Options.FloatABIType, TM.Options.EABIVersion,
934 TM.Options.MCOptions.getABIName(), TM.Options.VecLib),
935 Libcalls(RuntimeLibcallInfo, STI) {
936 initActions();
937
938 // Perform these initializations only once.
944 HasExtractBitsInsn = false;
945 JumpIsExpensive = JumpIsExpensiveOverride;
947 EnableExtLdPromotion = false;
948 StackPointerRegisterToSaveRestore = 0;
949 BooleanContents = UndefinedBooleanContent;
950 BooleanFloatContents = UndefinedBooleanContent;
951 BooleanVectorContents = UndefinedBooleanContent;
952 SchedPreferenceInfo = Sched::ILP;
955 MaxBytesForAlignment = 0;
956 MaxAtomicSizeInBitsSupported = 0;
957
958 // Assume that even with libcalls, no target supports wider than 128 bit
959 // division.
960 MaxDivRemBitWidthSupported = 128;
961
962 MaxLargeFPConvertBitWidthSupported = llvm::IntegerType::MAX_INT_BITS;
963
964 MinCmpXchgSizeInBits = 0;
965 SupportsUnalignedAtomics = false;
966
967 MinimumBitTestCmps = MinimumBitTestCmpsOverride;
968}
969
970// Define the virtual destructor out-of-line to act as a key method to anchor
971// debug info (see coding standards).
973
975 // All operations default to being supported.
976 memset(OpActions, 0, sizeof(OpActions));
977 memset(LoadExtActions, 0, sizeof(LoadExtActions));
978 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
979 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
980 memset(CondCodeActions, 0, sizeof(CondCodeActions));
981 llvm::fill(RegClassForVT, nullptr);
982 llvm::fill(TargetDAGCombineArray, 0);
983
984 // Let extending atomic loads be unsupported by default.
985 for (MVT ValVT : MVT::all_valuetypes())
986 for (MVT MemVT : MVT::all_valuetypes())
988 Expand);
989
990 // We're somewhat special casing MVT::i2 and MVT::i4. Ideally we want to
991 // remove this and targets should individually set these types if not legal.
994 for (MVT VT : {MVT::i2, MVT::i4})
995 OpActions[(unsigned)VT.SimpleTy][NT] = Expand;
996 }
997 for (MVT AVT : MVT::all_valuetypes()) {
998 for (MVT VT : {MVT::i2, MVT::i4, MVT::v128i2, MVT::v64i4}) {
999 setTruncStoreAction(AVT, VT, Expand);
1002 }
1003 }
1004 for (unsigned IM = (unsigned)ISD::PRE_INC;
1005 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
1006 for (MVT VT : {MVT::i2, MVT::i4}) {
1011 }
1012 }
1013
1014 for (MVT VT : MVT::fp_valuetypes()) {
1015 MVT IntVT = MVT::getIntegerVT(VT.getFixedSizeInBits());
1016 if (IntVT.isValid()) {
1017 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote);
1018 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT);
1019 }
1020 }
1021
1022 // Set default actions for various operations.
1023 for (MVT VT : MVT::all_valuetypes()) {
1024 // Default all indexed load / store to expand.
1025 for (unsigned IM = (unsigned)ISD::PRE_INC;
1026 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
1031 }
1032
1033 // Most backends expect to see the node which just returns the value loaded.
1034 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
1035
1036 // These operations default to expand.
1038 ISD::FMINNUM, ISD::FMAXNUM,
1039 ISD::FMINNUM_IEEE, ISD::FMAXNUM_IEEE,
1040 ISD::FMINIMUM, ISD::FMAXIMUM,
1041 ISD::FMINIMUMNUM, ISD::FMAXIMUMNUM,
1054 ISD::IS_FPCLASS, ISD::FCBRT,
1055 ISD::FLOG, ISD::FLOG2,
1056 ISD::FLOG10, ISD::FEXP,
1057 ISD::FEXP2, ISD::FEXP10,
1058 ISD::FFLOOR, ISD::FNEARBYINT,
1059 ISD::FCEIL, ISD::FRINT,
1060 ISD::FTRUNC, ISD::FROUNDEVEN,
1061 ISD::FTAN, ISD::FACOS,
1062 ISD::FASIN, ISD::FATAN,
1063 ISD::FCOSH, ISD::FSINH,
1064 ISD::FTANH, ISD::FATAN2,
1065 ISD::FMULADD},
1066 VT, Expand);
1067
1068 // Overflow operations default to expand
1071 VT, Expand);
1072
1073 // Carry-using overflow operations default to expand.
1076 VT, Expand);
1077
1078 // ADDC/ADDE/SUBC/SUBE default to expand.
1080 Expand);
1081
1082 // [US]CMP default to expand
1084
1085 // Halving adds
1088 Expand);
1089
1090 // Absolute difference
1092
1093 // Saturated trunc
1097
1098 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
1100 Expand);
1101
1103
1104 // These library functions default to expand.
1105 setOperationAction({ISD::FROUND, ISD::FPOWI, ISD::FLDEXP, ISD::FFREXP,
1106 ISD::FSINCOS, ISD::FSINCOSPI, ISD::FMODF},
1107 VT, Expand);
1108
1109 // These operations default to expand for vector types.
1110 if (VT.isVector())
1115 ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND},
1116 VT, Expand);
1117
1118 // Constrained floating-point operations default to expand.
1119#define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
1120 setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
1121#include "llvm/IR/ConstrainedOps.def"
1122
1123 // For most targets @llvm.get.dynamic.area.offset just returns 0.
1124 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
1125
1126 // Vector reduction default to expand.
1128 {ISD::VECREDUCE_FADD, ISD::VECREDUCE_FMUL, ISD::VECREDUCE_ADD,
1129 ISD::VECREDUCE_MUL, ISD::VECREDUCE_AND, ISD::VECREDUCE_OR,
1130 ISD::VECREDUCE_XOR, ISD::VECREDUCE_SMAX, ISD::VECREDUCE_SMIN,
1131 ISD::VECREDUCE_UMAX, ISD::VECREDUCE_UMIN, ISD::VECREDUCE_FMAX,
1132 ISD::VECREDUCE_FMIN, ISD::VECREDUCE_FMAXIMUM, ISD::VECREDUCE_FMINIMUM,
1133 ISD::VECREDUCE_SEQ_FADD, ISD::VECREDUCE_SEQ_FMUL},
1134 VT, Expand);
1135
1136 // Named vector shuffles default to expand.
1138
1139 // Only some target support this vector operation. Most need to expand it.
1141
1142 // VP operations default to expand.
1143#define BEGIN_REGISTER_VP_SDNODE(SDOPC, ...) \
1144 setOperationAction(ISD::SDOPC, VT, Expand);
1145#include "llvm/IR/VPIntrinsics.def"
1146
1147 // Masked vector extracts default to expand.
1148 setOperationAction(ISD::VECTOR_FIND_LAST_ACTIVE, VT, Expand);
1149
1152
1153 // FP environment operations default to expand.
1154 setOperationAction(ISD::GET_FPENV, VT, Expand);
1155 setOperationAction(ISD::SET_FPENV, VT, Expand);
1156 setOperationAction(ISD::RESET_FPENV, VT, Expand);
1157
1158 setOperationAction(ISD::MSTORE, VT, Expand);
1159 }
1160
1161 // Most targets ignore the @llvm.prefetch intrinsic.
1162 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
1163
1164 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
1165 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
1166
1167 // Most targets also ignore the @llvm.readsteadycounter intrinsic.
1168 setOperationAction(ISD::READSTEADYCOUNTER, MVT::i64, Expand);
1169
1170 // ConstantFP nodes default to expand. Targets can either change this to
1171 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
1172 // to optimize expansions for certain constants.
1174 {MVT::bf16, MVT::f16, MVT::f32, MVT::f64, MVT::f80, MVT::f128},
1175 Expand);
1176
1177 // Insert custom handling default for llvm.canonicalize.*.
1179 {MVT::f16, MVT::f32, MVT::f64, MVT::f128}, Expand);
1180
1181 // FIXME: Query RuntimeLibCalls to make the decision.
1182 setOperationAction({ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND},
1183 {MVT::f32, MVT::f64, MVT::f128}, LibCall);
1184
1185 setOperationAction({ISD::FTAN, ISD::FACOS, ISD::FASIN, ISD::FATAN, ISD::FCOSH,
1186 ISD::FSINH, ISD::FTANH, ISD::FATAN2},
1187 MVT::f16, Promote);
1188 // Default ISD::TRAP to expand (which turns it into abort).
1189 setOperationAction(ISD::TRAP, MVT::Other, Expand);
1190
1191 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
1192 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
1193 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
1194
1195 setOperationAction(ISD::UBSANTRAP, MVT::Other, Expand);
1196
1197 setOperationAction(ISD::GET_FPENV_MEM, MVT::Other, Expand);
1198 setOperationAction(ISD::SET_FPENV_MEM, MVT::Other, Expand);
1199
1200 for (MVT VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) {
1201 setOperationAction(ISD::GET_FPMODE, VT, Expand);
1202 setOperationAction(ISD::SET_FPMODE, VT, Expand);
1203 }
1204 setOperationAction(ISD::RESET_FPMODE, MVT::Other, Expand);
1205
1206 // This one by default will call __clear_cache unless the target
1207 // wants something different.
1209}
1210
1212 EVT) const {
1213 return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
1214}
1215
1217 const DataLayout &DL) const {
1218 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
1219 if (LHSTy.isVector())
1220 return LHSTy;
1221 MVT ShiftVT = getScalarShiftAmountTy(DL, LHSTy);
1222 // If any possible shift value won't fit in the prefered type, just use
1223 // something safe. Assume it will be legalized when the shift is expanded.
1224 if (ShiftVT.getSizeInBits() < Log2_32_Ceil(LHSTy.getSizeInBits()))
1225 ShiftVT = MVT::i32;
1226 assert(ShiftVT.getSizeInBits() >= Log2_32_Ceil(LHSTy.getSizeInBits()) &&
1227 "ShiftVT is still too small!");
1228 return ShiftVT;
1229}
1230
1231bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
1232 assert(isTypeLegal(VT));
1233 switch (Op) {
1234 default:
1235 return false;
1236 case ISD::SDIV:
1237 case ISD::UDIV:
1238 case ISD::SREM:
1239 case ISD::UREM:
1240 return true;
1241 }
1242}
1243
1245 unsigned DestAS) const {
1246 return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
1247}
1248
1250 Type *RetTy, ElementCount EC, bool ZeroIsPoison,
1251 const ConstantRange *VScaleRange) const {
1252 // Find the smallest "sensible" element type to use for the expansion.
1253 ConstantRange CR(APInt(64, EC.getKnownMinValue()));
1254 if (EC.isScalable())
1255 CR = CR.umul_sat(*VScaleRange);
1256
1257 if (ZeroIsPoison)
1258 CR = CR.subtract(APInt(64, 1));
1259
1260 unsigned EltWidth = RetTy->getScalarSizeInBits();
1261 EltWidth = std::min(EltWidth, CR.getActiveBits());
1262 EltWidth = std::max(llvm::bit_ceil(EltWidth), (unsigned)8);
1263
1264 return EltWidth;
1265}
1266
1268 // If the command-line option was specified, ignore this request.
1269 if (!JumpIsExpensiveOverride.getNumOccurrences())
1270 JumpIsExpensive = isExpensive;
1271}
1272
1275 // If this is a simple type, use the ComputeRegisterProp mechanism.
1276 if (VT.isSimple()) {
1277 MVT SVT = VT.getSimpleVT();
1278 assert((unsigned)SVT.SimpleTy < std::size(TransformToType));
1279 MVT NVT = TransformToType[SVT.SimpleTy];
1280 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
1281
1282 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
1283 LA == TypeSoftPromoteHalf ||
1284 (NVT.isVector() ||
1285 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
1286 "Promote may not follow Expand or Promote");
1287
1288 if (LA == TypeSplitVector)
1289 return LegalizeKind(LA, EVT(SVT).getHalfNumVectorElementsVT(Context));
1290 if (LA == TypeScalarizeVector)
1291 return LegalizeKind(LA, SVT.getVectorElementType());
1292 return LegalizeKind(LA, NVT);
1293 }
1294
1295 // Handle Extended Scalar Types.
1296 if (!VT.isVector()) {
1297 assert(VT.isInteger() && "Float types must be simple");
1298 unsigned BitSize = VT.getSizeInBits();
1299 // First promote to a power-of-two size, then expand if necessary.
1300 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
1301 EVT NVT = VT.getRoundIntegerType(Context);
1302 assert(NVT != VT && "Unable to round integer VT");
1303 LegalizeKind NextStep = getTypeConversion(Context, NVT);
1304 // Avoid multi-step promotion.
1305 if (NextStep.first == TypePromoteInteger)
1306 return NextStep;
1307 // Return rounded integer type.
1308 return LegalizeKind(TypePromoteInteger, NVT);
1309 }
1310
1312 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
1313 }
1314
1315 // Handle vector types.
1316 ElementCount NumElts = VT.getVectorElementCount();
1317 EVT EltVT = VT.getVectorElementType();
1318
1319 // Vectors with only one element are always scalarized.
1320 if (NumElts.isScalar())
1321 return LegalizeKind(TypeScalarizeVector, EltVT);
1322
1323 // Try to widen vector elements until the element type is a power of two and
1324 // promote it to a legal type later on, for example:
1325 // <3 x i8> -> <4 x i8> -> <4 x i32>
1326 if (EltVT.isInteger()) {
1327 // Vectors with a number of elements that is not a power of two are always
1328 // widened, for example <3 x i8> -> <4 x i8>.
1329 if (!VT.isPow2VectorType()) {
1330 NumElts = NumElts.coefficientNextPowerOf2();
1331 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
1332 return LegalizeKind(TypeWidenVector, NVT);
1333 }
1334
1335 // Examine the element type.
1336 LegalizeKind LK = getTypeConversion(Context, EltVT);
1337
1338 // If type is to be expanded, split the vector.
1339 // <4 x i140> -> <2 x i140>
1340 if (LK.first == TypeExpandInteger) {
1341 if (NumElts.isScalable() && NumElts.getKnownMinValue() == 1)
1344 VT.getHalfNumVectorElementsVT(Context));
1345 }
1346
1347 // Promote the integer element types until a legal vector type is found
1348 // or until the element integer type is too big. If a legal type was not
1349 // found, fallback to the usual mechanism of widening/splitting the
1350 // vector.
1351 EVT OldEltVT = EltVT;
1352 while (true) {
1353 // Increase the bitwidth of the element to the next pow-of-two
1354 // (which is greater than 8 bits).
1355 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
1356 .getRoundIntegerType(Context);
1357
1358 // Stop trying when getting a non-simple element type.
1359 // Note that vector elements may be greater than legal vector element
1360 // types. Example: X86 XMM registers hold 64bit element on 32bit
1361 // systems.
1362 if (!EltVT.isSimple())
1363 break;
1364
1365 // Build a new vector type and check if it is legal.
1366 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1367 // Found a legal promoted vector type.
1368 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
1370 EVT::getVectorVT(Context, EltVT, NumElts));
1371 }
1372
1373 // Reset the type to the unexpanded type if we did not find a legal vector
1374 // type with a promoted vector element type.
1375 EltVT = OldEltVT;
1376 }
1377
1378 // Try to widen the vector until a legal type is found.
1379 // If there is no wider legal type, split the vector.
1380 while (true) {
1381 // Round up to the next power of 2.
1382 NumElts = NumElts.coefficientNextPowerOf2();
1383
1384 // If there is no simple vector type with this many elements then there
1385 // cannot be a larger legal vector type. Note that this assumes that
1386 // there are no skipped intermediate vector types in the simple types.
1387 if (!EltVT.isSimple())
1388 break;
1389 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1390 if (LargerVector == MVT())
1391 break;
1392
1393 // If this type is legal then widen the vector.
1394 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1395 return LegalizeKind(TypeWidenVector, LargerVector);
1396 }
1397
1398 // Widen odd vectors to next power of two.
1399 if (!VT.isPow2VectorType()) {
1400 EVT NVT = VT.getPow2VectorType(Context);
1401 return LegalizeKind(TypeWidenVector, NVT);
1402 }
1403
1406
1407 // Vectors with illegal element types are expanded.
1408 EVT NVT = EVT::getVectorVT(Context, EltVT,
1410 return LegalizeKind(TypeSplitVector, NVT);
1411}
1412
1413static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
1414 unsigned &NumIntermediates,
1415 MVT &RegisterVT,
1416 TargetLoweringBase *TLI) {
1417 // Figure out the right, legal destination reg to copy into.
1419 MVT EltTy = VT.getVectorElementType();
1420
1421 unsigned NumVectorRegs = 1;
1422
1423 // Scalable vectors cannot be scalarized, so splitting or widening is
1424 // required.
1425 if (VT.isScalableVector() && !isPowerOf2_32(EC.getKnownMinValue()))
1427 "Splitting or widening of non-power-of-2 MVTs is not implemented.");
1428
1429 // FIXME: We don't support non-power-of-2-sized vectors for now.
1430 // Ideally we could break down into LHS/RHS like LegalizeDAG does.
1431 if (!isPowerOf2_32(EC.getKnownMinValue())) {
1432 // Split EC to unit size (scalable property is preserved).
1433 NumVectorRegs = EC.getKnownMinValue();
1434 EC = ElementCount::getFixed(1);
1435 }
1436
1437 // Divide the input until we get to a supported size. This will
1438 // always end up with an EC that represent a scalar or a scalable
1439 // scalar.
1440 while (EC.getKnownMinValue() > 1 &&
1441 !TLI->isTypeLegal(MVT::getVectorVT(EltTy, EC))) {
1442 EC = EC.divideCoefficientBy(2);
1443 NumVectorRegs <<= 1;
1444 }
1445
1446 NumIntermediates = NumVectorRegs;
1447
1448 MVT NewVT = MVT::getVectorVT(EltTy, EC);
1449 if (!TLI->isTypeLegal(NewVT))
1450 NewVT = EltTy;
1451 IntermediateVT = NewVT;
1452
1453 unsigned LaneSizeInBits = NewVT.getScalarSizeInBits();
1454
1455 // Convert sizes such as i33 to i64.
1456 LaneSizeInBits = llvm::bit_ceil(LaneSizeInBits);
1457
1458 MVT DestVT = TLI->getRegisterType(NewVT);
1459 RegisterVT = DestVT;
1460 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1461 return NumVectorRegs * (LaneSizeInBits / DestVT.getScalarSizeInBits());
1462
1463 // Otherwise, promotion or legal types use the same number of registers as
1464 // the vector decimated to the appropriate level.
1465 return NumVectorRegs;
1466}
1467
1468/// isLegalRC - Return true if the value types that can be represented by the
1469/// specified register class are all legal.
1471 const TargetRegisterClass &RC) const {
1472 for (const auto *I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1473 if (isTypeLegal(*I))
1474 return true;
1475 return false;
1476}
1477
1478/// Replace/modify any TargetFrameIndex operands with a targte-dependent
1479/// sequence of memory operands that is recognized by PrologEpilogInserter.
1482 MachineBasicBlock *MBB) const {
1483 MachineInstr *MI = &InitialMI;
1484 MachineFunction &MF = *MI->getMF();
1485 MachineFrameInfo &MFI = MF.getFrameInfo();
1486
1487 // We're handling multiple types of operands here:
1488 // PATCHPOINT MetaArgs - live-in, read only, direct
1489 // STATEPOINT Deopt Spill - live-through, read only, indirect
1490 // STATEPOINT Deopt Alloca - live-through, read only, direct
1491 // (We're currently conservative and mark the deopt slots read/write in
1492 // practice.)
1493 // STATEPOINT GC Spill - live-through, read/write, indirect
1494 // STATEPOINT GC Alloca - live-through, read/write, direct
1495 // The live-in vs live-through is handled already (the live through ones are
1496 // all stack slots), but we need to handle the different type of stackmap
1497 // operands and memory effects here.
1498
1499 if (llvm::none_of(MI->operands(),
1500 [](MachineOperand &Operand) { return Operand.isFI(); }))
1501 return MBB;
1502
1503 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1504
1505 // Inherit previous memory operands.
1506 MIB.cloneMemRefs(*MI);
1507
1508 for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1509 MachineOperand &MO = MI->getOperand(i);
1510 if (!MO.isFI()) {
1511 // Index of Def operand this Use it tied to.
1512 // Since Defs are coming before Uses, if Use is tied, then
1513 // index of Def must be smaller that index of that Use.
1514 // Also, Defs preserve their position in new MI.
1515 unsigned TiedTo = i;
1516 if (MO.isReg() && MO.isTied())
1517 TiedTo = MI->findTiedOperandIdx(i);
1518 MIB.add(MO);
1519 if (TiedTo < i)
1520 MIB->tieOperands(TiedTo, MIB->getNumOperands() - 1);
1521 continue;
1522 }
1523
1524 // foldMemoryOperand builds a new MI after replacing a single FI operand
1525 // with the canonical set of five x86 addressing-mode operands.
1526 int FI = MO.getIndex();
1527
1528 // Add frame index operands recognized by stackmaps.cpp
1530 // indirect-mem-ref tag, size, #FI, offset.
1531 // Used for spills inserted by StatepointLowering. This codepath is not
1532 // used for patchpoints/stackmaps at all, for these spilling is done via
1533 // foldMemoryOperand callback only.
1534 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1535 MIB.addImm(StackMaps::IndirectMemRefOp);
1536 MIB.addImm(MFI.getObjectSize(FI));
1537 MIB.add(MO);
1538 MIB.addImm(0);
1539 } else {
1540 // direct-mem-ref tag, #FI, offset.
1541 // Used by patchpoint, and direct alloca arguments to statepoints
1542 MIB.addImm(StackMaps::DirectMemRefOp);
1543 MIB.add(MO);
1544 MIB.addImm(0);
1545 }
1546
1547 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1548
1549 // Add a new memory operand for this FI.
1550 assert(MFI.getObjectOffset(FI) != -1);
1551
1552 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and
1553 // PATCHPOINT should be updated to do the same. (TODO)
1554 if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1555 auto Flags = MachineMemOperand::MOLoad;
1557 MachinePointerInfo::getFixedStack(MF, FI), Flags,
1559 MIB->addMemOperand(MF, MMO);
1560 }
1561 }
1562 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1563 MI->eraseFromParent();
1564 return MBB;
1565}
1566
1567/// findRepresentativeClass - Return the largest legal super-reg register class
1568/// of the register class for the specified type and its associated "cost".
1569// This function is in TargetLowering because it uses RegClassForVT which would
1570// need to be moved to TargetRegisterInfo and would necessitate moving
1571// isTypeLegal over as well - a massive change that would just require
1572// TargetLowering having a TargetRegisterInfo class member that it would use.
1573std::pair<const TargetRegisterClass *, uint8_t>
1575 MVT VT) const {
1576 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1577 if (!RC)
1578 return std::make_pair(RC, 0);
1579
1580 // Compute the set of all super-register classes.
1581 BitVector SuperRegRC(TRI->getNumRegClasses());
1582 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1583 SuperRegRC.setBitsInMask(RCI.getMask());
1584
1585 // Find the first legal register class with the largest spill size.
1586 const TargetRegisterClass *BestRC = RC;
1587 for (unsigned i : SuperRegRC.set_bits()) {
1588 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1589 // We want the largest possible spill size.
1590 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1591 continue;
1592 if (!isLegalRC(*TRI, *SuperRC))
1593 continue;
1594 BestRC = SuperRC;
1595 }
1596 return std::make_pair(BestRC, 1);
1597}
1598
1599/// computeRegisterProperties - Once all of the register classes are added,
1600/// this allows us to compute derived properties we expose.
1602 const TargetRegisterInfo *TRI) {
1603 // Everything defaults to needing one register.
1604 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1605 NumRegistersForVT[i] = 1;
1606 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1607 }
1608 // ...except isVoid, which doesn't need any registers.
1609 NumRegistersForVT[MVT::isVoid] = 0;
1610
1611 // Find the largest integer register class.
1612 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1613 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1614 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1615
1616 // Every integer value type larger than this largest register takes twice as
1617 // many registers to represent as the previous ValueType.
1618 for (unsigned ExpandedReg = LargestIntReg + 1;
1619 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1620 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1621 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1622 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1623 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1625 }
1626
1627 // Inspect all of the ValueType's smaller than the largest integer
1628 // register to see which ones need promotion.
1629 unsigned LegalIntReg = LargestIntReg;
1630 for (unsigned IntReg = LargestIntReg - 1;
1631 IntReg >= (unsigned)MVT::i1; --IntReg) {
1632 MVT IVT = (MVT::SimpleValueType)IntReg;
1633 if (isTypeLegal(IVT)) {
1634 LegalIntReg = IntReg;
1635 } else {
1636 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1637 (MVT::SimpleValueType)LegalIntReg;
1638 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1639 }
1640 }
1641
1642 // ppcf128 type is really two f64's.
1643 if (!isTypeLegal(MVT::ppcf128)) {
1644 if (isTypeLegal(MVT::f64)) {
1645 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1646 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1647 TransformToType[MVT::ppcf128] = MVT::f64;
1648 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1649 } else {
1650 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1651 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1652 TransformToType[MVT::ppcf128] = MVT::i128;
1653 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1654 }
1655 }
1656
1657 // Decide how to handle f128. If the target does not have native f128 support,
1658 // expand it to i128 and we will be generating soft float library calls.
1659 if (!isTypeLegal(MVT::f128)) {
1660 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1661 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1662 TransformToType[MVT::f128] = MVT::i128;
1663 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1664 }
1665
1666 // Decide how to handle f80. If the target does not have native f80 support,
1667 // expand it to i96 and we will be generating soft float library calls.
1668 if (!isTypeLegal(MVT::f80)) {
1669 NumRegistersForVT[MVT::f80] = 3*NumRegistersForVT[MVT::i32];
1670 RegisterTypeForVT[MVT::f80] = RegisterTypeForVT[MVT::i32];
1671 TransformToType[MVT::f80] = MVT::i32;
1672 ValueTypeActions.setTypeAction(MVT::f80, TypeSoftenFloat);
1673 }
1674
1675 // Decide how to handle f64. If the target does not have native f64 support,
1676 // expand it to i64 and we will be generating soft float library calls.
1677 if (!isTypeLegal(MVT::f64)) {
1678 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1679 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1680 TransformToType[MVT::f64] = MVT::i64;
1681 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1682 }
1683
1684 // Decide how to handle f32. If the target does not have native f32 support,
1685 // expand it to i32 and we will be generating soft float library calls.
1686 if (!isTypeLegal(MVT::f32)) {
1687 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1688 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1689 TransformToType[MVT::f32] = MVT::i32;
1690 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1691 }
1692
1693 // Decide how to handle f16. If the target does not have native f16 support,
1694 // promote it to f32, because there are no f16 library calls (except for
1695 // conversions).
1696 if (!isTypeLegal(MVT::f16)) {
1697 // Allow targets to control how we legalize half.
1698 bool SoftPromoteHalfType = softPromoteHalfType();
1699 bool UseFPRegsForHalfType = !SoftPromoteHalfType || useFPRegsForHalfType();
1700
1701 if (!UseFPRegsForHalfType) {
1702 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1703 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1704 } else {
1705 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1706 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1707 }
1708 TransformToType[MVT::f16] = MVT::f32;
1709 if (SoftPromoteHalfType) {
1710 ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1711 } else {
1712 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1713 }
1714 }
1715
1716 // Decide how to handle bf16. If the target does not have native bf16 support,
1717 // promote it to f32, because there are no bf16 library calls (except for
1718 // converting from f32 to bf16).
1719 if (!isTypeLegal(MVT::bf16)) {
1720 NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32];
1721 RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32];
1722 TransformToType[MVT::bf16] = MVT::f32;
1723 ValueTypeActions.setTypeAction(MVT::bf16, TypeSoftPromoteHalf);
1724 }
1725
1726 // Loop over all of the vector value types to see which need transformations.
1727 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1728 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1729 MVT VT = (MVT::SimpleValueType) i;
1730 if (isTypeLegal(VT))
1731 continue;
1732
1733 MVT EltVT = VT.getVectorElementType();
1735 bool IsLegalWiderType = false;
1736 bool IsScalable = VT.isScalableVector();
1737 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1738 switch (PreferredAction) {
1739 case TypePromoteInteger: {
1740 MVT::SimpleValueType EndVT = IsScalable ?
1741 MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1742 MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1743 // Try to promote the elements of integer vectors. If no legal
1744 // promotion was found, fall through to the widen-vector method.
1745 for (unsigned nVT = i + 1;
1746 (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1747 MVT SVT = (MVT::SimpleValueType) nVT;
1748 // Promote vectors of integers to vectors with the same number
1749 // of elements, with a wider element type.
1750 if (SVT.getScalarSizeInBits() > EltVT.getFixedSizeInBits() &&
1751 SVT.getVectorElementCount() == EC && isTypeLegal(SVT)) {
1752 TransformToType[i] = SVT;
1753 RegisterTypeForVT[i] = SVT;
1754 NumRegistersForVT[i] = 1;
1755 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1756 IsLegalWiderType = true;
1757 break;
1758 }
1759 }
1760 if (IsLegalWiderType)
1761 break;
1762 [[fallthrough]];
1763 }
1764
1765 case TypeWidenVector:
1766 if (isPowerOf2_32(EC.getKnownMinValue())) {
1767 // Try to widen the vector.
1768 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1769 MVT SVT = (MVT::SimpleValueType) nVT;
1770 if (SVT.getVectorElementType() == EltVT &&
1771 SVT.isScalableVector() == IsScalable &&
1773 EC.getKnownMinValue() &&
1774 isTypeLegal(SVT)) {
1775 TransformToType[i] = SVT;
1776 RegisterTypeForVT[i] = SVT;
1777 NumRegistersForVT[i] = 1;
1778 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1779 IsLegalWiderType = true;
1780 break;
1781 }
1782 }
1783 if (IsLegalWiderType)
1784 break;
1785 } else {
1786 // Only widen to the next power of 2 to keep consistency with EVT.
1787 MVT NVT = VT.getPow2VectorType();
1788 if (isTypeLegal(NVT)) {
1789 TransformToType[i] = NVT;
1790 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1791 RegisterTypeForVT[i] = NVT;
1792 NumRegistersForVT[i] = 1;
1793 break;
1794 }
1795 }
1796 [[fallthrough]];
1797
1798 case TypeSplitVector:
1799 case TypeScalarizeVector: {
1800 MVT IntermediateVT;
1801 MVT RegisterVT;
1802 unsigned NumIntermediates;
1803 unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1804 NumIntermediates, RegisterVT, this);
1805 NumRegistersForVT[i] = NumRegisters;
1806 assert(NumRegistersForVT[i] == NumRegisters &&
1807 "NumRegistersForVT size cannot represent NumRegisters!");
1808 RegisterTypeForVT[i] = RegisterVT;
1809
1810 MVT NVT = VT.getPow2VectorType();
1811 if (NVT == VT) {
1812 // Type is already a power of 2. The default action is to split.
1813 TransformToType[i] = MVT::Other;
1814 if (PreferredAction == TypeScalarizeVector)
1815 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1816 else if (PreferredAction == TypeSplitVector)
1817 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1818 else if (EC.getKnownMinValue() > 1)
1819 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1820 else
1821 ValueTypeActions.setTypeAction(VT, EC.isScalable()
1824 } else {
1825 TransformToType[i] = NVT;
1826 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1827 }
1828 break;
1829 }
1830 default:
1831 llvm_unreachable("Unknown vector legalization action!");
1832 }
1833 }
1834
1835 // Determine the 'representative' register class for each value type.
1836 // An representative register class is the largest (meaning one which is
1837 // not a sub-register class / subreg register class) legal register class for
1838 // a group of value types. For example, on i386, i8, i16, and i32
1839 // representative would be GR32; while on x86_64 it's GR64.
1840 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1841 const TargetRegisterClass* RRC;
1842 uint8_t Cost;
1844 RepRegClassForVT[i] = RRC;
1845 RepRegClassCostForVT[i] = Cost;
1846 }
1847}
1848
1850 EVT VT) const {
1851 assert(!VT.isVector() && "No default SetCC type for vectors!");
1852 return getPointerTy(DL).SimpleTy;
1853}
1854
1856 return MVT::i32; // return the default value
1857}
1858
1859/// getVectorTypeBreakdown - Vector types are broken down into some number of
1860/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1861/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1862/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1863///
1864/// This method returns the number of registers needed, and the VT for each
1865/// register. It also returns the VT and quantity of the intermediate values
1866/// before they are promoted/expanded.
1868 EVT VT, EVT &IntermediateVT,
1869 unsigned &NumIntermediates,
1870 MVT &RegisterVT) const {
1871 ElementCount EltCnt = VT.getVectorElementCount();
1872
1873 // If there is a wider vector type with the same element type as this one,
1874 // or a promoted vector type that has the same number of elements which
1875 // are wider, then we should convert to that legal vector type.
1876 // This handles things like <2 x float> -> <4 x float> and
1877 // <4 x i1> -> <4 x i32>.
1878 LegalizeTypeAction TA = getTypeAction(Context, VT);
1879 if (!EltCnt.isScalar() &&
1880 (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1881 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1882 if (isTypeLegal(RegisterEVT)) {
1883 IntermediateVT = RegisterEVT;
1884 RegisterVT = RegisterEVT.getSimpleVT();
1885 NumIntermediates = 1;
1886 return 1;
1887 }
1888 }
1889
1890 // Figure out the right, legal destination reg to copy into.
1891 EVT EltTy = VT.getVectorElementType();
1892
1893 unsigned NumVectorRegs = 1;
1894
1895 // Scalable vectors cannot be scalarized, so handle the legalisation of the
1896 // types like done elsewhere in SelectionDAG.
1897 if (EltCnt.isScalable()) {
1898 LegalizeKind LK;
1899 EVT PartVT = VT;
1900 do {
1901 // Iterate until we've found a legal (part) type to hold VT.
1902 LK = getTypeConversion(Context, PartVT);
1903 PartVT = LK.second;
1904 } while (LK.first != TypeLegal);
1905
1906 if (!PartVT.isVector()) {
1908 "Don't know how to legalize this scalable vector type");
1909 }
1910
1911 NumIntermediates =
1914 IntermediateVT = PartVT;
1915 RegisterVT = getRegisterType(Context, IntermediateVT);
1916 return NumIntermediates;
1917 }
1918
1919 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally
1920 // we could break down into LHS/RHS like LegalizeDAG does.
1921 if (!isPowerOf2_32(EltCnt.getKnownMinValue())) {
1922 NumVectorRegs = EltCnt.getKnownMinValue();
1923 EltCnt = ElementCount::getFixed(1);
1924 }
1925
1926 // Divide the input until we get to a supported size. This will always
1927 // end with a scalar if the target doesn't support vectors.
1928 while (EltCnt.getKnownMinValue() > 1 &&
1929 !isTypeLegal(EVT::getVectorVT(Context, EltTy, EltCnt))) {
1930 EltCnt = EltCnt.divideCoefficientBy(2);
1931 NumVectorRegs <<= 1;
1932 }
1933
1934 NumIntermediates = NumVectorRegs;
1935
1936 EVT NewVT = EVT::getVectorVT(Context, EltTy, EltCnt);
1937 if (!isTypeLegal(NewVT))
1938 NewVT = EltTy;
1939 IntermediateVT = NewVT;
1940
1941 MVT DestVT = getRegisterType(Context, NewVT);
1942 RegisterVT = DestVT;
1943
1944 if (EVT(DestVT).bitsLT(NewVT)) { // Value is expanded, e.g. i64 -> i16.
1945 TypeSize NewVTSize = NewVT.getSizeInBits();
1946 // Convert sizes such as i33 to i64.
1948 NewVTSize = NewVTSize.coefficientNextPowerOf2();
1949 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1950 }
1951
1952 // Otherwise, promotion or legal types use the same number of registers as
1953 // the vector decimated to the appropriate level.
1954 return NumVectorRegs;
1955}
1956
1958 uint64_t NumCases,
1960 ProfileSummaryInfo *PSI,
1961 BlockFrequencyInfo *BFI) const {
1962 // FIXME: This function check the maximum table size and density, but the
1963 // minimum size is not checked. It would be nice if the minimum size is
1964 // also combined within this function. Currently, the minimum size check is
1965 // performed in findJumpTable() in SelectionDAGBuiler and
1966 // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
1967 const bool OptForSize =
1968 llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
1969 const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
1970 const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
1971
1972 // Check whether the number of cases is small enough and
1973 // the range is dense enough for a jump table.
1974 return (OptForSize || Range <= MaxJumpTableSize) &&
1975 (NumCases * 100 >= Range * MinDensity);
1976}
1977
1979 EVT ConditionVT) const {
1980 return getRegisterType(Context, ConditionVT);
1981}
1982
1983/// Get the EVTs and ArgFlags collections that represent the legalized return
1984/// type of the given function. This does not require a DAG or a return value,
1985/// and is suitable for use before any DAGs for the function are constructed.
1986/// TODO: Move this out of TargetLowering.cpp.
1988 AttributeList attr,
1990 const TargetLowering &TLI, const DataLayout &DL) {
1992 ComputeValueTypes(DL, ReturnType, Types);
1993 unsigned NumValues = Types.size();
1994 if (NumValues == 0) return;
1995
1996 for (Type *Ty : Types) {
1997 EVT VT = TLI.getValueType(DL, Ty);
1998 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1999
2000 if (attr.hasRetAttr(Attribute::SExt))
2001 ExtendKind = ISD::SIGN_EXTEND;
2002 else if (attr.hasRetAttr(Attribute::ZExt))
2003 ExtendKind = ISD::ZERO_EXTEND;
2004
2005 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
2006 VT = TLI.getTypeForExtReturn(ReturnType->getContext(), VT, ExtendKind);
2007
2008 unsigned NumParts =
2009 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
2010 MVT PartVT =
2011 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
2012
2013 // 'inreg' on function refers to return value
2015 if (attr.hasRetAttr(Attribute::InReg))
2016 Flags.setInReg();
2017
2018 // Propagate extension type if any
2019 if (attr.hasRetAttr(Attribute::SExt))
2020 Flags.setSExt();
2021 else if (attr.hasRetAttr(Attribute::ZExt))
2022 Flags.setZExt();
2023
2024 for (unsigned i = 0; i < NumParts; ++i)
2025 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, Ty, 0, 0));
2026 }
2027}
2028
2030 const DataLayout &DL) const {
2031 return DL.getABITypeAlign(Ty);
2032}
2033
2035 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
2036 Align Alignment, MachineMemOperand::Flags Flags, unsigned *Fast) const {
2037 // Check if the specified alignment is sufficient based on the data layout.
2038 // TODO: While using the data layout works in practice, a better solution
2039 // would be to implement this check directly (make this a virtual function).
2040 // For example, the ABI alignment may change based on software platform while
2041 // this function should only be affected by hardware implementation.
2042 Type *Ty = VT.getTypeForEVT(Context);
2043 if (VT.isZeroSized() || Alignment >= DL.getABITypeAlign(Ty)) {
2044 // Assume that an access that meets the ABI-specified alignment is fast.
2045 if (Fast != nullptr)
2046 *Fast = 1;
2047 return true;
2048 }
2049
2050 // This is a misaligned access.
2051 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast);
2052}
2053
2055 LLVMContext &Context, const DataLayout &DL, EVT VT,
2056 const MachineMemOperand &MMO, unsigned *Fast) const {
2057 return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
2058 MMO.getAlign(), MMO.getFlags(), Fast);
2059}
2060
2062 const DataLayout &DL, EVT VT,
2063 unsigned AddrSpace, Align Alignment,
2065 unsigned *Fast) const {
2066 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
2067 Flags, Fast);
2068}
2069
2071 const DataLayout &DL, EVT VT,
2072 const MachineMemOperand &MMO,
2073 unsigned *Fast) const {
2074 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
2075 MMO.getFlags(), Fast);
2076}
2077
2079 const DataLayout &DL, LLT Ty,
2080 const MachineMemOperand &MMO,
2081 unsigned *Fast) const {
2082 EVT VT = getApproximateEVTForLLT(Ty, Context);
2083 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
2084 MMO.getFlags(), Fast);
2085}
2086
2087//===----------------------------------------------------------------------===//
2088// TargetTransformInfo Helpers
2089//===----------------------------------------------------------------------===//
2090
2092 enum InstructionOpcodes {
2093#define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
2094#define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
2095#include "llvm/IR/Instruction.def"
2096 };
2097 switch (static_cast<InstructionOpcodes>(Opcode)) {
2098 case Ret: return 0;
2099 case Br: return 0;
2100 case Switch: return 0;
2101 case IndirectBr: return 0;
2102 case Invoke: return 0;
2103 case CallBr: return 0;
2104 case Resume: return 0;
2105 case Unreachable: return 0;
2106 case CleanupRet: return 0;
2107 case CatchRet: return 0;
2108 case CatchPad: return 0;
2109 case CatchSwitch: return 0;
2110 case CleanupPad: return 0;
2111 case FNeg: return ISD::FNEG;
2112 case Add: return ISD::ADD;
2113 case FAdd: return ISD::FADD;
2114 case Sub: return ISD::SUB;
2115 case FSub: return ISD::FSUB;
2116 case Mul: return ISD::MUL;
2117 case FMul: return ISD::FMUL;
2118 case UDiv: return ISD::UDIV;
2119 case SDiv: return ISD::SDIV;
2120 case FDiv: return ISD::FDIV;
2121 case URem: return ISD::UREM;
2122 case SRem: return ISD::SREM;
2123 case FRem: return ISD::FREM;
2124 case Shl: return ISD::SHL;
2125 case LShr: return ISD::SRL;
2126 case AShr: return ISD::SRA;
2127 case And: return ISD::AND;
2128 case Or: return ISD::OR;
2129 case Xor: return ISD::XOR;
2130 case Alloca: return 0;
2131 case Load: return ISD::LOAD;
2132 case Store: return ISD::STORE;
2133 case GetElementPtr: return 0;
2134 case Fence: return 0;
2135 case AtomicCmpXchg: return 0;
2136 case AtomicRMW: return 0;
2137 case Trunc: return ISD::TRUNCATE;
2138 case ZExt: return ISD::ZERO_EXTEND;
2139 case SExt: return ISD::SIGN_EXTEND;
2140 case FPToUI: return ISD::FP_TO_UINT;
2141 case FPToSI: return ISD::FP_TO_SINT;
2142 case UIToFP: return ISD::UINT_TO_FP;
2143 case SIToFP: return ISD::SINT_TO_FP;
2144 case FPTrunc: return ISD::FP_ROUND;
2145 case FPExt: return ISD::FP_EXTEND;
2146 case PtrToAddr: return ISD::BITCAST;
2147 case PtrToInt: return ISD::BITCAST;
2148 case IntToPtr: return ISD::BITCAST;
2149 case BitCast: return ISD::BITCAST;
2150 case AddrSpaceCast: return ISD::ADDRSPACECAST;
2151 case ICmp: return ISD::SETCC;
2152 case FCmp: return ISD::SETCC;
2153 case PHI: return 0;
2154 case Call: return 0;
2155 case Select: return ISD::SELECT;
2156 case UserOp1: return 0;
2157 case UserOp2: return 0;
2158 case VAArg: return 0;
2159 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
2160 case InsertElement: return ISD::INSERT_VECTOR_ELT;
2161 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
2162 case ExtractValue: return ISD::MERGE_VALUES;
2163 case InsertValue: return ISD::MERGE_VALUES;
2164 case LandingPad: return 0;
2165 case Freeze: return ISD::FREEZE;
2166 }
2167
2168 llvm_unreachable("Unknown instruction type encountered!");
2169}
2170
2172 switch (ID) {
2173 case Intrinsic::exp:
2174 return ISD::FEXP;
2175 case Intrinsic::exp2:
2176 return ISD::FEXP2;
2177 case Intrinsic::log:
2178 return ISD::FLOG;
2179 default:
2180 return ISD::DELETED_NODE;
2181 }
2182}
2183
2184Value *
2186 bool UseTLS) const {
2187 // compiler-rt provides a variable with a magic name. Targets that do not
2188 // link with compiler-rt may also provide such a variable.
2189 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
2190 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
2191 auto UnsafeStackPtr =
2192 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
2193
2194 const DataLayout &DL = M->getDataLayout();
2195 PointerType *StackPtrTy = DL.getAllocaPtrType(M->getContext());
2196
2197 if (!UnsafeStackPtr) {
2198 auto TLSModel = UseTLS ?
2201 // The global variable is not defined yet, define it ourselves.
2202 // We use the initial-exec TLS model because we do not support the
2203 // variable living anywhere other than in the main executable.
2204 UnsafeStackPtr = new GlobalVariable(
2205 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
2206 UnsafeStackPtrVar, nullptr, TLSModel);
2207 } else {
2208 // The variable exists, check its type and attributes.
2209 //
2210 // FIXME: Move to IR verifier.
2211 if (UnsafeStackPtr->getValueType() != StackPtrTy)
2212 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
2213 if (UseTLS != UnsafeStackPtr->isThreadLocal())
2214 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
2215 (UseTLS ? "" : "not ") + "be thread-local");
2216 }
2217 return UnsafeStackPtr;
2218}
2219
2220Value *
2222 // FIXME: Can this triple check be replaced with SAFESTACK_POINTER_ADDRESS
2223 // being available?
2224 if (!TM.getTargetTriple().isAndroid())
2225 return getDefaultSafeStackPointerLocation(IRB, true);
2226
2227 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
2228 auto *PtrTy = PointerType::getUnqual(M->getContext());
2229
2230 const char *SafestackPointerAddressName =
2231 getLibcallName(RTLIB::SAFESTACK_POINTER_ADDRESS);
2232 if (!SafestackPointerAddressName) {
2233 M->getContext().emitError(
2234 "no libcall available for safestack pointer address");
2235 return PoisonValue::get(PtrTy);
2236 }
2237
2238 // Android provides a libc function to retrieve the address of the current
2239 // thread's unsafe stack pointer.
2240 FunctionCallee Fn =
2241 M->getOrInsertFunction(SafestackPointerAddressName, PtrTy);
2242 return IRB.CreateCall(Fn);
2243}
2244
2245//===----------------------------------------------------------------------===//
2246// Loop Strength Reduction hooks
2247//===----------------------------------------------------------------------===//
2248
2249/// isLegalAddressingMode - Return true if the addressing mode represented
2250/// by AM is legal for this target, for a load/store of the specified type.
2252 const AddrMode &AM, Type *Ty,
2253 unsigned AS, Instruction *I) const {
2254 // The default implementation of this implements a conservative RISCy, r+r and
2255 // r+i addr mode.
2256
2257 // Scalable offsets not supported
2258 if (AM.ScalableOffset)
2259 return false;
2260
2261 // Allows a sign-extended 16-bit immediate field.
2262 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2263 return false;
2264
2265 // No global is ever allowed as a base.
2266 if (AM.BaseGV)
2267 return false;
2268
2269 // Only support r+r,
2270 switch (AM.Scale) {
2271 case 0: // "r+i" or just "i", depending on HasBaseReg.
2272 break;
2273 case 1:
2274 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2275 return false;
2276 // Otherwise we have r+r or r+i.
2277 break;
2278 case 2:
2279 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2280 return false;
2281 // Allow 2*r as r+r.
2282 break;
2283 default: // Don't allow n * r
2284 return false;
2285 }
2286
2287 return true;
2288}
2289
2290//===----------------------------------------------------------------------===//
2291// Stack Protector
2292//===----------------------------------------------------------------------===//
2293
2294// For OpenBSD return its special guard variable. Otherwise return nullptr,
2295// so that SelectionDAG handle SSP.
2297 RTLIB::LibcallImpl GuardLocalImpl = getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2298 if (GuardLocalImpl != RTLIB::impl___guard_local)
2299 return nullptr;
2300
2301 Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
2302 const DataLayout &DL = M.getDataLayout();
2303 PointerType *PtrTy =
2304 PointerType::get(M.getContext(), DL.getDefaultGlobalsAddressSpace());
2305 GlobalVariable *G =
2306 M.getOrInsertGlobal(getLibcallImplName(GuardLocalImpl), PtrTy);
2307 G->setVisibility(GlobalValue::HiddenVisibility);
2308 return G;
2309}
2310
2311// Currently only support "standard" __stack_chk_guard.
2312// TODO: add LOAD_STACK_GUARD support.
2314 RTLIB::LibcallImpl StackGuardImpl = getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2315 if (StackGuardImpl == RTLIB::Unsupported)
2316 return;
2317
2318 StringRef StackGuardVarName = getLibcallImplName(StackGuardImpl);
2319 M.getOrInsertGlobal(
2320 StackGuardVarName, PointerType::getUnqual(M.getContext()), [=, &M]() {
2321 auto *GV = new GlobalVariable(M, PointerType::getUnqual(M.getContext()),
2322 false, GlobalVariable::ExternalLinkage,
2323 nullptr, StackGuardVarName);
2324
2325 // FreeBSD has "__stack_chk_guard" defined externally on libc.so
2326 if (M.getDirectAccessExternalData() &&
2327 !TM.getTargetTriple().isOSCygMing() &&
2328 !(TM.getTargetTriple().isPPC64() &&
2329 TM.getTargetTriple().isOSFreeBSD()) &&
2330 (!TM.getTargetTriple().isOSDarwin() ||
2331 TM.getRelocationModel() == Reloc::Static))
2332 GV->setDSOLocal(true);
2333
2334 return GV;
2335 });
2336}
2337
2338// Currently only support "standard" __stack_chk_guard.
2339// TODO: add LOAD_STACK_GUARD support.
2341 RTLIB::LibcallImpl GuardVarImpl = getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2342 if (GuardVarImpl == RTLIB::Unsupported)
2343 return nullptr;
2344 return M.getNamedValue(getLibcallImplName(GuardVarImpl));
2345}
2346
2348 // MSVC CRT has a function to validate security cookie.
2349 RTLIB::LibcallImpl SecurityCheckCookieLibcall =
2350 getLibcallImpl(RTLIB::SECURITY_CHECK_COOKIE);
2351 if (SecurityCheckCookieLibcall != RTLIB::Unsupported)
2352 return M.getFunction(getLibcallImplName(SecurityCheckCookieLibcall));
2353 return nullptr;
2354}
2355
2359
2363
2364unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
2365 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
2366}
2367
2371
2375
2379
2381 return MinimumBitTestCmps;
2382}
2383
2385 MinimumBitTestCmps = Val;
2386}
2387
2389 if (TM.Options.LoopAlignment)
2390 return Align(TM.Options.LoopAlignment);
2391 return PrefLoopAlignment;
2392}
2393
2395 MachineBasicBlock *MBB) const {
2396 return MaxBytesForAlignment;
2397}
2398
2399//===----------------------------------------------------------------------===//
2400// Reciprocal Estimates
2401//===----------------------------------------------------------------------===//
2402
2403/// Get the reciprocal estimate attribute string for a function that will
2404/// override the target defaults.
2406 const Function &F = MF.getFunction();
2407 return F.getFnAttribute("reciprocal-estimates").getValueAsString();
2408}
2409
2410/// Construct a string for the given reciprocal operation of the given type.
2411/// This string should match the corresponding option to the front-end's
2412/// "-mrecip" flag assuming those strings have been passed through in an
2413/// attribute string. For example, "vec-divf" for a division of a vXf32.
2414static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
2415 std::string Name = VT.isVector() ? "vec-" : "";
2416
2417 Name += IsSqrt ? "sqrt" : "div";
2418
2419 // TODO: Handle other float types?
2420 if (VT.getScalarType() == MVT::f64) {
2421 Name += "d";
2422 } else if (VT.getScalarType() == MVT::f16) {
2423 Name += "h";
2424 } else {
2425 assert(VT.getScalarType() == MVT::f32 &&
2426 "Unexpected FP type for reciprocal estimate");
2427 Name += "f";
2428 }
2429
2430 return Name;
2431}
2432
2433/// Return the character position and value (a single numeric character) of a
2434/// customized refinement operation in the input string if it exists. Return
2435/// false if there is no customized refinement step count.
2436static bool parseRefinementStep(StringRef In, size_t &Position,
2437 uint8_t &Value) {
2438 const char RefStepToken = ':';
2439 Position = In.find(RefStepToken);
2440 if (Position == StringRef::npos)
2441 return false;
2442
2443 StringRef RefStepString = In.substr(Position + 1);
2444 // Allow exactly one numeric character for the additional refinement
2445 // step parameter.
2446 if (RefStepString.size() == 1) {
2447 char RefStepChar = RefStepString[0];
2448 if (isDigit(RefStepChar)) {
2449 Value = RefStepChar - '0';
2450 return true;
2451 }
2452 }
2453 report_fatal_error("Invalid refinement step for -recip.");
2454}
2455
2456/// For the input attribute string, return one of the ReciprocalEstimate enum
2457/// status values (enabled, disabled, or not specified) for this operation on
2458/// the specified data type.
2459static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
2460 if (Override.empty())
2462
2463 SmallVector<StringRef, 4> OverrideVector;
2464 Override.split(OverrideVector, ',');
2465 unsigned NumArgs = OverrideVector.size();
2466
2467 // Check if "all", "none", or "default" was specified.
2468 if (NumArgs == 1) {
2469 // Look for an optional setting of the number of refinement steps needed
2470 // for this type of reciprocal operation.
2471 size_t RefPos;
2472 uint8_t RefSteps;
2473 if (parseRefinementStep(Override, RefPos, RefSteps)) {
2474 // Split the string for further processing.
2475 Override = Override.substr(0, RefPos);
2476 }
2477
2478 // All reciprocal types are enabled.
2479 if (Override == "all")
2481
2482 // All reciprocal types are disabled.
2483 if (Override == "none")
2485
2486 // Target defaults for enablement are used.
2487 if (Override == "default")
2489 }
2490
2491 // The attribute string may omit the size suffix ('f'/'d').
2492 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2493 std::string VTNameNoSize = VTName;
2494 VTNameNoSize.pop_back();
2495 static const char DisabledPrefix = '!';
2496
2497 for (StringRef RecipType : OverrideVector) {
2498 size_t RefPos;
2499 uint8_t RefSteps;
2500 if (parseRefinementStep(RecipType, RefPos, RefSteps))
2501 RecipType = RecipType.substr(0, RefPos);
2502
2503 // Ignore the disablement token for string matching.
2504 bool IsDisabled = RecipType[0] == DisabledPrefix;
2505 if (IsDisabled)
2506 RecipType = RecipType.substr(1);
2507
2508 if (RecipType == VTName || RecipType == VTNameNoSize)
2511 }
2512
2514}
2515
2516/// For the input attribute string, return the customized refinement step count
2517/// for this operation on the specified data type. If the step count does not
2518/// exist, return the ReciprocalEstimate enum value for unspecified.
2519static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2520 if (Override.empty())
2522
2523 SmallVector<StringRef, 4> OverrideVector;
2524 Override.split(OverrideVector, ',');
2525 unsigned NumArgs = OverrideVector.size();
2526
2527 // Check if "all", "default", or "none" was specified.
2528 if (NumArgs == 1) {
2529 // Look for an optional setting of the number of refinement steps needed
2530 // for this type of reciprocal operation.
2531 size_t RefPos;
2532 uint8_t RefSteps;
2533 if (!parseRefinementStep(Override, RefPos, RefSteps))
2535
2536 // Split the string for further processing.
2537 Override = Override.substr(0, RefPos);
2538 assert(Override != "none" &&
2539 "Disabled reciprocals, but specifed refinement steps?");
2540
2541 // If this is a general override, return the specified number of steps.
2542 if (Override == "all" || Override == "default")
2543 return RefSteps;
2544 }
2545
2546 // The attribute string may omit the size suffix ('f'/'d').
2547 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2548 std::string VTNameNoSize = VTName;
2549 VTNameNoSize.pop_back();
2550
2551 for (StringRef RecipType : OverrideVector) {
2552 size_t RefPos;
2553 uint8_t RefSteps;
2554 if (!parseRefinementStep(RecipType, RefPos, RefSteps))
2555 continue;
2556
2557 RecipType = RecipType.substr(0, RefPos);
2558 if (RecipType == VTName || RecipType == VTNameNoSize)
2559 return RefSteps;
2560 }
2561
2563}
2564
2569
2574
2579
2584
2586 EVT LoadVT, EVT BitcastVT, const SelectionDAG &DAG,
2587 const MachineMemOperand &MMO) const {
2588 // Single-element vectors are scalarized, so we should generally avoid having
2589 // any memory operations on such types, as they would get scalarized too.
2590 if (LoadVT.isFixedLengthVector() && BitcastVT.isFixedLengthVector() &&
2591 BitcastVT.getVectorNumElements() == 1)
2592 return false;
2593
2594 // Don't do if we could do an indexed load on the original type, but not on
2595 // the new one.
2596 if (!LoadVT.isSimple() || !BitcastVT.isSimple())
2597 return true;
2598
2599 MVT LoadMVT = LoadVT.getSimpleVT();
2600
2601 // Don't bother doing this if it's just going to be promoted again later, as
2602 // doing so might interfere with other combines.
2603 if (getOperationAction(ISD::LOAD, LoadMVT) == Promote &&
2604 getTypeToPromoteTo(ISD::LOAD, LoadMVT) == BitcastVT.getSimpleVT())
2605 return false;
2606
2607 unsigned Fast = 0;
2608 return allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), BitcastVT,
2609 MMO, &Fast) &&
2610 Fast;
2611}
2612
2616
2618 const LoadInst &LI, const DataLayout &DL, AssumptionCache *AC,
2619 const TargetLibraryInfo *LibInfo) const {
2621 if (LI.isVolatile())
2623
2624 if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2626
2627 if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2629
2631 LI.getAlign(), DL, &LI, AC,
2632 /*DT=*/nullptr, LibInfo))
2634
2635 Flags |= getTargetMMOFlags(LI);
2636 return Flags;
2637}
2638
2641 const DataLayout &DL) const {
2643
2644 if (SI.isVolatile())
2646
2647 if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2649
2650 // FIXME: Not preserving dereferenceable
2651 Flags |= getTargetMMOFlags(SI);
2652 return Flags;
2653}
2654
2657 const DataLayout &DL) const {
2659
2660 if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2661 if (RMW->isVolatile())
2663 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2664 if (CmpX->isVolatile())
2666 } else
2667 llvm_unreachable("not an atomic instruction");
2668
2669 // FIXME: Not preserving dereferenceable
2670 Flags |= getTargetMMOFlags(AI);
2671 return Flags;
2672}
2673
2675 const VPIntrinsic &VPIntrin) const {
2677 Intrinsic::ID IntrinID = VPIntrin.getIntrinsicID();
2678
2679 switch (IntrinID) {
2680 default:
2681 llvm_unreachable("unexpected intrinsic. Existing code may be appropriate "
2682 "for it, but support must be explicitly enabled");
2683 case Intrinsic::vp_load:
2684 case Intrinsic::vp_gather:
2685 case Intrinsic::experimental_vp_strided_load:
2687 break;
2688 case Intrinsic::vp_store:
2689 case Intrinsic::vp_scatter:
2690 case Intrinsic::experimental_vp_strided_store:
2692 break;
2693 }
2694
2695 if (VPIntrin.hasMetadata(LLVMContext::MD_nontemporal))
2697
2698 Flags |= getTargetMMOFlags(VPIntrin);
2699 return Flags;
2700}
2701
2703 Instruction *Inst,
2704 AtomicOrdering Ord) const {
2705 if (isReleaseOrStronger(Ord) && Inst->hasAtomicStore())
2706 return Builder.CreateFence(Ord);
2707 else
2708 return nullptr;
2709}
2710
2712 Instruction *Inst,
2713 AtomicOrdering Ord) const {
2714 if (isAcquireOrStronger(Ord))
2715 return Builder.CreateFence(Ord);
2716 else
2717 return nullptr;
2718}
2719
2720//===----------------------------------------------------------------------===//
2721// GlobalISel Hooks
2722//===----------------------------------------------------------------------===//
2723
2725 const TargetTransformInfo *TTI) const {
2726 auto &MF = *MI.getMF();
2727 auto &MRI = MF.getRegInfo();
2728 // Assuming a spill and reload of a value has a cost of 1 instruction each,
2729 // this helper function computes the maximum number of uses we should consider
2730 // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2731 // break even in terms of code size when the original MI has 2 users vs
2732 // choosing to potentially spill. Any more than 2 users we we have a net code
2733 // size increase. This doesn't take into account register pressure though.
2734 auto maxUses = [](unsigned RematCost) {
2735 // A cost of 1 means remats are basically free.
2736 if (RematCost == 1)
2737 return std::numeric_limits<unsigned>::max();
2738 if (RematCost == 2)
2739 return 2U;
2740
2741 // Remat is too expensive, only sink if there's one user.
2742 if (RematCost > 2)
2743 return 1U;
2744 llvm_unreachable("Unexpected remat cost");
2745 };
2746
2747 switch (MI.getOpcode()) {
2748 default:
2749 return false;
2750 // Constants-like instructions should be close to their users.
2751 // We don't want long live-ranges for them.
2752 case TargetOpcode::G_CONSTANT:
2753 case TargetOpcode::G_FCONSTANT:
2754 case TargetOpcode::G_FRAME_INDEX:
2755 case TargetOpcode::G_INTTOPTR:
2756 return true;
2757 case TargetOpcode::G_GLOBAL_VALUE: {
2758 unsigned RematCost = TTI->getGISelRematGlobalCost();
2759 Register Reg = MI.getOperand(0).getReg();
2760 unsigned MaxUses = maxUses(RematCost);
2761 if (MaxUses == UINT_MAX)
2762 return true; // Remats are "free" so always localize.
2763 return MRI.hasAtMostUserInstrs(Reg, MaxUses);
2764 }
2765 }
2766}
unsigned const MachineRegisterInfo * MRI
assert(UImm &&(UImm !=~static_cast< T >(0)) &&"Invalid immediate!")
AMDGPU Register Bank Select
Rewrite undef for PHI
MachineBasicBlock & MBB
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
This file contains the simple types necessary to represent the attributes associated with functions a...
This file implements the BitVector class.
#define LLVM_ABI
Definition Compiler.h:213
This file defines the DenseMap class.
IRTranslator LLVM IR MI
Module.h This file contains the declarations for the Module class.
static LVOptions Options
Definition LVOptions.cpp:25
#define F(x, y, z)
Definition MD5.cpp:54
#define I(x, y, z)
Definition MD5.cpp:57
#define G(x, y, z)
Definition MD5.cpp:55
Register const TargetRegisterInfo * TRI
ConstantRange Range(APInt(BitWidth, Low), APInt(BitWidth, High))
This file contains some templates that are useful if you are working with the STL at all.
This file defines the SmallVector class.
This file contains some functions that are useful when dealing with strings.
static cl::opt< unsigned > MinimumBitTestCmpsOverride("min-bit-test-cmps", cl::init(2), cl::Hidden, cl::desc("Set minimum of largest number of comparisons " "to use bit test for switch."))
static cl::opt< bool > JumpIsExpensiveOverride("jump-is-expensive", cl::init(false), cl::desc("Do not create extra branches to split comparison logic."), cl::Hidden)
#define OP_TO_LIBCALL(Name, Enum)
static cl::opt< unsigned > MinimumJumpTableEntries("min-jump-table-entries", cl::init(4), cl::Hidden, cl::desc("Set minimum number of entries to use a jump table."))
static cl::opt< bool > DisableStrictNodeMutation("disable-strictnode-mutation", cl::desc("Don't mutate strict-float node to a legalize node"), cl::init(false), cl::Hidden)
static bool parseRefinementStep(StringRef In, size_t &Position, uint8_t &Value)
Return the character position and value (a single numeric character) of a customized refinement opera...
static cl::opt< unsigned > MaximumJumpTableSize("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden, cl::desc("Set maximum size of jump tables."))
static cl::opt< unsigned > JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, cl::desc("Minimum density for building a jump table in " "a normal function"))
Minimum jump table density for normal functions.
static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, unsigned &NumIntermediates, MVT &RegisterVT, TargetLoweringBase *TLI)
static std::string getReciprocalOpName(bool IsSqrt, EVT VT)
Construct a string for the given reciprocal operation of the given type.
#define LCALL5(A)
static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override)
For the input attribute string, return the customized refinement step count for this operation on the...
static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override)
For the input attribute string, return one of the ReciprocalEstimate enum status values (enabled,...
static StringRef getRecipEstimateForFunc(MachineFunction &MF)
Get the reciprocal estimate attribute string for a function that will override the target defaults.
static cl::opt< unsigned > OptsizeJumpTableDensity("optsize-jump-table-density", cl::init(40), cl::Hidden, cl::desc("Minimum density for building a jump table in " "an optsize function"))
Minimum jump table density for -Os or -Oz functions.
This file describes how to lower LLVM code to machine code.
This pass exposes codegen information to IR-level passes.
Class for arbitrary precision integers.
Definition APInt.h:78
A cache of @llvm.assume calls within a function.
An instruction that atomically checks whether a specified value is in a memory location,...
an instruction that atomically reads a memory location, combines it with another value,...
const Function * getParent() const
Return the enclosing method, or null if none.
Definition BasicBlock.h:213
void setBitsInMask(const uint32_t *Mask, unsigned MaskWords=~0u)
setBitsInMask - Add '1' bits from Mask to this vector.
Definition BitVector.h:723
iterator_range< const_set_bits_iterator > set_bits() const
Definition BitVector.h:159
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
This class represents a range of values.
LLVM_ABI unsigned getActiveBits() const
Compute the maximal number of active bits needed to represent every value in this range.
LLVM_ABI ConstantRange umul_sat(const ConstantRange &Other) const
Perform an unsigned saturating multiplication of two constant ranges.
LLVM_ABI ConstantRange subtract(const APInt &CI) const
Subtract the specified constant from the endpoints of this constant range.
A parsed version of the target data layout string in and methods for querying it.
Definition DataLayout.h:63
LLVM_ABI unsigned getPointerSize(unsigned AS=0) const
The pointer representation size in bytes, rounded up to a whole number of bytes.
static constexpr ElementCount getScalable(ScalarTy MinVal)
Definition TypeSize.h:312
static constexpr ElementCount getFixed(ScalarTy MinVal)
Definition TypeSize.h:309
constexpr bool isScalar() const
Exactly one element.
Definition TypeSize.h:320
A handy container for a FunctionType+Callee-pointer pair, which can be passed around as a single enti...
Module * getParent()
Get the module that this global value is contained inside of...
@ HiddenVisibility
The GV is hidden.
Definition GlobalValue.h:69
@ ExternalLinkage
Externally visible function.
Definition GlobalValue.h:53
Common base class shared among various IRBuilders.
Definition IRBuilder.h:114
BasicBlock * GetInsertBlock() const
Definition IRBuilder.h:201
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args={}, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition IRBuilder.h:2511
LLVM_ABI bool hasAtomicStore() const LLVM_READONLY
Return true if this atomic instruction stores to memory.
bool hasMetadata() const
Return true if this instruction has any metadata attached to it.
@ MAX_INT_BITS
Maximum number of bits that can be specified.
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
This is an important class for using LLVM in a threaded context.
Definition LLVMContext.h:68
An instruction for reading from memory.
Value * getPointerOperand()
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Align getAlign() const
Return the alignment of the access that is being performed.
Machine Value Type.
SimpleValueType SimpleTy
uint64_t getScalarSizeInBits() const
bool isVector() const
Return true if this is a vector value type.
bool isScalableVector() const
Return true if this is a vector value type where the runtime length is machine dependent.
static auto all_valuetypes()
SimpleValueType Iteration.
TypeSize getSizeInBits() const
Returns the size of the specified MVT in bits.
uint64_t getFixedSizeInBits() const
Return the size of the specified fixed width value type in bits.
ElementCount getVectorElementCount() const
bool isScalarInteger() const
Return true if this is an integer, not including vectors.
static MVT getVectorVT(MVT VT, unsigned NumElements)
MVT getVectorElementType() const
bool isValid() const
Return true if this is a valid simple valuetype.
static MVT getIntegerVT(unsigned BitWidth)
static auto fp_valuetypes()
MVT getPow2VectorType() const
Widens the length of the given vector MVT up to the nearest power of 2 and returns that type.
MachineInstrBundleIterator< MachineInstr > iterator
The MachineFrameInfo class represents an abstract stack frame until prolog/epilog code is inserted.
bool isStatepointSpillSlotObjectIndex(int ObjectIdx) const
Align getObjectAlign(int ObjectIdx) const
Return the alignment of the specified stack object.
int64_t getObjectSize(int ObjectIdx) const
Return the size of the specified object.
int64_t getObjectOffset(int ObjectIdx) const
Return the assigned stack offset of the specified object from the incoming stack pointer.
MachineMemOperand * getMachineMemOperand(MachinePointerInfo PtrInfo, MachineMemOperand::Flags f, LLT MemTy, Align base_alignment, const AAMDNodes &AAInfo=AAMDNodes(), const MDNode *Ranges=nullptr, SyncScope::ID SSID=SyncScope::System, AtomicOrdering Ordering=AtomicOrdering::NotAtomic, AtomicOrdering FailureOrdering=AtomicOrdering::NotAtomic)
getMachineMemOperand - Allocate a new MachineMemOperand.
MachineFrameInfo & getFrameInfo()
getFrameInfo - Return the frame info object for the current function.
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
const DataLayout & getDataLayout() const
Return the DataLayout attached to the Module associated to this MF.
Function & getFunction()
Return the LLVM function that this machine code represents.
const MachineInstrBuilder & addImm(int64_t Val) const
Add a new immediate operand.
const MachineInstrBuilder & add(const MachineOperand &MO) const
const MachineInstrBuilder & cloneMemRefs(const MachineInstr &OtherMI) const
Representation of each machine instruction.
unsigned getNumOperands() const
Retuns the total number of operands.
bool mayLoad(QueryType Type=AnyInBundle) const
Return true if this instruction could possibly read memory.
LLVM_ABI void tieOperands(unsigned DefIdx, unsigned UseIdx)
Add a tie between the register operands at DefIdx and UseIdx.
LLVM_ABI void addMemOperand(MachineFunction &MF, MachineMemOperand *MO)
Add a MachineMemOperand to the machine instruction.
A description of a memory reference used in the backend.
unsigned getAddrSpace() const
Flags
Flags values. These may be or'd together.
@ MOVolatile
The memory access is volatile.
@ MODereferenceable
The memory access is dereferenceable (i.e., doesn't trap).
@ MOLoad
The memory access reads data.
@ MONonTemporal
The memory access is non-temporal.
@ MOInvariant
The memory access always returns the same value (or traps).
@ MOStore
The memory access writes data.
Flags getFlags() const
Return the raw flags of the source value,.
LLVM_ABI Align getAlign() const
Return the minimum known alignment in bytes of the actual memory reference.
MachineOperand class - Representation of each machine instruction operand.
bool isReg() const
isReg - Tests if this is a MO_Register operand.
bool isFI() const
isFI - Tests if this is a MO_FrameIndex operand.
LLVM_ABI void freezeReservedRegs()
freezeReservedRegs - Called by the register allocator to freeze the set of reserved registers before ...
A Module instance is used to store all the information related to an LLVM module.
Definition Module.h:67
Class to represent pointers.
static PointerType * getUnqual(Type *ElementType)
This constructs a pointer to an object of the specified type in the default address space (address sp...
static LLVM_ABI PointerType * get(Type *ElementType, unsigned AddressSpace)
This constructs a pointer to an object of the specified type in a numbered address space.
static LLVM_ABI PoisonValue * get(Type *T)
Static factory methods - Return an 'poison' object of the specified type.
Analysis providing profile information.
Wrapper class representing virtual and physical registers.
Definition Register.h:20
This is used to represent a portion of an LLVM function in a low-level Data Dependence DAG representa...
const DataLayout & getDataLayout() const
LLVMContext * getContext() const
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
void push_back(const T &Elt)
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
An instruction for storing to memory.
StringRef - Represent a constant reference to a string, i.e.
Definition StringRef.h:55
std::pair< StringRef, StringRef > split(char Separator) const
Split into two substrings around the first occurrence of a separator character.
Definition StringRef.h:702
static constexpr size_t npos
Definition StringRef.h:57
constexpr StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition StringRef.h:573
constexpr bool empty() const
empty - Check if the string is empty.
Definition StringRef.h:143
constexpr size_t size() const
size - Get the string size.
Definition StringRef.h:146
bool isValid() const
Returns true if this iterator is still pointing at a valid entry.
Multiway switch.
Provides information about what library functions are available for the current target.
This base class for TargetLowering contains the SelectionDAG-independent parts that can be used from ...
virtual Align getByValTypeAlignment(Type *Ty, const DataLayout &DL) const
Returns the desired alignment for ByVal or InAlloca aggregate function arguments in the caller parame...
int InstructionOpcodeToISD(unsigned Opcode) const
Get the ISD node that corresponds to the Instruction class opcode.
void setOperationAction(unsigned Op, MVT VT, LegalizeAction Action)
Indicate that the specified operation does not work with the specified type and indicate what to do a...
virtual void finalizeLowering(MachineFunction &MF) const
Execute target specific actions to finalize target lowering.
void initActions()
Initialize all of the actions to default values.
bool PredictableSelectIsExpensive
Tells the code generator that select is more expensive than a branch if the branch is usually predict...
EVT getValueType(const DataLayout &DL, Type *Ty, bool AllowUnknown=false) const
Return the EVT corresponding to this LLVM type.
void setMinimumBitTestCmps(unsigned Val)
Set the minimum of largest of number of comparisons to generate BitTest.
unsigned MaxStoresPerMemcpyOptSize
Likewise for functions with the OptSize attribute.
MachineBasicBlock * emitPatchPoint(MachineInstr &MI, MachineBasicBlock *MBB) const
Replace/modify any TargetFrameIndex operands with a targte-dependent sequence of memory operands that...
virtual Value * getSafeStackPointerLocation(IRBuilderBase &IRB) const
Returns the target-specific address of the unsafe stack pointer.
int getRecipEstimateSqrtEnabled(EVT VT, MachineFunction &MF) const
Return a ReciprocalEstimate enum value for a square root of the given type based on the function's at...
virtual bool canOpTrap(unsigned Op, EVT VT) const
Returns true if the operation can trap for the value type.
virtual bool shouldLocalize(const MachineInstr &MI, const TargetTransformInfo *TTI) const
Check whether or not MI needs to be moved close to its uses.
virtual unsigned getMaxPermittedBytesForAlignment(MachineBasicBlock *MBB) const
Return the maximum amount of bytes allowed to be emitted when padding for alignment.
void setMaximumJumpTableSize(unsigned)
Indicate the maximum number of entries in jump tables.
virtual unsigned getMinimumJumpTableEntries() const
Return lower limit for number of blocks in a jump table.
const TargetMachine & getTargetMachine() const
unsigned MaxLoadsPerMemcmp
Specify maximum number of load instructions per memcmp call.
virtual unsigned getNumRegistersForCallingConv(LLVMContext &Context, CallingConv::ID CC, EVT VT) const
Certain targets require unusual breakdowns of certain types.
virtual MachineMemOperand::Flags getTargetMMOFlags(const Instruction &I) const
This callback is used to inspect load/store instructions and add target-specific MachineMemOperand fl...
unsigned MaxGluedStoresPerMemcpy
Specify max number of store instructions to glue in inlined memcpy.
virtual MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC, EVT VT) const
Certain combinations of ABIs, Targets and features require that types are legal for some operations a...
LegalizeTypeAction
This enum indicates whether a types are legal for a target, and if not, what action should be used to...
virtual bool isSuitableForJumpTable(const SwitchInst *SI, uint64_t NumCases, uint64_t Range, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) const
Return true if lowering to a jump table is suitable for a set of case clusters which may contain NumC...
void setIndexedMaskedLoadAction(unsigned IdxMode, MVT VT, LegalizeAction Action)
Indicate that the specified indexed masked load does or does not work with the specified type and ind...
virtual Value * getSDagStackGuard(const Module &M) const
Return the variable that's previously inserted by insertSSPDeclarations, if any, otherwise return nul...
unsigned getMinimumBitTestCmps() const
Retuen the minimum of largest number of comparisons in BitTest.
virtual bool useFPRegsForHalfType() const
virtual bool isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT, const SelectionDAG &DAG, const MachineMemOperand &MMO) const
Return true if the following transform is beneficial: fold (conv (load x)) -> (load (conv*)x) On arch...
void setIndexedLoadAction(ArrayRef< unsigned > IdxModes, MVT VT, LegalizeAction Action)
Indicate that the specified indexed load does or does not work with the specified type and indicate w...
virtual bool softPromoteHalfType() const
unsigned getMaximumJumpTableSize() const
Return upper limit for number of entries in a jump table.
virtual MVT::SimpleValueType getCmpLibcallReturnType() const
Return the ValueType for comparison libcalls.
unsigned getBitWidthForCttzElements(Type *RetTy, ElementCount EC, bool ZeroIsPoison, const ConstantRange *VScaleRange) const
Return the minimum number of bits required to hold the maximum possible number of trailing zero vecto...
bool isLegalRC(const TargetRegisterInfo &TRI, const TargetRegisterClass &RC) const
Return true if the value types that can be represented by the specified register class are all legal.
virtual TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(MVT VT) const
Return the preferred vector type legalization action.
void setAtomicLoadExtAction(unsigned ExtType, MVT ValVT, MVT MemVT, LegalizeAction Action)
Let target indicate that an extending atomic load of the specified type is legal.
Value * getDefaultSafeStackPointerLocation(IRBuilderBase &IRB, bool UseTLS) const
Function * getSSPStackGuardCheck(const Module &M) const
If the target has a standard stack protection check function that performs validation and error handl...
MachineMemOperand::Flags getAtomicMemOperandFlags(const Instruction &AI, const DataLayout &DL) const
virtual bool allowsMisalignedMemoryAccesses(EVT, unsigned AddrSpace=0, Align Alignment=Align(1), MachineMemOperand::Flags Flags=MachineMemOperand::MONone, unsigned *=nullptr) const
Determine if the target supports unaligned memory accesses.
unsigned MaxStoresPerMemsetOptSize
Likewise for functions with the OptSize attribute.
EVT getShiftAmountTy(EVT LHSTy, const DataLayout &DL) const
Returns the type for the shift amount of a shift opcode.
unsigned MaxStoresPerMemmove
Specify maximum number of store instructions per memmove call.
virtual Align getPrefLoopAlignment(MachineLoop *ML=nullptr) const
Return the preferred loop alignment.
void computeRegisterProperties(const TargetRegisterInfo *TRI)
Once all of the register classes are added, this allows us to compute derived properties we expose.
MachineMemOperand::Flags getVPIntrinsicMemOperandFlags(const VPIntrinsic &VPIntrin) const
int getDivRefinementSteps(EVT VT, MachineFunction &MF) const
Return the refinement step count for a division of the given type based on the function's attributes.
virtual EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, EVT VT) const
Return the ValueType of the result of SETCC operations.
MachineMemOperand::Flags getLoadMemOperandFlags(const LoadInst &LI, const DataLayout &DL, AssumptionCache *AC=nullptr, const TargetLibraryInfo *LibInfo=nullptr) const
virtual EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const
For types supported by the target, this is an identity function.
unsigned MaxStoresPerMemmoveOptSize
Likewise for functions with the OptSize attribute.
virtual Value * getIRStackGuard(IRBuilderBase &IRB) const
If the target has a standard location for the stack protector guard, returns the address of that loca...
virtual MVT getPreferredSwitchConditionType(LLVMContext &Context, EVT ConditionVT) const
Returns preferred type for switch condition.
bool isTypeLegal(EVT VT) const
Return true if the target has native support for the specified value type.
int getRecipEstimateDivEnabled(EVT VT, MachineFunction &MF) const
Return a ReciprocalEstimate enum value for a division of the given type based on the function's attri...
void setIndexedStoreAction(ArrayRef< unsigned > IdxModes, MVT VT, LegalizeAction Action)
Indicate that the specified indexed store does or does not work with the specified type and indicate ...
virtual bool isJumpTableRelative() const
virtual MVT getScalarShiftAmountTy(const DataLayout &, EVT) const
Return the type to use for a scalar shift opcode, given the shifted amount type.
virtual MVT getPointerTy(const DataLayout &DL, uint32_t AS=0) const
Return the pointer type for the given address space, defaults to the pointer type from the data layou...
virtual bool isFreeAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const
Returns true if a cast from SrcAS to DestAS is "cheap", such that e.g.
ISD::CondCode getSoftFloatCmpLibcallPredicate(RTLIB::LibcallImpl Call) const
Get the comparison predicate that's to be used to test the result of the comparison libcall against z...
void setIndexedMaskedStoreAction(unsigned IdxMode, MVT VT, LegalizeAction Action)
Indicate that the specified indexed masked store does or does not work with the specified type and in...
TargetLoweringBase(const TargetMachine &TM, const TargetSubtargetInfo &STI)
NOTE: The TargetMachine owns TLOF.
unsigned MaxStoresPerMemset
Specify maximum number of store instructions per memset call.
void setMinimumJumpTableEntries(unsigned Val)
Indicate the minimum number of blocks to generate jump tables.
void setTruncStoreAction(MVT ValVT, MVT MemVT, LegalizeAction Action)
Indicate that the specified truncating store does not work with the specified type and indicate what ...
virtual bool allowsMemoryAccess(LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace=0, Align Alignment=Align(1), MachineMemOperand::Flags Flags=MachineMemOperand::MONone, unsigned *Fast=nullptr) const
Return true if the target supports a memory access of this type for the given address space and align...
unsigned MaxLoadsPerMemcmpOptSize
Likewise for functions with the OptSize attribute.
MachineMemOperand::Flags getStoreMemOperandFlags(const StoreInst &SI, const DataLayout &DL) const
void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT)
If Opc/OrigVT is specified as being promoted, the promotion code defaults to trying a larger integer/...
unsigned getMinimumJumpTableDensity(bool OptForSize) const
Return lower limit of the density in a jump table.
virtual std::pair< const TargetRegisterClass *, uint8_t > findRepresentativeClass(const TargetRegisterInfo *TRI, MVT VT) const
Return the largest legal super-reg register class of the register class for the specified type and it...
RTLIB::LibcallImpl getLibcallImpl(RTLIB::Libcall Call) const
Get the libcall impl routine name for the specified libcall.
static StringRef getLibcallImplName(RTLIB::LibcallImpl Call)
Get the libcall routine name for the specified libcall implementation.
LegalizeKind getTypeConversion(LLVMContext &Context, EVT VT) const
Return pair that represents the legalization kind (first) that needs to happen to EVT (second) in ord...
void setLoadExtAction(unsigned ExtType, MVT ValVT, MVT MemVT, LegalizeAction Action)
Indicate that the specified load with extension does not work with the specified type and indicate wh...
unsigned GatherAllAliasesMaxDepth
Depth that GatherAllAliases should continue looking for chain dependencies when trying to find a more...
int IntrinsicIDToISD(Intrinsic::ID ID) const
Get the ISD node that corresponds to the Intrinsic ID.
LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const
Return how we should legalize values of this type, either it is already legal (return 'Legal') or we ...
int getSqrtRefinementSteps(EVT VT, MachineFunction &MF) const
Return the refinement step count for a square root of the given type based on the function's attribut...
const char * getLibcallName(RTLIB::Libcall Call) const
Get the libcall routine name for the specified libcall.
bool allowsMemoryAccessForAlignment(LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace=0, Align Alignment=Align(1), MachineMemOperand::Flags Flags=MachineMemOperand::MONone, unsigned *Fast=nullptr) const
This function returns true if the memory access is aligned or if the target allows this specific unal...
virtual Instruction * emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst, AtomicOrdering Ord) const
virtual Instruction * emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst, AtomicOrdering Ord) const
Inserts in the IR a target-specific intrinsic specifying a fence.
unsigned MaxStoresPerMemcpy
Specify maximum number of store instructions per memcpy call.
MVT getRegisterType(MVT VT) const
Return the type of registers that this ValueType will eventually require.
virtual void insertSSPDeclarations(Module &M) const
Inserts necessary declarations for SSP (stack protection) purpose.
void setJumpIsExpensive(bool isExpensive=true)
Tells the code generator not to expand logic operations on comparison predicates into separate sequen...
LegalizeAction getOperationAction(unsigned Op, EVT VT) const
Return how this operation should be treated: either it is legal, needs to be promoted to a larger siz...
MVT getTypeToPromoteTo(unsigned Op, MVT VT) const
If the action for this operation is to promote, this method returns the ValueType to promote to.
virtual bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty, unsigned AddrSpace, Instruction *I=nullptr) const
Return true if the addressing mode represented by AM is legal for this target, for a load/store of th...
unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT, EVT &IntermediateVT, unsigned &NumIntermediates, MVT &RegisterVT) const
Vector types are broken down into some number of legal first class types.
std::pair< LegalizeTypeAction, EVT > LegalizeKind
LegalizeKind holds the legalization kind that needs to happen to EVT in order to type-legalize it.
This class defines information used to lower LLVM code to legal SelectionDAG operators that the targe...
virtual EVT getTypeForExtReturn(LLVMContext &Context, EVT VT, ISD::NodeType) const
Return the type that should be used to zero or sign extend a zeroext/signext integer return value.
Primary interface to the complete machine description for the target machine.
bool isPositionIndependent() const
TargetRegisterInfo base class - We assume that the target defines a static array of TargetRegisterDes...
TargetSubtargetInfo - Generic base class for all target subtargets.
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition Twine.h:82
The instances of the Type class are immutable: once they are created, they are never changed.
Definition Type.h:45
LLVM_ABI unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type.
Definition Type.cpp:230
This is the common base class for vector predication intrinsics.
LLVM Value Representation.
Definition Value.h:75
Type * getType() const
All values are typed, get the type of this value.
Definition Value.h:256
constexpr LeafTy coefficientNextPowerOf2() const
Definition TypeSize.h:260
constexpr bool isScalable() const
Returns whether the quantity is scaled by a runtime quantity (vscale).
Definition TypeSize.h:168
constexpr ScalarTy getKnownMinValue() const
Returns the minimum value this quantity can represent.
Definition TypeSize.h:165
constexpr LeafTy divideCoefficientBy(ScalarTy RHS) const
We do not provide the '/' operator here because division for polynomial types does not work in the sa...
Definition TypeSize.h:252
CallInst * Call
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition CallingConv.h:24
@ Fast
Attempts to make calls as fast as possible (e.g.
Definition CallingConv.h:41
NodeType
ISD::NodeType enum - This enum defines the target-independent operators for a SelectionDAG.
Definition ISDOpcodes.h:41
@ SETCC
SetCC operator - This evaluates to a true value iff the condition is true.
Definition ISDOpcodes.h:807
@ MERGE_VALUES
MERGE_VALUES - This node takes multiple discrete operands and returns them all as its individual resu...
Definition ISDOpcodes.h:256
@ CTLZ_ZERO_UNDEF
Definition ISDOpcodes.h:780
@ DELETED_NODE
DELETED_NODE - This is an illegal value that is used to catch errors.
Definition ISDOpcodes.h:45
@ LOOP_DEPENDENCE_RAW_MASK
@ FGETSIGN
INT = FGETSIGN(FP) - Return the sign bit of the specified floating point value as an integer 0/1 valu...
Definition ISDOpcodes.h:531
@ SMULFIX
RESULT = [US]MULFIX(LHS, RHS, SCALE) - Perform fixed point multiplication on 2 integers with the same...
Definition ISDOpcodes.h:387
@ ADDC
Carry-setting nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:289
@ FMAD
FMAD - Perform a * b + c, while getting the same result as the separately rounded operations.
Definition ISDOpcodes.h:515
@ ADD
Simple integer binary arithmetic operators.
Definition ISDOpcodes.h:259
@ SMULFIXSAT
Same as the corresponding unsaturated fixed point instructions, but the result is clamped between the...
Definition ISDOpcodes.h:393
@ ANY_EXTEND
ANY_EXTEND - Used for integer types. The high bits are undefined.
Definition ISDOpcodes.h:841
@ SINT_TO_FP
[SU]INT_TO_FP - These operators convert integers (whose interpreted sign depends on the first letter)...
Definition ISDOpcodes.h:868
@ CONCAT_VECTORS
CONCAT_VECTORS(VECTOR0, VECTOR1, ...) - Given a number of values of vector type with the same length ...
Definition ISDOpcodes.h:577
@ FADD
Simple binary floating point operators.
Definition ISDOpcodes.h:410
@ ABS
ABS - Determine the unsigned absolute value of a signed integer value of the same bitwidth.
Definition ISDOpcodes.h:744
@ SIGN_EXTEND_VECTOR_INREG
SIGN_EXTEND_VECTOR_INREG(Vector) - This operator represents an in-register sign-extension of the low ...
Definition ISDOpcodes.h:898
@ FMULADD
FMULADD - Performs a * b + c, with, or without, intermediate rounding.
Definition ISDOpcodes.h:521
@ SDIVFIX
RESULT = [US]DIVFIX(LHS, RHS, SCALE) - Perform fixed point division on 2 integers with the same width...
Definition ISDOpcodes.h:400
@ BUILTIN_OP_END
BUILTIN_OP_END - This must be the last enum value in this list.
@ SIGN_EXTEND
Conversion operators.
Definition ISDOpcodes.h:832
@ AVGCEILS
AVGCEILS/AVGCEILU - Rounding averaging add - Add two integers using an integer of type i[N+2],...
Definition ISDOpcodes.h:712
@ CTTZ_ZERO_UNDEF
Bit counting operators with an undefined result for zero inputs.
Definition ISDOpcodes.h:779
@ TRUNCATE_SSAT_U
Definition ISDOpcodes.h:861
@ SETCCCARRY
Like SetCC, ops #0 and #1 are the LHS and RHS operands to compare, but op #2 is a boolean indicating ...
Definition ISDOpcodes.h:815
@ SSUBO
Same for subtraction.
Definition ISDOpcodes.h:347
@ FCANONICALIZE
Returns platform specific canonical encoding of a floating point number.
Definition ISDOpcodes.h:534
@ IS_FPCLASS
Performs a check of floating point class property, defined by IEEE-754.
Definition ISDOpcodes.h:541
@ SSUBSAT
RESULT = [US]SUBSAT(LHS, RHS) - Perform saturation subtraction on 2 integers with the same bit width ...
Definition ISDOpcodes.h:369
@ SELECT
Select(COND, TRUEVAL, FALSEVAL).
Definition ISDOpcodes.h:784
@ SPLAT_VECTOR
SPLAT_VECTOR(VAL) - Returns a vector with the scalar value VAL duplicated in all lanes.
Definition ISDOpcodes.h:669
@ SADDO
RESULT, BOOL = [SU]ADDO(LHS, RHS) - Overflow-aware nodes for addition.
Definition ISDOpcodes.h:343
@ SHL
Shift and rotation operations.
Definition ISDOpcodes.h:762
@ VECTOR_SHUFFLE
VECTOR_SHUFFLE(VEC1, VEC2) - Returns a vector, of the same type as VEC1/VEC2.
Definition ISDOpcodes.h:642
@ EXTRACT_VECTOR_ELT
EXTRACT_VECTOR_ELT(VECTOR, IDX) - Returns a single element from VECTOR identified by the (potentially...
Definition ISDOpcodes.h:569
@ ZERO_EXTEND
ZERO_EXTEND - Used for integer types, zeroing the new bits.
Definition ISDOpcodes.h:838
@ SSHLSAT
RESULT = [US]SHLSAT(LHS, RHS) - Perform saturation left shift.
Definition ISDOpcodes.h:379
@ SMULO
Same for multiplication.
Definition ISDOpcodes.h:351
@ ANY_EXTEND_VECTOR_INREG
ANY_EXTEND_VECTOR_INREG(Vector) - This operator represents an in-register any-extension of the low la...
Definition ISDOpcodes.h:887
@ SIGN_EXTEND_INREG
SIGN_EXTEND_INREG - This operator atomically performs a SHL/SRA pair to sign extend a small value in ...
Definition ISDOpcodes.h:876
@ SMIN
[US]{MIN/MAX} - Binary minimum or maximum of signed or unsigned integers.
Definition ISDOpcodes.h:724
@ SDIVFIXSAT
Same as the corresponding unsaturated fixed point instructions, but the result is clamped between the...
Definition ISDOpcodes.h:406
@ UADDO_CARRY
Carry-using nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:323
@ FP_TO_SINT
FP_TO_[US]INT - Convert a floating point value to a signed or unsigned integer.
Definition ISDOpcodes.h:914
@ AND
Bitwise operators - logical and, logical or, logical xor.
Definition ISDOpcodes.h:736
@ SCMP
[US]CMP - 3-way comparison of signed or unsigned integers.
Definition ISDOpcodes.h:732
@ AVGFLOORS
AVGFLOORS/AVGFLOORU - Averaging add - Add two integers using an integer of type i[N+1],...
Definition ISDOpcodes.h:707
@ ADDE
Carry-using nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:299
@ FREEZE
FREEZE - FREEZE(VAL) returns an arbitrary value if VAL is UNDEF (or is evaluated to UNDEF),...
Definition ISDOpcodes.h:236
@ INSERT_VECTOR_ELT
INSERT_VECTOR_ELT(VECTOR, VAL, IDX) - Returns VECTOR with the element at IDX replaced with VAL.
Definition ISDOpcodes.h:558
@ VECTOR_SPLICE
VECTOR_SPLICE(VEC1, VEC2, IMM) - Returns a subvector of the same type as VEC1/VEC2 from CONCAT_VECTOR...
Definition ISDOpcodes.h:654
@ FP_ROUND
X = FP_ROUND(Y, TRUNC) - Rounding 'Y' from a larger floating point type down to the precision of the ...
Definition ISDOpcodes.h:947
@ VECTOR_COMPRESS
VECTOR_COMPRESS(Vec, Mask, Passthru) consecutively place vector elements based on mask e....
Definition ISDOpcodes.h:696
@ ZERO_EXTEND_VECTOR_INREG
ZERO_EXTEND_VECTOR_INREG(Vector) - This operator represents an in-register zero-extension of the low ...
Definition ISDOpcodes.h:909
@ FP_TO_SINT_SAT
FP_TO_[US]INT_SAT - Convert floating point value in operand 0 to a signed or unsigned scalar integer ...
Definition ISDOpcodes.h:933
@ TRUNCATE
TRUNCATE - Completely drop the high bits.
Definition ISDOpcodes.h:844
@ FCOPYSIGN
FCOPYSIGN(X, Y) - Return the value of X with the sign of Y.
Definition ISDOpcodes.h:527
@ SADDSAT
RESULT = [US]ADDSAT(LHS, RHS) - Perform saturation addition on 2 integers with the same bit width (W)...
Definition ISDOpcodes.h:360
@ TRUNCATE_SSAT_S
TRUNCATE_[SU]SAT_[SU] - Truncate for saturated operand [SU] located in middle, prefix for SAT means i...
Definition ISDOpcodes.h:859
@ ABDS
ABDS/ABDU - Absolute difference - Return the absolute difference between two numbers interpreted as s...
Definition ISDOpcodes.h:719
@ TRUNCATE_USAT_U
Definition ISDOpcodes.h:863
@ SADDO_CARRY
Carry-using overflow-aware nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:333
@ LOOP_DEPENDENCE_WAR_MASK
Set rounding mode.
CondCode
ISD::CondCode enum - These are ordered carefully to make the bitfields below work out,...
static const int LAST_INDEXED_MODE
LLVM_ABI Libcall getPOWI(EVT RetVT)
getPOWI - Return the POWI_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getSINTTOFP(EVT OpVT, EVT RetVT)
getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getUREM(EVT VT)
LLVM_ABI Libcall getSHL(EVT VT)
LLVM_ABI Libcall getSYNC(unsigned Opc, MVT VT)
Return the SYNC_FETCH_AND_* value for the given opcode and type, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getLDEXP(EVT RetVT)
getLDEXP - Return the LDEXP_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getUINTTOFP(EVT OpVT, EVT RetVT)
getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getFREXP(EVT RetVT)
getFREXP - Return the FREXP_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getSINCOSPI(EVT RetVT)
getSINCOSPI - Return the SINCOSPI_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getSDIV(EVT VT)
LLVM_ABI Libcall getSRL(EVT VT)
LLVM_ABI Libcall getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize)
getMEMCPY_ELEMENT_UNORDERED_ATOMIC - Return MEMCPY_ELEMENT_UNORDERED_ATOMIC_* value for the given ele...
LLVM_ABI Libcall getSRA(EVT VT)
LLVM_ABI Libcall getUDIV(EVT VT)
LLVM_ABI Libcall getFPLibCall(EVT VT, Libcall Call_F32, Libcall Call_F64, Libcall Call_F80, Libcall Call_F128, Libcall Call_PPCF128)
GetFPLibCall - Helper to return the right libcall for the given floating point type,...
LLVM_ABI Libcall getFPTOUINT(EVT OpVT, EVT RetVT)
getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getLLROUND(EVT VT)
LLVM_ABI Libcall getCOS(EVT RetVT)
Return the COS_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getLROUND(EVT VT)
LLVM_ABI Libcall getMODF(EVT VT)
getMODF - Return the MODF_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getFPTOSINT(EVT OpVT, EVT RetVT)
getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getLRINT(EVT RetVT)
LLVM_ABI Libcall getOUTLINE_ATOMIC(unsigned Opc, AtomicOrdering Order, MVT VT)
Return the outline atomics value for the given opcode, atomic ordering and type, or UNKNOWN_LIBCALL i...
LLVM_ABI Libcall getLLRINT(EVT RetVT)
LLVM_ABI Libcall getFPEXT(EVT OpVT, EVT RetVT)
getFPEXT - Return the FPEXT_*_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getFPROUND(EVT OpVT, EVT RetVT)
getFPROUND - Return the FPROUND_*_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getSREM(EVT VT)
LLVM_ABI Libcall getSIN(EVT RetVT)
Return the SIN_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize)
getMEMSET_ELEMENT_UNORDERED_ATOMIC - Return MEMSET_ELEMENT_UNORDERED_ATOMIC_* value for the given ele...
LLVM_ABI Libcall getSINCOS_STRET(EVT RetVT)
Return the SINCOS_STRET_ value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getPOW(EVT RetVT)
getPOW - Return the POW_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getOutlineAtomicHelper(const Libcall(&LC)[5][4], AtomicOrdering Order, uint64_t MemSize)
Return the outline atomics value for the given atomic ordering, access size and set of libcalls for a...
LLVM_ABI Libcall getMUL(EVT VT)
LLVM_ABI Libcall getCTPOP(EVT VT)
LLVM_ABI Libcall getSINCOS(EVT RetVT)
getSINCOS - Return the SINCOS_* value for the given types, or UNKNOWN_LIBCALL if there is none.
LLVM_ABI Libcall getMULO(EVT VT)
LLVM_ABI Libcall getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize)
getMEMMOVE_ELEMENT_UNORDERED_ATOMIC - Return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_* value for the given e...
initializer< Ty > init(const Ty &Val)
This is an optimization pass for GlobalISel generic memory operations.
unsigned Log2_32_Ceil(uint32_t Value)
Return the ceil log base 2 of the specified value, 32 if the value is zero.
Definition MathExtras.h:344
void fill(R &&Range, T &&Value)
Provide wrappers to std::fill which take ranges instead of having to pass begin/end explicitly.
Definition STLExtras.h:1757
LLVM_ABI void GetReturnInfo(CallingConv::ID CC, Type *ReturnType, AttributeList attr, SmallVectorImpl< ISD::OutputArg > &Outs, const TargetLowering &TLI, const DataLayout &DL)
Given an LLVM IR type and return type attributes, compute the return value EVTs and flags,...
MachineInstrBuilder BuildMI(MachineFunction &MF, const MIMetadata &MIMD, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
InstructionCost Cost
auto enum_seq(EnumT Begin, EnumT End)
Iterate over an enum type from Begin up to - but not including - End.
Definition Sequence.h:337
decltype(auto) dyn_cast(const From &Val)
dyn_cast<X> - Return the argument parameter cast to the specified type.
Definition Casting.h:643
LLVM_ABI bool isDereferenceableAndAlignedPointer(const Value *V, Type *Ty, Align Alignment, const DataLayout &DL, const Instruction *CtxI=nullptr, AssumptionCache *AC=nullptr, const DominatorTree *DT=nullptr, const TargetLibraryInfo *TLI=nullptr)
Returns true if V is always a dereferenceable pointer with alignment greater or equal than requested.
Definition Loads.cpp:229
LLVM_ABI bool shouldOptimizeForSize(const MachineFunction *MF, ProfileSummaryInfo *PSI, const MachineBlockFrequencyInfo *BFI, PGSOQueryType QueryType=PGSOQueryType::Other)
Returns true if machine function MF is suggested to be size-optimized based on the profile.
constexpr force_iteration_on_noniterable_enum_t force_iteration_on_noniterable_enum
Definition Sequence.h:109
T bit_ceil(T Value)
Returns the smallest integral power of two no smaller than Value if Value is nonzero.
Definition bit.h:345
void ComputeValueTypes(const DataLayout &DL, Type *Ty, SmallVectorImpl< Type * > &Types, SmallVectorImpl< TypeSize > *Offsets=nullptr, TypeSize StartingOffset=TypeSize::getZero())
Given an LLVM IR type, compute non-aggregate subtypes.
Definition Analysis.cpp:72
bool isReleaseOrStronger(AtomicOrdering AO)
auto dyn_cast_or_null(const Y &Val)
Definition Casting.h:753
constexpr bool has_single_bit(T Value) noexcept
Definition bit.h:147
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition MathExtras.h:279
bool none_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::none_of which take ranges instead of having to pass begin/end explicitly.
Definition STLExtras.h:1751
LLVM_ABI void report_fatal_error(Error Err, bool gen_crash_diag=true)
Definition Error.cpp:167
bool isDigit(char C)
Checks if character C is one of the 10 decimal digits.
AtomicOrdering
Atomic ordering for LLVM's memory model.
LLVM_ABI EVT getApproximateEVTForLLT(LLT Ty, LLVMContext &Ctx)
constexpr T divideCeil(U Numerator, V Denominator)
Returns the integer ceil(Numerator / Denominator).
Definition MathExtras.h:394
TargetTransformInfo TTI
@ Mul
Product of integers.
@ Xor
Bitwise or logical XOR of integers.
@ FMul
Product of floats.
@ Sub
Subtraction of integers.
@ Add
Sum of integers.
@ FAdd
Sum of floats.
DWARFExpression::Operation Op
bool isAcquireOrStronger(AtomicOrdering AO)
This struct is a compact representation of a valid (non-zero power of two) alignment.
Definition Alignment.h:39
Extended Value Type.
Definition ValueTypes.h:35
EVT getPow2VectorType(LLVMContext &Context) const
Widens the length of the given vector EVT up to the nearest power of 2 and returns that type.
Definition ValueTypes.h:477
bool isSimple() const
Test if the given EVT is simple (as opposed to being extended).
Definition ValueTypes.h:137
static EVT getVectorVT(LLVMContext &Context, EVT VT, unsigned NumElements, bool IsScalable=false)
Returns the EVT that represents a vector NumElements in length, where each element is of type VT.
Definition ValueTypes.h:74
ElementCount getVectorElementCount() const
Definition ValueTypes.h:350
TypeSize getSizeInBits() const
Return the size of the specified value type in bits.
Definition ValueTypes.h:373
bool isPow2VectorType() const
Returns true if the given vector is a power of 2.
Definition ValueTypes.h:470
MVT getSimpleVT() const
Return the SimpleValueType held in the specified simple EVT.
Definition ValueTypes.h:316
static EVT getIntegerVT(LLVMContext &Context, unsigned BitWidth)
Returns the EVT that represents an integer with the given number of bits.
Definition ValueTypes.h:65
bool isFixedLengthVector() const
Definition ValueTypes.h:181
EVT getRoundIntegerType(LLVMContext &Context) const
Rounds the bit-width of the given integer EVT up to the nearest power of two (and at least to eight),...
Definition ValueTypes.h:419
bool isVector() const
Return true if this is a vector value type.
Definition ValueTypes.h:168
EVT getScalarType() const
If this is a vector type, return the element type, otherwise return this.
Definition ValueTypes.h:323
LLVM_ABI Type * getTypeForEVT(LLVMContext &Context) const
This method returns an LLVM type corresponding to the specified EVT.
EVT getVectorElementType() const
Given a vector type, return the type of each element.
Definition ValueTypes.h:328
unsigned getVectorNumElements() const
Given a vector type, return the number of elements it contains.
Definition ValueTypes.h:336
bool isZeroSized() const
Test if the given EVT has zero size, this will fail if called on a scalable type.
Definition ValueTypes.h:132
EVT getHalfNumVectorElementsVT(LLVMContext &Context) const
Definition ValueTypes.h:453
bool isInteger() const
Return true if this is an integer or a vector integer type.
Definition ValueTypes.h:152
OutputArg - This struct carries flags and a value for a single outgoing (actual) argument or outgoing...
Matching combinators.
static LLVM_ABI MachinePointerInfo getFixedStack(MachineFunction &MF, int FI, int64_t Offset=0)
Return a MachinePointerInfo record that refers to the specified FrameIndex.
static RTLIB::Libcall getLibcallFromImpl(RTLIB::LibcallImpl Impl)
Return the libcall provided by Impl.
This represents an addressing mode of: BaseGV + BaseOffs + BaseReg + Scale*ScaleReg + ScalableOffset*...