LLVM 23.0.0git
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 = 128;
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(AtomicLoadExtActions, 0, sizeof(AtomicLoadExtActions));
1019 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
1020 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
1021 memset(CondCodeActions, 0, sizeof(CondCodeActions));
1022 llvm::fill(RegClassForVT, nullptr);
1023 llvm::fill(TargetDAGCombineArray, 0);
1024
1025 // Let extending atomic loads be unsupported by default.
1026 for (MVT ValVT : MVT::all_valuetypes())
1027 for (MVT MemVT : MVT::all_valuetypes())
1029 Expand);
1030
1031 // We're somewhat special casing MVT::i2 and MVT::i4. Ideally we want to
1032 // remove this and targets should individually set these types if not legal.
1035 for (MVT VT : {MVT::i2, MVT::i4})
1036 OpActions[(unsigned)VT.SimpleTy][NT] = Expand;
1037 }
1038 for (MVT AVT : MVT::all_valuetypes()) {
1039 for (MVT VT : {MVT::i2, MVT::i4, MVT::v128i2, MVT::v64i4}) {
1040 setTruncStoreAction(AVT, VT, Expand);
1043 }
1044 }
1045 for (unsigned IM = (unsigned)ISD::PRE_INC;
1046 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
1047 for (MVT VT : {MVT::i2, MVT::i4}) {
1052 }
1053 }
1054
1055 for (MVT VT : MVT::fp_valuetypes()) {
1056 MVT IntVT = MVT::getIntegerVT(VT.getFixedSizeInBits());
1057 if (IntVT.isValid()) {
1060 }
1061 }
1062
1063 // If f16 fma is not natively supported, the value must be promoted to an f64
1064 // (and not to f32!) to prevent double rounding issues.
1065 AddPromotedToType(ISD::FMA, MVT::f16, MVT::f64);
1066 AddPromotedToType(ISD::STRICT_FMA, MVT::f16, MVT::f64);
1067
1068 // Set default actions for various operations.
1069 for (MVT VT : MVT::all_valuetypes()) {
1070 // Default all indexed load / store to expand.
1071 for (unsigned IM = (unsigned)ISD::PRE_INC;
1072 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
1077 }
1078
1079 // Most backends expect to see the node which just returns the value loaded.
1081
1082 // These operations default to expand.
1111 ISD::FMULADD},
1112 VT, Expand);
1113
1114 // Overflow operations default to expand
1117 VT, Expand);
1118
1119 // Carry-using overflow operations default to expand.
1122 VT, Expand);
1123
1124 // ADDC/ADDE/SUBC/SUBE default to expand.
1126 Expand);
1127
1128 // [US]CMP default to expand
1130
1131 // Halving adds
1134 Expand);
1135
1136 // Absolute difference
1138
1139 // Carry-less multiply
1141
1142 // Saturated trunc
1146
1147 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
1149 Expand);
1151
1153
1154 // These library functions default to expand.
1157 VT, Expand);
1158
1159 // These operations default to expand for vector types.
1160 if (VT.isVector())
1166 VT, Expand);
1167
1168 // Constrained floating-point operations default to expand.
1169#define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
1170 setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
1171#include "llvm/IR/ConstrainedOps.def"
1172
1173 // For most targets @llvm.get.dynamic.area.offset just returns 0.
1175
1176 // Vector reduction default to expand.
1184 VT, Expand);
1185
1186 // Named vector shuffles default to expand.
1188 Expand);
1189
1190 // Only some target support this vector operation. Most need to expand it.
1192
1193 // VP operations default to expand.
1194#define BEGIN_REGISTER_VP_SDNODE(SDOPC, ...) \
1195 setOperationAction(ISD::SDOPC, VT, Expand);
1196#include "llvm/IR/VPIntrinsics.def"
1197
1198 // Masked vector extracts default to expand.
1200
1203
1204 // FP environment operations default to expand.
1208
1210 }
1211
1212 // Most targets ignore the @llvm.prefetch intrinsic.
1214
1215 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
1217
1218 // Most targets also ignore the @llvm.readsteadycounter intrinsic.
1220
1221 // ConstantFP nodes default to expand. Targets can either change this to
1222 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
1223 // to optimize expansions for certain constants.
1225 {MVT::bf16, MVT::f16, MVT::f32, MVT::f64, MVT::f80, MVT::f128},
1226 Expand);
1227
1228 // Insert custom handling default for llvm.canonicalize.*.
1230 {MVT::f16, MVT::f32, MVT::f64, MVT::f128}, Expand);
1231
1232 // FIXME: Query RuntimeLibCalls to make the decision.
1234 {MVT::f32, MVT::f64, MVT::f128}, LibCall);
1235
1238 MVT::f16, Promote);
1239 // Default ISD::TRAP to expand (which turns it into abort).
1240 setOperationAction(ISD::TRAP, MVT::Other, Expand);
1241
1242 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
1243 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
1245
1247
1250
1251 for (MVT VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) {
1254 }
1256
1257 // This one by default will call __clear_cache unless the target
1258 // wants something different.
1260
1261 // By default, STACKADDRESS nodes are expanded like STACKSAVE nodes.
1262 // On SPARC targets, custom lowering is required.
1264}
1265
1267 EVT) const {
1268 return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
1269}
1270
1272 const DataLayout &DL) const {
1273 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
1274 if (LHSTy.isVector())
1275 return LHSTy;
1276 MVT ShiftVT = getScalarShiftAmountTy(DL, LHSTy);
1277 // If any possible shift value won't fit in the prefered type, just use
1278 // something safe. Assume it will be legalized when the shift is expanded.
1279 if (ShiftVT.getSizeInBits() < Log2_32_Ceil(LHSTy.getSizeInBits()))
1280 ShiftVT = MVT::i32;
1281 assert(ShiftVT.getSizeInBits() >= Log2_32_Ceil(LHSTy.getSizeInBits()) &&
1282 "ShiftVT is still too small!");
1283 return ShiftVT;
1284}
1285
1286bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
1287 assert(isTypeLegal(VT));
1288 switch (Op) {
1289 default:
1290 return false;
1291 case ISD::SDIV:
1292 case ISD::UDIV:
1293 case ISD::SREM:
1294 case ISD::UREM:
1295 return true;
1296 }
1297}
1298
1300 unsigned DestAS) const {
1301 return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
1302}
1303
1305 Type *RetTy, ElementCount EC, bool ZeroIsPoison,
1306 const ConstantRange *VScaleRange) const {
1307 // Find the smallest "sensible" element type to use for the expansion.
1308 ConstantRange CR(APInt(64, EC.getKnownMinValue()));
1309 if (EC.isScalable())
1310 CR = CR.umul_sat(*VScaleRange);
1311
1312 if (ZeroIsPoison)
1313 CR = CR.subtract(APInt(64, 1));
1314
1315 unsigned EltWidth = RetTy->getScalarSizeInBits();
1316 EltWidth = std::min(EltWidth, CR.getActiveBits());
1317 EltWidth = std::max(llvm::bit_ceil(EltWidth), (unsigned)8);
1318
1319 return EltWidth;
1320}
1321
1323 // If the command-line option was specified, ignore this request.
1324 if (!JumpIsExpensiveOverride.getNumOccurrences())
1325 JumpIsExpensive = isExpensive;
1326}
1327
1330 // If this is a simple type, use the ComputeRegisterProp mechanism.
1331 if (VT.isSimple()) {
1332 MVT SVT = VT.getSimpleVT();
1333 assert((unsigned)SVT.SimpleTy < std::size(TransformToType));
1334 MVT NVT = TransformToType[SVT.SimpleTy];
1335 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
1336
1337 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
1338 LA == TypeSoftPromoteHalf ||
1339 (NVT.isVector() ||
1340 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
1341 "Promote may not follow Expand or Promote");
1342
1343 if (LA == TypeSplitVector)
1344 return LegalizeKind(LA, EVT(SVT).getHalfNumVectorElementsVT(Context));
1345 if (LA == TypeScalarizeVector)
1346 return LegalizeKind(LA, SVT.getVectorElementType());
1347 return LegalizeKind(LA, NVT);
1348 }
1349
1350 // Handle Extended Scalar Types.
1351 if (!VT.isVector()) {
1352 assert(VT.isInteger() && "Float types must be simple");
1353 unsigned BitSize = VT.getSizeInBits();
1354 // First promote to a power-of-two size, then expand if necessary.
1355 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
1356 EVT NVT = VT.getRoundIntegerType(Context);
1357 assert(NVT != VT && "Unable to round integer VT");
1358 LegalizeKind NextStep = getTypeConversion(Context, NVT);
1359 // Avoid multi-step promotion.
1360 if (NextStep.first == TypePromoteInteger)
1361 return NextStep;
1362 // Return rounded integer type.
1363 return LegalizeKind(TypePromoteInteger, NVT);
1364 }
1365
1367 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
1368 }
1369
1370 // Handle vector types.
1371 ElementCount NumElts = VT.getVectorElementCount();
1372 EVT EltVT = VT.getVectorElementType();
1373
1374 // Vectors with only one element are always scalarized.
1375 if (NumElts.isScalar())
1376 return LegalizeKind(TypeScalarizeVector, EltVT);
1377
1378 // Try to widen vector elements until the element type is a power of two and
1379 // promote it to a legal type later on, for example:
1380 // <3 x i8> -> <4 x i8> -> <4 x i32>
1381 if (EltVT.isInteger()) {
1382 // Vectors with a number of elements that is not a power of two are always
1383 // widened, for example <3 x i8> -> <4 x i8>.
1384 if (!VT.isPow2VectorType()) {
1385 NumElts = NumElts.coefficientNextPowerOf2();
1386 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
1387 return LegalizeKind(TypeWidenVector, NVT);
1388 }
1389
1390 // Examine the element type.
1391 LegalizeKind LK = getTypeConversion(Context, EltVT);
1392
1393 // If type is to be expanded, split the vector.
1394 // <4 x i140> -> <2 x i140>
1395 if (LK.first == TypeExpandInteger) {
1396 if (NumElts.isScalable() && NumElts.getKnownMinValue() == 1)
1399 VT.getHalfNumVectorElementsVT(Context));
1400 }
1401
1402 // Promote the integer element types until a legal vector type is found
1403 // or until the element integer type is too big. If a legal type was not
1404 // found, fallback to the usual mechanism of widening/splitting the
1405 // vector.
1406 EVT OldEltVT = EltVT;
1407 while (true) {
1408 // Increase the bitwidth of the element to the next pow-of-two
1409 // (which is greater than 8 bits).
1410 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
1411 .getRoundIntegerType(Context);
1412
1413 // Stop trying when getting a non-simple element type.
1414 // Note that vector elements may be greater than legal vector element
1415 // types. Example: X86 XMM registers hold 64bit element on 32bit
1416 // systems.
1417 if (!EltVT.isSimple())
1418 break;
1419
1420 // Build a new vector type and check if it is legal.
1421 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1422 // Found a legal promoted vector type.
1423 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
1425 EVT::getVectorVT(Context, EltVT, NumElts));
1426 }
1427
1428 // Reset the type to the unexpanded type if we did not find a legal vector
1429 // type with a promoted vector element type.
1430 EltVT = OldEltVT;
1431 }
1432
1433 // Try to widen the vector until a legal type is found.
1434 // If there is no wider legal type, split the vector.
1435 while (true) {
1436 // Round up to the next power of 2.
1437 NumElts = NumElts.coefficientNextPowerOf2();
1438
1439 // If there is no simple vector type with this many elements then there
1440 // cannot be a larger legal vector type. Note that this assumes that
1441 // there are no skipped intermediate vector types in the simple types.
1442 if (!EltVT.isSimple())
1443 break;
1444 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1445 if (LargerVector == MVT())
1446 break;
1447
1448 // If this type is legal then widen the vector.
1449 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1450 return LegalizeKind(TypeWidenVector, LargerVector);
1451 }
1452
1453 // Widen odd vectors to next power of two.
1454 if (!VT.isPow2VectorType()) {
1455 EVT NVT = VT.getPow2VectorType(Context);
1456 return LegalizeKind(TypeWidenVector, NVT);
1457 }
1458
1461
1462 // Vectors with illegal element types are expanded.
1463 EVT NVT = EVT::getVectorVT(Context, EltVT,
1465 return LegalizeKind(TypeSplitVector, NVT);
1466}
1467
1468static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
1469 unsigned &NumIntermediates,
1470 MVT &RegisterVT,
1471 TargetLoweringBase *TLI) {
1472 // Figure out the right, legal destination reg to copy into.
1474 MVT EltTy = VT.getVectorElementType();
1475
1476 unsigned NumVectorRegs = 1;
1477
1478 // Scalable vectors cannot be scalarized, so splitting or widening is
1479 // required.
1480 if (VT.isScalableVector() && !isPowerOf2_32(EC.getKnownMinValue()))
1482 "Splitting or widening of non-power-of-2 MVTs is not implemented.");
1483
1484 // FIXME: We don't support non-power-of-2-sized vectors for now.
1485 // Ideally we could break down into LHS/RHS like LegalizeDAG does.
1486 if (!isPowerOf2_32(EC.getKnownMinValue())) {
1487 // Split EC to unit size (scalable property is preserved).
1488 NumVectorRegs = EC.getKnownMinValue();
1489 EC = ElementCount::getFixed(1);
1490 }
1491
1492 // Divide the input until we get to a supported size. This will
1493 // always end up with an EC that represent a scalar or a scalable
1494 // scalar.
1495 while (EC.getKnownMinValue() > 1 &&
1496 !TLI->isTypeLegal(MVT::getVectorVT(EltTy, EC))) {
1497 EC = EC.divideCoefficientBy(2);
1498 NumVectorRegs <<= 1;
1499 }
1500
1501 NumIntermediates = NumVectorRegs;
1502
1503 MVT NewVT = MVT::getVectorVT(EltTy, EC);
1504 if (!TLI->isTypeLegal(NewVT))
1505 NewVT = EltTy;
1506 IntermediateVT = NewVT;
1507
1508 unsigned LaneSizeInBits = NewVT.getScalarSizeInBits();
1509
1510 // Convert sizes such as i33 to i64.
1511 LaneSizeInBits = llvm::bit_ceil(LaneSizeInBits);
1512
1513 MVT DestVT = TLI->getRegisterType(NewVT);
1514 RegisterVT = DestVT;
1515 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1516 return NumVectorRegs * (LaneSizeInBits / DestVT.getScalarSizeInBits());
1517
1518 // Otherwise, promotion or legal types use the same number of registers as
1519 // the vector decimated to the appropriate level.
1520 return NumVectorRegs;
1521}
1522
1523/// isLegalRC - Return true if the value types that can be represented by the
1524/// specified register class are all legal.
1526 const TargetRegisterClass &RC) const {
1527 for (const auto *I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1528 if (isTypeLegal(*I))
1529 return true;
1530 return false;
1531}
1532
1533/// Replace/modify any TargetFrameIndex operands with a targte-dependent
1534/// sequence of memory operands that is recognized by PrologEpilogInserter.
1537 MachineBasicBlock *MBB) const {
1538 MachineInstr *MI = &InitialMI;
1539 MachineFunction &MF = *MI->getMF();
1540 MachineFrameInfo &MFI = MF.getFrameInfo();
1541
1542 // We're handling multiple types of operands here:
1543 // PATCHPOINT MetaArgs - live-in, read only, direct
1544 // STATEPOINT Deopt Spill - live-through, read only, indirect
1545 // STATEPOINT Deopt Alloca - live-through, read only, direct
1546 // (We're currently conservative and mark the deopt slots read/write in
1547 // practice.)
1548 // STATEPOINT GC Spill - live-through, read/write, indirect
1549 // STATEPOINT GC Alloca - live-through, read/write, direct
1550 // The live-in vs live-through is handled already (the live through ones are
1551 // all stack slots), but we need to handle the different type of stackmap
1552 // operands and memory effects here.
1553
1554 if (llvm::none_of(MI->operands(),
1555 [](MachineOperand &Operand) { return Operand.isFI(); }))
1556 return MBB;
1557
1558 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1559
1560 // Inherit previous memory operands.
1561 MIB.cloneMemRefs(*MI);
1562
1563 for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1564 MachineOperand &MO = MI->getOperand(i);
1565 if (!MO.isFI()) {
1566 // Index of Def operand this Use it tied to.
1567 // Since Defs are coming before Uses, if Use is tied, then
1568 // index of Def must be smaller that index of that Use.
1569 // Also, Defs preserve their position in new MI.
1570 unsigned TiedTo = i;
1571 if (MO.isReg() && MO.isTied())
1572 TiedTo = MI->findTiedOperandIdx(i);
1573 MIB.add(MO);
1574 if (TiedTo < i)
1575 MIB->tieOperands(TiedTo, MIB->getNumOperands() - 1);
1576 continue;
1577 }
1578
1579 // foldMemoryOperand builds a new MI after replacing a single FI operand
1580 // with the canonical set of five x86 addressing-mode operands.
1581 int FI = MO.getIndex();
1582
1583 // Add frame index operands recognized by stackmaps.cpp
1585 // indirect-mem-ref tag, size, #FI, offset.
1586 // Used for spills inserted by StatepointLowering. This codepath is not
1587 // used for patchpoints/stackmaps at all, for these spilling is done via
1588 // foldMemoryOperand callback only.
1589 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1590 MIB.addImm(StackMaps::IndirectMemRefOp);
1591 MIB.addImm(MFI.getObjectSize(FI));
1592 MIB.add(MO);
1593 MIB.addImm(0);
1594 } else {
1595 // direct-mem-ref tag, #FI, offset.
1596 // Used by patchpoint, and direct alloca arguments to statepoints
1597 MIB.addImm(StackMaps::DirectMemRefOp);
1598 MIB.add(MO);
1599 MIB.addImm(0);
1600 }
1601
1602 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1603
1604 // Add a new memory operand for this FI.
1605 assert(MFI.getObjectOffset(FI) != -1);
1606
1607 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and
1608 // PATCHPOINT should be updated to do the same. (TODO)
1609 if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1610 auto Flags = MachineMemOperand::MOLoad;
1612 MachinePointerInfo::getFixedStack(MF, FI), Flags,
1614 MIB->addMemOperand(MF, MMO);
1615 }
1616 }
1617 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1618 MI->eraseFromParent();
1619 return MBB;
1620}
1621
1622/// findRepresentativeClass - Return the largest legal super-reg register class
1623/// of the register class for the specified type and its associated "cost".
1624// This function is in TargetLowering because it uses RegClassForVT which would
1625// need to be moved to TargetRegisterInfo and would necessitate moving
1626// isTypeLegal over as well - a massive change that would just require
1627// TargetLowering having a TargetRegisterInfo class member that it would use.
1628std::pair<const TargetRegisterClass *, uint8_t>
1630 MVT VT) const {
1631 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1632 if (!RC)
1633 return std::make_pair(RC, 0);
1634
1635 // Compute the set of all super-register classes.
1636 BitVector SuperRegRC(TRI->getNumRegClasses());
1637 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1638 SuperRegRC.setBitsInMask(RCI.getMask());
1639
1640 // Find the first legal register class with the largest spill size.
1641 const TargetRegisterClass *BestRC = RC;
1642 for (unsigned i : SuperRegRC.set_bits()) {
1643 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1644 // We want the largest possible spill size.
1645 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1646 continue;
1647 if (!isLegalRC(*TRI, *SuperRC))
1648 continue;
1649 BestRC = SuperRC;
1650 }
1651 return std::make_pair(BestRC, 1);
1652}
1653
1654/// computeRegisterProperties - Once all of the register classes are added,
1655/// this allows us to compute derived properties we expose.
1657 const TargetRegisterInfo *TRI) {
1658 // Everything defaults to needing one register.
1659 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1660 NumRegistersForVT[i] = 1;
1661 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1662 }
1663 // ...except isVoid, which doesn't need any registers.
1664 NumRegistersForVT[MVT::isVoid] = 0;
1665
1666 // Find the largest integer register class.
1667 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1668 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1669 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1670
1671 // Every integer value type larger than this largest register takes twice as
1672 // many registers to represent as the previous ValueType.
1673 for (unsigned ExpandedReg = LargestIntReg + 1;
1674 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1675 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1676 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1677 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1678 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1680 }
1681
1682 // Inspect all of the ValueType's smaller than the largest integer
1683 // register to see which ones need promotion.
1684 unsigned LegalIntReg = LargestIntReg;
1685 for (unsigned IntReg = LargestIntReg - 1;
1686 IntReg >= (unsigned)MVT::i1; --IntReg) {
1687 MVT IVT = (MVT::SimpleValueType)IntReg;
1688 if (isTypeLegal(IVT)) {
1689 LegalIntReg = IntReg;
1690 } else {
1691 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1692 (MVT::SimpleValueType)LegalIntReg;
1693 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1694 }
1695 }
1696
1697 // ppcf128 type is really two f64's.
1698 if (!isTypeLegal(MVT::ppcf128)) {
1699 if (isTypeLegal(MVT::f64)) {
1700 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1701 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1702 TransformToType[MVT::ppcf128] = MVT::f64;
1703 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1704 } else {
1705 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1706 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1707 TransformToType[MVT::ppcf128] = MVT::i128;
1708 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1709 }
1710 }
1711
1712 // Decide how to handle f128. If the target does not have native f128 support,
1713 // expand it to i128 and we will be generating soft float library calls.
1714 if (!isTypeLegal(MVT::f128)) {
1715 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1716 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1717 TransformToType[MVT::f128] = MVT::i128;
1718 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1719 }
1720
1721 // Decide how to handle f80. If the target does not have native f80 support,
1722 // expand it to i96 and we will be generating soft float library calls.
1723 if (!isTypeLegal(MVT::f80)) {
1724 NumRegistersForVT[MVT::f80] = 3*NumRegistersForVT[MVT::i32];
1725 RegisterTypeForVT[MVT::f80] = RegisterTypeForVT[MVT::i32];
1726 TransformToType[MVT::f80] = MVT::i32;
1727 ValueTypeActions.setTypeAction(MVT::f80, TypeSoftenFloat);
1728 }
1729
1730 // Decide how to handle f64. If the target does not have native f64 support,
1731 // expand it to i64 and we will be generating soft float library calls.
1732 if (!isTypeLegal(MVT::f64)) {
1733 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1734 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1735 TransformToType[MVT::f64] = MVT::i64;
1736 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1737 }
1738
1739 // Decide how to handle f32. If the target does not have native f32 support,
1740 // expand it to i32 and we will be generating soft float library calls.
1741 if (!isTypeLegal(MVT::f32)) {
1742 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1743 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1744 TransformToType[MVT::f32] = MVT::i32;
1745 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1746 }
1747
1748 // Decide how to handle f16. If the target does not have native f16 support,
1749 // promote it to f32, because there are no f16 library calls (except for
1750 // conversions).
1751 if (!isTypeLegal(MVT::f16)) {
1752 // Allow targets to control how we legalize half.
1753 bool UseFPRegsForHalfType = 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 ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1764 }
1765
1766 // Decide how to handle bf16. If the target does not have native bf16 support,
1767 // promote it to f32, because there are no bf16 library calls (except for
1768 // converting from f32 to bf16).
1769 if (!isTypeLegal(MVT::bf16)) {
1770 NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32];
1771 RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32];
1772 TransformToType[MVT::bf16] = MVT::f32;
1773 ValueTypeActions.setTypeAction(MVT::bf16, TypeSoftPromoteHalf);
1774 }
1775
1776 // Loop over all of the vector value types to see which need transformations.
1777 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1778 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1779 MVT VT = (MVT::SimpleValueType) i;
1780 if (isTypeLegal(VT))
1781 continue;
1782
1783 MVT EltVT = VT.getVectorElementType();
1785 bool IsLegalWiderType = false;
1786 bool IsScalable = VT.isScalableVector();
1787 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1788 switch (PreferredAction) {
1789 case TypePromoteInteger: {
1790 MVT::SimpleValueType EndVT = IsScalable ?
1791 MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1792 MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1793 // Try to promote the elements of integer vectors. If no legal
1794 // promotion was found, fall through to the widen-vector method.
1795 for (unsigned nVT = i + 1;
1796 (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1797 MVT SVT = (MVT::SimpleValueType) nVT;
1798 // Promote vectors of integers to vectors with the same number
1799 // of elements, with a wider element type.
1800 if (SVT.getScalarSizeInBits() > EltVT.getFixedSizeInBits() &&
1801 SVT.getVectorElementCount() == EC && isTypeLegal(SVT)) {
1802 TransformToType[i] = SVT;
1803 RegisterTypeForVT[i] = SVT;
1804 NumRegistersForVT[i] = 1;
1805 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1806 IsLegalWiderType = true;
1807 break;
1808 }
1809 }
1810 if (IsLegalWiderType)
1811 break;
1812 [[fallthrough]];
1813 }
1814
1815 case TypeWidenVector:
1816 if (isPowerOf2_32(EC.getKnownMinValue())) {
1817 // Try to widen the vector.
1818 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1819 MVT SVT = (MVT::SimpleValueType) nVT;
1820 if (SVT.getVectorElementType() == EltVT &&
1821 SVT.isScalableVector() == IsScalable &&
1823 EC.getKnownMinValue() &&
1824 isTypeLegal(SVT)) {
1825 TransformToType[i] = SVT;
1826 RegisterTypeForVT[i] = SVT;
1827 NumRegistersForVT[i] = 1;
1828 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1829 IsLegalWiderType = true;
1830 break;
1831 }
1832 }
1833 if (IsLegalWiderType)
1834 break;
1835 } else {
1836 // Only widen to the next power of 2 to keep consistency with EVT.
1837 MVT NVT = VT.getPow2VectorType();
1838 if (isTypeLegal(NVT)) {
1839 TransformToType[i] = NVT;
1840 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1841 RegisterTypeForVT[i] = NVT;
1842 NumRegistersForVT[i] = 1;
1843 break;
1844 }
1845 }
1846 [[fallthrough]];
1847
1848 case TypeSplitVector:
1849 case TypeScalarizeVector: {
1850 MVT IntermediateVT;
1851 MVT RegisterVT;
1852 unsigned NumIntermediates;
1853 unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1854 NumIntermediates, RegisterVT, this);
1855 NumRegistersForVT[i] = NumRegisters;
1856 assert(NumRegistersForVT[i] == NumRegisters &&
1857 "NumRegistersForVT size cannot represent NumRegisters!");
1858 RegisterTypeForVT[i] = RegisterVT;
1859
1860 MVT NVT = VT.getPow2VectorType();
1861 if (NVT == VT) {
1862 // Type is already a power of 2. The default action is to split.
1863 TransformToType[i] = MVT::Other;
1864 if (PreferredAction == TypeScalarizeVector)
1865 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1866 else if (PreferredAction == TypeSplitVector)
1867 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1868 else if (EC.getKnownMinValue() > 1)
1869 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1870 else
1871 ValueTypeActions.setTypeAction(VT, EC.isScalable()
1874 } else {
1875 TransformToType[i] = NVT;
1876 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1877 }
1878 break;
1879 }
1880 default:
1881 llvm_unreachable("Unknown vector legalization action!");
1882 }
1883 }
1884
1885 // Determine the 'representative' register class for each value type.
1886 // An representative register class is the largest (meaning one which is
1887 // not a sub-register class / subreg register class) legal register class for
1888 // a group of value types. For example, on i386, i8, i16, and i32
1889 // representative would be GR32; while on x86_64 it's GR64.
1890 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1891 const TargetRegisterClass* RRC;
1892 uint8_t Cost;
1894 RepRegClassForVT[i] = RRC;
1895 RepRegClassCostForVT[i] = Cost;
1896 }
1897}
1898
1900 EVT VT) const {
1901 assert(!VT.isVector() && "No default SetCC type for vectors!");
1902 return getPointerTy(DL).SimpleTy;
1903}
1904
1906 return MVT::i32; // return the default value
1907}
1908
1909/// getVectorTypeBreakdown - Vector types are broken down into some number of
1910/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1911/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1912/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1913///
1914/// This method returns the number of registers needed, and the VT for each
1915/// register. It also returns the VT and quantity of the intermediate values
1916/// before they are promoted/expanded.
1918 EVT VT, EVT &IntermediateVT,
1919 unsigned &NumIntermediates,
1920 MVT &RegisterVT) const {
1921 ElementCount EltCnt = VT.getVectorElementCount();
1922
1923 // If there is a wider vector type with the same element type as this one,
1924 // or a promoted vector type that has the same number of elements which
1925 // are wider, then we should convert to that legal vector type.
1926 // This handles things like <2 x float> -> <4 x float> and
1927 // <4 x i1> -> <4 x i32>.
1928 LegalizeTypeAction TA = getTypeAction(Context, VT);
1929 if (!EltCnt.isScalar() &&
1930 (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1931 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1932 if (isTypeLegal(RegisterEVT)) {
1933 IntermediateVT = RegisterEVT;
1934 RegisterVT = RegisterEVT.getSimpleVT();
1935 NumIntermediates = 1;
1936 return 1;
1937 }
1938 }
1939
1940 // Figure out the right, legal destination reg to copy into.
1941 EVT EltTy = VT.getVectorElementType();
1942
1943 unsigned NumVectorRegs = 1;
1944
1945 // Scalable vectors cannot be scalarized, so handle the legalisation of the
1946 // types like done elsewhere in SelectionDAG.
1947 if (EltCnt.isScalable()) {
1948 LegalizeKind LK;
1949 EVT PartVT = VT;
1950 do {
1951 // Iterate until we've found a legal (part) type to hold VT.
1952 LK = getTypeConversion(Context, PartVT);
1953 PartVT = LK.second;
1954 } while (LK.first != TypeLegal);
1955
1956 if (!PartVT.isVector()) {
1958 "Don't know how to legalize this scalable vector type");
1959 }
1960
1961 NumIntermediates =
1964 IntermediateVT = PartVT;
1965 RegisterVT = getRegisterType(Context, IntermediateVT);
1966 return NumIntermediates;
1967 }
1968
1969 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally
1970 // we could break down into LHS/RHS like LegalizeDAG does.
1971 if (!isPowerOf2_32(EltCnt.getKnownMinValue())) {
1972 NumVectorRegs = EltCnt.getKnownMinValue();
1973 EltCnt = ElementCount::getFixed(1);
1974 }
1975
1976 // Divide the input until we get to a supported size. This will always
1977 // end with a scalar if the target doesn't support vectors.
1978 while (EltCnt.getKnownMinValue() > 1 &&
1979 !isTypeLegal(EVT::getVectorVT(Context, EltTy, EltCnt))) {
1980 EltCnt = EltCnt.divideCoefficientBy(2);
1981 NumVectorRegs <<= 1;
1982 }
1983
1984 NumIntermediates = NumVectorRegs;
1985
1986 EVT NewVT = EVT::getVectorVT(Context, EltTy, EltCnt);
1987 if (!isTypeLegal(NewVT))
1988 NewVT = EltTy;
1989 IntermediateVT = NewVT;
1990
1991 MVT DestVT = getRegisterType(Context, NewVT);
1992 RegisterVT = DestVT;
1993
1994 if (EVT(DestVT).bitsLT(NewVT)) { // Value is expanded, e.g. i64 -> i16.
1995 TypeSize NewVTSize = NewVT.getSizeInBits();
1996 // Convert sizes such as i33 to i64.
1998 NewVTSize = NewVTSize.coefficientNextPowerOf2();
1999 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
2000 }
2001
2002 // Otherwise, promotion or legal types use the same number of registers as
2003 // the vector decimated to the appropriate level.
2004 return NumVectorRegs;
2005}
2006
2008 uint64_t NumCases,
2010 ProfileSummaryInfo *PSI,
2011 BlockFrequencyInfo *BFI) const {
2012 // FIXME: This function check the maximum table size and density, but the
2013 // minimum size is not checked. It would be nice if the minimum size is
2014 // also combined within this function. Currently, the minimum size check is
2015 // performed in findJumpTable() in SelectionDAGBuiler and
2016 // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
2017 const bool OptForSize =
2018 llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
2019 const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
2020 const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
2021
2022 // Check whether the number of cases is small enough and
2023 // the range is dense enough for a jump table.
2024 return (OptForSize || Range <= MaxJumpTableSize) &&
2025 (NumCases * 100 >= Range * MinDensity);
2026}
2027
2029 EVT ConditionVT) const {
2030 return getRegisterType(Context, ConditionVT);
2031}
2032
2033/// Get the EVTs and ArgFlags collections that represent the legalized return
2034/// type of the given function. This does not require a DAG or a return value,
2035/// and is suitable for use before any DAGs for the function are constructed.
2036/// TODO: Move this out of TargetLowering.cpp.
2038 AttributeList attr,
2040 const TargetLowering &TLI, const DataLayout &DL) {
2042 ComputeValueTypes(DL, ReturnType, Types);
2043 unsigned NumValues = Types.size();
2044 if (NumValues == 0) return;
2045
2046 for (Type *Ty : Types) {
2047 EVT VT = TLI.getValueType(DL, Ty);
2048 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
2049
2050 if (attr.hasRetAttr(Attribute::SExt))
2051 ExtendKind = ISD::SIGN_EXTEND;
2052 else if (attr.hasRetAttr(Attribute::ZExt))
2053 ExtendKind = ISD::ZERO_EXTEND;
2054
2055 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
2056 VT = TLI.getTypeForExtReturn(ReturnType->getContext(), VT, ExtendKind);
2057
2058 unsigned NumParts =
2059 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
2060 MVT PartVT =
2061 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
2062
2063 // 'inreg' on function refers to return value
2065 if (attr.hasRetAttr(Attribute::InReg))
2066 Flags.setInReg();
2067
2068 // Propagate extension type if any
2069 if (attr.hasRetAttr(Attribute::SExt))
2070 Flags.setSExt();
2071 else if (attr.hasRetAttr(Attribute::ZExt))
2072 Flags.setZExt();
2073
2074 for (unsigned i = 0; i < NumParts; ++i)
2075 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, Ty, 0, 0));
2076 }
2077}
2078
2080 const DataLayout &DL) const {
2081 return DL.getABITypeAlign(Ty);
2082}
2083
2085 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
2086 Align Alignment, MachineMemOperand::Flags Flags, unsigned *Fast) const {
2087 // Check if the specified alignment is sufficient based on the data layout.
2088 // TODO: While using the data layout works in practice, a better solution
2089 // would be to implement this check directly (make this a virtual function).
2090 // For example, the ABI alignment may change based on software platform while
2091 // this function should only be affected by hardware implementation.
2092 Type *Ty = VT.getTypeForEVT(Context);
2093 if (VT.isZeroSized() || Alignment >= DL.getABITypeAlign(Ty)) {
2094 // Assume that an access that meets the ABI-specified alignment is fast.
2095 if (Fast != nullptr)
2096 *Fast = 1;
2097 return true;
2098 }
2099
2100 // This is a misaligned access.
2101 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast);
2102}
2103
2105 LLVMContext &Context, const DataLayout &DL, EVT VT,
2106 const MachineMemOperand &MMO, unsigned *Fast) const {
2107 return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
2108 MMO.getAlign(), MMO.getFlags(), Fast);
2109}
2110
2112 const DataLayout &DL, EVT VT,
2113 unsigned AddrSpace, Align Alignment,
2115 unsigned *Fast) const {
2116 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
2117 Flags, Fast);
2118}
2119
2121 const DataLayout &DL, EVT VT,
2122 const MachineMemOperand &MMO,
2123 unsigned *Fast) const {
2124 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
2125 MMO.getFlags(), Fast);
2126}
2127
2129 const DataLayout &DL, LLT Ty,
2130 const MachineMemOperand &MMO,
2131 unsigned *Fast) const {
2132 EVT VT = getApproximateEVTForLLT(Ty, Context);
2133 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
2134 MMO.getFlags(), Fast);
2135}
2136
2137unsigned TargetLoweringBase::getMaxStoresPerMemset(bool OptSize) const {
2140
2142}
2143
2144unsigned TargetLoweringBase::getMaxStoresPerMemcpy(bool OptSize) const {
2147
2149}
2150
2154
2156}
2157
2158//===----------------------------------------------------------------------===//
2159// TargetTransformInfo Helpers
2160//===----------------------------------------------------------------------===//
2161
2163 enum InstructionOpcodes {
2164#define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
2165#define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
2166#include "llvm/IR/Instruction.def"
2167 };
2168 switch (static_cast<InstructionOpcodes>(Opcode)) {
2169 case Ret: return 0;
2170 case Br: return 0;
2171 case Switch: return 0;
2172 case IndirectBr: return 0;
2173 case Invoke: return 0;
2174 case CallBr: return 0;
2175 case Resume: return 0;
2176 case Unreachable: return 0;
2177 case CleanupRet: return 0;
2178 case CatchRet: return 0;
2179 case CatchPad: return 0;
2180 case CatchSwitch: return 0;
2181 case CleanupPad: return 0;
2182 case FNeg: return ISD::FNEG;
2183 case Add: return ISD::ADD;
2184 case FAdd: return ISD::FADD;
2185 case Sub: return ISD::SUB;
2186 case FSub: return ISD::FSUB;
2187 case Mul: return ISD::MUL;
2188 case FMul: return ISD::FMUL;
2189 case UDiv: return ISD::UDIV;
2190 case SDiv: return ISD::SDIV;
2191 case FDiv: return ISD::FDIV;
2192 case URem: return ISD::UREM;
2193 case SRem: return ISD::SREM;
2194 case FRem: return ISD::FREM;
2195 case Shl: return ISD::SHL;
2196 case LShr: return ISD::SRL;
2197 case AShr: return ISD::SRA;
2198 case And: return ISD::AND;
2199 case Or: return ISD::OR;
2200 case Xor: return ISD::XOR;
2201 case Alloca: return 0;
2202 case Load: return ISD::LOAD;
2203 case Store: return ISD::STORE;
2204 case GetElementPtr: return 0;
2205 case Fence: return 0;
2206 case AtomicCmpXchg: return 0;
2207 case AtomicRMW: return 0;
2208 case Trunc: return ISD::TRUNCATE;
2209 case ZExt: return ISD::ZERO_EXTEND;
2210 case SExt: return ISD::SIGN_EXTEND;
2211 case FPToUI: return ISD::FP_TO_UINT;
2212 case FPToSI: return ISD::FP_TO_SINT;
2213 case UIToFP: return ISD::UINT_TO_FP;
2214 case SIToFP: return ISD::SINT_TO_FP;
2215 case FPTrunc: return ISD::FP_ROUND;
2216 case FPExt: return ISD::FP_EXTEND;
2217 case PtrToAddr: return ISD::BITCAST;
2218 case PtrToInt: return ISD::BITCAST;
2219 case IntToPtr: return ISD::BITCAST;
2220 case BitCast: return ISD::BITCAST;
2221 case AddrSpaceCast: return ISD::ADDRSPACECAST;
2222 case ICmp: return ISD::SETCC;
2223 case FCmp: return ISD::SETCC;
2224 case PHI: return 0;
2225 case Call: return 0;
2226 case Select: return ISD::SELECT;
2227 case UserOp1: return 0;
2228 case UserOp2: return 0;
2229 case VAArg: return 0;
2230 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
2231 case InsertElement: return ISD::INSERT_VECTOR_ELT;
2232 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
2233 case ExtractValue: return ISD::MERGE_VALUES;
2234 case InsertValue: return ISD::MERGE_VALUES;
2235 case LandingPad: return 0;
2236 case Freeze: return ISD::FREEZE;
2237 }
2238
2239 llvm_unreachable("Unknown instruction type encountered!");
2240}
2241
2243 switch (ID) {
2244 case Intrinsic::exp:
2245 return ISD::FEXP;
2246 case Intrinsic::exp2:
2247 return ISD::FEXP2;
2248 case Intrinsic::log:
2249 return ISD::FLOG;
2250 default:
2251 return ISD::DELETED_NODE;
2252 }
2253}
2254
2255Value *
2257 bool UseTLS) const {
2258 // compiler-rt provides a variable with a magic name. Targets that do not
2259 // link with compiler-rt may also provide such a variable.
2260 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
2261
2262 RTLIB::LibcallImpl UnsafeStackPtrImpl =
2263 Libcalls.getLibcallImpl(RTLIB::SAFESTACK_UNSAFE_STACK_PTR);
2264 if (UnsafeStackPtrImpl == RTLIB::Unsupported)
2265 return nullptr;
2266
2267 StringRef UnsafeStackPtrVar =
2269 auto UnsafeStackPtr =
2270 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
2271
2272 const DataLayout &DL = M->getDataLayout();
2273 PointerType *StackPtrTy = DL.getAllocaPtrType(M->getContext());
2274
2275 if (!UnsafeStackPtr) {
2276 auto TLSModel = UseTLS ?
2279 // The global variable is not defined yet, define it ourselves.
2280 // We use the initial-exec TLS model because we do not support the
2281 // variable living anywhere other than in the main executable.
2282 UnsafeStackPtr = new GlobalVariable(
2283 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
2284 UnsafeStackPtrVar, nullptr, TLSModel);
2285 } else {
2286 // The variable exists, check its type and attributes.
2287 //
2288 // FIXME: Move to IR verifier.
2289 if (UnsafeStackPtr->getValueType() != StackPtrTy)
2290 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
2291 if (UseTLS != UnsafeStackPtr->isThreadLocal())
2292 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
2293 (UseTLS ? "" : "not ") + "be thread-local");
2294 }
2295 return UnsafeStackPtr;
2296}
2297
2299 IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const {
2300 RTLIB::LibcallImpl SafestackPointerAddressImpl =
2301 Libcalls.getLibcallImpl(RTLIB::SAFESTACK_POINTER_ADDRESS);
2302 if (SafestackPointerAddressImpl == RTLIB::Unsupported)
2303 return getDefaultSafeStackPointerLocation(IRB, true);
2304
2305 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
2306 auto *PtrTy = PointerType::getUnqual(M->getContext());
2307
2308 // Android provides a libc function to retrieve the address of the current
2309 // thread's unsafe stack pointer.
2310 FunctionCallee Fn =
2312 SafestackPointerAddressImpl),
2313 PtrTy);
2314 return IRB.CreateCall(Fn);
2315}
2316
2317//===----------------------------------------------------------------------===//
2318// Loop Strength Reduction hooks
2319//===----------------------------------------------------------------------===//
2320
2321/// isLegalAddressingMode - Return true if the addressing mode represented
2322/// by AM is legal for this target, for a load/store of the specified type.
2324 const AddrMode &AM, Type *Ty,
2325 unsigned AS, Instruction *I) const {
2326 // The default implementation of this implements a conservative RISCy, r+r and
2327 // r+i addr mode.
2328
2329 // Scalable offsets not supported
2330 if (AM.ScalableOffset)
2331 return false;
2332
2333 // Allows a sign-extended 16-bit immediate field.
2334 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2335 return false;
2336
2337 // No global is ever allowed as a base.
2338 if (AM.BaseGV)
2339 return false;
2340
2341 // Only support r+r,
2342 switch (AM.Scale) {
2343 case 0: // "r+i" or just "i", depending on HasBaseReg.
2344 break;
2345 case 1:
2346 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2347 return false;
2348 // Otherwise we have r+r or r+i.
2349 break;
2350 case 2:
2351 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2352 return false;
2353 // Allow 2*r as r+r.
2354 break;
2355 default: // Don't allow n * r
2356 return false;
2357 }
2358
2359 return true;
2360}
2361
2362//===----------------------------------------------------------------------===//
2363// Stack Protector
2364//===----------------------------------------------------------------------===//
2365
2366// For OpenBSD return its special guard variable. Otherwise return nullptr,
2367// so that SelectionDAG handle SSP.
2368Value *
2370 const LibcallLoweringInfo &Libcalls) const {
2371 RTLIB::LibcallImpl GuardLocalImpl =
2372 Libcalls.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 Module &M, const LibcallLoweringInfo &Libcalls) const {
2390 RTLIB::LibcallImpl StackGuardImpl =
2391 Libcalls.getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2392 if (StackGuardImpl == RTLIB::Unsupported)
2393 return;
2394
2395 StringRef StackGuardVarName = getLibcallImplName(StackGuardImpl);
2396 M.getOrInsertGlobal(
2397 StackGuardVarName, PointerType::getUnqual(M.getContext()), [=, &M]() {
2398 auto *GV = new GlobalVariable(M, PointerType::getUnqual(M.getContext()),
2399 false, GlobalVariable::ExternalLinkage,
2400 nullptr, StackGuardVarName);
2401
2402 // FreeBSD has "__stack_chk_guard" defined externally on libc.so
2403 if (M.getDirectAccessExternalData() &&
2404 !TM.getTargetTriple().isOSCygMing() &&
2405 !(TM.getTargetTriple().isPPC64() &&
2406 TM.getTargetTriple().isOSFreeBSD()) &&
2407 (!TM.getTargetTriple().isOSDarwin() ||
2408 TM.getRelocationModel() == Reloc::Static))
2409 GV->setDSOLocal(true);
2410
2411 return GV;
2412 });
2413}
2414
2415// Currently only support "standard" __stack_chk_guard.
2416// TODO: add LOAD_STACK_GUARD support.
2418 const Module &M, const LibcallLoweringInfo &Libcalls) const {
2419 RTLIB::LibcallImpl GuardVarImpl =
2420 Libcalls.getLibcallImpl(RTLIB::STACK_CHECK_GUARD);
2421 if (GuardVarImpl == RTLIB::Unsupported)
2422 return nullptr;
2423 return M.getNamedValue(getLibcallImplName(GuardVarImpl));
2424}
2425
2427 const Module &M, const LibcallLoweringInfo &Libcalls) const {
2428 // MSVC CRT has a function to validate security cookie.
2429 RTLIB::LibcallImpl SecurityCheckCookieLibcall =
2430 Libcalls.getLibcallImpl(RTLIB::SECURITY_CHECK_COOKIE);
2431 if (SecurityCheckCookieLibcall != RTLIB::Unsupported)
2432 return M.getFunction(getLibcallImplName(SecurityCheckCookieLibcall));
2433 return nullptr;
2434}
2435
2439
2443
2444unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
2445 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
2446}
2447
2451
2455
2459
2461 return MinimumBitTestCmps;
2462}
2463
2465 MinimumBitTestCmps = Val;
2466}
2467
2469 if (TM.Options.LoopAlignment)
2470 return Align(TM.Options.LoopAlignment);
2471 return PrefLoopAlignment;
2472}
2473
2475 MachineBasicBlock *MBB) const {
2476 return MaxBytesForAlignment;
2477}
2478
2479//===----------------------------------------------------------------------===//
2480// Reciprocal Estimates
2481//===----------------------------------------------------------------------===//
2482
2483/// Get the reciprocal estimate attribute string for a function that will
2484/// override the target defaults.
2486 const Function &F = MF.getFunction();
2487 return F.getFnAttribute("reciprocal-estimates").getValueAsString();
2488}
2489
2490/// Construct a string for the given reciprocal operation of the given type.
2491/// This string should match the corresponding option to the front-end's
2492/// "-mrecip" flag assuming those strings have been passed through in an
2493/// attribute string. For example, "vec-divf" for a division of a vXf32.
2494static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
2495 std::string Name = VT.isVector() ? "vec-" : "";
2496
2497 Name += IsSqrt ? "sqrt" : "div";
2498
2499 // TODO: Handle other float types?
2500 if (VT.getScalarType() == MVT::f64) {
2501 Name += "d";
2502 } else if (VT.getScalarType() == MVT::f16) {
2503 Name += "h";
2504 } else {
2505 assert(VT.getScalarType() == MVT::f32 &&
2506 "Unexpected FP type for reciprocal estimate");
2507 Name += "f";
2508 }
2509
2510 return Name;
2511}
2512
2513/// Return the character position and value (a single numeric character) of a
2514/// customized refinement operation in the input string if it exists. Return
2515/// false if there is no customized refinement step count.
2516static bool parseRefinementStep(StringRef In, size_t &Position,
2517 uint8_t &Value) {
2518 const char RefStepToken = ':';
2519 Position = In.find(RefStepToken);
2520 if (Position == StringRef::npos)
2521 return false;
2522
2523 StringRef RefStepString = In.substr(Position + 1);
2524 // Allow exactly one numeric character for the additional refinement
2525 // step parameter.
2526 if (RefStepString.size() == 1) {
2527 char RefStepChar = RefStepString[0];
2528 if (isDigit(RefStepChar)) {
2529 Value = RefStepChar - '0';
2530 return true;
2531 }
2532 }
2533 report_fatal_error("Invalid refinement step for -recip.");
2534}
2535
2536/// For the input attribute string, return one of the ReciprocalEstimate enum
2537/// status values (enabled, disabled, or not specified) for this operation on
2538/// the specified data type.
2539static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
2540 if (Override.empty())
2542
2543 SmallVector<StringRef, 4> OverrideVector;
2544 Override.split(OverrideVector, ',');
2545 unsigned NumArgs = OverrideVector.size();
2546
2547 // Check if "all", "none", or "default" was specified.
2548 if (NumArgs == 1) {
2549 // Look for an optional setting of the number of refinement steps needed
2550 // for this type of reciprocal operation.
2551 size_t RefPos;
2552 uint8_t RefSteps;
2553 if (parseRefinementStep(Override, RefPos, RefSteps)) {
2554 // Split the string for further processing.
2555 Override = Override.substr(0, RefPos);
2556 }
2557
2558 // All reciprocal types are enabled.
2559 if (Override == "all")
2561
2562 // All reciprocal types are disabled.
2563 if (Override == "none")
2565
2566 // Target defaults for enablement are used.
2567 if (Override == "default")
2569 }
2570
2571 // The attribute string may omit the size suffix ('f'/'d').
2572 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2573 std::string VTNameNoSize = VTName;
2574 VTNameNoSize.pop_back();
2575 static const char DisabledPrefix = '!';
2576
2577 for (StringRef RecipType : OverrideVector) {
2578 size_t RefPos;
2579 uint8_t RefSteps;
2580 if (parseRefinementStep(RecipType, RefPos, RefSteps))
2581 RecipType = RecipType.substr(0, RefPos);
2582
2583 // Ignore the disablement token for string matching.
2584 bool IsDisabled = RecipType[0] == DisabledPrefix;
2585 if (IsDisabled)
2586 RecipType = RecipType.substr(1);
2587
2588 if (RecipType == VTName || RecipType == VTNameNoSize)
2591 }
2592
2594}
2595
2596/// For the input attribute string, return the customized refinement step count
2597/// for this operation on the specified data type. If the step count does not
2598/// exist, return the ReciprocalEstimate enum value for unspecified.
2599static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2600 if (Override.empty())
2602
2603 SmallVector<StringRef, 4> OverrideVector;
2604 Override.split(OverrideVector, ',');
2605 unsigned NumArgs = OverrideVector.size();
2606
2607 // Check if "all", "default", or "none" was specified.
2608 if (NumArgs == 1) {
2609 // Look for an optional setting of the number of refinement steps needed
2610 // for this type of reciprocal operation.
2611 size_t RefPos;
2612 uint8_t RefSteps;
2613 if (!parseRefinementStep(Override, RefPos, RefSteps))
2615
2616 // Split the string for further processing.
2617 Override = Override.substr(0, RefPos);
2618 assert(Override != "none" &&
2619 "Disabled reciprocals, but specifed refinement steps?");
2620
2621 // If this is a general override, return the specified number of steps.
2622 if (Override == "all" || Override == "default")
2623 return RefSteps;
2624 }
2625
2626 // The attribute string may omit the size suffix ('f'/'d').
2627 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2628 std::string VTNameNoSize = VTName;
2629 VTNameNoSize.pop_back();
2630
2631 for (StringRef RecipType : OverrideVector) {
2632 size_t RefPos;
2633 uint8_t RefSteps;
2634 if (!parseRefinementStep(RecipType, RefPos, RefSteps))
2635 continue;
2636
2637 RecipType = RecipType.substr(0, RefPos);
2638 if (RecipType == VTName || RecipType == VTNameNoSize)
2639 return RefSteps;
2640 }
2641
2643}
2644
2649
2654
2659
2664
2666 EVT LoadVT, EVT BitcastVT, const SelectionDAG &DAG,
2667 const MachineMemOperand &MMO) const {
2668 // Single-element vectors are scalarized, so we should generally avoid having
2669 // any memory operations on such types, as they would get scalarized too.
2670 if (LoadVT.isFixedLengthVector() && BitcastVT.isFixedLengthVector() &&
2671 BitcastVT.getVectorNumElements() == 1)
2672 return false;
2673
2674 // Don't do if we could do an indexed load on the original type, but not on
2675 // the new one.
2676 if (!LoadVT.isSimple() || !BitcastVT.isSimple())
2677 return true;
2678
2679 MVT LoadMVT = LoadVT.getSimpleVT();
2680
2681 // Don't bother doing this if it's just going to be promoted again later, as
2682 // doing so might interfere with other combines.
2683 if (getOperationAction(ISD::LOAD, LoadMVT) == Promote &&
2684 getTypeToPromoteTo(ISD::LOAD, LoadMVT) == BitcastVT.getSimpleVT())
2685 return false;
2686
2687 unsigned Fast = 0;
2688 return allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), BitcastVT,
2689 MMO, &Fast) &&
2690 Fast;
2691}
2692
2696
2698 const LoadInst &LI, const DataLayout &DL, AssumptionCache *AC,
2699 const TargetLibraryInfo *LibInfo) const {
2701 if (LI.isVolatile())
2703
2704 if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2706
2707 if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2709
2711 LI.getAlign(), DL, &LI, AC,
2712 /*DT=*/nullptr, LibInfo))
2714
2715 Flags |= getTargetMMOFlags(LI);
2716 return Flags;
2717}
2718
2721 const DataLayout &DL) const {
2723
2724 if (SI.isVolatile())
2726
2727 if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2729
2730 // FIXME: Not preserving dereferenceable
2731 Flags |= getTargetMMOFlags(SI);
2732 return Flags;
2733}
2734
2737 const DataLayout &DL) const {
2739
2740 if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2741 if (RMW->isVolatile())
2743 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2744 if (CmpX->isVolatile())
2746 } else
2747 llvm_unreachable("not an atomic instruction");
2748
2749 // FIXME: Not preserving dereferenceable
2750 Flags |= getTargetMMOFlags(AI);
2751 return Flags;
2752}
2753
2755 const VPIntrinsic &VPIntrin) const {
2757 Intrinsic::ID IntrinID = VPIntrin.getIntrinsicID();
2758
2759 switch (IntrinID) {
2760 default:
2761 llvm_unreachable("unexpected intrinsic. Existing code may be appropriate "
2762 "for it, but support must be explicitly enabled");
2763 case Intrinsic::vp_load:
2764 case Intrinsic::vp_gather:
2765 case Intrinsic::experimental_vp_strided_load:
2767 break;
2768 case Intrinsic::vp_store:
2769 case Intrinsic::vp_scatter:
2770 case Intrinsic::experimental_vp_strided_store:
2772 break;
2773 }
2774
2775 if (VPIntrin.hasMetadata(LLVMContext::MD_nontemporal))
2777
2778 Flags |= getTargetMMOFlags(VPIntrin);
2779 return Flags;
2780}
2781
2783 Instruction *Inst,
2784 AtomicOrdering Ord) const {
2785 if (isReleaseOrStronger(Ord) && Inst->hasAtomicStore())
2786 return Builder.CreateFence(Ord);
2787 else
2788 return nullptr;
2789}
2790
2792 Instruction *Inst,
2793 AtomicOrdering Ord) const {
2794 if (isAcquireOrStronger(Ord))
2795 return Builder.CreateFence(Ord);
2796 else
2797 return nullptr;
2798}
2799
2800//===----------------------------------------------------------------------===//
2801// GlobalISel Hooks
2802//===----------------------------------------------------------------------===//
2803
2805 const TargetTransformInfo *TTI) const {
2806 auto &MF = *MI.getMF();
2807 auto &MRI = MF.getRegInfo();
2808 // Assuming a spill and reload of a value has a cost of 1 instruction each,
2809 // this helper function computes the maximum number of uses we should consider
2810 // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2811 // break even in terms of code size when the original MI has 2 users vs
2812 // choosing to potentially spill. Any more than 2 users we we have a net code
2813 // size increase. This doesn't take into account register pressure though.
2814 auto maxUses = [](unsigned RematCost) {
2815 // A cost of 1 means remats are basically free.
2816 if (RematCost == 1)
2817 return std::numeric_limits<unsigned>::max();
2818 if (RematCost == 2)
2819 return 2U;
2820
2821 // Remat is too expensive, only sink if there's one user.
2822 if (RematCost > 2)
2823 return 1U;
2824 llvm_unreachable("Unexpected remat cost");
2825 };
2826
2827 switch (MI.getOpcode()) {
2828 default:
2829 return false;
2830 // Constants-like instructions should be close to their users.
2831 // We don't want long live-ranges for them.
2832 case TargetOpcode::G_CONSTANT:
2833 case TargetOpcode::G_FCONSTANT:
2834 case TargetOpcode::G_FRAME_INDEX:
2835 case TargetOpcode::G_INTTOPTR:
2836 return true;
2837 case TargetOpcode::G_GLOBAL_VALUE: {
2838 unsigned RematCost = TTI->getGISelRematGlobalCost();
2839 Register Reg = MI.getOperand(0).getReg();
2840 unsigned MaxUses = maxUses(RematCost);
2841 if (MaxUses == UINT_MAX)
2842 return true; // Remats are "free" so always localize.
2843 return MRI.hasAtMostUserInstrs(Reg, MaxUses);
2844 }
2845 }
2846}
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...
const Function & getFunction() const
Definition Function.h:166
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:2487
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.
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
Tracks which library functions to use for a particular subtarget.
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.
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...
Function * getSSPStackGuardCheck(const Module &M, const LibcallLoweringInfo &Libcalls) const
If the target has a standard stack protection check function that performs validation and error handl...
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...
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 Value * getIRStackGuard(IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const
If the target has a standard location for the stack protector guard, returns the address of that loca...
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 void insertSSPDeclarations(Module &M, const LibcallLoweringInfo &Libcalls) const
Inserts necessary declarations for SSP (stack protection) purpose.
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.
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...
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
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 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 Value * getSDagStackGuard(const Module &M, const LibcallLoweringInfo &Libcalls) const
Return the variable that's previously inserted by insertSSPDeclarations, if any, otherwise return nul...
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...
static StringRef getLibcallImplName(RTLIB::LibcallImpl Call)
Get the libcall routine name for the specified libcall implementation.
virtual Value * getSafeStackPointerLocation(IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const
Returns the target-specific address of the unsafe stack pointer.
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...
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.
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:819
@ 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:788
@ 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:853
@ 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:880
@ 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:747
@ 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:910
@ 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:993
@ CLMUL
Carry-less multiplication operations.
Definition ISDOpcodes.h:774
@ 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:844
@ AVGCEILS
AVGCEILS/AVGCEILU - Rounding averaging add - Add two integers using an integer of type i[N+2],...
Definition ISDOpcodes.h:715
@ 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:787
@ PREFETCH
PREFETCH - This corresponds to a prefetch intrinsic.
@ TRUNCATE_SSAT_U
Definition ISDOpcodes.h:873
@ 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:827
@ 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:796
@ SPLAT_VECTOR
SPLAT_VECTOR(VAL) - Returns a vector with the scalar value VAL duplicated in all lanes.
Definition ISDOpcodes.h:672
@ 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:792
@ 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:765
@ 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:850
@ 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, OFFSET) - Shifts CONCAT_VECTORS(VEC1, VEC2) left by OFFSET elements an...
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:899
@ SIGN_EXTEND_INREG
SIGN_EXTEND_INREG - This operator atomically performs a SHL/SRA pair to sign extend a small value in ...
Definition ISDOpcodes.h:888
@ SMIN
[US]{MIN/MAX} - Binary minimum or maximum of signed or unsigned integers.
Definition ISDOpcodes.h:727
@ 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:978
@ 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:926
@ READCYCLECOUNTER
READCYCLECOUNTER - This corresponds to the readcyclecounter intrinsic.
@ AND
Bitwise operators - logical and, logical or, logical xor.
Definition ISDOpcodes.h:739
@ 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:735
@ AVGFLOORS
AVGFLOORS/AVGFLOORU - Averaging add - Add two integers using an integer of type i[N+1],...
Definition ISDOpcodes.h:710
@ VECTOR_SPLICE_RIGHT
VECTOR_SPLICE_RIGHT(VEC1, VEC2, OFFSET) - Shifts CONCAT_VECTORS(VEC1,VEC2) right by OFFSET elements a...
Definition ISDOpcodes.h:657
@ 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:959
@ VECTOR_COMPRESS
VECTOR_COMPRESS(Vec, Mask, Passthru) consecutively place vector elements based on mask e....
Definition ISDOpcodes.h:699
@ 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:921
@ ADDRSPACECAST
ADDRSPACECAST - This operator converts between pointers of different address spaces.
Definition ISDOpcodes.h:997
@ 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:945
@ VECREDUCE_FMINIMUM
@ TRUNCATE
TRUNCATE - Completely drop the high bits.
Definition ISDOpcodes.h:856
@ 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:871
@ ABDS
ABDS/ABDU - Absolute difference - Return the absolute difference between two numbers interpreted as s...
Definition ISDOpcodes.h:722
@ TRUNCATE_USAT_U
Definition ISDOpcodes.h:875
@ 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:1759
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:1753
LLVM_ABI void report_fatal_error(Error Err, bool gen_crash_diag=true)
Definition Error.cpp:163
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 StringRef getLibcallImplName(RTLIB::LibcallImpl CallImpl)
Get the libcall routine name for the specified libcall implementation.
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*...