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/STLExtras.h"
15#include "llvm/ADT/SmallVector.h"
16#include "llvm/ADT/StringExtras.h"
17#include "llvm/ADT/StringRef.h"
18#include "llvm/ADT/Twine.h"
19#include "llvm/Analysis/Loads.h"
20#include "llvm/Analysis/TargetTransformInfo.h"
21#include "llvm/CodeGen/Analysis.h"
22#include "llvm/CodeGen/ISDOpcodes.h"
23#include "llvm/CodeGen/MachineBasicBlock.h"
24#include "llvm/CodeGen/MachineFrameInfo.h"
25#include "llvm/CodeGen/MachineFunction.h"
26#include "llvm/CodeGen/MachineInstr.h"
27#include "llvm/CodeGen/MachineInstrBuilder.h"
28#include "llvm/CodeGen/MachineMemOperand.h"
29#include "llvm/CodeGen/MachineOperand.h"
30#include "llvm/CodeGen/MachineRegisterInfo.h"
31#include "llvm/CodeGen/RuntimeLibcallUtil.h"
32#include "llvm/CodeGen/StackMaps.h"
33#include "llvm/CodeGen/TargetLowering.h"
34#include "llvm/CodeGen/TargetOpcodes.h"
35#include "llvm/CodeGen/TargetRegisterInfo.h"
36#include "llvm/CodeGen/ValueTypes.h"
37#include "llvm/CodeGenTypes/MachineValueType.h"
38#include "llvm/IR/Attributes.h"
39#include "llvm/IR/CallingConv.h"
40#include "llvm/IR/DataLayout.h"
41#include "llvm/IR/DerivedTypes.h"
42#include "llvm/IR/Function.h"
43#include "llvm/IR/GlobalValue.h"
44#include "llvm/IR/GlobalVariable.h"
45#include "llvm/IR/IRBuilder.h"
46#include "llvm/IR/Module.h"
47#include "llvm/IR/Type.h"
48#include "llvm/Support/Casting.h"
49#include "llvm/Support/CommandLine.h"
50#include "llvm/Support/Compiler.h"
51#include "llvm/Support/ErrorHandling.h"
52#include "llvm/Support/FormatVariadic.h"
53#include "llvm/Support/MathExtras.h"
54#include "llvm/Target/TargetMachine.h"
55#include "llvm/Target/TargetOptions.h"
56#include "llvm/TargetParser/Triple.h"
57#include "llvm/Transforms/Utils/SizeOpts.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
68static cl::opt<bool> JumpIsExpensiveOverride(
69 "jump-is-expensive", cl::init(Val: false),
70 cl::desc("Do not create extra branches to split comparison logic."),
71 cl::Hidden);
72
73static cl::opt<unsigned> MinimumJumpTableEntries
74 ("min-jump-table-entries", cl::init(Val: 4), cl::Hidden,
75 cl::desc("Set minimum number of entries to use a jump table."));
76
77static cl::opt<unsigned> MaximumJumpTableSize
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.
82static cl::opt<unsigned>
83 JumpTableDensity("jump-table-density", cl::init(Val: 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.
88static cl::opt<unsigned> OptsizeJumpTableDensity(
89 "optsize-jump-table-density", cl::init(Val: 40), cl::Hidden,
90 cl::desc("Minimum density for building a jump table in "
91 "an optsize function"));
92
93static cl::opt<unsigned> MinimumBitTestCmpsOverride(
94 "min-bit-test-cmps", cl::init(Val: 2), cl::Hidden,
95 cl::desc("Set minimum of largest number of comparisons "
96 "to use bit test for switch."));
97
98static cl::opt<unsigned> MaxStoresPerMemsetOverride(
99 "max-store-memset", cl::init(Val: 0), cl::Hidden,
100 cl::desc("Override target's MaxStoresPerMemset and "
101 "MaxStoresPerMemsetOptSize. "
102 "Set to 0 to use the target default."));
103
104static cl::opt<unsigned> MaxStoresPerMemcpyOverride(
105 "max-store-memcpy", cl::init(Val: 0), cl::Hidden,
106 cl::desc("Override target's MaxStoresPerMemcpy and "
107 "MaxStoresPerMemcpyOptSize. "
108 "Set to 0 to use the target default."));
109
110static cl::opt<unsigned> MaxStoresPerMemmoveOverride(
111 "max-store-memmove", cl::init(Val: 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(Val: 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
536// The floating-point RTLIB::getXXX(EVT) selectors are generated from the
537// RuntimeLibcallFamily table in RuntimeLibcalls.td.
538#define GET_RUNTIME_LIBCALL_FP_SELECTORS
539#include "llvm/IR/RuntimeLibcalls.inc"
540
541static RTLIB::Libcall getOutlineAtomicHelper(const RTLIB::Libcall (&LC)[5][4],
542 AtomicOrdering Order,
543 uint64_t MemSize) {
544 using namespace RTLIB;
545 unsigned ModeN, ModelN;
546 switch (MemSize) {
547 case 1:
548 ModeN = 0;
549 break;
550 case 2:
551 ModeN = 1;
552 break;
553 case 4:
554 ModeN = 2;
555 break;
556 case 8:
557 ModeN = 3;
558 break;
559 case 16:
560 ModeN = 4;
561 break;
562 default:
563 return RTLIB::UNKNOWN_LIBCALL;
564 }
565
566 switch (Order) {
567 case AtomicOrdering::Monotonic:
568 ModelN = 0;
569 break;
570 case AtomicOrdering::Acquire:
571 ModelN = 1;
572 break;
573 case AtomicOrdering::Release:
574 ModelN = 2;
575 break;
576 case AtomicOrdering::AcquireRelease:
577 case AtomicOrdering::SequentiallyConsistent:
578 ModelN = 3;
579 break;
580 default:
581 return UNKNOWN_LIBCALL;
582 }
583
584 return LC[ModeN][ModelN];
585}
586
587RTLIB::Libcall RTLIB::getOUTLINE_ATOMIC(unsigned Opc, AtomicOrdering Order,
588 MVT VT) {
589 if (!VT.isScalarInteger())
590 return UNKNOWN_LIBCALL;
591 uint64_t MemSize = VT.getScalarSizeInBits() / 8;
592
593#define LCALLS(A, B) \
594 { A##B##_RELAX, A##B##_ACQ, A##B##_REL, A##B##_ACQ_REL }
595#define LCALL5(A) \
596 LCALLS(A, 1), LCALLS(A, 2), LCALLS(A, 4), LCALLS(A, 8), LCALLS(A, 16)
597 switch (Opc) {
598 case ISD::ATOMIC_CMP_SWAP: {
599 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_CAS)};
600 return getOutlineAtomicHelper(LC, Order, MemSize);
601 }
602 case ISD::ATOMIC_SWAP: {
603 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_SWP)};
604 return getOutlineAtomicHelper(LC, Order, MemSize);
605 }
606 case ISD::ATOMIC_LOAD_ADD: {
607 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDADD)};
608 return getOutlineAtomicHelper(LC, Order, MemSize);
609 }
610 case ISD::ATOMIC_LOAD_OR: {
611 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDSET)};
612 return getOutlineAtomicHelper(LC, Order, MemSize);
613 }
614 case ISD::ATOMIC_LOAD_CLR: {
615 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDCLR)};
616 return getOutlineAtomicHelper(LC, Order, MemSize);
617 }
618 case ISD::ATOMIC_LOAD_XOR: {
619 const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDEOR)};
620 return getOutlineAtomicHelper(LC, Order, MemSize);
621 }
622 default:
623 return UNKNOWN_LIBCALL;
624 }
625#undef LCALLS
626#undef LCALL5
627}
628
629RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
630#define OP_TO_LIBCALL(Name, Enum) \
631 case Name: \
632 switch (VT.SimpleTy) { \
633 default: \
634 return UNKNOWN_LIBCALL; \
635 case MVT::i8: \
636 return Enum##_1; \
637 case MVT::i16: \
638 return Enum##_2; \
639 case MVT::i32: \
640 return Enum##_4; \
641 case MVT::i64: \
642 return Enum##_8; \
643 case MVT::i128: \
644 return Enum##_16; \
645 }
646
647 switch (Opc) {
648 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
649 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
650 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
651 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
652 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
653 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
654 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
655 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
656 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
657 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
658 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
659 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
660 }
661
662#undef OP_TO_LIBCALL
663
664 return UNKNOWN_LIBCALL;
665}
666
667RTLIB::AtomicLibcall
668RTLIB::selectAtomicLibcall(const LibcallLoweringInfo &Libcalls, unsigned Opc,
669 AtomicOrdering Order, MVT VT) {
670 LibcallImpl Impl = Libcalls.getLibcallImpl(Call: getOUTLINE_ATOMIC(Opc, Order, VT));
671 if (Impl != Unsupported)
672 return {.Impl: Impl, /*IsOutline=*/true};
673 return {.Impl: Libcalls.getLibcallImpl(Call: getSYNC(Opc, VT)), /*IsOutline=*/false};
674}
675
676RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
677 switch (ElementSize) {
678 case 1:
679 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
680 case 2:
681 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
682 case 4:
683 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
684 case 8:
685 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
686 case 16:
687 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
688 default:
689 return UNKNOWN_LIBCALL;
690 }
691}
692
693RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
694 switch (ElementSize) {
695 case 1:
696 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
697 case 2:
698 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
699 case 4:
700 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
701 case 8:
702 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
703 case 16:
704 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
705 default:
706 return UNKNOWN_LIBCALL;
707 }
708}
709
710RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
711 switch (ElementSize) {
712 case 1:
713 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
714 case 2:
715 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
716 case 4:
717 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
718 case 8:
719 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
720 case 16:
721 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
722 default:
723 return UNKNOWN_LIBCALL;
724 }
725}
726
727/// NOTE: The TargetMachine owns TLOF.
728TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm,
729 const TargetSubtargetInfo &STI)
730 : TM(tm),
731 RuntimeLibcallInfo(TM.getTargetTriple(), ExceptionHandling::Default,
732 TM.getTargetTriple().getDefaultWCharSize(),
733 TM.getTargetTriple().getDefaultFloatABI(),
734 TM.getTargetTriple().getDefaultLongDoubleFormat(),
735 TM.Options.MCOptions.getABIName(), TM.Options.VecLib),
736 Libcalls(RuntimeLibcallInfo, [&STI](LibcallLoweringInfo &Info) {
737 STI.initLibcallLoweringInfo(Info);
738 }) {
739 initActions();
740
741 // Perform these initializations only once.
742 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
743 MaxLoadsPerMemcmp = 8;
744 MaxGluedStoresPerMemcpy = 0;
745 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
746 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
747 HasExtractBitsInsn = false;
748 JumpIsExpensive = JumpIsExpensiveOverride;
749 PredictableSelectIsExpensive = false;
750 EnableExtLdPromotion = false;
751 StackPointerRegisterToSaveRestore = 0;
752 BooleanContents = UndefinedBooleanContent;
753 BooleanFloatContents = UndefinedBooleanContent;
754 BooleanVectorContents = UndefinedBooleanContent;
755 SchedPreferenceInfo = Sched::ILP;
756 GatherAllAliasesMaxDepth = 18;
757 IsStrictFPEnabled = DisableStrictNodeMutation;
758 MaxBytesForAlignment = 0;
759 MaxAtomicSizeInBitsSupported = 0;
760
761 // Assume that even with libcalls, no target supports wider than 128 bit
762 // division.
763 MaxDivRemBitWidthSupported = 128;
764
765 MaxLargeFPConvertBitWidthSupported = 128;
766
767 MinCmpXchgSizeInBits = 0;
768 SupportsUnalignedAtomics = false;
769
770 MinimumBitTestCmps = MinimumBitTestCmpsOverride;
771}
772
773// Define the virtual destructor out-of-line to act as a key method to anchor
774// debug info (see coding standards).
775TargetLoweringBase::~TargetLoweringBase() = default;
776
777void TargetLoweringBase::initActions() {
778 // All operations default to being supported.
779 memset(s: OpActions, c: 0, n: sizeof(OpActions));
780 memset(s: LoadExtActions, c: 0, n: sizeof(LoadExtActions));
781 memset(s: AtomicLoadExtActions, c: 0, n: sizeof(AtomicLoadExtActions));
782 memset(s: TruncStoreActions, c: 0, n: sizeof(TruncStoreActions));
783 memset(s: IndexedModeActions, c: 0, n: sizeof(IndexedModeActions));
784 memset(s: CondCodeActions, c: 0, n: sizeof(CondCodeActions));
785 llvm::fill(Range&: RegClassForVT, Value: nullptr);
786 llvm::fill(Range&: TargetDAGCombineArray, Value: 0);
787
788 // Let extending atomic loads be unsupported by default.
789 for (MVT ValVT : MVT::all_valuetypes())
790 for (MVT MemVT : MVT::all_valuetypes())
791 setAtomicLoadExtAction(ExtTypes: {ISD::SEXTLOAD, ISD::ZEXTLOAD}, ValVT, MemVT,
792 Action: Expand);
793
794 // We're somewhat special casing MVT::i2 and MVT::i4. Ideally we want to
795 // remove this and targets should individually set these types if not legal.
796 for (ISD::NodeType NT : enum_seq(Begin: ISD::DELETED_NODE, End: ISD::BUILTIN_OP_END,
797 force_iteration_on_noniterable_enum)) {
798 for (MVT VT : {MVT::i2, MVT::i4})
799 OpActions[(unsigned)VT.SimpleTy][NT] = Expand;
800 }
801 for (MVT AVT : MVT::all_valuetypes()) {
802 for (MVT VT : {MVT::i2, MVT::i4, MVT::v128i2, MVT::v64i4}) {
803 setTruncStoreAction(ValVT: AVT, MemVT: VT, Action: Expand);
804 setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: AVT, MemVT: VT, Action: Expand);
805 setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: AVT, MemVT: VT, Action: Expand);
806 }
807 }
808 for (unsigned IM = (unsigned)ISD::PRE_INC;
809 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
810 for (MVT VT : {MVT::i2, MVT::i4}) {
811 setIndexedLoadAction(IdxModes: IM, VT, Action: Expand);
812 setIndexedStoreAction(IdxModes: IM, VT, Action: Expand);
813 setIndexedMaskedLoadAction(IdxMode: IM, VT, Action: Expand);
814 setIndexedMaskedStoreAction(IdxMode: IM, VT, Action: Expand);
815 }
816 }
817
818 for (MVT VT : MVT::fp_valuetypes()) {
819 MVT IntVT = MVT::getIntegerVT(BitWidth: VT.getFixedSizeInBits());
820 if (IntVT.isValid()) {
821 setOperationAction(Op: ISD::ATOMIC_SWAP, VT, Action: Promote);
822 AddPromotedToType(Opc: ISD::ATOMIC_SWAP, OrigVT: VT, DestVT: IntVT);
823 }
824 }
825
826 // If f16 fma is not natively supported, the value must be promoted to an f64
827 // (and not to f32!) to prevent double rounding issues.
828 AddPromotedToType(Opc: ISD::FMA, OrigVT: MVT::f16, DestVT: MVT::f64);
829 AddPromotedToType(Opc: ISD::STRICT_FMA, OrigVT: MVT::f16, DestVT: MVT::f64);
830
831 // Set default actions for various operations.
832 for (MVT VT : MVT::all_valuetypes()) {
833 // Default all indexed load / store to expand.
834 for (unsigned IM = (unsigned)ISD::PRE_INC;
835 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
836 setIndexedLoadAction(IdxModes: IM, VT, Action: Expand);
837 setIndexedStoreAction(IdxModes: IM, VT, Action: Expand);
838 setIndexedMaskedLoadAction(IdxMode: IM, VT, Action: Expand);
839 setIndexedMaskedStoreAction(IdxMode: IM, VT, Action: Expand);
840 }
841
842 // Most backends expect to see the node which just returns the value loaded.
843 setOperationAction(Op: ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Action: Expand);
844
845 // clang-format off
846 // These operations default to expand.
847 setOperationAction(Ops: {ISD::FGETSIGN, ISD::CONCAT_VECTORS,
848 ISD::FMINNUM, ISD::FMAXNUM,
849 ISD::FMINNUM_IEEE, ISD::FMAXNUM_IEEE,
850 ISD::FMINIMUM, ISD::FMAXIMUM,
851 ISD::FMINIMUMNUM, ISD::FMAXIMUMNUM,
852 ISD::FMAD, ISD::SMIN,
853 ISD::SMAX, ISD::UMIN,
854 ISD::UMAX, ISD::ABS,
855 ISD::FSHL, ISD::FSHR,
856 ISD::SADDSAT, ISD::UADDSAT,
857 ISD::SSUBSAT, ISD::USUBSAT,
858 ISD::SSHLSAT, ISD::USHLSAT,
859 ISD::SMULFIX, ISD::SMULFIXSAT,
860 ISD::UMULFIX, ISD::UMULFIXSAT,
861 ISD::SDIVFIX, ISD::SDIVFIXSAT,
862 ISD::UDIVFIX, ISD::UDIVFIXSAT,
863 ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT,
864 ISD::IS_FPCLASS, ISD::FCBRT,
865 ISD::FLOG, ISD::FLOG2,
866 ISD::FLOG10, ISD::FEXP,
867 ISD::FEXP2, ISD::FEXP10,
868 ISD::FFLOOR, ISD::FNEARBYINT,
869 ISD::FCEIL, ISD::FRINT,
870 ISD::FTRUNC, ISD::FROUNDEVEN,
871 ISD::FTAN, ISD::FACOS,
872 ISD::FASIN, ISD::FATAN,
873 ISD::FCOSH, ISD::FSINH,
874 ISD::FTANH, ISD::FATAN2,
875 ISD::FMULADD, ISD::CONVERT_FROM_ARBITRARY_FP,
876 ISD::CONVERT_TO_ARBITRARY_FP,
877 ISD::PSEUDO_FMIN, ISD::PSEUDO_FMAX},
878 VT, Action: Expand);
879 // clang-format on
880
881 // Overflow operations default to expand
882 setOperationAction(Ops: {ISD::SADDO, ISD::SSUBO, ISD::UADDO, ISD::USUBO,
883 ISD::SMULO, ISD::UMULO},
884 VT, Action: Expand);
885
886 // Carry-using overflow operations default to expand.
887 setOperationAction(Ops: {ISD::UADDO_CARRY, ISD::USUBO_CARRY, ISD::SETCCCARRY,
888 ISD::SADDO_CARRY, ISD::SSUBO_CARRY},
889 VT, Action: Expand);
890
891 // ADDC/ADDE/SUBC/SUBE default to expand.
892 setOperationAction(Ops: {ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}, VT,
893 Action: Expand);
894
895 // [US]CMP default to expand
896 setOperationAction(Ops: {ISD::UCMP, ISD::SCMP}, VT, Action: Expand);
897
898 // Halving adds and subtracts
899 setOperationAction(Ops: {ISD::AVGFLOORS, ISD::AVGFLOORU, ISD::AVGCEILS,
900 ISD::AVGCEILU, ISD::AVGSUBFLOORS, ISD::AVGSUBFLOORU},
901 VT, Action: Expand);
902
903 // Absolute difference
904 setOperationAction(Ops: {ISD::ABDS, ISD::ABDU}, VT, Action: Expand);
905
906 // Carry-less multiply
907 setOperationAction(Ops: {ISD::CLMUL, ISD::CLMULR, ISD::CLMULH}, VT, Action: Expand);
908
909 // Bit extract/deposit (compress/expand)
910 setOperationAction(Ops: {ISD::PEXT, ISD::PDEP}, VT, Action: Expand);
911
912 // Saturated trunc
913 setOperationAction(Op: ISD::TRUNCATE_SSAT_S, VT, Action: Expand);
914 setOperationAction(Op: ISD::TRUNCATE_SSAT_U, VT, Action: Expand);
915 setOperationAction(Op: ISD::TRUNCATE_USAT_U, VT, Action: Expand);
916
917 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
918 setOperationAction(Ops: {ISD::CTLZ_ZERO_POISON, ISD::CTTZ_ZERO_POISON}, VT,
919 Action: Expand);
920
921 // This defaults to Expand so it will be expanded to ABS by default.
922 setOperationAction(Op: ISD::ABS_MIN_POISON, VT, Action: Expand);
923 setOperationAction(Op: ISD::CTLS, VT, Action: Expand);
924
925 setOperationAction(Ops: {ISD::BITREVERSE, ISD::PARITY}, VT, Action: Expand);
926
927 // These library functions default to expand.
928 setOperationAction(Ops: {ISD::FROUND, ISD::FPOWI, ISD::FLDEXP, ISD::FFREXP,
929 ISD::FSINCOS, ISD::FSINCOSPI, ISD::FMODF},
930 VT, Action: Expand);
931
932 // These operations default to expand for vector types.
933 if (VT.isVector())
934 setOperationAction(Ops: {ISD::FCOPYSIGN, ISD::SIGN_EXTEND_INREG,
935 ISD::ANY_EXTEND_VECTOR_INREG,
936 ISD::SIGN_EXTEND_VECTOR_INREG,
937 ISD::ZERO_EXTEND_VECTOR_INREG, ISD::SPLAT_VECTOR,
938 ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND},
939 VT, Action: Expand);
940
941 // Constrained floating-point operations default to expand.
942#define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
943 setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
944#include "llvm/IR/ConstrainedOps.def"
945 setOperationAction(Op: ISD::STRICT_PSEUDO_FMIN, VT, Action: Expand);
946 setOperationAction(Op: ISD::STRICT_PSEUDO_FMAX, VT, Action: Expand);
947
948 // For most targets @llvm.get.dynamic.area.offset just returns 0.
949 setOperationAction(Op: ISD::GET_DYNAMIC_AREA_OFFSET, VT, Action: Expand);
950
951 // Vector reduction default to expand.
952 setOperationAction(
953 Ops: {ISD::VECREDUCE_FADD, ISD::VECREDUCE_FMUL, ISD::VECREDUCE_ADD,
954 ISD::VECREDUCE_MUL, ISD::VECREDUCE_AND, ISD::VECREDUCE_OR,
955 ISD::VECREDUCE_XOR, ISD::VECREDUCE_SMAX, ISD::VECREDUCE_SMIN,
956 ISD::VECREDUCE_UMAX, ISD::VECREDUCE_UMIN, ISD::VECREDUCE_FMAX,
957 ISD::VECREDUCE_FMIN, ISD::VECREDUCE_FMAXIMUM, ISD::VECREDUCE_FMINIMUM,
958 ISD::VECREDUCE_FMAXIMUMNUM, ISD::VECREDUCE_FMINIMUMNUM,
959 ISD::VECREDUCE_SEQ_FADD, ISD::VECREDUCE_SEQ_FMUL},
960 VT, Action: Expand);
961
962 // Named vector shuffles default to expand.
963 setOperationAction(Ops: {ISD::VECTOR_SPLICE_LEFT, ISD::VECTOR_SPLICE_RIGHT}, VT,
964 Action: Expand);
965
966 // Only some target support these vector operations. Default them to Expand.
967 setOperationAction(
968 Ops: {ISD::VECTOR_COMPRESS, ISD::VECTOR_MATCH, ISD::VECTOR_SHUFFLE_VAR}, VT,
969 Action: Expand);
970 setOperationAction(Ops: {ISD::CTTZ_ELTS, ISD::CTTZ_ELTS_ZERO_POISON}, VT,
971 Action: Expand);
972 setOperationAction(Op: ISD::GET_ACTIVE_LANE_MASK, VT, Action: Expand);
973
974 // VP operations default to expand.
975#define BEGIN_REGISTER_VP_SDNODE(SDOPC, ...) \
976 setOperationAction(ISD::SDOPC, VT, Expand);
977#include "llvm/IR/VPIntrinsics.def"
978
979 // Masked vector extracts default to expand.
980 setOperationAction(Op: ISD::VECTOR_FIND_LAST_ACTIVE, VT, Action: Expand);
981
982 setOperationAction(Op: ISD::LOOP_DEPENDENCE_RAW_MASK, VT, Action: Expand);
983 setOperationAction(Op: ISD::LOOP_DEPENDENCE_WAR_MASK, VT, Action: Expand);
984
985 setOperationAction(Op: ISD::MASK_BEFOREFIRST, VT, Action: Expand);
986
987 // FP environment operations default to expand.
988 setOperationAction(Op: ISD::GET_FPENV, VT, Action: Expand);
989 setOperationAction(Op: ISD::SET_FPENV, VT, Action: Expand);
990 setOperationAction(Op: ISD::RESET_FPENV, VT, Action: Expand);
991
992 setOperationAction(Op: ISD::MSTORE, VT, Action: Expand);
993
994 setOperationAction(Op: ISD::MASKED_UDIV, VT, Action: Expand);
995 setOperationAction(Op: ISD::MASKED_SDIV, VT, Action: Expand);
996 setOperationAction(Op: ISD::MASKED_UREM, VT, Action: Expand);
997 setOperationAction(Op: ISD::MASKED_SREM, VT, Action: Expand);
998 }
999
1000 // Most targets ignore the @llvm.prefetch intrinsic.
1001 setOperationAction(Op: ISD::PREFETCH, VT: MVT::Other, Action: Expand);
1002
1003 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
1004 setOperationAction(Op: ISD::READCYCLECOUNTER, VT: MVT::i64, Action: Expand);
1005
1006 // Most targets also ignore the @llvm.readsteadycounter intrinsic.
1007 setOperationAction(Op: ISD::READSTEADYCOUNTER, VT: MVT::i64, Action: Expand);
1008
1009 // ConstantFP nodes default to expand. Targets can either change this to
1010 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
1011 // to optimize expansions for certain constants.
1012 setOperationAction(Ops: ISD::ConstantFP,
1013 VTs: {MVT::bf16, MVT::f16, MVT::f32, MVT::f64, MVT::f80, MVT::f128},
1014 Action: Expand);
1015
1016 // Insert custom handling default for llvm.canonicalize.*.
1017 setOperationAction(Ops: ISD::FCANONICALIZE,
1018 VTs: {MVT::f16, MVT::f32, MVT::f64, MVT::f128}, Action: Expand);
1019
1020 // FIXME: Query RuntimeLibCalls to make the decision.
1021 setOperationAction(Ops: {ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND},
1022 VTs: {MVT::f32, MVT::f64, MVT::f128}, Action: LibCall);
1023
1024 setOperationAction(Ops: {ISD::FTAN, ISD::FACOS, ISD::FASIN, ISD::FATAN, ISD::FCOSH,
1025 ISD::FSINH, ISD::FTANH, ISD::FATAN2},
1026 VT: MVT::f16, Action: Promote);
1027 // Default ISD::TRAP to expand (which turns it into abort).
1028 setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Expand);
1029
1030 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
1031 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
1032 setOperationAction(Op: ISD::DEBUGTRAP, VT: MVT::Other, Action: Expand);
1033
1034 setOperationAction(Op: ISD::UBSANTRAP, VT: MVT::Other, Action: Expand);
1035
1036 setOperationAction(Op: ISD::GET_FPENV_MEM, VT: MVT::Other, Action: Expand);
1037 setOperationAction(Op: ISD::SET_FPENV_MEM, VT: MVT::Other, Action: Expand);
1038
1039 for (MVT VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) {
1040 setOperationAction(Op: ISD::GET_FPMODE, VT, Action: Expand);
1041 setOperationAction(Op: ISD::SET_FPMODE, VT, Action: Expand);
1042 }
1043 setOperationAction(Op: ISD::RESET_FPMODE, VT: MVT::Other, Action: Expand);
1044
1045 // This one by default will call __clear_cache unless the target
1046 // wants something different.
1047 setOperationAction(Op: ISD::CLEAR_CACHE, VT: MVT::Other, Action: LibCall);
1048
1049 // By default, STACKADDRESS nodes are expanded like STACKSAVE nodes.
1050 // On SPARC targets, custom lowering is required.
1051 setOperationAction(Op: ISD::STACKADDRESS, VT: MVT::Other, Action: Expand);
1052}
1053
1054MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
1055 EVT) const {
1056 return MVT::getIntegerVT(BitWidth: DL.getPointerSizeInBits(AS: 0));
1057}
1058
1059EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy,
1060 const DataLayout &DL) const {
1061 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
1062 if (LHSTy.isVector())
1063 return LHSTy;
1064 MVT ShiftVT = getScalarShiftAmountTy(DL, LHSTy);
1065 // If any possible shift value won't fit in the prefered type, just use
1066 // something safe. Assume it will be legalized when the shift is expanded.
1067 if (ShiftVT.getSizeInBits() < Log2_32_Ceil(Value: LHSTy.getSizeInBits()))
1068 ShiftVT = MVT::i32;
1069 assert(ShiftVT.getSizeInBits() >= Log2_32_Ceil(LHSTy.getSizeInBits()) &&
1070 "ShiftVT is still too small!");
1071 return ShiftVT;
1072}
1073
1074bool TargetLoweringBase::isFreeAddrSpaceCast(const DataLayout &DL,
1075 unsigned SrcAS,
1076 unsigned DestAS) const {
1077 return TM.isNoopAddrSpaceCast(DL, SrcAS, DestAS);
1078}
1079
1080unsigned TargetLoweringBase::getBitWidthForCttzElements(
1081 EVT RetVT, ElementCount EC, bool ZeroIsPoison,
1082 const ConstantRange *VScaleRange) const {
1083 // Find the smallest "sensible" element type to use for the expansion.
1084 ConstantRange CR(APInt(64, EC.getKnownMinValue()));
1085 if (EC.isScalable())
1086 CR = CR.umul_sat(Other: *VScaleRange);
1087
1088 if (ZeroIsPoison)
1089 CR = CR.subtract(CI: APInt(64, 1));
1090
1091 unsigned EltWidth = RetVT.getScalarSizeInBits();
1092 EltWidth = std::min(a: EltWidth, b: CR.getActiveBits());
1093 EltWidth = std::max(a: llvm::bit_ceil(Value: EltWidth), b: (unsigned)8);
1094
1095 return EltWidth;
1096}
1097
1098void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
1099 // If the command-line option was specified, ignore this request.
1100 if (!JumpIsExpensiveOverride.getNumOccurrences())
1101 JumpIsExpensive = isExpensive;
1102}
1103
1104TargetLoweringBase::LegalizeKind
1105TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
1106 // If this is a simple type, use the ComputeRegisterProp mechanism.
1107 if (VT.isSimple()) {
1108 MVT SVT = VT.getSimpleVT();
1109 assert((unsigned)SVT.SimpleTy < std::size(TransformToType));
1110 MVT NVT = TransformToType[SVT.SimpleTy];
1111 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(VT: SVT);
1112
1113 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
1114 LA == TypeSoftPromoteHalf ||
1115 (NVT.isVector() ||
1116 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
1117 "Promote may not follow Expand or Promote");
1118
1119 if (LA == TypeSplitVector)
1120 return LegalizeKind(LA, EVT(SVT).getHalfNumVectorElementsVT(Context));
1121 if (LA == TypeScalarizeVector)
1122 return LegalizeKind(LA, SVT.getVectorElementType());
1123 return LegalizeKind(LA, NVT);
1124 }
1125
1126 // Handle Extended Scalar Types.
1127 if (!VT.isVector()) {
1128 assert(VT.isInteger() && "Float types must be simple");
1129 unsigned BitSize = VT.getSizeInBits();
1130 // First promote to a power-of-two size, then expand if necessary.
1131 if (BitSize < 8 || !isPowerOf2_32(Value: BitSize)) {
1132 EVT NVT = VT.getRoundIntegerType(Context);
1133 assert(NVT != VT && "Unable to round integer VT");
1134 LegalizeKind NextStep = getTypeConversion(Context, VT: NVT);
1135 // Avoid multi-step promotion.
1136 if (NextStep.first == TypePromoteInteger)
1137 return NextStep;
1138 // Return rounded integer type.
1139 return LegalizeKind(TypePromoteInteger, NVT);
1140 }
1141
1142 return LegalizeKind(TypeExpandInteger,
1143 EVT::getIntegerVT(Context, BitWidth: VT.getSizeInBits() / 2));
1144 }
1145
1146 // Handle vector types.
1147 ElementCount NumElts = VT.getVectorElementCount();
1148 EVT EltVT = VT.getVectorElementType();
1149
1150 // Vectors with only one element are always scalarized.
1151 if (NumElts.isScalar())
1152 return LegalizeKind(TypeScalarizeVector, EltVT);
1153
1154 // Try to widen vector elements until the element type is a power of two and
1155 // promote it to a legal type later on, for example:
1156 // <3 x i8> -> <4 x i8> -> <4 x i32>
1157 if (EltVT.isInteger()) {
1158 // Vectors with a number of elements that is not a power of two are always
1159 // widened, for example <3 x i8> -> <4 x i8>.
1160 if (!VT.isPow2VectorType()) {
1161 NumElts = NumElts.coefficientNextPowerOf2();
1162 EVT NVT = EVT::getVectorVT(Context, VT: EltVT, EC: NumElts);
1163 return LegalizeKind(TypeWidenVector, NVT);
1164 }
1165
1166 // Examine the element type.
1167 LegalizeKind LK = getTypeConversion(Context, VT: EltVT);
1168
1169 // If type is to be expanded, split the vector.
1170 // <4 x i140> -> <2 x i140>
1171 if (LK.first == TypeExpandInteger) {
1172 if (NumElts.isScalable() && NumElts.getKnownMinValue() == 1)
1173 return LegalizeKind(TypeScalarizeScalableVector, EltVT);
1174 return LegalizeKind(TypeSplitVector,
1175 VT.getHalfNumVectorElementsVT(Context));
1176 }
1177
1178 // Promote the integer element types until a legal vector type is found
1179 // or until the element integer type is too big. If a legal type was not
1180 // found, fallback to the usual mechanism of widening/splitting the
1181 // vector.
1182 EVT OldEltVT = EltVT;
1183 while (true) {
1184 // Increase the bitwidth of the element to the next pow-of-two
1185 // (which is greater than 8 bits).
1186 EltVT = EVT::getIntegerVT(Context, BitWidth: 1 + EltVT.getSizeInBits())
1187 .getRoundIntegerType(Context);
1188
1189 // Stop trying when getting a non-simple element type.
1190 // Note that vector elements may be greater than legal vector element
1191 // types. Example: X86 XMM registers hold 64bit element on 32bit
1192 // systems.
1193 if (!EltVT.isSimple())
1194 break;
1195
1196 // Build a new vector type and check if it is legal.
1197 MVT NVT = MVT::getVectorVT(VT: EltVT.getSimpleVT(), EC: NumElts);
1198 // Found a legal promoted vector type.
1199 if (NVT != MVT() && ValueTypeActions.getTypeAction(VT: NVT) == TypeLegal)
1200 return LegalizeKind(TypePromoteInteger,
1201 EVT::getVectorVT(Context, VT: EltVT, EC: NumElts));
1202 }
1203
1204 // Reset the type to the unexpanded type if we did not find a legal vector
1205 // type with a promoted vector element type.
1206 EltVT = OldEltVT;
1207 }
1208
1209 // Try to widen the vector until a legal type is found.
1210 // If there is no wider legal type, split the vector.
1211 while (true) {
1212 // Round up to the next power of 2.
1213 NumElts = NumElts.coefficientNextPowerOf2();
1214
1215 // If there is no simple vector type with this many elements then there
1216 // cannot be a larger legal vector type. Note that this assumes that
1217 // there are no skipped intermediate vector types in the simple types.
1218 if (!EltVT.isSimple())
1219 break;
1220 MVT LargerVector = MVT::getVectorVT(VT: EltVT.getSimpleVT(), EC: NumElts);
1221 if (LargerVector == MVT())
1222 break;
1223
1224 // If this type is legal then widen the vector.
1225 if (ValueTypeActions.getTypeAction(VT: LargerVector) == TypeLegal)
1226 return LegalizeKind(TypeWidenVector, LargerVector);
1227 }
1228
1229 // Widen odd vectors to next power of two.
1230 if (!VT.isPow2VectorType()) {
1231 EVT NVT = VT.getPow2VectorType(Context);
1232 return LegalizeKind(TypeWidenVector, NVT);
1233 }
1234
1235 if (VT.getVectorElementCount() == ElementCount::getScalable(MinVal: 1))
1236 return LegalizeKind(TypeScalarizeScalableVector, EltVT);
1237
1238 // Vectors with illegal element types are expanded.
1239 EVT NVT = EVT::getVectorVT(Context, VT: EltVT,
1240 EC: VT.getVectorElementCount().divideCoefficientBy(RHS: 2));
1241 return LegalizeKind(TypeSplitVector, NVT);
1242}
1243
1244unsigned TargetLoweringBase::getVectorTypeBreakdownMVT(
1245 MVT VT, MVT &IntermediateVT, unsigned &NumIntermediates, MVT &RegisterVT) {
1246 // Figure out the right, legal destination reg to copy into.
1247 ElementCount EC = VT.getVectorElementCount();
1248 MVT EltTy = VT.getVectorElementType();
1249
1250 unsigned NumVectorRegs = 1;
1251
1252 // Scalable vectors cannot be scalarized, so splitting or widening is
1253 // required.
1254 if (VT.isScalableVector() && !EC.isPowerOf2())
1255 llvm_unreachable(
1256 "Splitting or widening of non-power-of-2 MVTs is not implemented.");
1257
1258 // FIXME: We don't support non-power-of-2-sized vectors for now.
1259 // Ideally we could break down into LHS/RHS like LegalizeDAG does.
1260 if (!EC.isPowerOf2()) {
1261 // Split EC to unit size (scalable property is preserved).
1262 NumVectorRegs = EC.getKnownMinValue();
1263 EC = ElementCount::getFixed(MinVal: 1);
1264 }
1265
1266 // Divide the input until we get to a supported size. This will
1267 // always end up with an EC that represent a scalar or a scalable
1268 // scalar.
1269 while (EC.getKnownMinValue() > 1 &&
1270 !isTypeLegal(VT: MVT::getVectorVT(VT: EltTy, EC))) {
1271 EC = EC.divideCoefficientBy(RHS: 2);
1272 NumVectorRegs <<= 1;
1273 }
1274
1275 NumIntermediates = NumVectorRegs;
1276
1277 MVT NewVT = MVT::getVectorVT(VT: EltTy, EC);
1278 if (!isTypeLegal(VT: NewVT))
1279 NewVT = EltTy;
1280 IntermediateVT = NewVT;
1281
1282 unsigned LaneSizeInBits = NewVT.getScalarSizeInBits();
1283
1284 // Convert sizes such as i33 to i64.
1285 LaneSizeInBits = llvm::bit_ceil(Value: LaneSizeInBits);
1286
1287 MVT DestVT = getCachedRegisterType(VT: NewVT);
1288 RegisterVT = DestVT;
1289 if (EVT(DestVT).bitsLT(VT: NewVT)) // Value is expanded, e.g. i64 -> i16.
1290 return NumVectorRegs * (LaneSizeInBits / DestVT.getScalarSizeInBits());
1291
1292 // Otherwise, promotion or legal types use the same number of registers as
1293 // the vector decimated to the appropriate level.
1294 return NumVectorRegs;
1295}
1296
1297/// isLegalRC - Return true if the value types that can be represented by the
1298/// specified register class are all legal.
1299bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
1300 const TargetRegisterClass &RC) const {
1301 for (const auto *I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1302 if (isTypeLegal(VT: *I))
1303 return true;
1304 return false;
1305}
1306
1307/// Replace/modify any TargetFrameIndex operands with a targte-dependent
1308/// sequence of memory operands that is recognized by PrologEpilogInserter.
1309MachineBasicBlock *
1310TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
1311 MachineBasicBlock *MBB) const {
1312 MachineInstr *MI = &InitialMI;
1313 MachineFunction &MF = *MI->getMF();
1314 MachineFrameInfo &MFI = MF.getFrameInfo();
1315
1316 // We're handling multiple types of operands here:
1317 // PATCHPOINT MetaArgs - live-in, read only, direct
1318 // STATEPOINT Deopt Spill - live-through, read only, indirect
1319 // STATEPOINT Deopt Alloca - live-through, read only, direct
1320 // (We're currently conservative and mark the deopt slots read/write in
1321 // practice.)
1322 // STATEPOINT GC Spill - live-through, read/write, indirect
1323 // STATEPOINT GC Alloca - live-through, read/write, direct
1324 // The live-in vs live-through is handled already (the live through ones are
1325 // all stack slots), but we need to handle the different type of stackmap
1326 // operands and memory effects here.
1327
1328 if (llvm::none_of(Range: MI->operands(),
1329 P: [](MachineOperand &Operand) { return Operand.isFI(); }))
1330 return MBB;
1331
1332 MachineInstrBuilder MIB = BuildMI(MF, MIMD: MI->getDebugLoc(), MCID: MI->getDesc());
1333
1334 // Inherit previous memory operands.
1335 MIB.cloneMemRefs(OtherMI: *MI);
1336
1337 for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1338 MachineOperand &MO = MI->getOperand(i);
1339 if (!MO.isFI()) {
1340 // Index of Def operand this Use it tied to.
1341 // Since Defs are coming before Uses, if Use is tied, then
1342 // index of Def must be smaller that index of that Use.
1343 // Also, Defs preserve their position in new MI.
1344 unsigned TiedTo = i;
1345 if (MO.isReg() && MO.isTied())
1346 TiedTo = MI->findTiedOperandIdx(OpIdx: i);
1347 MIB.add(MO);
1348 if (TiedTo < i)
1349 MIB->tieOperands(DefIdx: TiedTo, UseIdx: MIB->getNumOperands() - 1);
1350 continue;
1351 }
1352
1353 // foldMemoryOperand builds a new MI after replacing a single FI operand
1354 // with the canonical set of five x86 addressing-mode operands.
1355 int FI = MO.getIndex();
1356
1357 // Add frame index operands recognized by stackmaps.cpp
1358 if (MFI.isStatepointSpillSlotObjectIndex(ObjectIdx: FI)) {
1359 // indirect-mem-ref tag, size, #FI, offset.
1360 // Used for spills inserted by StatepointLowering. This codepath is not
1361 // used for patchpoints/stackmaps at all, for these spilling is done via
1362 // foldMemoryOperand callback only.
1363 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1364 MIB.addImm(Val: StackMaps::IndirectMemRefOp);
1365 MIB.addImm(Val: MFI.getObjectSize(ObjectIdx: FI));
1366 MIB.add(MO);
1367 MIB.addImm(Val: 0);
1368 } else {
1369 // direct-mem-ref tag, #FI, offset.
1370 // Used by patchpoint, and direct alloca arguments to statepoints
1371 MIB.addImm(Val: StackMaps::DirectMemRefOp);
1372 MIB.add(MO);
1373 MIB.addImm(Val: 0);
1374 }
1375
1376 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1377
1378 // Add a new memory operand for this FI.
1379 assert(MFI.getObjectOffset(FI) != -1);
1380
1381 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and
1382 // PATCHPOINT should be updated to do the same. (TODO)
1383 if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1384 auto Flags = MachineMemOperand::MOLoad;
1385 MachineMemOperand *MMO = MF.getMachineMemOperand(
1386 PtrInfo: MachinePointerInfo::getFixedStack(MF, FI), F: Flags,
1387 Size: MF.getDataLayout().getPointerSize(), BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI));
1388 MIB->addMemOperand(MF, MO: MMO);
1389 }
1390 }
1391 MBB->insert(I: MachineBasicBlock::iterator(MI), MI: MIB);
1392 MI->eraseFromParent();
1393 return MBB;
1394}
1395
1396/// findRepresentativeClass - Return the largest legal super-reg register class
1397/// of the register class for the specified type and its associated "cost".
1398// This function is in TargetLowering because it uses RegClassForVT which would
1399// need to be moved to TargetRegisterInfo and would necessitate moving
1400// isTypeLegal over as well - a massive change that would just require
1401// TargetLowering having a TargetRegisterInfo class member that it would use.
1402std::pair<const TargetRegisterClass *, uint8_t>
1403TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1404 MVT VT) const {
1405 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1406 if (!RC)
1407 return std::make_pair(x&: RC, y: 0);
1408
1409 // Compute the set of all super-register classes.
1410 BitVector SuperRegRC(TRI->getNumRegClasses());
1411 for (SuperRegClassIterator RCI(RC); RCI.isValid(); ++RCI)
1412 SuperRegRC.setBitsInMask(Mask: RCI.getMask());
1413
1414 // Find the first legal register class with the largest spill size.
1415 const TargetRegisterClass *BestRC = RC;
1416 for (unsigned i : SuperRegRC.set_bits()) {
1417 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1418 // We want the largest possible spill size.
1419 if (TRI->getSpillSize(RC: *SuperRC) <= TRI->getSpillSize(RC: *BestRC))
1420 continue;
1421 if (!isLegalRC(TRI: *TRI, RC: *SuperRC))
1422 continue;
1423 BestRC = SuperRC;
1424 }
1425 return std::make_pair(x&: BestRC, y: 1);
1426}
1427
1428/// computeRegisterProperties - Once all of the register classes are added,
1429/// this allows us to compute derived properties we expose.
1430void TargetLoweringBase::computeRegisterProperties(
1431 const TargetRegisterInfo *TRI) {
1432 // Everything defaults to needing one register.
1433 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1434 NumRegistersForVT[i] = 1;
1435 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1436 }
1437 // ...except isVoid, which doesn't need any registers.
1438 NumRegistersForVT[MVT::isVoid] = 0;
1439
1440 // Find the largest integer register class.
1441 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1442 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1443 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1444
1445 // Every integer value type larger than this largest register takes twice as
1446 // many registers to represent as the previous ValueType.
1447 for (unsigned ExpandedReg = LargestIntReg + 1;
1448 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1449 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1450 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1451 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1452 ValueTypeActions.setTypeAction(VT: (MVT::SimpleValueType)ExpandedReg,
1453 Action: TypeExpandInteger);
1454 }
1455
1456 // Inspect all of the ValueType's smaller than the largest integer
1457 // register to see which ones need promotion.
1458 unsigned LegalIntReg = LargestIntReg;
1459 for (unsigned IntReg = LargestIntReg - 1;
1460 IntReg >= (unsigned)MVT::i1; --IntReg) {
1461 MVT IVT = (MVT::SimpleValueType)IntReg;
1462 if (isTypeLegal(VT: IVT)) {
1463 LegalIntReg = IntReg;
1464 } else {
1465 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1466 (MVT::SimpleValueType)LegalIntReg;
1467 ValueTypeActions.setTypeAction(VT: IVT, Action: TypePromoteInteger);
1468 }
1469 }
1470
1471 // ppcf128 type is really two f64's.
1472 if (!isTypeLegal(VT: MVT::ppcf128)) {
1473 if (isTypeLegal(VT: MVT::f64)) {
1474 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1475 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1476 TransformToType[MVT::ppcf128] = MVT::f64;
1477 ValueTypeActions.setTypeAction(VT: MVT::ppcf128, Action: TypeExpandFloat);
1478 } else {
1479 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1480 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1481 TransformToType[MVT::ppcf128] = MVT::i128;
1482 ValueTypeActions.setTypeAction(VT: MVT::ppcf128, Action: TypeSoftenFloat);
1483 }
1484 }
1485
1486 // Decide how to handle f128. If the target does not have native f128 support,
1487 // expand it to i128 and we will be generating soft float library calls.
1488 if (!isTypeLegal(VT: MVT::f128)) {
1489 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1490 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1491 TransformToType[MVT::f128] = MVT::i128;
1492 ValueTypeActions.setTypeAction(VT: MVT::f128, Action: TypeSoftenFloat);
1493 }
1494
1495 // Decide how to handle f80. If the target does not have native f80 support,
1496 // expand it to i96 and we will be generating soft float library calls.
1497 if (!isTypeLegal(VT: MVT::f80)) {
1498 NumRegistersForVT[MVT::f80] = 3*NumRegistersForVT[MVT::i32];
1499 RegisterTypeForVT[MVT::f80] = RegisterTypeForVT[MVT::i32];
1500 TransformToType[MVT::f80] = MVT::i32;
1501 ValueTypeActions.setTypeAction(VT: MVT::f80, Action: TypeSoftenFloat);
1502 }
1503
1504 // Decide how to handle f64. If the target does not have native f64 support,
1505 // expand it to i64 and we will be generating soft float library calls.
1506 if (!isTypeLegal(VT: MVT::f64)) {
1507 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1508 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1509 TransformToType[MVT::f64] = MVT::i64;
1510 ValueTypeActions.setTypeAction(VT: MVT::f64, Action: TypeSoftenFloat);
1511 }
1512
1513 // Decide how to handle f32. If the target does not have native f32 support,
1514 // expand it to i32 and we will be generating soft float library calls.
1515 if (!isTypeLegal(VT: MVT::f32)) {
1516 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1517 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1518 TransformToType[MVT::f32] = MVT::i32;
1519 ValueTypeActions.setTypeAction(VT: MVT::f32, Action: TypeSoftenFloat);
1520 }
1521
1522 // Decide how to handle f16. If the target does not have native f16 support,
1523 // promote it to f32, because there are no f16 library calls (except for
1524 // conversions).
1525 if (!isTypeLegal(VT: MVT::f16)) {
1526 // Allow targets to control how we legalize half.
1527 bool UseFPRegsForHalfType = useFPRegsForHalfType();
1528
1529 if (!UseFPRegsForHalfType) {
1530 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1531 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1532 } else {
1533 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1534 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1535 }
1536 TransformToType[MVT::f16] = MVT::f32;
1537 ValueTypeActions.setTypeAction(VT: MVT::f16, Action: TypeSoftPromoteHalf);
1538 }
1539
1540 // Decide how to handle bf16. If the target does not have native bf16 support,
1541 // promote it to f32, because there are no bf16 library calls (except for
1542 // converting from f32 to bf16).
1543 if (!isTypeLegal(VT: MVT::bf16)) {
1544 NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32];
1545 RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32];
1546 TransformToType[MVT::bf16] = MVT::f32;
1547 ValueTypeActions.setTypeAction(VT: MVT::bf16, Action: TypeSoftPromoteHalf);
1548 }
1549
1550 // Loop over all of the vector value types to see which need transformations.
1551 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1552 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1553 MVT VT = (MVT::SimpleValueType) i;
1554 if (isTypeLegal(VT))
1555 continue;
1556
1557 MVT EltVT = VT.getVectorElementType();
1558 ElementCount EC = VT.getVectorElementCount();
1559 bool IsLegalWiderType = false;
1560 bool IsScalable = VT.isScalableVector();
1561 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1562 switch (PreferredAction) {
1563 case TypePromoteInteger: {
1564 MVT::SimpleValueType EndVT = IsScalable ?
1565 MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1566 MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1567 // Try to promote the elements of integer vectors. If no legal
1568 // promotion was found, fall through to the widen-vector method.
1569 for (unsigned nVT = i + 1;
1570 (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1571 MVT SVT = (MVT::SimpleValueType) nVT;
1572 // Promote vectors of integers to vectors with the same number
1573 // of elements, with a wider element type.
1574 if (SVT.getScalarSizeInBits() > EltVT.getFixedSizeInBits() &&
1575 SVT.getVectorElementCount() == EC && isTypeLegal(VT: SVT)) {
1576 TransformToType[i] = SVT;
1577 RegisterTypeForVT[i] = SVT;
1578 NumRegistersForVT[i] = 1;
1579 ValueTypeActions.setTypeAction(VT, Action: TypePromoteInteger);
1580 IsLegalWiderType = true;
1581 break;
1582 }
1583 }
1584 if (IsLegalWiderType)
1585 break;
1586 [[fallthrough]];
1587 }
1588
1589 case TypeWidenVector:
1590 if (EC.isPowerOf2()) {
1591 // Try to widen the vector.
1592 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1593 MVT SVT = (MVT::SimpleValueType) nVT;
1594 if (SVT.getVectorElementType() == EltVT &&
1595 SVT.isScalableVector() == IsScalable &&
1596 SVT.getVectorElementCount().getKnownMinValue() >
1597 EC.getKnownMinValue() &&
1598 isTypeLegal(VT: SVT)) {
1599 TransformToType[i] = SVT;
1600 RegisterTypeForVT[i] = SVT;
1601 NumRegistersForVT[i] = 1;
1602 ValueTypeActions.setTypeAction(VT, Action: TypeWidenVector);
1603 IsLegalWiderType = true;
1604 break;
1605 }
1606 }
1607 if (IsLegalWiderType)
1608 break;
1609 } else {
1610 // Only widen to the next power of 2 to keep consistency with EVT.
1611 MVT NVT = VT.getPow2VectorType();
1612 if (isTypeLegal(VT: NVT)) {
1613 TransformToType[i] = NVT;
1614 ValueTypeActions.setTypeAction(VT, Action: TypeWidenVector);
1615 RegisterTypeForVT[i] = NVT;
1616 NumRegistersForVT[i] = 1;
1617 break;
1618 }
1619 }
1620 [[fallthrough]];
1621
1622 case TypeSplitVector:
1623 case TypeScalarizeVector: {
1624 MVT IntermediateVT;
1625 MVT RegisterVT;
1626 unsigned NumIntermediates;
1627 unsigned NumRegisters = getVectorTypeBreakdownMVT(
1628 VT, IntermediateVT, NumIntermediates, RegisterVT);
1629 NumRegistersForVT[i] = NumRegisters;
1630 assert(NumRegistersForVT[i] == NumRegisters &&
1631 "NumRegistersForVT size cannot represent NumRegisters!");
1632 RegisterTypeForVT[i] = RegisterVT;
1633
1634 MVT NVT = VT.getPow2VectorType();
1635 if (NVT == VT) {
1636 // Type is already a power of 2. The default action is to split.
1637 TransformToType[i] = MVT::Other;
1638 if (PreferredAction == TypeScalarizeVector)
1639 ValueTypeActions.setTypeAction(VT, Action: TypeScalarizeVector);
1640 else if (PreferredAction == TypeSplitVector)
1641 ValueTypeActions.setTypeAction(VT, Action: TypeSplitVector);
1642 else if (EC.getKnownMinValue() > 1)
1643 ValueTypeActions.setTypeAction(VT, Action: TypeSplitVector);
1644 else
1645 ValueTypeActions.setTypeAction(VT, Action: EC.isScalable()
1646 ? TypeScalarizeScalableVector
1647 : TypeScalarizeVector);
1648 } else {
1649 TransformToType[i] = NVT;
1650 ValueTypeActions.setTypeAction(VT, Action: TypeWidenVector);
1651 }
1652 break;
1653 }
1654 default:
1655 llvm_unreachable("Unknown vector legalization action!");
1656 }
1657 }
1658
1659 // Determine the 'representative' register class for each value type.
1660 // An representative register class is the largest (meaning one which is
1661 // not a sub-register class / subreg register class) legal register class for
1662 // a group of value types. For example, on i386, i8, i16, and i32
1663 // representative would be GR32; while on x86_64 it's GR64.
1664 for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) {
1665 const TargetRegisterClass* RRC;
1666 uint8_t Cost;
1667 std::tie(args&: RRC, args&: Cost) = findRepresentativeClass(TRI, VT: (MVT::SimpleValueType)i);
1668 RepRegClassForVT[i] = RRC;
1669 RepRegClassCostForVT[i] = Cost;
1670 }
1671
1672 // Compute minimum known-legal store size.
1673 MaximumLegalStoreInBits = 0;
1674 for (MVT VT : MVT::all_valuetypes())
1675 if (VT != MVT::Other && isTypeLegal(VT) &&
1676 VT.getSizeInBits().getKnownMinValue() >= MaximumLegalStoreInBits)
1677 MaximumLegalStoreInBits = VT.getSizeInBits().getKnownMinValue();
1678}
1679
1680EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1681 EVT VT) const {
1682 assert(!VT.isVector() && "No default SetCC type for vectors!");
1683 return getPointerTy(DL).SimpleTy;
1684}
1685
1686/// getVectorTypeBreakdown - Vector types are broken down into some number of
1687/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1688/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1689/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1690///
1691/// This method returns the number of registers needed, and the VT for each
1692/// register. It also returns the VT and quantity of the intermediate values
1693/// before they are promoted/expanded.
1694unsigned TargetLoweringBase::getVectorTypeBreakdownImpl(
1695 LLVMContext &Context, EVT VT, EVT &IntermediateVT,
1696 unsigned &NumIntermediates, MVT &RegisterVT, bool ForCallingConv) const {
1697 ElementCount EltCnt = VT.getVectorElementCount();
1698
1699 // If there is a wider vector type with the same element type as this one,
1700 // or a promoted vector type that has the same number of elements which
1701 // are wider, then we should convert to that legal vector type.
1702 // This handles things like <2 x float> -> <4 x float> and
1703 // <4 x i1> -> <4 x i32>.
1704 LegalizeTypeAction TA = getTypeAction(Context, VT);
1705 if (!EltCnt.isScalar() &&
1706 (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1707 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1708 if (isTypeLegal(VT: RegisterEVT)) {
1709 IntermediateVT = RegisterEVT;
1710 RegisterVT = RegisterEVT.getSimpleVT();
1711 NumIntermediates = 1;
1712 return 1;
1713 }
1714 }
1715
1716 // Figure out the right, legal destination reg to copy into.
1717 EVT EltTy = VT.getVectorElementType();
1718
1719 unsigned NumVectorRegs = 1;
1720
1721 auto GetLegalVectorBreakdown = [&]() -> std::optional<unsigned> {
1722 LegalizeKind LK;
1723 EVT PartVT = VT;
1724 do {
1725 // Iterate until we've found a legal (part) type to hold VT.
1726 LK = getTypeConversion(Context, VT: PartVT);
1727 PartVT = LK.second;
1728 } while (LK.first != TypeLegal);
1729
1730 if (!PartVT.isVector())
1731 return std::nullopt;
1732
1733 assert(PartVT.isScalableVector() == VT.isScalableVector() &&
1734 "Vector legalization changed scalability");
1735 NumIntermediates =
1736 divideCeil(Numerator: VT.getVectorElementCount().getKnownMinValue(),
1737 Denominator: PartVT.getVectorElementCount().getKnownMinValue());
1738 IntermediateVT = PartVT;
1739 RegisterVT = getRegisterType(Context, VT: IntermediateVT);
1740 return NumIntermediates;
1741 };
1742
1743 // Scalable vectors cannot be scalarized, so handle the legalisation of the
1744 // types like done elsewhere in SelectionDAG.
1745 if (EltCnt.isScalable()) {
1746 if (std::optional<unsigned> NumRegs = GetLegalVectorBreakdown())
1747 return *NumRegs;
1748 report_fatal_error(reason: "Don't know how to legalize this scalable vector type");
1749 }
1750
1751 // FIXME: We don't generically support non-power-of-2-sized vectors for now.
1752 // Ideally we could break down into LHS/RHS like LegalizeDAG does.
1753 if (!EltCnt.isPowerOf2()) {
1754 assert(VT.isFixedLengthVector() && "Expected a fixed-length vector VT");
1755 unsigned NumElts = EltCnt.getKnownMinValue();
1756
1757 if (!ForCallingConv && preferVectorizedNonPowerOfTwoTypeBreakdown())
1758 if (std::optional<unsigned> NumRegs = GetLegalVectorBreakdown())
1759 return *NumRegs;
1760
1761 // Fall back to scalars if there is no legal vector decomposition.
1762 NumVectorRegs = NumElts;
1763 EltCnt = ElementCount::getFixed(MinVal: 1);
1764 }
1765
1766 // Divide the input until we get to a supported size. This will always
1767 // end with a scalar if the target doesn't support vectors.
1768 while (EltCnt.getKnownMinValue() > 1 &&
1769 !isTypeLegal(VT: EVT::getVectorVT(Context, VT: EltTy, EC: EltCnt))) {
1770 EltCnt = EltCnt.divideCoefficientBy(RHS: 2);
1771 NumVectorRegs <<= 1;
1772 }
1773
1774 NumIntermediates = NumVectorRegs;
1775
1776 EVT NewVT = EVT::getVectorVT(Context, VT: EltTy, EC: EltCnt);
1777 if (!isTypeLegal(VT: NewVT))
1778 NewVT = EltTy;
1779 IntermediateVT = NewVT;
1780
1781 MVT DestVT = getRegisterType(Context, VT: NewVT);
1782 RegisterVT = DestVT;
1783
1784 if (EVT(DestVT).bitsLT(VT: NewVT)) { // Value is expanded, e.g. i64 -> i16.
1785 TypeSize NewVTSize = NewVT.getSizeInBits();
1786 // Convert sizes such as i33 to i64.
1787 if (!llvm::has_single_bit<uint32_t>(Value: NewVTSize.getKnownMinValue()))
1788 NewVTSize = NewVTSize.coefficientNextPowerOf2();
1789 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1790 }
1791
1792 // Otherwise, promotion or legal types use the same number of registers as
1793 // the vector decimated to the appropriate level.
1794 return NumVectorRegs;
1795}
1796
1797bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI,
1798 uint64_t NumCases,
1799 uint64_t Range,
1800 ProfileSummaryInfo *PSI,
1801 BlockFrequencyInfo *BFI) const {
1802 // FIXME: This function check the maximum table size and density, but the
1803 // minimum size is not checked. It would be nice if the minimum size is
1804 // also combined within this function. Currently, the minimum size check is
1805 // performed in findJumpTable() in SelectionDAGBuiler and
1806 // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
1807 const bool OptForSize =
1808 llvm::shouldOptimizeForSize(BB: SI->getParent(), PSI, BFI);
1809 const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
1810 const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
1811
1812 // Check whether the number of cases is small enough and
1813 // the range is dense enough for a jump table.
1814 return (OptForSize || Range <= MaxJumpTableSize) &&
1815 (NumCases * 100 >= Range * MinDensity);
1816}
1817
1818MVT TargetLoweringBase::getPreferredSwitchConditionType(LLVMContext &Context,
1819 EVT ConditionVT) const {
1820 return getRegisterType(Context, VT: ConditionVT);
1821}
1822
1823/// Get the EVTs and ArgFlags collections that represent the legalized return
1824/// type of the given function. This does not require a DAG or a return value,
1825/// and is suitable for use before any DAGs for the function are constructed.
1826/// TODO: Move this out of TargetLowering.cpp.
1827void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1828 AttributeList attr,
1829 SmallVectorImpl<ISD::OutputArg> &Outs,
1830 const TargetLowering &TLI, const DataLayout &DL) {
1831 SmallVector<Type *, 4> Types;
1832 ComputeValueTypes(DL, Ty: ReturnType, Types);
1833 unsigned NumValues = Types.size();
1834 if (NumValues == 0) return;
1835
1836 for (Type *Ty : Types) {
1837 EVT VT = TLI.getValueType(DL, Ty);
1838 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1839
1840 if (attr.hasRetAttr(Kind: Attribute::SExt))
1841 ExtendKind = ISD::SIGN_EXTEND;
1842 else if (attr.hasRetAttr(Kind: Attribute::ZExt))
1843 ExtendKind = ISD::ZERO_EXTEND;
1844
1845 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1846 VT = TLI.getTypeForExtReturn(Context&: ReturnType->getContext(), VT, ExtendKind);
1847
1848 unsigned NumParts =
1849 TLI.getNumRegistersForCallingConv(Context&: ReturnType->getContext(), CC, VT);
1850 MVT PartVT =
1851 TLI.getRegisterTypeForCallingConv(Context&: ReturnType->getContext(), CC, VT);
1852
1853 // 'inreg' on function refers to return value
1854 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1855 if (attr.hasRetAttr(Kind: Attribute::InReg))
1856 Flags.setInReg();
1857
1858 // Propagate extension type if any
1859 if (attr.hasRetAttr(Kind: Attribute::SExt))
1860 Flags.setSExt();
1861 else if (attr.hasRetAttr(Kind: Attribute::ZExt))
1862 Flags.setZExt();
1863
1864 for (unsigned i = 0; i < NumParts; ++i)
1865 Outs.push_back(Elt: ISD::OutputArg(Flags, PartVT, VT, Ty, 0, 0));
1866 }
1867}
1868
1869Align TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1870 const DataLayout &DL) const {
1871 return DL.getABITypeAlign(Ty);
1872}
1873
1874bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1875 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
1876 Align Alignment, MachineMemOperand::Flags Flags, unsigned *Fast) const {
1877 // Check if the specified alignment is sufficient based on the data layout.
1878 // TODO: While using the data layout works in practice, a better solution
1879 // would be to implement this check directly (make this a virtual function).
1880 // For example, the ABI alignment may change based on software platform while
1881 // this function should only be affected by hardware implementation.
1882 Type *Ty = VT.getTypeForEVT(Context);
1883 if (VT.isZeroSized() || Alignment >= DL.getABITypeAlign(Ty)) {
1884 // Assume that an access that meets the ABI-specified alignment is fast.
1885 if (Fast != nullptr)
1886 *Fast = 1;
1887 return true;
1888 }
1889
1890 // This is a misaligned access.
1891 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast);
1892}
1893
1894bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1895 LLVMContext &Context, const DataLayout &DL, EVT VT,
1896 const MachineMemOperand &MMO, unsigned *Fast) const {
1897 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace: MMO.getAddrSpace(),
1898 Alignment: MMO.getAlign(), Flags: MMO.getFlags(), Fast);
1899}
1900
1901bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1902 const DataLayout &DL, EVT VT,
1903 unsigned AddrSpace, Align Alignment,
1904 MachineMemOperand::Flags Flags,
1905 unsigned *Fast) const {
1906 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
1907 Flags, Fast);
1908}
1909
1910bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1911 const DataLayout &DL, EVT VT,
1912 const MachineMemOperand &MMO,
1913 unsigned *Fast) const {
1914 return allowsMemoryAccess(Context, DL, VT, AddrSpace: MMO.getAddrSpace(), Alignment: MMO.getAlign(),
1915 Flags: MMO.getFlags(), Fast);
1916}
1917
1918bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1919 const DataLayout &DL, LLT Ty,
1920 const MachineMemOperand &MMO,
1921 unsigned *Fast) const {
1922 EVT VT = getApproximateEVTForLLT(Ty, Ctx&: Context);
1923 return allowsMemoryAccess(Context, DL, VT, AddrSpace: MMO.getAddrSpace(), Alignment: MMO.getAlign(),
1924 Flags: MMO.getFlags(), Fast);
1925}
1926
1927unsigned TargetLoweringBase::getMaxStoresPerMemset(bool OptSize) const {
1928 if (MaxStoresPerMemsetOverride > 0)
1929 return MaxStoresPerMemsetOverride;
1930
1931 return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset;
1932}
1933
1934unsigned TargetLoweringBase::getMaxStoresPerMemcpy(bool OptSize) const {
1935 if (MaxStoresPerMemcpyOverride > 0)
1936 return MaxStoresPerMemcpyOverride;
1937
1938 return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy;
1939}
1940
1941unsigned TargetLoweringBase::getMaxStoresPerMemmove(bool OptSize) const {
1942 if (MaxStoresPerMemmoveOverride > 0)
1943 return MaxStoresPerMemmoveOverride;
1944
1945 return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove;
1946}
1947
1948//===----------------------------------------------------------------------===//
1949// TargetTransformInfo Helpers
1950//===----------------------------------------------------------------------===//
1951
1952int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1953 enum InstructionOpcodes {
1954#define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1955#define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1956#include "llvm/IR/Instruction.def"
1957 };
1958 // clang-format off
1959 switch (static_cast<InstructionOpcodes>(Opcode)) {
1960 case Ret: return 0;
1961 case UncondBr: return 0;
1962 case CondBr: return 0;
1963 case Switch: return 0;
1964 case IndirectBr: return 0;
1965 case Invoke: return 0;
1966 case CallBr: return 0;
1967 case Resume: return 0;
1968 case Unreachable: return 0;
1969 case CleanupRet: return 0;
1970 case CatchRet: return 0;
1971 case CatchPad: return 0;
1972 case CatchSwitch: return 0;
1973 case CleanupPad: return 0;
1974 case FNeg: return ISD::FNEG;
1975 case Add: return ISD::ADD;
1976 case FAdd: return ISD::FADD;
1977 case Sub: return ISD::SUB;
1978 case FSub: return ISD::FSUB;
1979 case Mul: return ISD::MUL;
1980 case FMul: return ISD::FMUL;
1981 case UDiv: return ISD::UDIV;
1982 case SDiv: return ISD::SDIV;
1983 case FDiv: return ISD::FDIV;
1984 case URem: return ISD::UREM;
1985 case SRem: return ISD::SREM;
1986 case FRem: return ISD::FREM;
1987 case Shl: return ISD::SHL;
1988 case LShr: return ISD::SRL;
1989 case AShr: return ISD::SRA;
1990 case And: return ISD::AND;
1991 case Or: return ISD::OR;
1992 case Xor: return ISD::XOR;
1993 case Alloca: return 0;
1994 case Load: return ISD::LOAD;
1995 case Store: return ISD::STORE;
1996 case GetElementPtr: return 0;
1997 case Fence: return 0;
1998 case AtomicCmpXchg: return 0;
1999 case AtomicRMW: return 0;
2000 case Trunc: return ISD::TRUNCATE;
2001 case ZExt: return ISD::ZERO_EXTEND;
2002 case SExt: return ISD::SIGN_EXTEND;
2003 case FPToUI: return ISD::FP_TO_UINT;
2004 case FPToSI: return ISD::FP_TO_SINT;
2005 case UIToFP: return ISD::UINT_TO_FP;
2006 case SIToFP: return ISD::SINT_TO_FP;
2007 case FPTrunc: return ISD::FP_ROUND;
2008 case FPExt: return ISD::FP_EXTEND;
2009 case PtrToAddr: return ISD::BITCAST;
2010 case PtrToInt: return ISD::BITCAST;
2011 case IntToPtr: return ISD::BITCAST;
2012 case BitCast: return ISD::BITCAST;
2013 case AddrSpaceCast: return ISD::ADDRSPACECAST;
2014 case ICmp: return ISD::SETCC;
2015 case FCmp: return ISD::SETCC;
2016 case PHI: return 0;
2017 case Call: return 0;
2018 case Select: return ISD::SELECT;
2019 case UserOp1: return 0;
2020 case UserOp2: return 0;
2021 case VAArg: return 0;
2022 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
2023 case InsertElement: return ISD::INSERT_VECTOR_ELT;
2024 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
2025 case ExtractValue: return ISD::MERGE_VALUES;
2026 case InsertValue: return ISD::MERGE_VALUES;
2027 case LandingPad: return 0;
2028 case Freeze: return ISD::FREEZE;
2029 case BitInsert: return 0;
2030 case BitExtract: return 0;
2031 }
2032 // clang-format on
2033 llvm_unreachable("Unknown instruction type encountered!");
2034}
2035
2036int TargetLoweringBase::IntrinsicIDToISD(Intrinsic::ID ID) const {
2037 switch (ID) {
2038 case Intrinsic::acos:
2039 return ISD::FACOS;
2040 case Intrinsic::asin:
2041 return ISD::FASIN;
2042 case Intrinsic::atan:
2043 return ISD::FATAN;
2044 case Intrinsic::cos:
2045 return ISD::FCOS;
2046 case Intrinsic::cosh:
2047 return ISD::FCOSH;
2048 case Intrinsic::exp:
2049 return ISD::FEXP;
2050 case Intrinsic::exp2:
2051 return ISD::FEXP2;
2052 case Intrinsic::exp10:
2053 return ISD::FEXP10;
2054 case Intrinsic::log:
2055 return ISD::FLOG;
2056 case Intrinsic::log2:
2057 return ISD::FLOG2;
2058 case Intrinsic::log10:
2059 return ISD::FLOG10;
2060 case Intrinsic::modf:
2061 return ISD::FMODF;
2062 case Intrinsic::sin:
2063 return ISD::FSIN;
2064 case Intrinsic::sincos:
2065 return ISD::FSINCOS;
2066 case Intrinsic::sincospi:
2067 return ISD::FSINCOSPI;
2068 case Intrinsic::sinh:
2069 return ISD::FSINH;
2070 case Intrinsic::tan:
2071 return ISD::FTAN;
2072 case Intrinsic::tanh:
2073 return ISD::FTANH;
2074 default:
2075 return ISD::DELETED_NODE;
2076 }
2077}
2078
2079Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(
2080 IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const {
2081 // compiler-rt provides a variable with a magic name. Targets that do not
2082 // link with compiler-rt may also provide such a variable.
2083 Module *M = IRB.getModule();
2084
2085 RTLIB::LibcallImpl UnsafeStackPtrImpl =
2086 Libcalls.getLibcallImpl(Call: RTLIB::SAFESTACK_UNSAFE_STACK_PTR);
2087 if (UnsafeStackPtrImpl == RTLIB::Unsupported)
2088 return nullptr;
2089
2090 const RTLIB::RuntimeLibcallsInfo &RTLCI = Libcalls.getRuntimeLibcallsInfo();
2091 GlobalValue *GV = RTLCI.getOrInsertImplGlobal(M&: *M, LibcallImpl: UnsafeStackPtrImpl);
2092
2093 // The variable may already exist, check its type and attributes.
2094 LLVMContext &Ctx = M->getContext();
2095 StringRef UnsafeStackPtrVar = GV->getName();
2096 auto *UnsafeStackPtr = dyn_cast<GlobalVariable>(Val: GV);
2097 if (!UnsafeStackPtr) {
2098 Ctx.emitError(ErrorStr: Twine(UnsafeStackPtrVar) + " must be a global variable");
2099 return GV;
2100 }
2101
2102 Type *StackPtrTy = RTLCI.getVariableType(
2103 Ctx, TT: M->getTargetTriple(), DL: M->getDataLayout(), LibcallImpl: UnsafeStackPtrImpl);
2104 if (UnsafeStackPtr->getValueType() != StackPtrTy) {
2105 Ctx.emitError(
2106 ErrorStr: formatv(Fmt: "{0} must have type {1}", Vals&: UnsafeStackPtrVar, Vals&: *StackPtrTy));
2107 }
2108
2109 if (RTLIB::RuntimeLibcallsInfo::isThreadLocalVariable(LibcallImpl: UnsafeStackPtrImpl) &&
2110 !UnsafeStackPtr->isThreadLocal())
2111 Ctx.emitError(ErrorStr: Twine(UnsafeStackPtrVar) + " must be thread-local");
2112
2113 return UnsafeStackPtr;
2114}
2115
2116Value *TargetLoweringBase::getSafeStackPointerAddressCall(
2117 IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const {
2118 RTLIB::LibcallImpl SafestackPointerAddressImpl =
2119 Libcalls.getLibcallImpl(Call: RTLIB::SAFESTACK_POINTER_ADDRESS);
2120 if (SafestackPointerAddressImpl == RTLIB::Unsupported)
2121 return nullptr;
2122
2123 FunctionCallee Fn = Libcalls.getRuntimeLibcallsInfo().getOrInsertLibcallDecl(
2124 M&: *IRB.getModule(), LibcallImpl: SafestackPointerAddressImpl);
2125 CallInst *Call = IRB.CreateCall(Callee: Fn);
2126 Call->setCallingConv(
2127 Libcalls.getLibcallImplCallingConv(Call: SafestackPointerAddressImpl));
2128 return Call;
2129}
2130
2131Value *TargetLoweringBase::getSafeStackPointerLocation(
2132 IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const {
2133 // Android provides a libc function to retrieve the address of the current
2134 // thread's unsafe stack pointer.
2135 if (Value *Ptr = getSafeStackPointerAddressCall(IRB, Libcalls))
2136 return Ptr;
2137 return getDefaultSafeStackPointerLocation(IRB, Libcalls);
2138}
2139
2140//===----------------------------------------------------------------------===//
2141// Loop Strength Reduction hooks
2142//===----------------------------------------------------------------------===//
2143
2144/// isLegalAddressingMode - Return true if the addressing mode represented
2145/// by AM is legal for this target, for a load/store of the specified type.
2146bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
2147 const AddrMode &AM, Type *Ty,
2148 unsigned AS, Instruction *I) const {
2149 // The default implementation of this implements a conservative RISCy, r+r and
2150 // r+i addr mode.
2151
2152 // Scalable offsets not supported
2153 if (AM.ScalableOffset)
2154 return false;
2155
2156 // Allows a sign-extended 16-bit immediate field.
2157 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2158 return false;
2159
2160 // No global is ever allowed as a base.
2161 if (AM.BaseGV)
2162 return false;
2163
2164 // Only support r+r,
2165 switch (AM.Scale) {
2166 case 0: // "r+i" or just "i", depending on HasBaseReg.
2167 break;
2168 case 1:
2169 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2170 return false;
2171 // Otherwise we have r+r or r+i.
2172 break;
2173 case 2:
2174 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2175 return false;
2176 // Allow 2*r as r+r.
2177 break;
2178 default: // Don't allow n * r
2179 return false;
2180 }
2181
2182 return true;
2183}
2184
2185//===----------------------------------------------------------------------===//
2186// Stack Protector
2187//===----------------------------------------------------------------------===//
2188
2189// For OpenBSD return its special guard variable. Otherwise return nullptr,
2190// so that SelectionDAG handle SSP.
2191Value *
2192TargetLoweringBase::getIRStackGuard(IRBuilderBase &IRB,
2193 const LibcallLoweringInfo &Libcalls) const {
2194 RTLIB::LibcallImpl GuardLocalImpl =
2195 Libcalls.getLibcallImpl(Call: RTLIB::STACK_CHECK_GUARD);
2196 if (GuardLocalImpl != RTLIB::impl___guard_local)
2197 return nullptr;
2198
2199 GlobalValue *G = Libcalls.getRuntimeLibcallsInfo().getOrInsertImplGlobal(
2200 M&: *IRB.getModule(), LibcallImpl: GuardLocalImpl);
2201 G->setVisibility(GlobalValue::HiddenVisibility);
2202 return G;
2203}
2204
2205// Currently only support "standard" __stack_chk_guard.
2206// TODO: add LOAD_STACK_GUARD support.
2207void TargetLoweringBase::insertSSPDeclarations(
2208 Module &M, const LibcallLoweringInfo &Libcalls) const {
2209 RTLIB::LibcallImpl StackGuardImpl =
2210 Libcalls.getLibcallImpl(Call: RTLIB::STACK_CHECK_GUARD);
2211 if (StackGuardImpl == RTLIB::Unsupported)
2212 return;
2213
2214 StringRef StackGuardVarName = getLibcallImplName(Call: StackGuardImpl);
2215 M.getOrInsertGlobal(
2216 Name: StackGuardVarName, Ty: PointerType::getUnqual(C&: M.getContext()), CreateGlobalCallback: [=, &M]() {
2217 auto *GV = new GlobalVariable(M, PointerType::getUnqual(C&: M.getContext()),
2218 false, GlobalVariable::ExternalLinkage,
2219 nullptr, StackGuardVarName);
2220
2221 // FreeBSD has "__stack_chk_guard" defined externally on libc.so
2222 if (M.getDirectAccessExternalData() &&
2223 !M.getTargetTriple().isOSCygMing() &&
2224 !(M.getTargetTriple().isPPC64() &&
2225 M.getTargetTriple().isOSFreeBSD()) &&
2226 !M.getTargetTriple().isOSDarwin())
2227 GV->setDSOLocal(true);
2228
2229 return GV;
2230 });
2231}
2232
2233// Currently only support "standard" __stack_chk_guard.
2234// TODO: add LOAD_STACK_GUARD support.
2235Value *TargetLoweringBase::getSDagStackGuard(
2236 const Module &M, const LibcallLoweringInfo &Libcalls) const {
2237 RTLIB::LibcallImpl GuardVarImpl =
2238 Libcalls.getLibcallImpl(Call: RTLIB::STACK_CHECK_GUARD);
2239 if (GuardVarImpl == RTLIB::Unsupported)
2240 return nullptr;
2241 return M.getNamedValue(Name: getLibcallImplName(Call: GuardVarImpl));
2242}
2243
2244Function *TargetLoweringBase::getSSPStackGuardCheck(
2245 const Module &M, const LibcallLoweringInfo &Libcalls) const {
2246 // MSVC CRT has a function to validate security cookie.
2247 RTLIB::LibcallImpl SecurityCheckCookieLibcall =
2248 Libcalls.getLibcallImpl(Call: RTLIB::SECURITY_CHECK_COOKIE);
2249 if (SecurityCheckCookieLibcall != RTLIB::Unsupported)
2250 return M.getFunction(Name: getLibcallImplName(Call: SecurityCheckCookieLibcall));
2251 return nullptr;
2252}
2253
2254unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
2255 return MinimumJumpTableEntries;
2256}
2257
2258void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
2259 MinimumJumpTableEntries = Val;
2260}
2261
2262unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
2263 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
2264}
2265
2266unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
2267 return MaximumJumpTableSize;
2268}
2269
2270void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
2271 MaximumJumpTableSize = Val;
2272}
2273
2274bool TargetLoweringBase::isJumpTableRelative() const {
2275 return getTargetMachine().isPositionIndependent();
2276}
2277
2278unsigned TargetLoweringBase::getMinimumBitTestCmps() const {
2279 return MinimumBitTestCmps;
2280}
2281
2282void TargetLoweringBase::setMinimumBitTestCmps(unsigned Val) {
2283 MinimumBitTestCmps = Val;
2284}
2285
2286Align TargetLoweringBase::getPrefLoopAlignment(
2287 MachineLoop *ML, const MachineBasicBlock *BlockToAlign) const {
2288 if (TM.Options.LoopAlignment)
2289 return Align(TM.Options.LoopAlignment);
2290 return PrefLoopAlignment;
2291}
2292
2293unsigned TargetLoweringBase::getMaxPermittedBytesForAlignment(
2294 MachineBasicBlock *MBB) const {
2295 return MaxBytesForAlignment;
2296}
2297
2298//===----------------------------------------------------------------------===//
2299// Reciprocal Estimates
2300//===----------------------------------------------------------------------===//
2301
2302/// Get the reciprocal estimate attribute string for a function that will
2303/// override the target defaults.
2304static StringRef getRecipEstimateForFunc(const Function &F) {
2305 return F.getFnAttribute(Kind: "reciprocal-estimates").getValueAsString();
2306}
2307
2308/// Construct a string for the given reciprocal operation of the given type.
2309/// This string should match the corresponding option to the front-end's
2310/// "-mrecip" flag assuming those strings have been passed through in an
2311/// attribute string. For example, "vec-divf" for a division of a vXf32.
2312static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
2313 std::string Name = VT.isVector() ? "vec-" : "";
2314
2315 Name += IsSqrt ? "sqrt" : "div";
2316
2317 // TODO: Handle other float types?
2318 if (VT.getScalarType() == MVT::f64) {
2319 Name += "d";
2320 } else if (VT.getScalarType() == MVT::f16) {
2321 Name += "h";
2322 } else {
2323 assert(VT.getScalarType() == MVT::f32 &&
2324 "Unexpected FP type for reciprocal estimate");
2325 Name += "f";
2326 }
2327
2328 return Name;
2329}
2330
2331/// Return the character position and value (a single numeric character) of a
2332/// customized refinement operation in the input string if it exists. Return
2333/// false if there is no customized refinement step count.
2334static bool parseRefinementStep(StringRef In, size_t &Position,
2335 uint8_t &Value) {
2336 const char RefStepToken = ':';
2337 Position = In.find(C: RefStepToken);
2338 if (Position == StringRef::npos)
2339 return false;
2340
2341 StringRef RefStepString = In.substr(Start: Position + 1);
2342 // Allow exactly one numeric character for the additional refinement
2343 // step parameter.
2344 if (RefStepString.size() == 1) {
2345 char RefStepChar = RefStepString[0];
2346 if (isDigit(C: RefStepChar)) {
2347 Value = RefStepChar - '0';
2348 return true;
2349 }
2350 }
2351 report_fatal_error(reason: "Invalid refinement step for -recip.");
2352}
2353
2354/// For the input attribute string, return one of the ReciprocalEstimate enum
2355/// status values (enabled, disabled, or not specified) for this operation on
2356/// the specified data type.
2357static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
2358 if (Override.empty())
2359 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2360
2361 SmallVector<StringRef, 4> OverrideVector;
2362 Override.split(A&: OverrideVector, Separator: ',');
2363 unsigned NumArgs = OverrideVector.size();
2364
2365 // Check if "all", "none", or "default" was specified.
2366 if (NumArgs == 1) {
2367 // Look for an optional setting of the number of refinement steps needed
2368 // for this type of reciprocal operation.
2369 size_t RefPos;
2370 uint8_t RefSteps;
2371 if (parseRefinementStep(In: Override, Position&: RefPos, Value&: RefSteps)) {
2372 // Split the string for further processing.
2373 Override = Override.substr(Start: 0, N: RefPos);
2374 }
2375
2376 // All reciprocal types are enabled.
2377 if (Override == "all")
2378 return TargetLoweringBase::ReciprocalEstimate::Enabled;
2379
2380 // All reciprocal types are disabled.
2381 if (Override == "none")
2382 return TargetLoweringBase::ReciprocalEstimate::Disabled;
2383
2384 // Target defaults for enablement are used.
2385 if (Override == "default")
2386 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2387 }
2388
2389 // The attribute string may omit the size suffix ('f'/'d').
2390 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2391 std::string VTNameNoSize = VTName;
2392 VTNameNoSize.pop_back();
2393 static const char DisabledPrefix = '!';
2394
2395 for (StringRef RecipType : OverrideVector) {
2396 size_t RefPos;
2397 uint8_t RefSteps;
2398 if (parseRefinementStep(In: RecipType, Position&: RefPos, Value&: RefSteps))
2399 RecipType = RecipType.substr(Start: 0, N: RefPos);
2400
2401 // Ignore the disablement token for string matching.
2402 bool IsDisabled = RecipType[0] == DisabledPrefix;
2403 if (IsDisabled)
2404 RecipType = RecipType.substr(Start: 1);
2405
2406 if (RecipType == VTName || RecipType == VTNameNoSize)
2407 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
2408 : TargetLoweringBase::ReciprocalEstimate::Enabled;
2409 }
2410
2411 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2412}
2413
2414/// For the input attribute string, return the customized refinement step count
2415/// for this operation on the specified data type. If the step count does not
2416/// exist, return the ReciprocalEstimate enum value for unspecified.
2417static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2418 if (Override.empty())
2419 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2420
2421 SmallVector<StringRef, 4> OverrideVector;
2422 Override.split(A&: OverrideVector, Separator: ',');
2423 unsigned NumArgs = OverrideVector.size();
2424
2425 // Check if "all", "default", or "none" was specified.
2426 if (NumArgs == 1) {
2427 // Look for an optional setting of the number of refinement steps needed
2428 // for this type of reciprocal operation.
2429 size_t RefPos;
2430 uint8_t RefSteps;
2431 if (!parseRefinementStep(In: Override, Position&: RefPos, Value&: RefSteps))
2432 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2433
2434 // Split the string for further processing.
2435 Override = Override.substr(Start: 0, N: RefPos);
2436 assert(Override != "none" &&
2437 "Disabled reciprocals, but specifed refinement steps?");
2438
2439 // If this is a general override, return the specified number of steps.
2440 if (Override == "all" || Override == "default")
2441 return RefSteps;
2442 }
2443
2444 // The attribute string may omit the size suffix ('f'/'d').
2445 std::string VTName = getReciprocalOpName(IsSqrt, VT);
2446 std::string VTNameNoSize = VTName;
2447 VTNameNoSize.pop_back();
2448
2449 for (StringRef RecipType : OverrideVector) {
2450 size_t RefPos;
2451 uint8_t RefSteps;
2452 if (!parseRefinementStep(In: RecipType, Position&: RefPos, Value&: RefSteps))
2453 continue;
2454
2455 RecipType = RecipType.substr(Start: 0, N: RefPos);
2456 if (RecipType == VTName || RecipType == VTNameNoSize)
2457 return RefSteps;
2458 }
2459
2460 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2461}
2462
2463int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
2464 const Function &F) const {
2465 return getOpEnabled(IsSqrt: true, VT, Override: getRecipEstimateForFunc(F));
2466}
2467
2468int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
2469 const Function &F) const {
2470 return getOpEnabled(IsSqrt: false, VT, Override: getRecipEstimateForFunc(F));
2471}
2472
2473int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
2474 const Function &F) const {
2475 return getOpRefinementSteps(IsSqrt: true, VT, Override: getRecipEstimateForFunc(F));
2476}
2477
2478int TargetLoweringBase::getDivRefinementSteps(EVT VT, const Function &F) const {
2479 return getOpRefinementSteps(IsSqrt: false, VT, Override: getRecipEstimateForFunc(F));
2480}
2481
2482bool TargetLoweringBase::isLoadBitCastBeneficial(
2483 EVT LoadVT, EVT BitcastVT, const SelectionDAG &DAG,
2484 const MachineMemOperand &MMO) const {
2485 // Single-element vectors are scalarized, so we should generally avoid having
2486 // any memory operations on such types, as they would get scalarized too.
2487 if (LoadVT.isFixedLengthVector() && BitcastVT.isFixedLengthVector() &&
2488 BitcastVT.getVectorNumElements() == 1)
2489 return false;
2490
2491 // Don't do if we could do an indexed load on the original type, but not on
2492 // the new one.
2493 if (!LoadVT.isSimple() || !BitcastVT.isSimple())
2494 return true;
2495
2496 MVT LoadMVT = LoadVT.getSimpleVT();
2497
2498 // Don't bother doing this if it's just going to be promoted again later, as
2499 // doing so might interfere with other combines.
2500 if (getOperationAction(Op: ISD::LOAD, VT: LoadMVT) == Promote &&
2501 getTypeToPromoteTo(Op: ISD::LOAD, VT: LoadMVT) == BitcastVT.getSimpleVT())
2502 return false;
2503
2504 unsigned Fast = 0;
2505 return allowsMemoryAccess(Context&: *DAG.getContext(), DL: DAG.getDataLayout(), VT: BitcastVT,
2506 MMO, Fast: &Fast) &&
2507 Fast;
2508}
2509
2510void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
2511 MF.getRegInfo().freezeReservedRegs();
2512}
2513
2514MachineMemOperand::Flags TargetLoweringBase::getLoadMemOperandFlags(
2515 const LoadInst &LI, const DataLayout &DL, AssumptionCache *AC,
2516 const TargetLibraryInfo *LibInfo, CodeGenOptLevel OptLevel) const {
2517 MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad;
2518 if (LI.isVolatile())
2519 Flags |= MachineMemOperand::MOVolatile;
2520
2521 if (LI.hasMetadata(KindID: LLVMContext::MD_nontemporal))
2522 Flags |= MachineMemOperand::MONonTemporal;
2523
2524 if (LI.hasMetadata(KindID: LLVMContext::MD_invariant_load))
2525 Flags |= MachineMemOperand::MOInvariant;
2526
2527 // Dereferenceability analysis is expensive, skip at O0.
2528 if (OptLevel != CodeGenOptLevel::None &&
2529 isDereferenceableAndAlignedPointer(
2530 V: LI.getPointerOperand(), Ty: LI.getType(), Alignment: LI.getAlign(),
2531 Q: SimplifyQuery(DL, LibInfo, /*DT=*/nullptr, AC, &LI)))
2532 Flags |= MachineMemOperand::MODereferenceable;
2533
2534 Flags |= getTargetMMOFlags(I: LI);
2535 return Flags;
2536}
2537
2538MachineMemOperand::Flags
2539TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI,
2540 const DataLayout &DL) const {
2541 MachineMemOperand::Flags Flags = MachineMemOperand::MOStore;
2542
2543 if (SI.isVolatile())
2544 Flags |= MachineMemOperand::MOVolatile;
2545
2546 if (SI.hasMetadata(KindID: LLVMContext::MD_nontemporal))
2547 Flags |= MachineMemOperand::MONonTemporal;
2548
2549 // FIXME: Not preserving dereferenceable
2550 Flags |= getTargetMMOFlags(I: SI);
2551 return Flags;
2552}
2553
2554MachineMemOperand::Flags
2555TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI,
2556 const DataLayout &DL) const {
2557 auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
2558
2559 if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: &AI)) {
2560 if (RMW->isVolatile())
2561 Flags |= MachineMemOperand::MOVolatile;
2562 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: &AI)) {
2563 if (CmpX->isVolatile())
2564 Flags |= MachineMemOperand::MOVolatile;
2565 } else
2566 llvm_unreachable("not an atomic instruction");
2567
2568 // FIXME: Not preserving dereferenceable
2569 Flags |= getTargetMMOFlags(I: AI);
2570 return Flags;
2571}
2572
2573MachineMemOperand::Flags TargetLoweringBase::getVPIntrinsicMemOperandFlags(
2574 const VPIntrinsic &VPIntrin) const {
2575 MachineMemOperand::Flags Flags = MachineMemOperand::MONone;
2576 Intrinsic::ID IntrinID = VPIntrin.getIntrinsicID();
2577
2578 switch (IntrinID) {
2579 default:
2580 llvm_unreachable("unexpected intrinsic. Existing code may be appropriate "
2581 "for it, but support must be explicitly enabled");
2582 case Intrinsic::vp_load:
2583 case Intrinsic::vp_gather:
2584 case Intrinsic::experimental_vp_strided_load:
2585 Flags = MachineMemOperand::MOLoad;
2586 break;
2587 case Intrinsic::vp_store:
2588 case Intrinsic::vp_scatter:
2589 case Intrinsic::experimental_vp_strided_store:
2590 Flags = MachineMemOperand::MOStore;
2591 break;
2592 }
2593
2594 if (VPIntrin.hasMetadata(KindID: LLVMContext::MD_nontemporal))
2595 Flags |= MachineMemOperand::MONonTemporal;
2596
2597 Flags |= getTargetMMOFlags(I: VPIntrin);
2598 return Flags;
2599}
2600
2601Instruction *TargetLoweringBase::emitLeadingFence(IRBuilderBase &Builder,
2602 Instruction *Inst,
2603 AtomicOrdering Ord) const {
2604 if (isReleaseOrStronger(AO: Ord) && Inst->hasAtomicStore())
2605 return Builder.CreateFence(Ordering: Ord);
2606 else
2607 return nullptr;
2608}
2609
2610Instruction *TargetLoweringBase::emitTrailingFence(IRBuilderBase &Builder,
2611 Instruction *Inst,
2612 AtomicOrdering Ord) const {
2613 if (isAcquireOrStronger(AO: Ord))
2614 return Builder.CreateFence(Ordering: Ord);
2615 else
2616 return nullptr;
2617}
2618
2619//===----------------------------------------------------------------------===//
2620// GlobalISel Hooks
2621//===----------------------------------------------------------------------===//
2622
2623bool TargetLoweringBase::shouldLocalize(const MachineInstr &MI,
2624 const TargetTransformInfo *TTI) const {
2625 auto &MF = *MI.getMF();
2626 auto &MRI = MF.getRegInfo();
2627 // Assuming a spill and reload of a value has a cost of 1 instruction each,
2628 // this helper function computes the maximum number of uses we should consider
2629 // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2630 // break even in terms of code size when the original MI has 2 users vs
2631 // choosing to potentially spill. Any more than 2 users we we have a net code
2632 // size increase. This doesn't take into account register pressure though.
2633 auto maxUses = [](unsigned RematCost) {
2634 // A cost of 1 means remats are basically free.
2635 if (RematCost == 1)
2636 return std::numeric_limits<unsigned>::max();
2637 if (RematCost == 2)
2638 return 2U;
2639
2640 // Remat is too expensive, only sink if there's one user.
2641 if (RematCost > 2)
2642 return 1U;
2643 llvm_unreachable("Unexpected remat cost");
2644 };
2645
2646 switch (MI.getOpcode()) {
2647 default:
2648 return false;
2649 // Constants-like instructions should be close to their users.
2650 // We don't want long live-ranges for them.
2651 case TargetOpcode::G_CONSTANT:
2652 case TargetOpcode::G_FCONSTANT:
2653 case TargetOpcode::G_FRAME_INDEX:
2654 case TargetOpcode::G_INTTOPTR:
2655 return true;
2656 case TargetOpcode::G_GLOBAL_VALUE: {
2657 unsigned RematCost = TTI->getGISelRematGlobalCost();
2658 Register Reg = MI.getOperand(i: 0).getReg();
2659 unsigned MaxUses = maxUses(RematCost);
2660 if (MaxUses == UINT_MAX)
2661 return true; // Remats are "free" so always localize.
2662 return MRI.hasAtMostUserInstrs(Reg, MaxUsers: MaxUses);
2663 }
2664 }
2665}
2666