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
99 "max-store-memset", cl::init(0), cl::Hidden,
100 cl::desc("Override target's MaxStoresPerMemset and "
101 "MaxStoresPerMemsetOptSize. "
102 "Set to 0 to use the target default."));
103
105 "max-store-memcpy", cl::init(0), cl::Hidden,
106 cl::desc("Override target's MaxStoresPerMemcpy and "
107 "MaxStoresPerMemcpyOptSize. "
108 "Set to 0 to use the target default."));
109
111 "max-store-memmove", cl::init(0), cl::Hidden,
112 cl::desc("Override target's MaxStoresPerMemmove and "
113 "MaxStoresPerMemmoveOptSize. "
114 "Set to 0 to use the target default."));
115
116// FIXME: This option is only to test if the strict fp operation processed
117// correctly by preventing mutating strict fp operation to normal fp operation
118// during development. When the backend supports strict float operation, this
119// option will be meaningless.
120static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation",
121 cl::desc("Don't mutate strict-float node to a legalize node"),
122 cl::init(false), cl::Hidden);
123
124LLVM_ABI RTLIB::Libcall RTLIB::getSHL(EVT VT) {
125 if (VT == MVT::i16)
126 return RTLIB::SHL_I16;
127 if (VT == MVT::i32)
128 return RTLIB::SHL_I32;
129 if (VT == MVT::i64)
130 return RTLIB::SHL_I64;
131 if (VT == MVT::i128)
132 return RTLIB::SHL_I128;
133
134 return RTLIB::UNKNOWN_LIBCALL;
135}
136
137LLVM_ABI RTLIB::Libcall RTLIB::getSRL(EVT VT) {
138 if (VT == MVT::i16)
139 return RTLIB::SRL_I16;
140 if (VT == MVT::i32)
141 return RTLIB::SRL_I32;
142 if (VT == MVT::i64)
143 return RTLIB::SRL_I64;
144 if (VT == MVT::i128)
145 return RTLIB::SRL_I128;
146
147 return RTLIB::UNKNOWN_LIBCALL;
148}
149
150LLVM_ABI RTLIB::Libcall RTLIB::getSRA(EVT VT) {
151 if (VT == MVT::i16)
152 return RTLIB::SRA_I16;
153 if (VT == MVT::i32)
154 return RTLIB::SRA_I32;
155 if (VT == MVT::i64)
156 return RTLIB::SRA_I64;
157 if (VT == MVT::i128)
158 return RTLIB::SRA_I128;
159
160 return RTLIB::UNKNOWN_LIBCALL;
161}
162
163LLVM_ABI RTLIB::Libcall RTLIB::getMUL(EVT VT) {
164 if (VT == MVT::i16)
165 return RTLIB::MUL_I16;
166 if (VT == MVT::i32)
167 return RTLIB::MUL_I32;
168 if (VT == MVT::i64)
169 return RTLIB::MUL_I64;
170 if (VT == MVT::i128)
171 return RTLIB::MUL_I128;
172 return RTLIB::UNKNOWN_LIBCALL;
173}
174
175LLVM_ABI RTLIB::Libcall RTLIB::getMULO(EVT VT) {
176 if (VT == MVT::i32)
177 return RTLIB::MULO_I32;
178 if (VT == MVT::i64)
179 return RTLIB::MULO_I64;
180 if (VT == MVT::i128)
181 return RTLIB::MULO_I128;
182 return RTLIB::UNKNOWN_LIBCALL;
183}
184
185LLVM_ABI RTLIB::Libcall RTLIB::getSDIV(EVT VT) {
186 if (VT == MVT::i16)
187 return RTLIB::SDIV_I16;
188 if (VT == MVT::i32)
189 return RTLIB::SDIV_I32;
190 if (VT == MVT::i64)
191 return RTLIB::SDIV_I64;
192 if (VT == MVT::i128)
193 return RTLIB::SDIV_I128;
194 return RTLIB::UNKNOWN_LIBCALL;
195}
196
197LLVM_ABI RTLIB::Libcall RTLIB::getUDIV(EVT VT) {
198 if (VT == MVT::i16)
199 return RTLIB::UDIV_I16;
200 if (VT == MVT::i32)
201 return RTLIB::UDIV_I32;
202 if (VT == MVT::i64)
203 return RTLIB::UDIV_I64;
204 if (VT == MVT::i128)
205 return RTLIB::UDIV_I128;
206 return RTLIB::UNKNOWN_LIBCALL;
207}
208
209LLVM_ABI RTLIB::Libcall RTLIB::getSREM(EVT VT) {
210 if (VT == MVT::i16)
211 return RTLIB::SREM_I16;
212 if (VT == MVT::i32)
213 return RTLIB::SREM_I32;
214 if (VT == MVT::i64)
215 return RTLIB::SREM_I64;
216 if (VT == MVT::i128)
217 return RTLIB::SREM_I128;
218 return RTLIB::UNKNOWN_LIBCALL;
219}
220
221LLVM_ABI RTLIB::Libcall RTLIB::getUREM(EVT VT) {
222 if (VT == MVT::i16)
223 return RTLIB::UREM_I16;
224 if (VT == MVT::i32)
225 return RTLIB::UREM_I32;
226 if (VT == MVT::i64)
227 return RTLIB::UREM_I64;
228 if (VT == MVT::i128)
229 return RTLIB::UREM_I128;
230 return RTLIB::UNKNOWN_LIBCALL;
231}
232
233LLVM_ABI RTLIB::Libcall RTLIB::getCTPOP(EVT VT) {
234 if (VT == MVT::i32)
235 return RTLIB::CTPOP_I32;
236 if (VT == MVT::i64)
237 return RTLIB::CTPOP_I64;
238 if (VT == MVT::i128)
239 return RTLIB::CTPOP_I128;
240 return RTLIB::UNKNOWN_LIBCALL;
241}
242
243/// GetFPLibCall - Helper to return the right libcall for the given floating
244/// point type, or UNKNOWN_LIBCALL if there is none.
245RTLIB::Libcall RTLIB::getFPLibCall(EVT VT,
246 RTLIB::Libcall Call_F32,
247 RTLIB::Libcall Call_F64,
248 RTLIB::Libcall Call_F80,
249 RTLIB::Libcall Call_F128,
250 RTLIB::Libcall Call_PPCF128) {
251 return
252 VT == MVT::f32 ? Call_F32 :
253 VT == MVT::f64 ? Call_F64 :
254 VT == MVT::f80 ? Call_F80 :
255 VT == MVT::f128 ? Call_F128 :
256 VT == MVT::ppcf128 ? Call_PPCF128 :
257 RTLIB::UNKNOWN_LIBCALL;
258}
259
260/// getFPEXT - Return the FPEXT_*_* value for the given types, or
261/// UNKNOWN_LIBCALL if there is none.
262RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
263 if (OpVT == MVT::f16) {
264 if (RetVT == MVT::f32)
265 return FPEXT_F16_F32;
266 if (RetVT == MVT::f64)
267 return FPEXT_F16_F64;
268 if (RetVT == MVT::f80)
269 return FPEXT_F16_F80;
270 if (RetVT == MVT::f128)
271 return FPEXT_F16_F128;
272 } else if (OpVT == MVT::f32) {
273 if (RetVT == MVT::f64)
274 return FPEXT_F32_F64;
275 if (RetVT == MVT::f128)
276 return FPEXT_F32_F128;
277 if (RetVT == MVT::ppcf128)
278 return FPEXT_F32_PPCF128;
279 } else if (OpVT == MVT::f64) {
280 if (RetVT == MVT::f128)
281 return FPEXT_F64_F128;
282 else if (RetVT == MVT::ppcf128)
283 return FPEXT_F64_PPCF128;
284 } else if (OpVT == MVT::f80) {
285 if (RetVT == MVT::f128)
286 return FPEXT_F80_F128;
287 } else if (OpVT == MVT::bf16) {
288 if (RetVT == MVT::f32)
289 return FPEXT_BF16_F32;
290 }
291
292 return UNKNOWN_LIBCALL;
293}
294
295/// getFPROUND - Return the FPROUND_*_* value for the given types, or
296/// UNKNOWN_LIBCALL if there is none.
297RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
298 if (RetVT == MVT::f16) {
299 if (OpVT == MVT::f32)
300 return FPROUND_F32_F16;
301 if (OpVT == MVT::f64)
302 return FPROUND_F64_F16;
303 if (OpVT == MVT::f80)
304 return FPROUND_F80_F16;
305 if (OpVT == MVT::f128)
306 return FPROUND_F128_F16;
307 if (OpVT == MVT::ppcf128)
308 return FPROUND_PPCF128_F16;
309 } else if (RetVT == MVT::bf16) {
310 if (OpVT == MVT::f32)
311 return FPROUND_F32_BF16;
312 if (OpVT == MVT::f64)
313 return FPROUND_F64_BF16;
314 if (OpVT == MVT::f80)
315 return FPROUND_F80_BF16;
316 if (OpVT == MVT::f128)
317 return FPROUND_F128_BF16;
318 } else if (RetVT == MVT::f32) {
319 if (OpVT == MVT::f64)
320 return FPROUND_F64_F32;
321 if (OpVT == MVT::f80)
322 return FPROUND_F80_F32;
323 if (OpVT == MVT::f128)
324 return FPROUND_F128_F32;
325 if (OpVT == MVT::ppcf128)
326 return FPROUND_PPCF128_F32;
327 } else if (RetVT == MVT::f64) {
328 if (OpVT == MVT::f80)
329 return FPROUND_F80_F64;
330 if (OpVT == MVT::f128)
331 return FPROUND_F128_F64;
332 if (OpVT == MVT::ppcf128)
333 return FPROUND_PPCF128_F64;
334 } else if (RetVT == MVT::f80) {
335 if (OpVT == MVT::f128)
336 return FPROUND_F128_F80;
337 }
338
339 return UNKNOWN_LIBCALL;
340}
341
342/// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
343/// UNKNOWN_LIBCALL if there is none.
344RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
345 if (OpVT == MVT::f16) {
346 if (RetVT == MVT::i32)
347 return FPTOSINT_F16_I32;
348 if (RetVT == MVT::i64)
349 return FPTOSINT_F16_I64;
350 if (RetVT == MVT::i128)
351 return FPTOSINT_F16_I128;
352 } else if (OpVT == MVT::f32) {
353 if (RetVT == MVT::i32)
354 return FPTOSINT_F32_I32;
355 if (RetVT == MVT::i64)
356 return FPTOSINT_F32_I64;
357 if (RetVT == MVT::i128)
358 return FPTOSINT_F32_I128;
359 } else if (OpVT == MVT::f64) {
360 if (RetVT == MVT::i32)
361 return FPTOSINT_F64_I32;
362 if (RetVT == MVT::i64)
363 return FPTOSINT_F64_I64;
364 if (RetVT == MVT::i128)
365 return FPTOSINT_F64_I128;
366 } else if (OpVT == MVT::f80) {
367 if (RetVT == MVT::i32)
368 return FPTOSINT_F80_I32;
369 if (RetVT == MVT::i64)
370 return FPTOSINT_F80_I64;
371 if (RetVT == MVT::i128)
372 return FPTOSINT_F80_I128;
373 } else if (OpVT == MVT::f128) {
374 if (RetVT == MVT::i32)
375 return FPTOSINT_F128_I32;
376 if (RetVT == MVT::i64)
377 return FPTOSINT_F128_I64;
378 if (RetVT == MVT::i128)
379 return FPTOSINT_F128_I128;
380 } else if (OpVT == MVT::ppcf128) {
381 if (RetVT == MVT::i32)
382 return FPTOSINT_PPCF128_I32;
383 if (RetVT == MVT::i64)
384 return FPTOSINT_PPCF128_I64;
385 if (RetVT == MVT::i128)
386 return FPTOSINT_PPCF128_I128;
387 }
388 return UNKNOWN_LIBCALL;
389}
390
391/// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
392/// UNKNOWN_LIBCALL if there is none.
393RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
394 if (OpVT == MVT::f16) {
395 if (RetVT == MVT::i32)
396 return FPTOUINT_F16_I32;
397 if (RetVT == MVT::i64)
398 return FPTOUINT_F16_I64;
399 if (RetVT == MVT::i128)
400 return FPTOUINT_F16_I128;
401 } else if (OpVT == MVT::f32) {
402 if (RetVT == MVT::i32)
403 return FPTOUINT_F32_I32;
404 if (RetVT == MVT::i64)
405 return FPTOUINT_F32_I64;
406 if (RetVT == MVT::i128)
407 return FPTOUINT_F32_I128;
408 } else if (OpVT == MVT::f64) {
409 if (RetVT == MVT::i32)
410 return FPTOUINT_F64_I32;
411 if (RetVT == MVT::i64)
412 return FPTOUINT_F64_I64;
413 if (RetVT == MVT::i128)
414 return FPTOUINT_F64_I128;
415 } else if (OpVT == MVT::f80) {
416 if (RetVT == MVT::i32)
417 return FPTOUINT_F80_I32;
418 if (RetVT == MVT::i64)
419 return FPTOUINT_F80_I64;
420 if (RetVT == MVT::i128)
421 return FPTOUINT_F80_I128;
422 } else if (OpVT == MVT::f128) {
423 if (RetVT == MVT::i32)
424 return FPTOUINT_F128_I32;
425 if (RetVT == MVT::i64)
426 return FPTOUINT_F128_I64;
427 if (RetVT == MVT::i128)
428 return FPTOUINT_F128_I128;
429 } else if (OpVT == MVT::ppcf128) {
430 if (RetVT == MVT::i32)
431 return FPTOUINT_PPCF128_I32;
432 if (RetVT == MVT::i64)
433 return FPTOUINT_PPCF128_I64;
434 if (RetVT == MVT::i128)
435 return FPTOUINT_PPCF128_I128;
436 }
437 return UNKNOWN_LIBCALL;
438}
439
440/// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
441/// UNKNOWN_LIBCALL if there is none.
442RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
443 if (OpVT == MVT::i32) {
444 if (RetVT == MVT::f16)
445 return SINTTOFP_I32_F16;
446 if (RetVT == MVT::f32)
447 return SINTTOFP_I32_F32;
448 if (RetVT == MVT::f64)
449 return SINTTOFP_I32_F64;
450 if (RetVT == MVT::f80)
451 return SINTTOFP_I32_F80;
452 if (RetVT == MVT::f128)
453 return SINTTOFP_I32_F128;
454 if (RetVT == MVT::ppcf128)
455 return SINTTOFP_I32_PPCF128;
456 } else if (OpVT == MVT::i64) {
457 if (RetVT == MVT::bf16)
458 return SINTTOFP_I64_BF16;
459 if (RetVT == MVT::f16)
460 return SINTTOFP_I64_F16;
461 if (RetVT == MVT::f32)
462 return SINTTOFP_I64_F32;
463 if (RetVT == MVT::f64)
464 return SINTTOFP_I64_F64;
465 if (RetVT == MVT::f80)
466 return SINTTOFP_I64_F80;
467 if (RetVT == MVT::f128)
468 return SINTTOFP_I64_F128;
469 if (RetVT == MVT::ppcf128)
470 return SINTTOFP_I64_PPCF128;
471 } else if (OpVT == MVT::i128) {
472 if (RetVT == MVT::f16)
473 return SINTTOFP_I128_F16;
474 if (RetVT == MVT::f32)
475 return SINTTOFP_I128_F32;
476 if (RetVT == MVT::f64)
477 return SINTTOFP_I128_F64;
478 if (RetVT == MVT::f80)
479 return SINTTOFP_I128_F80;
480 if (RetVT == MVT::f128)
481 return SINTTOFP_I128_F128;
482 if (RetVT == MVT::ppcf128)
483 return SINTTOFP_I128_PPCF128;
484 }
485 return UNKNOWN_LIBCALL;
486}
487
488/// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
489/// UNKNOWN_LIBCALL if there is none.
490RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
491 if (OpVT == MVT::i32) {
492 if (RetVT == MVT::f16)
493 return UINTTOFP_I32_F16;
494 if (RetVT == MVT::f32)
495 return UINTTOFP_I32_F32;
496 if (RetVT == MVT::f64)
497 return UINTTOFP_I32_F64;
498 if (RetVT == MVT::f80)
499 return UINTTOFP_I32_F80;
500 if (RetVT == MVT::f128)
501 return UINTTOFP_I32_F128;
502 if (RetVT == MVT::ppcf128)
503 return UINTTOFP_I32_PPCF128;
504 } else if (OpVT == MVT::i64) {
505 if (RetVT == MVT::bf16)
506 return UINTTOFP_I64_BF16;
507 if (RetVT == MVT::f16)
508 return UINTTOFP_I64_F16;
509 if (RetVT == MVT::f32)
510 return UINTTOFP_I64_F32;
511 if (RetVT == MVT::f64)
512 return UINTTOFP_I64_F64;
513 if (RetVT == MVT::f80)
514 return UINTTOFP_I64_F80;
515 if (RetVT == MVT::f128)
516 return UINTTOFP_I64_F128;
517 if (RetVT == MVT::ppcf128)
518 return UINTTOFP_I64_PPCF128;
519 } else if (OpVT == MVT::i128) {
520 if (RetVT == MVT::f16)
521 return UINTTOFP_I128_F16;
522 if (RetVT == MVT::f32)
523 return UINTTOFP_I128_F32;
524 if (RetVT == MVT::f64)
525 return UINTTOFP_I128_F64;
526 if (RetVT == MVT::f80)
527 return UINTTOFP_I128_F80;
528 if (RetVT == MVT::f128)
529 return UINTTOFP_I128_F128;
530 if (RetVT == MVT::ppcf128)
531 return UINTTOFP_I128_PPCF128;
532 }
533 return UNKNOWN_LIBCALL;
534}
535
536RTLIB::Libcall RTLIB::getPOWI(EVT RetVT) {
537 return getFPLibCall(RetVT, POWI_F32, POWI_F64, POWI_F80, POWI_F128,
538 POWI_PPCF128);
539}
540
541RTLIB::Libcall RTLIB::getPOW(EVT RetVT) {
542 return getFPLibCall(RetVT, POW_F32, POW_F64, POW_F80, POW_F128, POW_PPCF128);
543}
544
545RTLIB::Libcall RTLIB::getLDEXP(EVT RetVT) {
546 return getFPLibCall(RetVT, LDEXP_F32, LDEXP_F64, LDEXP_F80, LDEXP_F128,
547 LDEXP_PPCF128);
548}
549
550RTLIB::Libcall RTLIB::getFREXP(EVT RetVT) {
551 return getFPLibCall(RetVT, FREXP_F32, FREXP_F64, FREXP_F80, FREXP_F128,
552 FREXP_PPCF128);
553}
554
555RTLIB::Libcall RTLIB::getSIN(EVT RetVT) {
556 return getFPLibCall(RetVT, SIN_F32, SIN_F64, SIN_F80, SIN_F128, SIN_PPCF128);
557}
558
559RTLIB::Libcall RTLIB::getCOS(EVT RetVT) {
560 return getFPLibCall(RetVT, COS_F32, COS_F64, COS_F80, COS_F128, COS_PPCF128);
561}
562
563RTLIB::Libcall RTLIB::getSINCOS(EVT RetVT) {
564 // TODO: Tablegen should generate this function
565 if (RetVT.isVector()) {
566 if (!RetVT.isSimple())
567 return RTLIB::UNKNOWN_LIBCALL;
568 switch (RetVT.getSimpleVT().SimpleTy) {
569 case MVT::v4f32:
570 return RTLIB::SINCOS_V4F32;
571 case MVT::v2f64:
572 return RTLIB::SINCOS_V2F64;
573 case MVT::nxv4f32:
574 return RTLIB::SINCOS_NXV4F32;
575 case MVT::nxv2f64:
576 return RTLIB::SINCOS_NXV2F64;
577 default:
578 return RTLIB::UNKNOWN_LIBCALL;
579 }
580 }
581
582 return getFPLibCall(RetVT, SINCOS_F32, SINCOS_F64, SINCOS_F80, SINCOS_F128,
583 SINCOS_PPCF128);
584}
585
586RTLIB::Libcall RTLIB::getSINCOSPI(EVT RetVT) {
587 // TODO: Tablegen should generate this function
588 if (RetVT.isVector()) {
589 if (!RetVT.isSimple())
590 return RTLIB::UNKNOWN_LIBCALL;
591 switch (RetVT.getSimpleVT().SimpleTy) {
592 case MVT::v4f32:
593 return RTLIB::SINCOSPI_V4F32;
594 case MVT::v2f64:
595 return RTLIB::SINCOSPI_V2F64;
596 case MVT::nxv4f32:
597 return RTLIB::SINCOSPI_NXV4F32;
598 case MVT::nxv2f64:
599 return RTLIB::SINCOSPI_NXV2F64;
600 default:
601 return RTLIB::UNKNOWN_LIBCALL;
602 }
603 }
604
605 return getFPLibCall(RetVT, SINCOSPI_F32, SINCOSPI_F64, SINCOSPI_F80,
606 SINCOSPI_F128, SINCOSPI_PPCF128);
607}
608
609RTLIB::Libcall RTLIB::getSINCOS_STRET(EVT RetVT) {
610 return getFPLibCall(RetVT, SINCOS_STRET_F32, SINCOS_STRET_F64,
611 UNKNOWN_LIBCALL, UNKNOWN_LIBCALL, UNKNOWN_LIBCALL);
612}
613
614RTLIB::Libcall RTLIB::getREM(EVT VT) {
615 // TODO: Tablegen should generate this function
616 if (VT.isVector()) {
617 if (!VT.isSimple())
618 return RTLIB::UNKNOWN_LIBCALL;
619 switch (VT.getSimpleVT().SimpleTy) {
620 case MVT::v4f32:
621 return RTLIB::REM_V4F32;
622 case MVT::v2f64:
623 return RTLIB::REM_V2F64;
624 case MVT::nxv4f32:
625 return RTLIB::REM_NXV4F32;
626 case MVT::nxv2f64:
627 return RTLIB::REM_NXV2F64;
628 default:
629 return RTLIB::UNKNOWN_LIBCALL;
630 }
631 }
632
633 return getFPLibCall(VT, REM_F32, REM_F64, REM_F80, REM_F128, REM_PPCF128);
634}
635
636RTLIB::Libcall RTLIB::getMODF(EVT RetVT) {
637 // TODO: Tablegen should generate this function
638 if (RetVT.isVector()) {
639 if (!RetVT.isSimple())
640 return RTLIB::UNKNOWN_LIBCALL;
641 switch (RetVT.getSimpleVT().SimpleTy) {
642 case MVT::v4f32:
643 return RTLIB::MODF_V4F32;
644 case MVT::v2f64:
645 return RTLIB::MODF_V2F64;
646 case MVT::nxv4f32:
647 return RTLIB::MODF_NXV4F32;
648 case MVT::nxv2f64:
649 return RTLIB::MODF_NXV2F64;
650 default:
651 return RTLIB::UNKNOWN_LIBCALL;
652 }
653 }
654
655 return getFPLibCall(RetVT, MODF_F32, MODF_F64, MODF_F80, MODF_F128,
656 MODF_PPCF128);
657}
658
659RTLIB::Libcall RTLIB::getLROUND(EVT VT) {
660 if (VT == MVT::f32)
661 return RTLIB::LROUND_F32;
662 if (VT == MVT::f64)
663 return RTLIB::LROUND_F64;
664 if (VT == MVT::f80)
665 return RTLIB::LROUND_F80;
666 if (VT == MVT::f128)
667 return RTLIB::LROUND_F128;
668 if (VT == MVT::ppcf128)
669 return RTLIB::LROUND_PPCF128;
670
671 return RTLIB::UNKNOWN_LIBCALL;
672}
673
674RTLIB::Libcall RTLIB::getLLROUND(EVT VT) {
675 if (VT == MVT::f32)
676 return RTLIB::LLROUND_F32;
677 if (VT == MVT::f64)
678 return RTLIB::LLROUND_F64;
679 if (VT == MVT::f80)
680 return RTLIB::LLROUND_F80;
681 if (VT == MVT::f128)
682 return RTLIB::LLROUND_F128;
683 if (VT == MVT::ppcf128)
684 return RTLIB::LLROUND_PPCF128;
685
686 return RTLIB::UNKNOWN_LIBCALL;
687}
688
689RTLIB::Libcall RTLIB::getLRINT(EVT VT) {
690 if (VT == MVT::f32)
691 return RTLIB::LRINT_F32;
692 if (VT == MVT::f64)
693 return RTLIB::LRINT_F64;
694 if (VT == MVT::f80)
695 return RTLIB::LRINT_F80;
696 if (VT == MVT::f128)
697 return RTLIB::LRINT_F128;
698 if (VT == MVT::ppcf128)
699 return RTLIB::LRINT_PPCF128;
700 return RTLIB::UNKNOWN_LIBCALL;
701}
702
703RTLIB::Libcall RTLIB::getLLRINT(EVT VT) {
704 if (VT == MVT::f32)
705 return RTLIB::LLRINT_F32;
706 if (VT == MVT::f64)
707 return RTLIB::LLRINT_F64;
708 if (VT == MVT::f80)
709 return RTLIB::LLRINT_F80;
710 if (VT == MVT::f128)
711 return RTLIB::LLRINT_F128;
712 if (VT == MVT::ppcf128)
713 return RTLIB::LLRINT_PPCF128;
714 return RTLIB::UNKNOWN_LIBCALL;
715}
716
717RTLIB::Libcall RTLIB::getOutlineAtomicHelper(const Libcall (&LC)[5][4],
718 AtomicOrdering Order,
719 uint64_t MemSize) {
720 unsigned ModeN, ModelN;
721 switch (MemSize) {
722 case 1:
723 ModeN = 0;
724 break;
725 case 2:
726 ModeN = 1;
727 break;
728 case 4:
729 ModeN = 2;
730 break;
731 case 8:
732 ModeN = 3;
733 break;
734 case 16:
735 ModeN = 4;
736 break;
737 default:
738 return RTLIB::UNKNOWN_LIBCALL;
739 }
740
741 switch (Order) {
743 ModelN = 0;
744 break;
746 ModelN = 1;
747 break;
749 ModelN = 2;
750 break;
753 ModelN = 3;
754 break;
755 default:
756 return UNKNOWN_LIBCALL;
757 }
758
759 return LC[ModeN][ModelN];
760}
761
762RTLIB::Libcall RTLIB::getOUTLINE_ATOMIC(unsigned Opc, AtomicOrdering Order,
763 MVT VT) {
764 if (!VT.isScalarInteger())
765 return UNKNOWN_LIBCALL;
766 uint64_t MemSize = VT.getScalarSizeInBits() / 8;
767
768#define LCALLS(A, B) \
769 { A##B##_RELAX, A##B##_ACQ, A##B##_REL, A##B##_ACQ_REL }
770#define LCALL5(A) \
771 LCALLS(A, 1), LCALLS(A, 2), LCALLS(A, 4), LCALLS(A, 8), LCALLS(A, 16)
772 switch (Opc) {
774 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_CAS)};
775 return getOutlineAtomicHelper(LC, Order, MemSize);
776 }
777 case ISD::ATOMIC_SWAP: {
778 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_SWP)};
779 return getOutlineAtomicHelper(LC, Order, MemSize);
780 }
782 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDADD)};
783 return getOutlineAtomicHelper(LC, Order, MemSize);
784 }
785 case ISD::ATOMIC_LOAD_OR: {
786 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDSET)};
787 return getOutlineAtomicHelper(LC, Order, MemSize);
788 }
790 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDCLR)};
791 return getOutlineAtomicHelper(LC, Order, MemSize);
792 }
794 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDEOR)};
795 return getOutlineAtomicHelper(LC, Order, MemSize);
796 }
797 default:
798 return UNKNOWN_LIBCALL;
799 }
800#undef LCALLS
801#undef LCALL5
802}
803
804RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
805#define OP_TO_LIBCALL(Name, Enum) \
806 case Name: \
807 switch (VT.SimpleTy) { \
808 default: \
809 return UNKNOWN_LIBCALL; \
810 case MVT::i8: \
811 return Enum##_1; \
812 case MVT::i16: \
813 return Enum##_2; \
814 case MVT::i32: \
815 return Enum##_4; \
816 case MVT::i64: \
817 return Enum##_8; \
818 case MVT::i128: \
819 return Enum##_16; \
820 }
821
822 switch (Opc) {
823 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
824 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
825 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
826 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
827 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
828 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
829 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
830 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
831 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
832 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
833 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
834 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
835 }
836
837#undef OP_TO_LIBCALL
838
839 return UNKNOWN_LIBCALL;
840}
841
843 switch (ElementSize) {
844 case 1:
845 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
846 case 2:
847 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
848 case 4:
849 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
850 case 8:
851 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
852 case 16:
853 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
854 default:
855 return UNKNOWN_LIBCALL;
856 }
857}
858
860 switch (ElementSize) {
861 case 1:
862 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
863 case 2:
864 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
865 case 4:
866 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
867 case 8:
868 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
869 case 16:
870 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
871 default:
872 return UNKNOWN_LIBCALL;
873 }
874}
875
877 switch (ElementSize) {
878 case 1:
879 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
880 case 2:
881 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
882 case 4:
883 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
884 case 8:
885 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
886 case 16:
887 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
888 default:
889 return UNKNOWN_LIBCALL;
890 }
891}
892
894 RTLIB::LibcallImpl Impl) const {
895 switch (Impl) {
896 case RTLIB::impl___aeabi_dcmpeq__une:
897 case RTLIB::impl___aeabi_fcmpeq__une:
898 // Usage in the eq case, so we have to invert the comparison.
899 return ISD::SETEQ;
900 case RTLIB::impl___aeabi_dcmpeq__oeq:
901 case RTLIB::impl___aeabi_fcmpeq__oeq:
902 // Normal comparison to boolean value.
903 return ISD::SETNE;
904 case RTLIB::impl___aeabi_dcmplt:
905 case RTLIB::impl___aeabi_dcmple:
906 case RTLIB::impl___aeabi_dcmpge:
907 case RTLIB::impl___aeabi_dcmpgt:
908 case RTLIB::impl___aeabi_dcmpun:
909 case RTLIB::impl___aeabi_fcmplt:
910 case RTLIB::impl___aeabi_fcmple:
911 case RTLIB::impl___aeabi_fcmpge:
912 case RTLIB::impl___aeabi_fcmpgt:
913 /// The AEABI versions return a typical boolean value, so we can compare
914 /// against the integer result as simply != 0.
915 return ISD::SETNE;
916 default:
917 break;
918 }
919
920 // Assume libgcc/compiler-rt behavior. Most of the cases are really aliases of
921 // each other, and return a 3-way comparison style result of -1, 0, or 1
922 // depending on lt/eq/gt.
923 //
924 // FIXME: It would be cleaner to directly express this as a 3-way comparison
925 // soft FP libcall instead of individual compares.
926 RTLIB::Libcall LC = RTLIB::RuntimeLibcallsInfo::getLibcallFromImpl(Impl);
927 switch (LC) {
928 case RTLIB::OEQ_F32:
929 case RTLIB::OEQ_F64:
930 case RTLIB::OEQ_F128:
931 case RTLIB::OEQ_PPCF128:
932 return ISD::SETEQ;
933 case RTLIB::UNE_F32:
934 case RTLIB::UNE_F64:
935 case RTLIB::UNE_F128:
936 case RTLIB::UNE_PPCF128:
937 return ISD::SETNE;
938 case RTLIB::OGE_F32:
939 case RTLIB::OGE_F64:
940 case RTLIB::OGE_F128:
941 case RTLIB::OGE_PPCF128:
942 return ISD::SETGE;
943 case RTLIB::OLT_F32:
944 case RTLIB::OLT_F64:
945 case RTLIB::OLT_F128:
946 case RTLIB::OLT_PPCF128:
947 return ISD::SETLT;
948 case RTLIB::OLE_F32:
949 case RTLIB::OLE_F64:
950 case RTLIB::OLE_F128:
951 case RTLIB::OLE_PPCF128:
952 return ISD::SETLE;
953 case RTLIB::OGT_F32:
954 case RTLIB::OGT_F64:
955 case RTLIB::OGT_F128:
956 case RTLIB::OGT_PPCF128:
957 return ISD::SETGT;
958 case RTLIB::UO_F32:
959 case RTLIB::UO_F64:
960 case RTLIB::UO_F128:
961 case RTLIB::UO_PPCF128:
962 return ISD::SETNE;
963 default:
964 llvm_unreachable("not a compare libcall");
965 }
966}
967
968/// NOTE: The TargetMachine owns TLOF.
970 const TargetSubtargetInfo &STI)
971 : TM(tm),
972 RuntimeLibcallInfo(TM.getTargetTriple(), TM.Options.ExceptionModel,
973 TM.Options.FloatABIType, TM.Options.EABIVersion,
974 TM.Options.MCOptions.getABIName(), TM.Options.VecLib),
975 Libcalls(RuntimeLibcallInfo, STI) {
976 initActions();
977
978 // Perform these initializations only once.
984 HasExtractBitsInsn = false;
985 JumpIsExpensive = JumpIsExpensiveOverride;
987 EnableExtLdPromotion = false;
988 StackPointerRegisterToSaveRestore = 0;
989 BooleanContents = UndefinedBooleanContent;
990 BooleanFloatContents = UndefinedBooleanContent;
991 BooleanVectorContents = UndefinedBooleanContent;
992 SchedPreferenceInfo = Sched::ILP;
995 MaxBytesForAlignment = 0;
996 MaxAtomicSizeInBitsSupported = 0;
997
998 // Assume that even with libcalls, no target supports wider than 128 bit
999 // division.
1000 MaxDivRemBitWidthSupported = 128;
1001
1002 MaxLargeFPConvertBitWidthSupported = llvm::IntegerType::MAX_INT_BITS;
1003
1004 MinCmpXchgSizeInBits = 0;
1005 SupportsUnalignedAtomics = false;
1006
1007 MinimumBitTestCmps = MinimumBitTestCmpsOverride;
1008}
1009
1010// Define the virtual destructor out-of-line to act as a key method to anchor
1011// debug info (see coding standards).
1013
1015 // All operations default to being supported.
1016 memset(OpActions, 0, sizeof(OpActions));
1017 memset(LoadExtActions, 0, sizeof(LoadExtActions));
1018 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
1019 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
1020 memset(CondCodeActions, 0, sizeof(CondCodeActions));
1021 llvm::fill(RegClassForVT, nullptr);
1022 llvm::fill(TargetDAGCombineArray, 0);
1023
1024 // Let extending atomic loads be unsupported by default.
1025 for (MVT ValVT : MVT::all_valuetypes())
1026 for (MVT MemVT : MVT::all_valuetypes())
1028 Expand);
1029
1030 // We're somewhat special casing MVT::i2 and MVT::i4. Ideally we want to
1031 // remove this and targets should individually set these types if not legal.
1034 for (MVT VT : {MVT::i2, MVT::i4})
1035 OpActions[(unsigned)VT.SimpleTy][NT] = Expand;
1036 }
1037 for (MVT AVT : MVT::all_valuetypes()) {
1038 for (MVT VT : {MVT::i2, MVT::i4, MVT::v128i2, MVT::v64i4}) {
1039 setTruncStoreAction(AVT, VT, Expand);
1042 }
1043 }
1044 for (unsigned IM = (unsigned)ISD::PRE_INC;
1045 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
1046 for (MVT VT : {MVT::i2, MVT::i4}) {
1051 }
1052 }
1053
1054 for (MVT VT : MVT::fp_valuetypes()) {
1055 MVT IntVT = MVT::getIntegerVT(VT.getFixedSizeInBits());
1056 if (IntVT.isValid()) {
1059 }
1060 }
1061
1062 // If f16 fma is not natively supported, the value must be promoted to an f64
1063 // (and not to f32!) to prevent double rounding issues.
1064 AddPromotedToType(ISD::FMA, MVT::f16, MVT::f64);
1065 AddPromotedToType(ISD::STRICT_FMA, MVT::f16, MVT::f64);
1066
1067 // Set default actions for various operations.
1068 for (MVT VT : MVT::all_valuetypes()) {
1069 // Default all indexed load / store to expand.
1070 for (unsigned IM = (unsigned)ISD::PRE_INC;
1071 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
1076 }
1077
1078 // Most backends expect to see the node which just returns the value loaded.
1080
1081 // These operations default to expand.
1110 ISD::FMULADD},
1111 VT, Expand);
1112
1113 // Overflow operations default to expand
1116 VT, Expand);
1117
1118 // Carry-using overflow operations default to expand.
1121 VT, Expand);
1122
1123 // ADDC/ADDE/SUBC/SUBE default to expand.
1125 Expand);
1126
1127 // [US]CMP default to expand
1129
1130 // Halving adds
1133 Expand);
1134
1135 // Absolute difference
1137
1138 // Carry-less multiply
1140
1141 // Saturated trunc
1145
1146 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
1148 Expand);
1150
1152
1153 // These library functions default to expand.
1156 VT, Expand);
1157
1158 // These operations default to expand for vector types.
1159 if (VT.isVector())
1165 VT, Expand);
1166
1167 // Constrained floating-point operations default to expand.
1168#define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
1169 setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
1170#include "llvm/IR/ConstrainedOps.def"
1171
1172 // For most targets @llvm.get.dynamic.area.offset just returns 0.
1174
1175 // Vector reduction default to expand.
1183 VT, Expand);
1184
1185 // Named vector shuffles default to expand.
1187 Expand);
1188
1189 // Only some target support this vector operation. Most need to expand it.
1191
1192 // VP operations default to expand.
1193#define BEGIN_REGISTER_VP_SDNODE(SDOPC, ...) \
1194 setOperationAction(ISD::SDOPC, VT, Expand);
1195#include "llvm/IR/VPIntrinsics.def"
1196
1197 // Masked vector extracts default to expand.
1199
1202
1203 // FP environment operations default to expand.
1207
1209 }
1210
1211 // Most targets ignore the @llvm.prefetch intrinsic.
1213
1214 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
1216
1217 // Most targets also ignore the @llvm.readsteadycounter intrinsic.
1219
1220 // ConstantFP nodes default to expand. Targets can either change this to
1221 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
1222 // to optimize expansions for certain constants.
1224 {MVT::bf16, MVT::f16, MVT::f32, MVT::f64, MVT::f80, MVT::f128},
1225 Expand);
1226
1227 // Insert custom handling default for llvm.canonicalize.*.
1229 {MVT::f16, MVT::f32, MVT::f64, MVT::f128}, Expand);
1230
1231 // FIXME: Query RuntimeLibCalls to make the decision.
1233 {MVT::f32, MVT::f64, MVT::f128}, LibCall);
1234
1237 MVT::f16, Promote);
1238 // Default ISD::TRAP to expand (which turns it into abort).
1239 setOperationAction(ISD::TRAP, MVT::Other, Expand);
1240
1241 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
1242 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
1244
1246
1249
1250 for (MVT VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) {
1253 }
1255
1256 // This one by default will call __clear_cache unless the target
1257 // wants something different.
1259
1260 // By default, STACKADDRESS nodes are expanded like STACKSAVE nodes.
1261 // On SPARC targets, custom lowering is required.
1263}
1264
1266 EVT) const {
1267 return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
1268}
1269
1271 const DataLayout &DL) const {
1272 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
1273 if (LHSTy.isVector())
1274 return LHSTy;
1275 MVT ShiftVT = getScalarShiftAmountTy(DL, LHSTy);
1276 // If any possible shift value won't fit in the prefered type, just use
1277 // something safe. Assume it will be legalized when the shift is expanded.
1278 if (ShiftVT.getSizeInBits() < Log2_32_Ceil(LHSTy.getSizeInBits()))
1279 ShiftVT = MVT::i32;
1280 assert(ShiftVT.getSizeInBits() >= Log2_32_Ceil(LHSTy.getSizeInBits()) &&
1281 "ShiftVT is still too small!");
1282 return ShiftVT;
1283}
1284
1285bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
1286 assert(isTypeLegal(VT));
1287 switch (Op) {
1288 default:
1289 return false;
1290 case ISD::SDIV:
1291 case ISD::UDIV:
1292 case ISD::SREM:
1293 case ISD::UREM:
1294 return true;
1295 }
1296}
1297
1299 unsigned DestAS) const {
1300 return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
1301}
1302
1304 Type *RetTy, ElementCount EC, bool ZeroIsPoison,
1305 const ConstantRange *VScaleRange) const {
1306 // Find the smallest "sensible" element type to use for the expansion.
1307 ConstantRange CR(APInt(64, EC.getKnownMinValue()));
1308 if (EC.isScalable())
1309 CR = CR.umul_sat(*VScaleRange);
1310
1311 if (ZeroIsPoison)
1312 CR = CR.subtract(APInt(64, 1));
1313
1314 unsigned EltWidth = RetTy->getScalarSizeInBits();
1315 EltWidth = std::min(EltWidth, CR.getActiveBits());
1316 EltWidth = std::max(llvm::bit_ceil(EltWidth), (unsigned)8);
1317
1318 return EltWidth;
1319}
1320
1322 // If the command-line option was specified, ignore this request.
1323 if (!JumpIsExpensiveOverride.getNumOccurrences())
1324 JumpIsExpensive = isExpensive;
1325}
1326
1329 // If this is a simple type, use the ComputeRegisterProp mechanism.
1330 if (VT.isSimple()) {
1331 MVT SVT = VT.getSimpleVT();
1332 assert((unsigned)SVT.SimpleTy < std::size(TransformToType));
1333 MVT NVT = TransformToType[SVT.SimpleTy];
1334 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
1335
1336 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
1337 LA == TypeSoftPromoteHalf ||
1338 (NVT.isVector() ||
1339 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
1340 "Promote may not follow Expand or Promote");
1341
1342 if (LA == TypeSplitVector)
1343 return LegalizeKind(LA, EVT(SVT).getHalfNumVectorElementsVT(Context));
1344 if (LA == TypeScalarizeVector)
1345 return LegalizeKind(LA, SVT.getVectorElementType());
1346 return LegalizeKind(LA, NVT);
1347 }
1348
1349 // Handle Extended Scalar Types.
1350 if (!VT.isVector()) {
1351 assert(VT.isInteger() && "Float types must be simple");
1352 unsigned BitSize = VT.getSizeInBits();
1353 // First promote to a power-of-two size, then expand if necessary.
1354 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
1355 EVT NVT = VT.getRoundIntegerType(Context);
1356 assert(NVT != VT && "Unable to round integer VT");
1357 LegalizeKind NextStep = getTypeConversion(Context, NVT);
1358 // Avoid multi-step promotion.
1359 if (NextStep.first == TypePromoteInteger)
1360 return NextStep;
1361 // Return rounded integer type.
1362 return LegalizeKind(TypePromoteInteger, NVT);
1363 }
1364
1366 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
1367 }
1368
1369 // Handle vector types.
1370 ElementCount NumElts = VT.getVectorElementCount();
1371 EVT EltVT = VT.getVectorElementType();
1372
1373 // Vectors with only one element are always scalarized.
1374 if (NumElts.isScalar())
1375 return LegalizeKind(TypeScalarizeVector, EltVT);
1376
1377 // Try to widen vector elements until the element type is a power of two and
1378 // promote it to a legal type later on, for example:
1379 // <3 x i8> -> <4 x i8> -> <4 x i32>
1380 if (EltVT.isInteger()) {
1381 // Vectors with a number of elements that is not a power of two are always
1382 // widened, for example <3 x i8> -> <4 x i8>.
1383 if (!VT.isPow2VectorType()) {
1384 NumElts = NumElts.coefficientNextPowerOf2();
1385 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
1386 return LegalizeKind(TypeWidenVector, NVT);
1387 }
1388
1389 // Examine the element type.
1390 LegalizeKind LK = getTypeConversion(Context, EltVT);
1391
1392 // If type is to be expanded, split the vector.
1393 // <4 x i140> -> <2 x i140>
1394 if (LK.first == TypeExpandInteger) {
1395 if (NumElts.isScalable() && NumElts.getKnownMinValue() == 1)
1398 VT.getHalfNumVectorElementsVT(Context));
1399 }
1400
1401 // Promote the integer element types until a legal vector type is found
1402 // or until the element integer type is too big. If a legal type was not
1403 // found, fallback to the usual mechanism of widening/splitting the
1404 // vector.
1405 EVT OldEltVT = EltVT;
1406 while (true) {
1407 // Increase the bitwidth of the element to the next pow-of-two
1408 // (which is greater than 8 bits).
1409 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
1410 .getRoundIntegerType(Context);
1411
1412 // Stop trying when getting a non-simple element type.
1413 // Note that vector elements may be greater than legal vector element
1414 // types. Example: X86 XMM registers hold 64bit element on 32bit
1415 // systems.
1416 if (!EltVT.isSimple())
1417 break;
1418
1419 // Build a new vector type and check if it is legal.
1420 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1421 // Found a legal promoted vector type.
1422 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
1424 EVT::getVectorVT(Context, EltVT, NumElts));
1425 }
1426
1427 // Reset the type to the unexpanded type if we did not find a legal vector
1428 // type with a promoted vector element type.
1429 EltVT = OldEltVT;
1430 }
1431
1432 // Try to widen the vector until a legal type is found.
1433 // If there is no wider legal type, split the vector.
1434 while (true) {
1435 // Round up to the next power of 2.
1436 NumElts = NumElts.coefficientNextPowerOf2();
1437
1438 // If there is no simple vector type with this many elements then there
1439 // cannot be a larger legal vector type. Note that this assumes that
1440 // there are no skipped intermediate vector types in the simple types.
1441 if (!EltVT.isSimple())
1442 break;
1443 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1444 if (LargerVector == MVT())
1445 break;
1446
1447 // If this type is legal then widen the vector.
1448 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1449 return LegalizeKind(TypeWidenVector, LargerVector);
1450 }
1451
1452 // Widen odd vectors to next power of two.
1453 if (!VT.isPow2VectorType()) {
1454 EVT NVT = VT.getPow2VectorType(Context);
1455 return LegalizeKind(TypeWidenVector, NVT);
1456 }
1457
1460
1461 // Vectors with illegal element types are expanded.
1462 EVT NVT = EVT::getVectorVT(Context, EltVT,
1464 return LegalizeKind(TypeSplitVector, NVT);
1465}
1466
1467static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
1468 unsigned &NumIntermediates,
1469 MVT &RegisterVT,
1470 TargetLoweringBase *TLI) {
1471 // Figure out the right, legal destination reg to copy into.
1473 MVT EltTy = VT.getVectorElementType();
1474
1475 unsigned NumVectorRegs = 1;
1476
1477 // Scalable vectors cannot be scalarized, so splitting or widening is
1478 // required.
1479 if (VT.isScalableVector() && !isPowerOf2_32(EC.getKnownMinValue()))
1481 "Splitting or widening of non-power-of-2 MVTs is not implemented.");
1482
1483 // FIXME: We don't support non-power-of-2-sized vectors for now.
1484 // Ideally we could break down into LHS/RHS like LegalizeDAG does.
1485 if (!isPowerOf2_32(EC.getKnownMinValue())) {
1486 // Split EC to unit size (scalable property is preserved).
1487 NumVectorRegs = EC.getKnownMinValue();
1488 EC = ElementCount::getFixed(1);
1489 }
1490
1491 // Divide the input until we get to a supported size. This will
1492 // always end up with an EC that represent a scalar or a scalable
1493 // scalar.
1494 while (EC.getKnownMinValue() > 1 &&
1495 !TLI->isTypeLegal(MVT::getVectorVT(EltTy, EC))) {
1496 EC = EC.divideCoefficientBy(2);
1497 NumVectorRegs <<= 1;
1498 }
1499
1500 NumIntermediates = NumVectorRegs;
1501
1502 MVT NewVT = MVT::getVectorVT(EltTy, EC);
1503 if (!TLI->isTypeLegal(NewVT))
1504 NewVT = EltTy;
1505 IntermediateVT = NewVT;
1506
1507 unsigned LaneSizeInBits = NewVT.getScalarSizeInBits();
1508
1509 // Convert sizes such as i33 to i64.
1510 LaneSizeInBits = llvm::bit_ceil(LaneSizeInBits);
1511
1512 MVT DestVT = TLI->getRegisterType(NewVT);
1513 RegisterVT = DestVT;
1514 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1515 return NumVectorRegs * (LaneSizeInBits / DestVT.getScalarSizeInBits());
1516
1517 // Otherwise, promotion or legal types use the same number of registers as
1518 // the vector decimated to the appropriate level.
1519 return NumVectorRegs;
1520}
1521
1522/// isLegalRC - Return true if the value types that can be represented by the
1523/// specified register class are all legal.
1525 const TargetRegisterClass &RC) const {
1526 for (const auto *I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1527 if (isTypeLegal(*I))
1528 return true;
1529 return false;
1530}
1531
1532/// Replace/modify any TargetFrameIndex operands with a targte-dependent
1533/// sequence of memory operands that is recognized by PrologEpilogInserter.
1536 MachineBasicBlock *MBB) const {
1537 MachineInstr *MI = &InitialMI;
1538 MachineFunction &MF = *MI->getMF();
1539 MachineFrameInfo &MFI = MF.getFrameInfo();
1540
1541 // We're handling multiple types of operands here:
1542 // PATCHPOINT MetaArgs - live-in, read only, direct
1543 // STATEPOINT Deopt Spill - live-through, read only, indirect
1544 // STATEPOINT Deopt Alloca - live-through, read only, direct
1545 // (We're currently conservative and mark the deopt slots read/write in
1546 // practice.)
1547 // STATEPOINT GC Spill - live-through, read/write, indirect
1548 // STATEPOINT GC Alloca - live-through, read/write, direct
1549 // The live-in vs live-through is handled already (the live through ones are
1550 // all stack slots), but we need to handle the different type of stackmap
1551 // operands and memory effects here.
1552
1553 if (llvm::none_of(MI->operands(),
1554 [](MachineOperand &Operand) { return Operand.isFI(); }))
1555 return MBB;
1556
1557 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1558
1559 // Inherit previous memory operands.
1560 MIB.cloneMemRefs(*MI);
1561
1562 for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1563 MachineOperand &MO = MI->getOperand(i);
1564 if (!MO.isFI()) {
1565 // Index of Def operand this Use it tied to.
1566 // Since Defs are coming before Uses, if Use is tied, then
1567 // index of Def must be smaller that index of that Use.
1568 // Also, Defs preserve their position in new MI.
1569 unsigned TiedTo = i;
1570 if (MO.isReg() && MO.isTied())
1571 TiedTo = MI->findTiedOperandIdx(i);
1572 MIB.add(MO);
1573 if (TiedTo < i)
1574 MIB->tieOperands(TiedTo, MIB->getNumOperands() - 1);
1575 continue;
1576 }
1577
1578 // foldMemoryOperand builds a new MI after replacing a single FI operand
1579 // with the canonical set of five x86 addressing-mode operands.
1580 int FI = MO.getIndex();
1581
1582 // Add frame index operands recognized by stackmaps.cpp
1584 // indirect-mem-ref tag, size, #FI, offset.
1585 // Used for spills inserted by StatepointLowering. This codepath is not
1586 // used for patchpoints/stackmaps at all, for these spilling is done via
1587 // foldMemoryOperand callback only.
1588 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1589 MIB.addImm(StackMaps::IndirectMemRefOp);
1590 MIB.addImm(MFI.getObjectSize(FI));
1591 MIB.add(MO);
1592 MIB.addImm(0);
1593 } else {
1594 // direct-mem-ref tag, #FI, offset.
1595 // Used by patchpoint, and direct alloca arguments to statepoints
1596 MIB.addImm(StackMaps::DirectMemRefOp);
1597 MIB.add(MO);
1598 MIB.addImm(0);
1599 }
1600
1601 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1602
1603 // Add a new memory operand for this FI.
1604 assert(MFI.getObjectOffset(FI) != -1);
1605
1606 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and
1607 // PATCHPOINT should be updated to do the same. (TODO)
1608 if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1609 auto Flags = MachineMemOperand::MOLoad;
1611 MachinePointerInfo::getFixedStack(MF, FI), Flags,
1613 MIB->addMemOperand(MF, MMO);
1614 }
1615 }
1616 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1617 MI->eraseFromParent();
1618 return MBB;
1619}
1620
1621/// findRepresentativeClass - Return the largest legal super-reg register class
1622/// of the register class for the specified type and its associated "cost".
1623// This function is in TargetLowering because it uses RegClassForVT which would
1624// need to be moved to TargetRegisterInfo and would necessitate moving
1625// isTypeLegal over as well - a massive change that would just require
1626// TargetLowering having a TargetRegisterInfo class member that it would use.
1627std::pair<const TargetRegisterClass *, uint8_t>
1629 MVT VT) const {
1630 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1631 if (!RC)
1632 return std::make_pair(RC, 0);
1633
1634 // Compute the set of all super-register classes.
1635 BitVector SuperRegRC(TRI->getNumRegClasses());
1636 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1637 SuperRegRC.setBitsInMask(RCI.getMask());
1638
1639 // Find the first legal register class with the largest spill size.
1640 const TargetRegisterClass *BestRC = RC;
1641 for (unsigned i : SuperRegRC.set_bits()) {
1642 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1643 // We want the largest possible spill size.
1644 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1645 continue;
1646 if (!isLegalRC(*TRI, *SuperRC))
1647 continue;
1648 BestRC = SuperRC;
1649 }
1650 return std::make_pair(BestRC, 1);
1651}
1652
1653/// computeRegisterProperties - Once all of the register classes are added,
1654/// this allows us to compute derived properties we expose.
1656 const TargetRegisterInfo *TRI) {
1657 // Everything defaults to needing one register.
1658 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1659 NumRegistersForVT[i] = 1;
1660 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1661 }
1662 // ...except isVoid, which doesn't need any registers.
1663 NumRegistersForVT[MVT::isVoid] = 0;
1664
1665 // Find the largest integer register class.
1666 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1667 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1668 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1669
1670 // Every integer value type larger than this largest register takes twice as
1671 // many registers to represent as the previous ValueType.
1672 for (unsigned ExpandedReg = LargestIntReg + 1;
1673 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1674 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1675 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1676 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1677 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1679 }
1680
1681 // Inspect all of the ValueType's smaller than the largest integer
1682 // register to see which ones need promotion.
1683 unsigned LegalIntReg = LargestIntReg;
1684 for (unsigned IntReg = LargestIntReg - 1;
1685 IntReg >= (unsigned)MVT::i1; --IntReg) {
1686 MVT IVT = (MVT::SimpleValueType)IntReg;
1687 if (isTypeLegal(IVT)) {
1688 LegalIntReg = IntReg;
1689 } else {
1690 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1691 (MVT::SimpleValueType)LegalIntReg;
1692 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1693 }
1694 }
1695
1696 // ppcf128 type is really two f64's.
1697 if (!isTypeLegal(MVT::ppcf128)) {
1698 if (isTypeLegal(MVT::f64)) {
1699 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1700 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1701 TransformToType[MVT::ppcf128] = MVT::f64;
1702 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1703 } else {
1704 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1705 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1706 TransformToType[MVT::ppcf128] = MVT::i128;
1707 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1708 }
1709 }
1710
1711 // Decide how to handle f128. If the target does not have native f128 support,
1712 // expand it to i128 and we will be generating soft float library calls.
1713 if (!isTypeLegal(MVT::f128)) {
1714 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1715 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1716 TransformToType[MVT::f128] = MVT::i128;
1717 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1718 }
1719
1720 // Decide how to handle f80. If the target does not have native f80 support,
1721 // expand it to i96 and we will be generating soft float library calls.
1722 if (!isTypeLegal(MVT::f80)) {
1723 NumRegistersForVT[MVT::f80] = 3*NumRegistersForVT[MVT::i32];
1724 RegisterTypeForVT[MVT::f80] = RegisterTypeForVT[MVT::i32];
1725 TransformToType[MVT::f80] = MVT::i32;
1726 ValueTypeActions.setTypeAction(MVT::f80, TypeSoftenFloat);
1727 }
1728
1729 // Decide how to handle f64. If the target does not have native f64 support,
1730 // expand it to i64 and we will be generating soft float library calls.
1731 if (!isTypeLegal(MVT::f64)) {
1732 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1733 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1734 TransformToType[MVT::f64] = MVT::i64;
1735 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1736 }
1737
1738 // Decide how to handle f32. If the target does not have native f32 support,
1739 // expand it to i32 and we will be generating soft float library calls.
1740 if (!isTypeLegal(MVT::f32)) {
1741 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1742 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1743 TransformToType[MVT::f32] = MVT::i32;
1744 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1745 }
1746
1747 // Decide how to handle f16. If the target does not have native f16 support,
1748 // promote it to f32, because there are no f16 library calls (except for
1749 // conversions).
1750 if (!isTypeLegal(MVT::f16)) {
1751 // Allow targets to control how we legalize half.
1752 bool SoftPromoteHalfType = softPromoteHalfType();
1753 bool UseFPRegsForHalfType = !SoftPromoteHalfType || useFPRegsForHalfType();
1754
1755 if (!UseFPRegsForHalfType) {
1756 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1757 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1758 } else {
1759 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1760 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1761 }
1762 TransformToType[MVT::f16] = MVT::f32;
1763 if (SoftPromoteHalfType) {
1764 ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1765 } else {
1766 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1767 }
1768 }
1769
1770 // Decide how to handle bf16. If the target does not have native bf16 support,
1771 // promote it to f32, because there are no bf16 library calls (except for
1772 // converting from f32 to bf16).
1773 if (!isTypeLegal(MVT::bf16)) {
1774 NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32];
1775 RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32];
1776 TransformToType[MVT::bf16] = MVT::f32;
1777 ValueTypeActions.setTypeAction(MVT::bf16, TypeSoftPromoteHalf);
1778 }
1779
1780 // Loop over all of the vector value types to see which need transformations.
1781 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1782 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1783 MVT VT = (MVT::SimpleValueType) i;
1784 if (isTypeLegal(VT))
1785 continue;
1786
1787 MVT EltVT = VT.getVectorElementType();
1789 bool IsLegalWiderType = false;
1790 bool IsScalable = VT.isScalableVector();
1791 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1792 switch (PreferredAction) {
1793 case TypePromoteInteger: {
1794 MVT::SimpleValueType EndVT = IsScalable ?
1795 MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1796 MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1797 // Try to promote the elements of integer vectors. If no legal
1798 // promotion was found, fall through to the widen-vector method.
1799 for (unsigned nVT = i + 1;
1800 (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1801 MVT SVT = (MVT::SimpleValueType) nVT;
1802 // Promote vectors of integers to vectors with the same number
1803 // of elements, with a wider element type.
1804 if (SVT.getScalarSizeInBits() > EltVT.getFixedSizeInBits() &&
1805 SVT.getVectorElementCount() == EC && isTypeLegal(SVT)) {
1806 TransformToType[i] = SVT;
1807 RegisterTypeForVT[i] = SVT;
1808 NumRegistersForVT[i] = 1;
1809 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1810 IsLegalWiderType = true;
1811 break;
1812 }
1813 }
1814 if (IsLegalWiderType)
1815 break;
1816 [[fallthrough]];
1817 }
1818
1819 case TypeWidenVector:
1820 if (isPowerOf2_32(EC.getKnownMinValue())) {
1821 // Try to widen the vector.
1822 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1823 MVT SVT = (MVT::SimpleValueType) nVT;
1824 if (SVT.getVectorElementType() == EltVT &&
1825 SVT.isScalableVector() == IsScalable &&
1827 EC.getKnownMinValue() &&
1828 isTypeLegal(SVT)) {
1829 TransformToType[i] = SVT;
1830 RegisterTypeForVT[i] = SVT;
1831 NumRegistersForVT[i] = 1;
1832 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1833 IsLegalWiderType = true;
1834 break;
1835 }
1836 }
1837 if (IsLegalWiderType)
1838 break;
1839 } else {
1840 // Only widen to the next power of 2 to keep consistency with EVT.
1841 MVT NVT = VT.getPow2VectorType();
1842 if (isTypeLegal(NVT)) {
1843 TransformToType[i] = NVT;
1844 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1845 RegisterTypeForVT[i] = NVT;
1846 NumRegistersForVT[i] = 1;
1847 break;
1848 }
1849 }
1850 [[fallthrough]];
1851
1852 case TypeSplitVector:
1853 case TypeScalarizeVector: {
1854 MVT IntermediateVT;
1855 MVT RegisterVT;
1856 unsigned NumIntermediates;
1857 unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1858 NumIntermediates, RegisterVT, this);
1859 NumRegistersForVT[i] = NumRegisters;
1860 assert(NumRegistersForVT[i] == NumRegisters &&
1861 "NumRegistersForVT size cannot represent NumRegisters!");
1862 RegisterTypeForVT[i] = RegisterVT;
1863
1864 MVT NVT = VT.getPow2VectorType();
1865 if (NVT == VT) {
1866 // Type is already a power of 2. The default action is to split.
1867 TransformToType[i] = MVT::Other;
1868 if (PreferredAction == TypeScalarizeVector)
1869 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1870 else if (PreferredAction == TypeSplitVector)
1871 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1872 else if (EC.getKnownMinValue() > 1)
1873 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1874 else
1875 ValueTypeActions.setTypeAction(VT, EC.isScalable()
1878 } else {
1879 TransformToType[i] = NVT;
1880 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1881 }
1882 break;
1883 }
1884 default:
1885 llvm_unreachable("Unknown vector legalization action!");
1886 }
1887 }
1888
1889 // Determine the 'representative' register class for each value type.
1890 // An representative register class is the largest (meaning one which is
1891 // not a sub-register class / subreg register class) legal register class for
1892 // a group of value types. For example, on i386, i8, i16, and i32
1893 // representative would be GR32; while on x86_64 it's GR64.
1894 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1895 const TargetRegisterClass* RRC;
1896 uint8_t Cost;
1898 RepRegClassForVT[i] = RRC;
1899 RepRegClassCostForVT[i] = Cost;
1900 }
1901}
1902
1904 EVT VT) const {
1905 assert(!VT.isVector() && "No default SetCC type for vectors!");
1906 return getPointerTy(DL).SimpleTy;
1907}
1908
1910 return MVT::i32; // return the default value
1911}
1912
1913/// getVectorTypeBreakdown - Vector types are broken down into some number of
1914/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1915/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1916/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1917///
1918/// This method returns the number of registers needed, and the VT for each
1919/// register. It also returns the VT and quantity of the intermediate values
1920/// before they are promoted/expanded.
1922 EVT VT, EVT &IntermediateVT,
1923 unsigned &NumIntermediates,
1924 MVT &RegisterVT) const {
1925 ElementCount EltCnt = VT.getVectorElementCount();
1926
1927 // If there is a wider vector type with the same element type as this one,
1928 // or a promoted vector type that has the same number of elements which
1929 // are wider, then we should convert to that legal vector type.
1930 // This handles things like <2 x float> -> <4 x float> and
1931 // <4 x i1> -> <4 x i32>.
1932 LegalizeTypeAction TA = getTypeAction(Context, VT);
1933 if (!EltCnt.isScalar() &&
1934 (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1935 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1936 if (isTypeLegal(RegisterEVT)) {
1937 IntermediateVT = RegisterEVT;
1938 RegisterVT = RegisterEVT.getSimpleVT();
1939 NumIntermediates = 1;
1940 return 1;
1941 }
1942 }
1943
1944 // Figure out the right, legal destination reg to copy into.
1945 EVT EltTy = VT.getVectorElementType();
1946
1947 unsigned NumVectorRegs = 1;
1948
1949 // Scalable vectors cannot be scalarized, so handle the legalisation of the
1950 // types like done elsewhere in SelectionDAG.
1951 if (EltCnt.isScalable()) {
1952 LegalizeKind LK;
1953 EVT PartVT = VT;
1954 do {
1955 // Iterate until we've found a legal (part) type to hold VT.
1956 LK = getTypeConversion(Context, PartVT);
1957 PartVT = LK.second;
1958 } while (LK.first != TypeLegal);
1959
1960 if (!PartVT.isVector()) {
1962 "Don't know how to legalize this scalable vector type");
1963 }
1964
1965 NumIntermediates =
1968 IntermediateVT = PartVT;
1969 RegisterVT = getRegisterType(Context, IntermediateVT);
1970 return NumIntermediates;
1971 }
1972
1973 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally
1974 // we could break down into LHS/RHS like LegalizeDAG does.
1975 if (!isPowerOf2_32(EltCnt.getKnownMinValue())) {
1976 NumVectorRegs = EltCnt.getKnownMinValue();
1977 EltCnt = ElementCount::getFixed(1);
1978 }
1979
1980 // Divide the input until we get to a supported size. This will always
1981 // end with a scalar if the target doesn't support vectors.
1982 while (EltCnt.getKnownMinValue() > 1 &&
1983 !isTypeLegal(EVT::getVectorVT(Context, EltTy, EltCnt))) {
1984 EltCnt = EltCnt.divideCoefficientBy(2);
1985 NumVectorRegs <<= 1;
1986 }
1987
1988 NumIntermediates = NumVectorRegs;
1989
1990 EVT NewVT = EVT::getVectorVT(Context, EltTy, EltCnt);
1991 if (!isTypeLegal(NewVT))
1992 NewVT = EltTy;
1993 IntermediateVT = NewVT;
1994
1995 MVT DestVT = getRegisterType(Context, NewVT);
1996 RegisterVT = DestVT;
1997
1998 if (EVT(DestVT).bitsLT(NewVT)) { // Value is expanded, e.g. i64 -> i16.
1999 TypeSize NewVTSize = NewVT.getSizeInBits();
2000 // Convert sizes such as i33 to i64.
2002 NewVTSize = NewVTSize.coefficientNextPowerOf2();
2003 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
2004 }
2005
2006 // Otherwise, promotion or legal types use the same number of registers as
2007 // the vector decimated to the appropriate level.
2008 return NumVectorRegs;
2009}
2010
2012 uint64_t NumCases,
2014 ProfileSummaryInfo *PSI,
2015 BlockFrequencyInfo *BFI) const {
2016 // FIXME: This function check the maximum table size and density, but the
2017 // minimum size is not checked. It would be nice if the minimum size is
2018 // also combined within this function. Currently, the minimum size check is
2019 // performed in findJumpTable() in SelectionDAGBuiler and
2020 // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
2021 const bool OptForSize =
2022 llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
2023 const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
2024 const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
2025
2026 // Check whether the number of cases is small enough and
2027 // the range is dense enough for a jump table.
2028 return (OptForSize || Range <= MaxJumpTableSize) &&
2029 (NumCases * 100 >= Range * MinDensity);
2030}
2031
2033 EVT ConditionVT) const {
2034 return getRegisterType(Context, ConditionVT);
2035}
2036
2037/// Get the EVTs and ArgFlags collections that represent the legalized return
2038/// type of the given function. This does not require a DAG or a return value,
2039/// and is suitable for use before any DAGs for the function are constructed.
2040/// TODO: Move this out of TargetLowering.cpp.
2042 AttributeList attr,
2044 const TargetLowering &TLI, const DataLayout &DL) {
2046 ComputeValueTypes(DL, ReturnType, Types);
2047 unsigned NumValues = Types.size();
2048 if (NumValues == 0) return;
2049
2050 for (Type *Ty : Types) {
2051 EVT VT = TLI.getValueType(DL, Ty);
2052 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
2053
2054 if (attr.hasRetAttr(Attribute::SExt))
2055 ExtendKind = ISD::SIGN_EXTEND;
2056 else if (attr.hasRetAttr(Attribute::ZExt))
2057 ExtendKind = ISD::ZERO_EXTEND;
2058
2059 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
2060 VT = TLI.getTypeForExtReturn(ReturnType->getContext(), VT, ExtendKind);
2061
2062 unsigned NumParts =
2063 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
2064 MVT PartVT =
2065 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
2066
2067 // 'inreg' on function refers to return value
2069 if (attr.hasRetAttr(Attribute::InReg))
2070 Flags.setInReg();
2071
2072 // Propagate extension type if any
2073 if (attr.hasRetAttr(Attribute::SExt))
2074 Flags.setSExt();
2075 else if (attr.hasRetAttr(Attribute::ZExt))
2076 Flags.setZExt();
2077
2078 for (unsigned i = 0; i < NumParts; ++i)
2079 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, Ty, 0, 0));
2080 }
2081}
2082
2084 const DataLayout &DL) const {
2085 return DL.getABITypeAlign(Ty);
2086}
2087
2089 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
2090 Align Alignment, MachineMemOperand::Flags Flags, unsigned *Fast) const {
2091 // Check if the specified alignment is sufficient based on the data layout.
2092 // TODO: While using the data layout works in practice, a better solution
2093 // would be to implement this check directly (make this a virtual function).
2094 // For example, the ABI alignment may change based on software platform while
2095 // this function should only be affected by hardware implementation.
2096 Type *Ty = VT.getTypeForEVT(Context);
2097 if (VT.isZeroSized() || Alignment >= DL.getABITypeAlign(Ty)) {
2098 // Assume that an access that meets the ABI-specified alignment is fast.
2099 if (Fast != nullptr)
2100 *Fast = 1;
2101 return true;
2102 }
2103
2104 // This is a misaligned access.
2105 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast);
2106}
2107
2109 LLVMContext &Context, const DataLayout &DL, EVT VT,
2110 const MachineMemOperand &MMO, unsigned *Fast) const {
2111 return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
2112 MMO.getAlign(), MMO.getFlags(), Fast);
2113}
2114
2116 const DataLayout &DL, EVT VT,
2117 unsigned AddrSpace, Align Alignment,
2119 unsigned *Fast) const {
2120 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
2121 Flags, Fast);
2122}
2123
2125 const DataLayout &DL, EVT VT,
2126 const MachineMemOperand &MMO,
2127 unsigned *Fast) const {
2128 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
2129 MMO.getFlags(), Fast);
2130}
2131
2133 const DataLayout &DL, LLT Ty,
2134 const MachineMemOperand &MMO,
2135 unsigned *Fast) const {
2136 EVT VT = getApproximateEVTForLLT(Ty, Context);
2137 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
2138 MMO.getFlags(), Fast);
2139}
2140
2141unsigned TargetLoweringBase::getMaxStoresPerMemset(bool OptSize) const {
2144
2146}
2147
2148unsigned TargetLoweringBase::getMaxStoresPerMemcpy(bool OptSize) const {
2151
2153}
2154
2158
2160}
2161
2162//===----------------------------------------------------------------------===//
2163// TargetTransformInfo Helpers
2164//===----------------------------------------------------------------------===//
2165
2167 enum InstructionOpcodes {
2168#define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
2169#define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
2170#include "llvm/IR/Instruction.def"
2171 };
2172 switch (static_cast<InstructionOpcodes>(Opcode)) {
2173 case Ret: return 0;
2174 case Br: return 0;
2175 case Switch: return 0;
2176 case IndirectBr: return 0;
2177 case Invoke: return 0;
2178 case CallBr: return 0;
2179 case Resume: return 0;
2180 case Unreachable: return 0;
2181 case CleanupRet: return 0;
2182 case CatchRet: return 0;
2183 case CatchPad: return 0;
2184 case CatchSwitch: return 0;
2185 case CleanupPad: return 0;
2186 case FNeg: return ISD::FNEG;
2187 case Add: return ISD::ADD;
2188 case FAdd: return ISD::FADD;
2189 case Sub: return ISD::SUB;
2190 case FSub: return ISD::FSUB;
2191 case Mul: return ISD::MUL;
2192 case FMul: return ISD::FMUL;
2193 case UDiv: return ISD::UDIV;
2194 case SDiv: return ISD::SDIV;
2195 case FDiv: return ISD::FDIV;
2196 case URem: return ISD::UREM;
2197 case SRem: return ISD::SREM;
2198 case FRem: return ISD::FREM;
2199 case Shl: return ISD::SHL;
2200 case LShr: return ISD::SRL;
2201 case AShr: return ISD::SRA;
2202 case And: return ISD::AND;
2203 case Or: return ISD::OR;
2204 case Xor: return ISD::XOR;
2205 case Alloca: return 0;
2206 case Load: return ISD::LOAD;
2207 case Store: return ISD::STORE;
2208 case GetElementPtr: return 0;
2209 case Fence: return 0;
2210 case AtomicCmpXchg: return 0;
2211 case AtomicRMW: return 0;
2212 case Trunc: return ISD::TRUNCATE;
2213 case ZExt: return ISD::ZERO_EXTEND;
2214 case SExt: return ISD::SIGN_EXTEND;
2215 case FPToUI: return ISD::FP_TO_UINT;
2216 case FPToSI: return ISD::FP_TO_SINT;
2217 case UIToFP: return ISD::UINT_TO_FP;
2218 case SIToFP: return ISD::SINT_TO_FP;
2219 case FPTrunc: return ISD::FP_ROUND;
2220 case FPExt: return ISD::FP_EXTEND;
2221 case PtrToAddr: return ISD::BITCAST;
2222 case PtrToInt: return ISD::BITCAST;
2223 case IntToPtr: return ISD::BITCAST;
2224 case BitCast: return ISD::BITCAST;
2225 case AddrSpaceCast: return ISD::ADDRSPACECAST;
2226 case ICmp: return ISD::SETCC;
2227 case FCmp: return ISD::SETCC;
2228 case PHI: return 0;
2229 case Call: return 0;
2230 case Select: return ISD::SELECT;
2231 case UserOp1: return 0;
2232 case UserOp2: return 0;
2233 case VAArg: return 0;
2234 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
2235 case InsertElement: return ISD::INSERT_VECTOR_ELT;
2236 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
2237 case ExtractValue: return ISD::MERGE_VALUES;
2238 case InsertValue: return ISD::MERGE_VALUES;
2239 case LandingPad: return 0;
2240 case Freeze: return ISD::FREEZE;
2241 }
2242
2243 llvm_unreachable("Unknown instruction type encountered!");
2244}
2245
2247 switch (ID) {
2248 case Intrinsic::exp:
2249 return ISD::FEXP;
2250 case Intrinsic::exp2:
2251 return ISD::FEXP2;
2252 case Intrinsic::log:
2253 return ISD::FLOG;
2254 default:
2255 return ISD::DELETED_NODE;
2256 }
2257}
2258
2259Value *
2261 bool UseTLS) const {
2262 // compiler-rt provides a variable with a magic name. Targets that do not
2263 // link with compiler-rt may also provide such a variable.
2264 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
2265 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
2266 auto UnsafeStackPtr =
2267 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
2268
2269 const DataLayout &DL = M->getDataLayout();
2270 PointerType *StackPtrTy = DL.getAllocaPtrType(M->getContext());
2271
2272 if (!UnsafeStackPtr) {
2273 auto TLSModel = UseTLS ?
2276 // The global variable is not defined yet, define it ourselves.
2277 // We use the initial-exec TLS model because we do not support the
2278 // variable living anywhere other than in the main executable.
2279 UnsafeStackPtr = new GlobalVariable(
2280 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
2281 UnsafeStackPtrVar, nullptr, TLSModel);
2282 } else {
2283 // The variable exists, check its type and attributes.
2284 //
2285 // FIXME: Move to IR verifier.
2286 if (UnsafeStackPtr->getValueType() != StackPtrTy)
2287 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
2288 if (UseTLS != UnsafeStackPtr->isThreadLocal())
2289 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
2290 (UseTLS ? "" : "not ") + "be thread-local");
2291 }
2292 return UnsafeStackPtr;
2293}
2294
2295Value *
2297 // FIXME: Can this triple check be replaced with SAFESTACK_POINTER_ADDRESS
2298 // being available?
2299 if (!TM.getTargetTriple().isAndroid())
2300 return getDefaultSafeStackPointerLocation(IRB, true);
2301
2302 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
2303 auto *PtrTy = PointerType::getUnqual(M->getContext());
2304
2305 const char *SafestackPointerAddressName =
2306 getLibcallName(RTLIB::SAFESTACK_POINTER_ADDRESS);
2307 if (!SafestackPointerAddressName) {
2308 M->getContext().emitError(
2309 "no libcall available for safestack pointer address");
2310 return PoisonValue::get(PtrTy);
2311 }
2312
2313 // Android provides a libc function to retrieve the address of the current
2314 // thread's unsafe stack pointer.
2315 FunctionCallee Fn =
2316 M->getOrInsertFunction(SafestackPointerAddressName, PtrTy);
2317 return IRB.CreateCall(Fn);
2318}
2319
2320//===----------------------------------------------------------------------===//
2321// Loop Strength Reduction hooks
2322//===----------------------------------------------------------------------===//
2323
2324/// isLegalAddressingMode - Return true if the addressing mode represented
2325/// by AM is legal for this target, for a load/store of the specified type.
2327 const AddrMode &AM, Type *Ty,
2328 unsigned AS, Instruction *I) const {
2329 // The default implementation of this implements a conservative RISCy, r+r and
2330 // r+i addr mode.
2331
2332 // Scalable offsets not supported
2333 if (AM.ScalableOffset)
2334 return false;
2335
2336 // Allows a sign-extended 16-bit immediate field.
2337 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2338 return false;
2339
2340 // No global is ever allowed as a base.
2341 if (AM.BaseGV)
2342 return false;
2343
2344 // Only support r+r,
2345 switch (AM.Scale) {
2346 case 0: // "r+i" or just "i", depending on HasBaseReg.
2347 break;
2348 case 1:
2349 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2350 return false;
2351 // Otherwise we have r+r or r+i.
2352 break;
2353 case 2:
2354 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2355 return false;
2356 // Allow 2*r as r+r.
2357 break;
2358 default: // Don't allow n * r
2359 return false;
2360 }
2361
2362 return true;
2363}
2364
2365//===----------------------------------------------------------------------===//
2366// Stack Protector
2367//===----------------------------------------------------------------------===//
2368
2369// For OpenBSD return its special guard variable. Otherwise return nullptr,
2370// so that SelectionDAG handle SSP.
2372 RTLIB::LibcallImpl GuardLocalImpl = getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2373 if (GuardLocalImpl != RTLIB::impl___guard_local)
2374 return nullptr;
2375
2376 Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
2377 const DataLayout &DL = M.getDataLayout();
2378 PointerType *PtrTy =
2379 PointerType::get(M.getContext(), DL.getDefaultGlobalsAddressSpace());
2380 GlobalVariable *G =
2381 M.getOrInsertGlobal(getLibcallImplName(GuardLocalImpl), PtrTy);
2382 G->setVisibility(GlobalValue::HiddenVisibility);
2383 return G;
2384}
2385
2386// Currently only support "standard" __stack_chk_guard.
2387// TODO: add LOAD_STACK_GUARD support.
2389 RTLIB::LibcallImpl StackGuardImpl = getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2390 if (StackGuardImpl == RTLIB::Unsupported)
2391 return;
2392
2393 StringRef StackGuardVarName = getLibcallImplName(StackGuardImpl);
2394 M.getOrInsertGlobal(
2395 StackGuardVarName, PointerType::getUnqual(M.getContext()), [=, &M]() {
2396 auto *GV = new GlobalVariable(M, PointerType::getUnqual(M.getContext()),
2397 false, GlobalVariable::ExternalLinkage,
2398 nullptr, StackGuardVarName);
2399
2400 // FreeBSD has "__stack_chk_guard" defined externally on libc.so
2401 if (M.getDirectAccessExternalData() &&
2402 !TM.getTargetTriple().isOSCygMing() &&
2403 !(TM.getTargetTriple().isPPC64() &&
2404 TM.getTargetTriple().isOSFreeBSD()) &&
2405 (!TM.getTargetTriple().isOSDarwin() ||
2406 TM.getRelocationModel() == Reloc::Static))
2407 GV->setDSOLocal(true);
2408
2409 return GV;
2410 });
2411}
2412
2413// Currently only support "standard" __stack_chk_guard.
2414// TODO: add LOAD_STACK_GUARD support.
2416 RTLIB::LibcallImpl GuardVarImpl = getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2417 if (GuardVarImpl == RTLIB::Unsupported)
2418 return nullptr;
2419 return M.getNamedValue(getLibcallImplName(GuardVarImpl));
2420}
2421
2423 // MSVC CRT has a function to validate security cookie.
2424 RTLIB::LibcallImpl SecurityCheckCookieLibcall =
2425 getLibcallImpl(RTLIB::SECURITY_CHECK_COOKIE);
2426 if (SecurityCheckCookieLibcall != RTLIB::Unsupported)
2427 return M.getFunction(getLibcallImplName(SecurityCheckCookieLibcall));
2428 return nullptr;
2429}
2430
2434
2438
2439unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
2440 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
2441}
2442
2446
2450
2454
2456 return MinimumBitTestCmps;
2457}
2458
2460 MinimumBitTestCmps = Val;
2461}
2462
2464 if (TM.Options.LoopAlignment)
2465 return Align(TM.Options.LoopAlignment);
2466 return PrefLoopAlignment;
2467}
2468
2470 MachineBasicBlock *MBB) const {
2471 return MaxBytesForAlignment;
2472}
2473
2474//===----------------------------------------------------------------------===//
2475// Reciprocal Estimates
2476//===----------------------------------------------------------------------===//
2477
2478/// Get the reciprocal estimate attribute string for a function that will
2479/// override the target defaults.
2481 const Function &F = MF.getFunction();
2482 return F.getFnAttribute("reciprocal-estimates").getValueAsString();
2483}
2484
2485/// Construct a string for the given reciprocal operation of the given type.
2486/// This string should match the corresponding option to the front-end's
2487/// "-mrecip" flag assuming those strings have been passed through in an
2488/// attribute string. For example, "vec-divf" for a division of a vXf32.
2489static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
2490 std::string Name = VT.isVector() ? "vec-" : "";
2491
2492 Name += IsSqrt ? "sqrt" : "div";
2493
2494 // TODO: Handle other float types?
2495 if (VT.getScalarType() == MVT::f64) {
2496 Name += "d";
2497 } else if (VT.getScalarType() == MVT::f16) {
2498 Name += "h";
2499 } else {
2500 assert(VT.getScalarType() == MVT::f32 &&
2501 "Unexpected FP type for reciprocal estimate");
2502 Name += "f";
2503 }
2504
2505 return Name;
2506}
2507
2508/// Return the character position and value (a single numeric character) of a
2509/// customized refinement operation in the input string if it exists. Return
2510/// false if there is no customized refinement step count.
2511static bool parseRefinementStep(StringRef In, size_t &Position,
2512 uint8_t &Value) {
2513 const char RefStepToken = ':';
2514 Position = In.find(RefStepToken);
2515 if (Position == StringRef::npos)
2516 return false;
2517
2518 StringRef RefStepString = In.substr(Position + 1);
2519 // Allow exactly one numeric character for the additional refinement
2520 // step parameter.
2521 if (RefStepString.size() == 1) {
2522 char RefStepChar = RefStepString[0];
2523 if (isDigit(RefStepChar)) {
2524 Value = RefStepChar - '0';
2525 return true;
2526 }
2527 }
2528 report_fatal_error("Invalid refinement step for -recip.");
2529}
2530
2531/// For the input attribute string, return one of the ReciprocalEstimate enum
2532/// status values (enabled, disabled, or not specified) for this operation on
2533/// the specified data type.
2534static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
2535 if (Override.empty())
2537
2538 SmallVector<StringRef, 4> OverrideVector;
2539 Override.split(OverrideVector, ',');
2540 unsigned NumArgs = OverrideVector.size();
2541
2542 // Check if "all", "none", or "default" was specified.
2543 if (NumArgs == 1) {
2544 // Look for an optional setting of the number of refinement steps needed
2545 // for this type of reciprocal operation.
2546 size_t RefPos;
2547 uint8_t RefSteps;
2548 if (parseRefinementStep(Override, RefPos, RefSteps)) {
2549 // Split the string for further processing.
2550 Override = Override.substr(0, RefPos);
2551 }
2552
2553 // All reciprocal types are enabled.
2554 if (Override == "all")
2556
2557 // All reciprocal types are disabled.
2558 if (Override == "none")
2560
2561 // Target defaults for enablement are used.
2562 if (Override == "default")
2564 }
2565
2566 // The attribute string may omit the size suffix ('f'/'d').
2567 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2568 std::string VTNameNoSize = VTName;
2569 VTNameNoSize.pop_back();
2570 static const char DisabledPrefix = '!';
2571
2572 for (StringRef RecipType : OverrideVector) {
2573 size_t RefPos;
2574 uint8_t RefSteps;
2575 if (parseRefinementStep(RecipType, RefPos, RefSteps))
2576 RecipType = RecipType.substr(0, RefPos);
2577
2578 // Ignore the disablement token for string matching.
2579 bool IsDisabled = RecipType[0] == DisabledPrefix;
2580 if (IsDisabled)
2581 RecipType = RecipType.substr(1);
2582
2583 if (RecipType == VTName || RecipType == VTNameNoSize)
2586 }
2587
2589}
2590
2591/// For the input attribute string, return the customized refinement step count
2592/// for this operation on the specified data type. If the step count does not
2593/// exist, return the ReciprocalEstimate enum value for unspecified.
2594static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2595 if (Override.empty())
2597
2598 SmallVector<StringRef, 4> OverrideVector;
2599 Override.split(OverrideVector, ',');
2600 unsigned NumArgs = OverrideVector.size();
2601
2602 // Check if "all", "default", or "none" was specified.
2603 if (NumArgs == 1) {
2604 // Look for an optional setting of the number of refinement steps needed
2605 // for this type of reciprocal operation.
2606 size_t RefPos;
2607 uint8_t RefSteps;
2608 if (!parseRefinementStep(Override, RefPos, RefSteps))
2610
2611 // Split the string for further processing.
2612 Override = Override.substr(0, RefPos);
2613 assert(Override != "none" &&
2614 "Disabled reciprocals, but specifed refinement steps?");
2615
2616 // If this is a general override, return the specified number of steps.
2617 if (Override == "all" || Override == "default")
2618 return RefSteps;
2619 }
2620
2621 // The attribute string may omit the size suffix ('f'/'d').
2622 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2623 std::string VTNameNoSize = VTName;
2624 VTNameNoSize.pop_back();
2625
2626 for (StringRef RecipType : OverrideVector) {
2627 size_t RefPos;
2628 uint8_t RefSteps;
2629 if (!parseRefinementStep(RecipType, RefPos, RefSteps))
2630 continue;
2631
2632 RecipType = RecipType.substr(0, RefPos);
2633 if (RecipType == VTName || RecipType == VTNameNoSize)
2634 return RefSteps;
2635 }
2636
2638}
2639
2644
2649
2654
2659
2661 EVT LoadVT, EVT BitcastVT, const SelectionDAG &DAG,
2662 const MachineMemOperand &MMO) const {
2663 // Single-element vectors are scalarized, so we should generally avoid having
2664 // any memory operations on such types, as they would get scalarized too.
2665 if (LoadVT.isFixedLengthVector() && BitcastVT.isFixedLengthVector() &&
2666 BitcastVT.getVectorNumElements() == 1)
2667 return false;
2668
2669 // Don't do if we could do an indexed load on the original type, but not on
2670 // the new one.
2671 if (!LoadVT.isSimple() || !BitcastVT.isSimple())
2672 return true;
2673
2674 MVT LoadMVT = LoadVT.getSimpleVT();
2675
2676 // Don't bother doing this if it's just going to be promoted again later, as
2677 // doing so might interfere with other combines.
2678 if (getOperationAction(ISD::LOAD, LoadMVT) == Promote &&
2679 getTypeToPromoteTo(ISD::LOAD, LoadMVT) == BitcastVT.getSimpleVT())
2680 return false;
2681
2682 unsigned Fast = 0;
2683 return allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), BitcastVT,
2684 MMO, &Fast) &&
2685 Fast;
2686}
2687
2691
2693 const LoadInst &LI, const DataLayout &DL, AssumptionCache *AC,
2694 const TargetLibraryInfo *LibInfo) const {
2696 if (LI.isVolatile())
2698
2699 if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2701
2702 if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2704
2706 LI.getAlign(), DL, &LI, AC,
2707 /*DT=*/nullptr, LibInfo))
2709
2710 Flags |= getTargetMMOFlags(LI);
2711 return Flags;
2712}
2713
2716 const DataLayout &DL) const {
2718
2719 if (SI.isVolatile())
2721
2722 if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2724
2725 // FIXME: Not preserving dereferenceable
2726 Flags |= getTargetMMOFlags(SI);
2727 return Flags;
2728}
2729
2732 const DataLayout &DL) const {
2734
2735 if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2736 if (RMW->isVolatile())
2738 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2739 if (CmpX->isVolatile())
2741 } else
2742 llvm_unreachable("not an atomic instruction");
2743
2744 // FIXME: Not preserving dereferenceable
2745 Flags |= getTargetMMOFlags(AI);
2746 return Flags;
2747}
2748
2750 const VPIntrinsic &VPIntrin) const {
2752 Intrinsic::ID IntrinID = VPIntrin.getIntrinsicID();
2753
2754 switch (IntrinID) {
2755 default:
2756 llvm_unreachable("unexpected intrinsic. Existing code may be appropriate "
2757 "for it, but support must be explicitly enabled");
2758 case Intrinsic::vp_load:
2759 case Intrinsic::vp_gather:
2760 case Intrinsic::experimental_vp_strided_load:
2762 break;
2763 case Intrinsic::vp_store:
2764 case Intrinsic::vp_scatter:
2765 case Intrinsic::experimental_vp_strided_store:
2767 break;
2768 }
2769
2770 if (VPIntrin.hasMetadata(LLVMContext::MD_nontemporal))
2772
2773 Flags |= getTargetMMOFlags(VPIntrin);
2774 return Flags;
2775}
2776
2778 Instruction *Inst,
2779 AtomicOrdering Ord) const {
2780 if (isReleaseOrStronger(Ord) && Inst->hasAtomicStore())
2781 return Builder.CreateFence(Ord);
2782 else
2783 return nullptr;
2784}
2785
2787 Instruction *Inst,
2788 AtomicOrdering Ord) const {
2789 if (isAcquireOrStronger(Ord))
2790 return Builder.CreateFence(Ord);
2791 else
2792 return nullptr;
2793}
2794
2795//===----------------------------------------------------------------------===//
2796// GlobalISel Hooks
2797//===----------------------------------------------------------------------===//
2798
2800 const TargetTransformInfo *TTI) const {
2801 auto &MF = *MI.getMF();
2802 auto &MRI = MF.getRegInfo();
2803 // Assuming a spill and reload of a value has a cost of 1 instruction each,
2804 // this helper function computes the maximum number of uses we should consider
2805 // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2806 // break even in terms of code size when the original MI has 2 users vs
2807 // choosing to potentially spill. Any more than 2 users we we have a net code
2808 // size increase. This doesn't take into account register pressure though.
2809 auto maxUses = [](unsigned RematCost) {
2810 // A cost of 1 means remats are basically free.
2811 if (RematCost == 1)
2812 return std::numeric_limits<unsigned>::max();
2813 if (RematCost == 2)
2814 return 2U;
2815
2816 // Remat is too expensive, only sink if there's one user.
2817 if (RematCost > 2)
2818 return 1U;
2819 llvm_unreachable("Unexpected remat cost");
2820 };
2821
2822 switch (MI.getOpcode()) {
2823 default:
2824 return false;
2825 // Constants-like instructions should be close to their users.
2826 // We don't want long live-ranges for them.
2827 case TargetOpcode::G_CONSTANT:
2828 case TargetOpcode::G_FCONSTANT:
2829 case TargetOpcode::G_FRAME_INDEX:
2830 case TargetOpcode::G_INTTOPTR:
2831 return true;
2832 case TargetOpcode::G_GLOBAL_VALUE: {
2833 unsigned RematCost = TTI->getGISelRematGlobalCost();
2834 Register Reg = MI.getOperand(0).getReg();
2835 unsigned MaxUses = maxUses(RematCost);
2836 if (MaxUses == UINT_MAX)
2837 return true; // Remats are "free" so always localize.
2838 return MRI.hasAtMostUserInstrs(Reg, MaxUses);
2839 }
2840 }
2841}
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 cl::opt< unsigned > MaxStoresPerMemmoveOverride("max-store-memmove", cl::init(0), cl::Hidden, cl::desc("Override target's MaxStoresPerMemmove and " "MaxStoresPerMemmoveOptSize. " "Set to 0 to use the target default."))
static std::string getReciprocalOpName(bool IsSqrt, EVT VT)
Construct a string for the given reciprocal operation of the given type.
#define LCALL5(A)
static cl::opt< unsigned > MaxStoresPerMemsetOverride("max-store-memset", cl::init(0), cl::Hidden, cl::desc("Override target's MaxStoresPerMemset and " "MaxStoresPerMemsetOptSize. " "Set to 0 to use the target default."))
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 > MaxStoresPerMemcpyOverride("max-store-memcpy", cl::init(0), cl::Hidden, cl::desc("Override target's MaxStoresPerMemcpy and " "MaxStoresPerMemcpyOptSize. " "Set to 0 to use the target default."))
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:726
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:64
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:712
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...
unsigned getMaxStoresPerMemcpy(bool OptSize) const
Get maximum # of store operations permitted for llvm.memcpy.
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
Warning: this option is problem-prone and tends to introduce float miscompilations:
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...
unsigned getMaxStoresPerMemset(bool OptSize) const
Get maximum # of store operations permitted for llvm.memset.
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.
unsigned getMaxStoresPerMemmove(bool OptSize) const
Get maximum # of store operations permitted for llvm.memmove.
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:818
@ MERGE_VALUES
MERGE_VALUES - This node takes multiple discrete operands and returns them all as its individual resu...
Definition ISDOpcodes.h:261
@ CTLZ_ZERO_UNDEF
Definition ISDOpcodes.h:787
@ DELETED_NODE
DELETED_NODE - This is an illegal value that is used to catch errors.
Definition ISDOpcodes.h:45
@ SET_FPENV
Sets the current floating-point environment.
@ LOOP_DEPENDENCE_RAW_MASK
@ VECREDUCE_SEQ_FADD
Generic reduction nodes.
@ FGETSIGN
INT = FGETSIGN(FP) - Return the sign bit of the specified floating point value as an integer 0/1 valu...
Definition ISDOpcodes.h:538
@ STACKADDRESS
STACKADDRESS - Represents the llvm.stackaddress intrinsic.
Definition ISDOpcodes.h:127
@ SMULFIX
RESULT = [US]MULFIX(LHS, RHS, SCALE) - Perform fixed point multiplication on 2 integers with the same...
Definition ISDOpcodes.h:394
@ ADDC
Carry-setting nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:294
@ RESET_FPENV
Set floating-point environment to default state.
@ FMAD
FMAD - Perform a * b + c, while getting the same result as the separately rounded operations.
Definition ISDOpcodes.h:522
@ ADD
Simple integer binary arithmetic operators.
Definition ISDOpcodes.h:264
@ LOAD
LOAD and STORE have token chains as their first operand, then the same operands as an LLVM load/store...
@ SMULFIXSAT
Same as the corresponding unsaturated fixed point instructions, but the result is clamped between the...
Definition ISDOpcodes.h:400
@ SET_FPMODE
Sets the current dynamic floating-point control modes.
@ ANY_EXTEND
ANY_EXTEND - Used for integer types. The high bits are undefined.
Definition ISDOpcodes.h:852
@ FMA
FMA - Perform a * b + c with no intermediate rounding step.
Definition ISDOpcodes.h:518
@ VECTOR_FIND_LAST_ACTIVE
Finds the index of the last active mask element Operands: Mask.
@ FMODF
FMODF - Decomposes the operand into integral and fractional parts, each having the same type and sign...
@ FATAN2
FATAN2 - atan2, inspired by libm.
@ FSINCOSPI
FSINCOSPI - Compute both the sine and cosine times pi more accurately than FSINCOS(pi*x),...
@ ATOMIC_CMP_SWAP_WITH_SUCCESS
Val, Success, OUTCHAIN = ATOMIC_CMP_SWAP_WITH_SUCCESS(INCHAIN, ptr, cmp, swap) N.b.
@ SINT_TO_FP
[SU]INT_TO_FP - These operators convert integers (whose interpreted sign depends on the first letter)...
Definition ISDOpcodes.h:879
@ CONCAT_VECTORS
CONCAT_VECTORS(VECTOR0, VECTOR1, ...) - Given a number of values of vector type with the same length ...
Definition ISDOpcodes.h:584
@ VECREDUCE_FMAX
FMIN/FMAX nodes can have flags, for NaN/NoNaN variants.
@ FADD
Simple binary floating point operators.
Definition ISDOpcodes.h:417
@ VECREDUCE_FMAXIMUM
FMINIMUM/FMAXIMUM nodes propatate NaNs and signed zeroes using the llvm.minimum and llvm....
@ ABS
ABS - Determine the unsigned absolute value of a signed integer value of the same bitwidth.
Definition ISDOpcodes.h:746
@ RESET_FPMODE
Sets default dynamic floating-point control modes.
@ SIGN_EXTEND_VECTOR_INREG
SIGN_EXTEND_VECTOR_INREG(Vector) - This operator represents an in-register sign-extension of the low ...
Definition ISDOpcodes.h:909
@ FMULADD
FMULADD - Performs a * b + c, with, or without, intermediate rounding.
Definition ISDOpcodes.h:528
@ BITCAST
BITCAST - This operator converts between integer, vector and FP values, as if the value was stored to...
Definition ISDOpcodes.h:992
@ CLMUL
Carry-less multiplication operations.
Definition ISDOpcodes.h:773
@ FLDEXP
FLDEXP - ldexp, inspired by libm (op0 * 2**op1).
@ SDIVFIX
RESULT = [US]DIVFIX(LHS, RHS, SCALE) - Perform fixed point division on 2 integers with the same width...
Definition ISDOpcodes.h:407
@ BUILTIN_OP_END
BUILTIN_OP_END - This must be the last enum value in this list.
@ SIGN_EXTEND
Conversion operators.
Definition ISDOpcodes.h:843
@ AVGCEILS
AVGCEILS/AVGCEILU - Rounding averaging add - Add two integers using an integer of type i[N+2],...
Definition ISDOpcodes.h:714
@ READSTEADYCOUNTER
READSTEADYCOUNTER - This corresponds to the readfixedcounter intrinsic.
@ VECREDUCE_FADD
These reductions have relaxed evaluation order semantics, and have a single vector operand.
@ CTTZ_ZERO_UNDEF
Bit counting operators with an undefined result for zero inputs.
Definition ISDOpcodes.h:786
@ PREFETCH
PREFETCH - This corresponds to a prefetch intrinsic.
@ TRUNCATE_SSAT_U
Definition ISDOpcodes.h:872
@ FSINCOS
FSINCOS - Compute both fsin and fcos as a single operation.
@ 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:826
@ FNEG
Perform various unary floating-point operations inspired by libm.
@ SSUBO
Same for subtraction.
Definition ISDOpcodes.h:352
@ FCANONICALIZE
Returns platform specific canonical encoding of a floating point number.
Definition ISDOpcodes.h:541
@ IS_FPCLASS
Performs a check of floating point class property, defined by IEEE-754.
Definition ISDOpcodes.h:548
@ SSUBSAT
RESULT = [US]SUBSAT(LHS, RHS) - Perform saturation subtraction on 2 integers with the same bit width ...
Definition ISDOpcodes.h:374
@ SELECT
Select(COND, TRUEVAL, FALSEVAL).
Definition ISDOpcodes.h:795
@ SPLAT_VECTOR
SPLAT_VECTOR(VAL) - Returns a vector with the scalar value VAL duplicated in all lanes.
Definition ISDOpcodes.h:671
@ SADDO
RESULT, BOOL = [SU]ADDO(LHS, RHS) - Overflow-aware nodes for addition.
Definition ISDOpcodes.h:348
@ CTLS
Count leading redundant sign bits.
Definition ISDOpcodes.h:791
@ VECREDUCE_ADD
Integer reductions may have a result type larger than the vector element type.
@ GET_FPMODE
Reads the current dynamic floating-point control modes.
@ GET_FPENV
Gets the current floating-point environment.
@ SHL
Shift and rotation operations.
Definition ISDOpcodes.h:764
@ VECTOR_SHUFFLE
VECTOR_SHUFFLE(VEC1, VEC2) - Returns a vector, of the same type as VEC1/VEC2.
Definition ISDOpcodes.h:649
@ FMINNUM_IEEE
FMINNUM_IEEE/FMAXNUM_IEEE - Perform floating-point minimumNumber or maximumNumber on two values,...
@ EXTRACT_VECTOR_ELT
EXTRACT_VECTOR_ELT(VECTOR, IDX) - Returns a single element from VECTOR identified by the (potentially...
Definition ISDOpcodes.h:576
@ ZERO_EXTEND
ZERO_EXTEND - Used for integer types, zeroing the new bits.
Definition ISDOpcodes.h:849
@ DEBUGTRAP
DEBUGTRAP - Trap intended to get the attention of a debugger.
@ ATOMIC_CMP_SWAP
Val, OUTCHAIN = ATOMIC_CMP_SWAP(INCHAIN, ptr, cmp, swap) For double-word atomic operations: ValLo,...
@ FMINNUM
FMINNUM/FMAXNUM - Perform floating-point minimum maximum on two values, following IEEE-754 definition...
@ UBSANTRAP
UBSANTRAP - Trap with an immediate describing the kind of sanitizer failure.
@ SSHLSAT
RESULT = [US]SHLSAT(LHS, RHS) - Perform saturation left shift.
Definition ISDOpcodes.h:386
@ SMULO
Same for multiplication.
Definition ISDOpcodes.h:356
@ VECTOR_SPLICE_LEFT
VECTOR_SPLICE_LEFT(VEC1, VEC2, IMM) - Shifts CONCAT_VECTORS(VEC1, VEC2) left by IMM elements and retu...
Definition ISDOpcodes.h:653
@ ANY_EXTEND_VECTOR_INREG
ANY_EXTEND_VECTOR_INREG(Vector) - This operator represents an in-register any-extension of the low la...
Definition ISDOpcodes.h:898
@ SIGN_EXTEND_INREG
SIGN_EXTEND_INREG - This operator atomically performs a SHL/SRA pair to sign extend a small value in ...
Definition ISDOpcodes.h:887
@ SMIN
[US]{MIN/MAX} - Binary minimum or maximum of signed or unsigned integers.
Definition ISDOpcodes.h:726
@ SDIVFIXSAT
Same as the corresponding unsaturated fixed point instructions, but the result is clamped between the...
Definition ISDOpcodes.h:413
@ FP_EXTEND
X = FP_EXTEND(Y) - Extend a smaller FP type into a larger FP type.
Definition ISDOpcodes.h:977
@ UADDO_CARRY
Carry-using nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:328
@ FMINIMUM
FMINIMUM/FMAXIMUM - NaN-propagating minimum/maximum that also treat -0.0 as less than 0....
@ FP_TO_SINT
FP_TO_[US]INT - Convert a floating point value to a signed or unsigned integer.
Definition ISDOpcodes.h:925
@ READCYCLECOUNTER
READCYCLECOUNTER - This corresponds to the readcyclecounter intrinsic.
@ AND
Bitwise operators - logical and, logical or, logical xor.
Definition ISDOpcodes.h:738
@ TRAP
TRAP - Trapping instruction.
@ GET_FPENV_MEM
Gets the current floating-point environment.
@ SCMP
[US]CMP - 3-way comparison of signed or unsigned integers.
Definition ISDOpcodes.h:734
@ AVGFLOORS
AVGFLOORS/AVGFLOORU - Averaging add - Add two integers using an integer of type i[N+1],...
Definition ISDOpcodes.h:709
@ VECTOR_SPLICE_RIGHT
VECTOR_SPLICE_RIGHT(VEC1, VEC2, IMM) - Shifts CONCAT_VECTORS(VEC1, VEC2) right by IMM elements and re...
Definition ISDOpcodes.h:656
@ ADDE
Carry-using nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:304
@ FREEZE
FREEZE - FREEZE(VAL) returns an arbitrary value if VAL is UNDEF (or is evaluated to UNDEF),...
Definition ISDOpcodes.h:241
@ INSERT_VECTOR_ELT
INSERT_VECTOR_ELT(VECTOR, VAL, IDX) - Returns VECTOR with the element at IDX replaced with VAL.
Definition ISDOpcodes.h:565
@ ATOMIC_SWAP
Val, OUTCHAIN = ATOMIC_SWAP(INCHAIN, ptr, amt) Val, OUTCHAIN = ATOMIC_LOAD_[OpName](INCHAIN,...
@ FFREXP
FFREXP - frexp, extract fractional and exponent component of a floating-point value.
@ FP_ROUND
X = FP_ROUND(Y, TRUNC) - Rounding 'Y' from a larger floating point type down to the precision of the ...
Definition ISDOpcodes.h:958
@ VECTOR_COMPRESS
VECTOR_COMPRESS(Vec, Mask, Passthru) consecutively place vector elements based on mask e....
Definition ISDOpcodes.h:698
@ CLEAR_CACHE
llvm.clear_cache intrinsic Operands: Input Chain, Start Addres, End Address Outputs: Output Chain
@ ZERO_EXTEND_VECTOR_INREG
ZERO_EXTEND_VECTOR_INREG(Vector) - This operator represents an in-register zero-extension of the low ...
Definition ISDOpcodes.h:920
@ ADDRSPACECAST
ADDRSPACECAST - This operator converts between pointers of different address spaces.
Definition ISDOpcodes.h:996
@ 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:944
@ VECREDUCE_FMINIMUM
@ TRUNCATE
TRUNCATE - Completely drop the high bits.
Definition ISDOpcodes.h:855
@ VECREDUCE_SEQ_FMUL
@ FCOPYSIGN
FCOPYSIGN(X, Y) - Return the value of X with the sign of Y.
Definition ISDOpcodes.h:534
@ SADDSAT
RESULT = [US]ADDSAT(LHS, RHS) - Perform saturation addition on 2 integers with the same bit width (W)...
Definition ISDOpcodes.h:365
@ GET_DYNAMIC_AREA_OFFSET
GET_DYNAMIC_AREA_OFFSET - get offset from native SP to the address of the most recent dynamic alloca.
@ SET_FPENV_MEM
Sets the current floating point environment.
@ FMINIMUMNUM
FMINIMUMNUM/FMAXIMUMNUM - minimumnum/maximumnum that is same with FMINNUM_IEEE and FMAXNUM_IEEE besid...
@ TRUNCATE_SSAT_S
TRUNCATE_[SU]SAT_[SU] - Truncate for saturated operand [SU] located in middle, prefix for SAT means i...
Definition ISDOpcodes.h:870
@ ABDS
ABDS/ABDU - Absolute difference - Return the absolute difference between two numbers interpreted as s...
Definition ISDOpcodes.h:721
@ TRUNCATE_USAT_U
Definition ISDOpcodes.h:874
@ SADDO_CARRY
Carry-using overflow-aware nodes for multiple precision addition and subtraction.
Definition ISDOpcodes.h:338
@ LOOP_DEPENDENCE_WAR_MASK
The llvm.loop.dependence.
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 getREM(EVT VT)
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.
Definition Types.h:26
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*...