1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
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
7 //===----------------------------------------------------------------------===//
9 // This implements the TargetLoweringBase class.
11 //===----------------------------------------------------------------------===//
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/Triple.h"
19 #include "llvm/ADT/Twine.h"
20 #include "llvm/CodeGen/Analysis.h"
21 #include "llvm/CodeGen/ISDOpcodes.h"
22 #include "llvm/CodeGen/MachineBasicBlock.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstr.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineOperand.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/RuntimeLibcalls.h"
31 #include "llvm/CodeGen/StackMaps.h"
32 #include "llvm/CodeGen/TargetLowering.h"
33 #include "llvm/CodeGen/TargetOpcodes.h"
34 #include "llvm/CodeGen/TargetRegisterInfo.h"
35 #include "llvm/CodeGen/ValueTypes.h"
36 #include "llvm/IR/Attributes.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalValue.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/IRBuilder.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/IR/Type.h"
46 #include "llvm/Support/BranchProbability.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Compiler.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MachineValueType.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetMachine.h"
66 static cl::opt
<bool> JumpIsExpensiveOverride(
67 "jump-is-expensive", cl::init(false),
68 cl::desc("Do not create extra branches to split comparison logic."),
71 static cl::opt
<unsigned> MinimumJumpTableEntries
72 ("min-jump-table-entries", cl::init(4), cl::Hidden
,
73 cl::desc("Set minimum number of entries to use a jump table."));
75 static cl::opt
<unsigned> MaximumJumpTableSize
76 ("max-jump-table-size", cl::init(UINT_MAX
), cl::Hidden
,
77 cl::desc("Set maximum size of jump tables."));
79 /// Minimum jump table density for normal functions.
80 static cl::opt
<unsigned>
81 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden
,
82 cl::desc("Minimum density for building a jump table in "
83 "a normal function"));
85 /// Minimum jump table density for -Os or -Oz functions.
86 static cl::opt
<unsigned> OptsizeJumpTableDensity(
87 "optsize-jump-table-density", cl::init(40), cl::Hidden
,
88 cl::desc("Minimum density for building a jump table in "
89 "an optsize function"));
91 static bool darwinHasSinCos(const Triple
&TT
) {
92 assert(TT
.isOSDarwin() && "should be called with darwin triple");
93 // Don't bother with 32 bit x86.
94 if (TT
.getArch() == Triple::x86
)
96 // Macos < 10.9 has no sincos_stret.
98 return !TT
.isMacOSXVersionLT(10, 9) && TT
.isArch64Bit();
99 // iOS < 7.0 has no sincos_stret.
101 return !TT
.isOSVersionLT(7, 0);
102 // Any other darwin such as WatchOS/TvOS is new enough.
106 // Although this default value is arbitrary, it is not random. It is assumed
107 // that a condition that evaluates the same way by a higher percentage than this
108 // is best represented as control flow. Therefore, the default value N should be
109 // set such that the win from N% correct executions is greater than the loss
110 // from (100 - N)% mispredicted executions for the majority of intended targets.
111 static cl::opt
<int> MinPercentageForPredictableBranch(
112 "min-predictable-branch", cl::init(99),
113 cl::desc("Minimum percentage (0-100) that a condition must be either true "
114 "or false to assume that the condition is predictable"),
117 void TargetLoweringBase::InitLibcalls(const Triple
&TT
) {
118 #define HANDLE_LIBCALL(code, name) \
119 setLibcallName(RTLIB::code, name);
120 #include "llvm/IR/RuntimeLibcalls.def"
121 #undef HANDLE_LIBCALL
122 // Initialize calling conventions to their default.
123 for (int LC
= 0; LC
< RTLIB::UNKNOWN_LIBCALL
; ++LC
)
124 setLibcallCallingConv((RTLIB::Libcall
)LC
, CallingConv::C
);
126 // A few names are different on particular architectures or environments.
127 if (TT
.isOSDarwin()) {
128 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
129 // of the gnueabi-style __gnu_*_ieee.
130 // FIXME: What about other targets?
131 setLibcallName(RTLIB::FPEXT_F16_F32
, "__extendhfsf2");
132 setLibcallName(RTLIB::FPROUND_F32_F16
, "__truncsfhf2");
134 // Some darwins have an optimized __bzero/bzero function.
135 switch (TT
.getArch()) {
138 if (TT
.isMacOSX() && !TT
.isMacOSXVersionLT(10, 6))
139 setLibcallName(RTLIB::BZERO
, "__bzero");
141 case Triple::aarch64
:
142 setLibcallName(RTLIB::BZERO
, "bzero");
148 if (darwinHasSinCos(TT
)) {
149 setLibcallName(RTLIB::SINCOS_STRET_F32
, "__sincosf_stret");
150 setLibcallName(RTLIB::SINCOS_STRET_F64
, "__sincos_stret");
151 if (TT
.isWatchABI()) {
152 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32
,
153 CallingConv::ARM_AAPCS_VFP
);
154 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64
,
155 CallingConv::ARM_AAPCS_VFP
);
159 setLibcallName(RTLIB::FPEXT_F16_F32
, "__gnu_h2f_ieee");
160 setLibcallName(RTLIB::FPROUND_F32_F16
, "__gnu_f2h_ieee");
163 if (TT
.isGNUEnvironment() || TT
.isOSFuchsia() ||
164 (TT
.isAndroid() && !TT
.isAndroidVersionLT(9))) {
165 setLibcallName(RTLIB::SINCOS_F32
, "sincosf");
166 setLibcallName(RTLIB::SINCOS_F64
, "sincos");
167 setLibcallName(RTLIB::SINCOS_F80
, "sincosl");
168 setLibcallName(RTLIB::SINCOS_F128
, "sincosl");
169 setLibcallName(RTLIB::SINCOS_PPCF128
, "sincosl");
172 if (TT
.isOSOpenBSD()) {
173 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL
, nullptr);
177 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
178 /// UNKNOWN_LIBCALL if there is none.
179 RTLIB::Libcall
RTLIB::getFPEXT(EVT OpVT
, EVT RetVT
) {
180 if (OpVT
== MVT::f16
) {
181 if (RetVT
== MVT::f32
)
182 return FPEXT_F16_F32
;
183 } else if (OpVT
== MVT::f32
) {
184 if (RetVT
== MVT::f64
)
185 return FPEXT_F32_F64
;
186 if (RetVT
== MVT::f128
)
187 return FPEXT_F32_F128
;
188 if (RetVT
== MVT::ppcf128
)
189 return FPEXT_F32_PPCF128
;
190 } else if (OpVT
== MVT::f64
) {
191 if (RetVT
== MVT::f128
)
192 return FPEXT_F64_F128
;
193 else if (RetVT
== MVT::ppcf128
)
194 return FPEXT_F64_PPCF128
;
195 } else if (OpVT
== MVT::f80
) {
196 if (RetVT
== MVT::f128
)
197 return FPEXT_F80_F128
;
200 return UNKNOWN_LIBCALL
;
203 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
204 /// UNKNOWN_LIBCALL if there is none.
205 RTLIB::Libcall
RTLIB::getFPROUND(EVT OpVT
, EVT RetVT
) {
206 if (RetVT
== MVT::f16
) {
207 if (OpVT
== MVT::f32
)
208 return FPROUND_F32_F16
;
209 if (OpVT
== MVT::f64
)
210 return FPROUND_F64_F16
;
211 if (OpVT
== MVT::f80
)
212 return FPROUND_F80_F16
;
213 if (OpVT
== MVT::f128
)
214 return FPROUND_F128_F16
;
215 if (OpVT
== MVT::ppcf128
)
216 return FPROUND_PPCF128_F16
;
217 } else if (RetVT
== MVT::f32
) {
218 if (OpVT
== MVT::f64
)
219 return FPROUND_F64_F32
;
220 if (OpVT
== MVT::f80
)
221 return FPROUND_F80_F32
;
222 if (OpVT
== MVT::f128
)
223 return FPROUND_F128_F32
;
224 if (OpVT
== MVT::ppcf128
)
225 return FPROUND_PPCF128_F32
;
226 } else if (RetVT
== MVT::f64
) {
227 if (OpVT
== MVT::f80
)
228 return FPROUND_F80_F64
;
229 if (OpVT
== MVT::f128
)
230 return FPROUND_F128_F64
;
231 if (OpVT
== MVT::ppcf128
)
232 return FPROUND_PPCF128_F64
;
233 } else if (RetVT
== MVT::f80
) {
234 if (OpVT
== MVT::f128
)
235 return FPROUND_F128_F80
;
238 return UNKNOWN_LIBCALL
;
241 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
242 /// UNKNOWN_LIBCALL if there is none.
243 RTLIB::Libcall
RTLIB::getFPTOSINT(EVT OpVT
, EVT RetVT
) {
244 if (OpVT
== MVT::f32
) {
245 if (RetVT
== MVT::i32
)
246 return FPTOSINT_F32_I32
;
247 if (RetVT
== MVT::i64
)
248 return FPTOSINT_F32_I64
;
249 if (RetVT
== MVT::i128
)
250 return FPTOSINT_F32_I128
;
251 } else if (OpVT
== MVT::f64
) {
252 if (RetVT
== MVT::i32
)
253 return FPTOSINT_F64_I32
;
254 if (RetVT
== MVT::i64
)
255 return FPTOSINT_F64_I64
;
256 if (RetVT
== MVT::i128
)
257 return FPTOSINT_F64_I128
;
258 } else if (OpVT
== MVT::f80
) {
259 if (RetVT
== MVT::i32
)
260 return FPTOSINT_F80_I32
;
261 if (RetVT
== MVT::i64
)
262 return FPTOSINT_F80_I64
;
263 if (RetVT
== MVT::i128
)
264 return FPTOSINT_F80_I128
;
265 } else if (OpVT
== MVT::f128
) {
266 if (RetVT
== MVT::i32
)
267 return FPTOSINT_F128_I32
;
268 if (RetVT
== MVT::i64
)
269 return FPTOSINT_F128_I64
;
270 if (RetVT
== MVT::i128
)
271 return FPTOSINT_F128_I128
;
272 } else if (OpVT
== MVT::ppcf128
) {
273 if (RetVT
== MVT::i32
)
274 return FPTOSINT_PPCF128_I32
;
275 if (RetVT
== MVT::i64
)
276 return FPTOSINT_PPCF128_I64
;
277 if (RetVT
== MVT::i128
)
278 return FPTOSINT_PPCF128_I128
;
280 return UNKNOWN_LIBCALL
;
283 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
284 /// UNKNOWN_LIBCALL if there is none.
285 RTLIB::Libcall
RTLIB::getFPTOUINT(EVT OpVT
, EVT RetVT
) {
286 if (OpVT
== MVT::f32
) {
287 if (RetVT
== MVT::i32
)
288 return FPTOUINT_F32_I32
;
289 if (RetVT
== MVT::i64
)
290 return FPTOUINT_F32_I64
;
291 if (RetVT
== MVT::i128
)
292 return FPTOUINT_F32_I128
;
293 } else if (OpVT
== MVT::f64
) {
294 if (RetVT
== MVT::i32
)
295 return FPTOUINT_F64_I32
;
296 if (RetVT
== MVT::i64
)
297 return FPTOUINT_F64_I64
;
298 if (RetVT
== MVT::i128
)
299 return FPTOUINT_F64_I128
;
300 } else if (OpVT
== MVT::f80
) {
301 if (RetVT
== MVT::i32
)
302 return FPTOUINT_F80_I32
;
303 if (RetVT
== MVT::i64
)
304 return FPTOUINT_F80_I64
;
305 if (RetVT
== MVT::i128
)
306 return FPTOUINT_F80_I128
;
307 } else if (OpVT
== MVT::f128
) {
308 if (RetVT
== MVT::i32
)
309 return FPTOUINT_F128_I32
;
310 if (RetVT
== MVT::i64
)
311 return FPTOUINT_F128_I64
;
312 if (RetVT
== MVT::i128
)
313 return FPTOUINT_F128_I128
;
314 } else if (OpVT
== MVT::ppcf128
) {
315 if (RetVT
== MVT::i32
)
316 return FPTOUINT_PPCF128_I32
;
317 if (RetVT
== MVT::i64
)
318 return FPTOUINT_PPCF128_I64
;
319 if (RetVT
== MVT::i128
)
320 return FPTOUINT_PPCF128_I128
;
322 return UNKNOWN_LIBCALL
;
325 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
326 /// UNKNOWN_LIBCALL if there is none.
327 RTLIB::Libcall
RTLIB::getSINTTOFP(EVT OpVT
, EVT RetVT
) {
328 if (OpVT
== MVT::i32
) {
329 if (RetVT
== MVT::f32
)
330 return SINTTOFP_I32_F32
;
331 if (RetVT
== MVT::f64
)
332 return SINTTOFP_I32_F64
;
333 if (RetVT
== MVT::f80
)
334 return SINTTOFP_I32_F80
;
335 if (RetVT
== MVT::f128
)
336 return SINTTOFP_I32_F128
;
337 if (RetVT
== MVT::ppcf128
)
338 return SINTTOFP_I32_PPCF128
;
339 } else if (OpVT
== MVT::i64
) {
340 if (RetVT
== MVT::f32
)
341 return SINTTOFP_I64_F32
;
342 if (RetVT
== MVT::f64
)
343 return SINTTOFP_I64_F64
;
344 if (RetVT
== MVT::f80
)
345 return SINTTOFP_I64_F80
;
346 if (RetVT
== MVT::f128
)
347 return SINTTOFP_I64_F128
;
348 if (RetVT
== MVT::ppcf128
)
349 return SINTTOFP_I64_PPCF128
;
350 } else if (OpVT
== MVT::i128
) {
351 if (RetVT
== MVT::f32
)
352 return SINTTOFP_I128_F32
;
353 if (RetVT
== MVT::f64
)
354 return SINTTOFP_I128_F64
;
355 if (RetVT
== MVT::f80
)
356 return SINTTOFP_I128_F80
;
357 if (RetVT
== MVT::f128
)
358 return SINTTOFP_I128_F128
;
359 if (RetVT
== MVT::ppcf128
)
360 return SINTTOFP_I128_PPCF128
;
362 return UNKNOWN_LIBCALL
;
365 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
366 /// UNKNOWN_LIBCALL if there is none.
367 RTLIB::Libcall
RTLIB::getUINTTOFP(EVT OpVT
, EVT RetVT
) {
368 if (OpVT
== MVT::i32
) {
369 if (RetVT
== MVT::f32
)
370 return UINTTOFP_I32_F32
;
371 if (RetVT
== MVT::f64
)
372 return UINTTOFP_I32_F64
;
373 if (RetVT
== MVT::f80
)
374 return UINTTOFP_I32_F80
;
375 if (RetVT
== MVT::f128
)
376 return UINTTOFP_I32_F128
;
377 if (RetVT
== MVT::ppcf128
)
378 return UINTTOFP_I32_PPCF128
;
379 } else if (OpVT
== MVT::i64
) {
380 if (RetVT
== MVT::f32
)
381 return UINTTOFP_I64_F32
;
382 if (RetVT
== MVT::f64
)
383 return UINTTOFP_I64_F64
;
384 if (RetVT
== MVT::f80
)
385 return UINTTOFP_I64_F80
;
386 if (RetVT
== MVT::f128
)
387 return UINTTOFP_I64_F128
;
388 if (RetVT
== MVT::ppcf128
)
389 return UINTTOFP_I64_PPCF128
;
390 } else if (OpVT
== MVT::i128
) {
391 if (RetVT
== MVT::f32
)
392 return UINTTOFP_I128_F32
;
393 if (RetVT
== MVT::f64
)
394 return UINTTOFP_I128_F64
;
395 if (RetVT
== MVT::f80
)
396 return UINTTOFP_I128_F80
;
397 if (RetVT
== MVT::f128
)
398 return UINTTOFP_I128_F128
;
399 if (RetVT
== MVT::ppcf128
)
400 return UINTTOFP_I128_PPCF128
;
402 return UNKNOWN_LIBCALL
;
405 RTLIB::Libcall
RTLIB::getSYNC(unsigned Opc
, MVT VT
) {
406 #define OP_TO_LIBCALL(Name, Enum) \
408 switch (VT.SimpleTy) { \
410 return UNKNOWN_LIBCALL; \
424 OP_TO_LIBCALL(ISD::ATOMIC_SWAP
, SYNC_LOCK_TEST_AND_SET
)
425 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP
, SYNC_VAL_COMPARE_AND_SWAP
)
426 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD
, SYNC_FETCH_AND_ADD
)
427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB
, SYNC_FETCH_AND_SUB
)
428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND
, SYNC_FETCH_AND_AND
)
429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR
, SYNC_FETCH_AND_OR
)
430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR
, SYNC_FETCH_AND_XOR
)
431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND
, SYNC_FETCH_AND_NAND
)
432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX
, SYNC_FETCH_AND_MAX
)
433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX
, SYNC_FETCH_AND_UMAX
)
434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN
, SYNC_FETCH_AND_MIN
)
435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN
, SYNC_FETCH_AND_UMIN
)
440 return UNKNOWN_LIBCALL
;
443 RTLIB::Libcall
RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize
) {
444 switch (ElementSize
) {
446 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1
;
448 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2
;
450 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4
;
452 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8
;
454 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16
;
456 return UNKNOWN_LIBCALL
;
460 RTLIB::Libcall
RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize
) {
461 switch (ElementSize
) {
463 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1
;
465 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2
;
467 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4
;
469 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8
;
471 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16
;
473 return UNKNOWN_LIBCALL
;
477 RTLIB::Libcall
RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize
) {
478 switch (ElementSize
) {
480 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1
;
482 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2
;
484 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4
;
486 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8
;
488 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16
;
490 return UNKNOWN_LIBCALL
;
494 /// InitCmpLibcallCCs - Set default comparison libcall CC.
495 static void InitCmpLibcallCCs(ISD::CondCode
*CCs
) {
496 memset(CCs
, ISD::SETCC_INVALID
, sizeof(ISD::CondCode
)*RTLIB::UNKNOWN_LIBCALL
);
497 CCs
[RTLIB::OEQ_F32
] = ISD::SETEQ
;
498 CCs
[RTLIB::OEQ_F64
] = ISD::SETEQ
;
499 CCs
[RTLIB::OEQ_F128
] = ISD::SETEQ
;
500 CCs
[RTLIB::OEQ_PPCF128
] = ISD::SETEQ
;
501 CCs
[RTLIB::UNE_F32
] = ISD::SETNE
;
502 CCs
[RTLIB::UNE_F64
] = ISD::SETNE
;
503 CCs
[RTLIB::UNE_F128
] = ISD::SETNE
;
504 CCs
[RTLIB::UNE_PPCF128
] = ISD::SETNE
;
505 CCs
[RTLIB::OGE_F32
] = ISD::SETGE
;
506 CCs
[RTLIB::OGE_F64
] = ISD::SETGE
;
507 CCs
[RTLIB::OGE_F128
] = ISD::SETGE
;
508 CCs
[RTLIB::OGE_PPCF128
] = ISD::SETGE
;
509 CCs
[RTLIB::OLT_F32
] = ISD::SETLT
;
510 CCs
[RTLIB::OLT_F64
] = ISD::SETLT
;
511 CCs
[RTLIB::OLT_F128
] = ISD::SETLT
;
512 CCs
[RTLIB::OLT_PPCF128
] = ISD::SETLT
;
513 CCs
[RTLIB::OLE_F32
] = ISD::SETLE
;
514 CCs
[RTLIB::OLE_F64
] = ISD::SETLE
;
515 CCs
[RTLIB::OLE_F128
] = ISD::SETLE
;
516 CCs
[RTLIB::OLE_PPCF128
] = ISD::SETLE
;
517 CCs
[RTLIB::OGT_F32
] = ISD::SETGT
;
518 CCs
[RTLIB::OGT_F64
] = ISD::SETGT
;
519 CCs
[RTLIB::OGT_F128
] = ISD::SETGT
;
520 CCs
[RTLIB::OGT_PPCF128
] = ISD::SETGT
;
521 CCs
[RTLIB::UO_F32
] = ISD::SETNE
;
522 CCs
[RTLIB::UO_F64
] = ISD::SETNE
;
523 CCs
[RTLIB::UO_F128
] = ISD::SETNE
;
524 CCs
[RTLIB::UO_PPCF128
] = ISD::SETNE
;
525 CCs
[RTLIB::O_F32
] = ISD::SETEQ
;
526 CCs
[RTLIB::O_F64
] = ISD::SETEQ
;
527 CCs
[RTLIB::O_F128
] = ISD::SETEQ
;
528 CCs
[RTLIB::O_PPCF128
] = ISD::SETEQ
;
531 /// NOTE: The TargetMachine owns TLOF.
532 TargetLoweringBase::TargetLoweringBase(const TargetMachine
&tm
) : TM(tm
) {
535 // Perform these initializations only once.
536 MaxStoresPerMemset
= MaxStoresPerMemcpy
= MaxStoresPerMemmove
=
537 MaxLoadsPerMemcmp
= 8;
538 MaxGluedStoresPerMemcpy
= 0;
539 MaxStoresPerMemsetOptSize
= MaxStoresPerMemcpyOptSize
=
540 MaxStoresPerMemmoveOptSize
= MaxLoadsPerMemcmpOptSize
= 4;
541 UseUnderscoreSetJmp
= false;
542 UseUnderscoreLongJmp
= false;
543 HasMultipleConditionRegisters
= false;
544 HasExtractBitsInsn
= false;
545 JumpIsExpensive
= JumpIsExpensiveOverride
;
546 PredictableSelectIsExpensive
= false;
547 EnableExtLdPromotion
= false;
548 StackPointerRegisterToSaveRestore
= 0;
549 BooleanContents
= UndefinedBooleanContent
;
550 BooleanFloatContents
= UndefinedBooleanContent
;
551 BooleanVectorContents
= UndefinedBooleanContent
;
552 SchedPreferenceInfo
= Sched::ILP
;
554 JumpBufAlignment
= 0;
555 MinFunctionAlignment
= 0;
556 PrefFunctionAlignment
= 0;
557 PrefLoopAlignment
= 0;
558 GatherAllAliasesMaxDepth
= 18;
559 MinStackArgumentAlignment
= 1;
560 // TODO: the default will be switched to 0 in the next commit, along
561 // with the Target-specific changes necessary.
562 MaxAtomicSizeInBitsSupported
= 1024;
564 MinCmpXchgSizeInBits
= 0;
565 SupportsUnalignedAtomics
= false;
567 std::fill(std::begin(LibcallRoutineNames
), std::end(LibcallRoutineNames
), nullptr);
569 InitLibcalls(TM
.getTargetTriple());
570 InitCmpLibcallCCs(CmpLibcallCCs
);
573 void TargetLoweringBase::initActions() {
574 // All operations default to being supported.
575 memset(OpActions
, 0, sizeof(OpActions
));
576 memset(LoadExtActions
, 0, sizeof(LoadExtActions
));
577 memset(TruncStoreActions
, 0, sizeof(TruncStoreActions
));
578 memset(IndexedModeActions
, 0, sizeof(IndexedModeActions
));
579 memset(CondCodeActions
, 0, sizeof(CondCodeActions
));
580 std::fill(std::begin(RegClassForVT
), std::end(RegClassForVT
), nullptr);
581 std::fill(std::begin(TargetDAGCombineArray
),
582 std::end(TargetDAGCombineArray
), 0);
584 for (MVT VT
: MVT::fp_valuetypes()) {
585 MVT IntVT
= MVT::getIntegerVT(VT
.getSizeInBits());
586 if (IntVT
.isValid()) {
587 setOperationAction(ISD::ATOMIC_SWAP
, VT
, Promote
);
588 AddPromotedToType(ISD::ATOMIC_SWAP
, VT
, IntVT
);
592 // Set default actions for various operations.
593 for (MVT VT
: MVT::all_valuetypes()) {
594 // Default all indexed load / store to expand.
595 for (unsigned IM
= (unsigned)ISD::PRE_INC
;
596 IM
!= (unsigned)ISD::LAST_INDEXED_MODE
; ++IM
) {
597 setIndexedLoadAction(IM
, VT
, Expand
);
598 setIndexedStoreAction(IM
, VT
, Expand
);
601 // Most backends expect to see the node which just returns the value loaded.
602 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS
, VT
, Expand
);
604 // These operations default to expand.
605 setOperationAction(ISD::FGETSIGN
, VT
, Expand
);
606 setOperationAction(ISD::CONCAT_VECTORS
, VT
, Expand
);
607 setOperationAction(ISD::FMINNUM
, VT
, Expand
);
608 setOperationAction(ISD::FMAXNUM
, VT
, Expand
);
609 setOperationAction(ISD::FMINNUM_IEEE
, VT
, Expand
);
610 setOperationAction(ISD::FMAXNUM_IEEE
, VT
, Expand
);
611 setOperationAction(ISD::FMINIMUM
, VT
, Expand
);
612 setOperationAction(ISD::FMAXIMUM
, VT
, Expand
);
613 setOperationAction(ISD::FMAD
, VT
, Expand
);
614 setOperationAction(ISD::SMIN
, VT
, Expand
);
615 setOperationAction(ISD::SMAX
, VT
, Expand
);
616 setOperationAction(ISD::UMIN
, VT
, Expand
);
617 setOperationAction(ISD::UMAX
, VT
, Expand
);
618 setOperationAction(ISD::ABS
, VT
, Expand
);
619 setOperationAction(ISD::FSHL
, VT
, Expand
);
620 setOperationAction(ISD::FSHR
, VT
, Expand
);
621 setOperationAction(ISD::SADDSAT
, VT
, Expand
);
622 setOperationAction(ISD::UADDSAT
, VT
, Expand
);
623 setOperationAction(ISD::SSUBSAT
, VT
, Expand
);
624 setOperationAction(ISD::USUBSAT
, VT
, Expand
);
625 setOperationAction(ISD::SMULFIX
, VT
, Expand
);
626 setOperationAction(ISD::UMULFIX
, VT
, Expand
);
628 // Overflow operations default to expand
629 setOperationAction(ISD::SADDO
, VT
, Expand
);
630 setOperationAction(ISD::SSUBO
, VT
, Expand
);
631 setOperationAction(ISD::UADDO
, VT
, Expand
);
632 setOperationAction(ISD::USUBO
, VT
, Expand
);
633 setOperationAction(ISD::SMULO
, VT
, Expand
);
634 setOperationAction(ISD::UMULO
, VT
, Expand
);
636 // ADDCARRY operations default to expand
637 setOperationAction(ISD::ADDCARRY
, VT
, Expand
);
638 setOperationAction(ISD::SUBCARRY
, VT
, Expand
);
639 setOperationAction(ISD::SETCCCARRY
, VT
, Expand
);
641 // ADDC/ADDE/SUBC/SUBE default to expand.
642 setOperationAction(ISD::ADDC
, VT
, Expand
);
643 setOperationAction(ISD::ADDE
, VT
, Expand
);
644 setOperationAction(ISD::SUBC
, VT
, Expand
);
645 setOperationAction(ISD::SUBE
, VT
, Expand
);
647 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
648 setOperationAction(ISD::CTLZ_ZERO_UNDEF
, VT
, Expand
);
649 setOperationAction(ISD::CTTZ_ZERO_UNDEF
, VT
, Expand
);
651 setOperationAction(ISD::BITREVERSE
, VT
, Expand
);
653 // These library functions default to expand.
654 setOperationAction(ISD::FROUND
, VT
, Expand
);
655 setOperationAction(ISD::FPOWI
, VT
, Expand
);
657 // These operations default to expand for vector types.
659 setOperationAction(ISD::FCOPYSIGN
, VT
, Expand
);
660 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG
, VT
, Expand
);
661 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG
, VT
, Expand
);
662 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG
, VT
, Expand
);
665 // For most targets @llvm.get.dynamic.area.offset just returns 0.
666 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET
, VT
, Expand
);
668 // Vector reduction default to expand.
669 setOperationAction(ISD::VECREDUCE_FADD
, VT
, Expand
);
670 setOperationAction(ISD::VECREDUCE_FMUL
, VT
, Expand
);
671 setOperationAction(ISD::VECREDUCE_ADD
, VT
, Expand
);
672 setOperationAction(ISD::VECREDUCE_MUL
, VT
, Expand
);
673 setOperationAction(ISD::VECREDUCE_AND
, VT
, Expand
);
674 setOperationAction(ISD::VECREDUCE_OR
, VT
, Expand
);
675 setOperationAction(ISD::VECREDUCE_XOR
, VT
, Expand
);
676 setOperationAction(ISD::VECREDUCE_SMAX
, VT
, Expand
);
677 setOperationAction(ISD::VECREDUCE_SMIN
, VT
, Expand
);
678 setOperationAction(ISD::VECREDUCE_UMAX
, VT
, Expand
);
679 setOperationAction(ISD::VECREDUCE_UMIN
, VT
, Expand
);
680 setOperationAction(ISD::VECREDUCE_FMAX
, VT
, Expand
);
681 setOperationAction(ISD::VECREDUCE_FMIN
, VT
, Expand
);
684 // Most targets ignore the @llvm.prefetch intrinsic.
685 setOperationAction(ISD::PREFETCH
, MVT::Other
, Expand
);
687 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
688 setOperationAction(ISD::READCYCLECOUNTER
, MVT::i64
, Expand
);
690 // ConstantFP nodes default to expand. Targets can either change this to
691 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
692 // to optimize expansions for certain constants.
693 setOperationAction(ISD::ConstantFP
, MVT::f16
, Expand
);
694 setOperationAction(ISD::ConstantFP
, MVT::f32
, Expand
);
695 setOperationAction(ISD::ConstantFP
, MVT::f64
, Expand
);
696 setOperationAction(ISD::ConstantFP
, MVT::f80
, Expand
);
697 setOperationAction(ISD::ConstantFP
, MVT::f128
, Expand
);
699 // These library functions default to expand.
700 for (MVT VT
: {MVT::f32
, MVT::f64
, MVT::f128
}) {
701 setOperationAction(ISD::FCBRT
, VT
, Expand
);
702 setOperationAction(ISD::FLOG
, VT
, Expand
);
703 setOperationAction(ISD::FLOG2
, VT
, Expand
);
704 setOperationAction(ISD::FLOG10
, VT
, Expand
);
705 setOperationAction(ISD::FEXP
, VT
, Expand
);
706 setOperationAction(ISD::FEXP2
, VT
, Expand
);
707 setOperationAction(ISD::FFLOOR
, VT
, Expand
);
708 setOperationAction(ISD::FNEARBYINT
, VT
, Expand
);
709 setOperationAction(ISD::FCEIL
, VT
, Expand
);
710 setOperationAction(ISD::FRINT
, VT
, Expand
);
711 setOperationAction(ISD::FTRUNC
, VT
, Expand
);
712 setOperationAction(ISD::FROUND
, VT
, Expand
);
715 // Default ISD::TRAP to expand (which turns it into abort).
716 setOperationAction(ISD::TRAP
, MVT::Other
, Expand
);
718 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
719 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
720 setOperationAction(ISD::DEBUGTRAP
, MVT::Other
, Expand
);
723 MVT
TargetLoweringBase::getScalarShiftAmountTy(const DataLayout
&DL
,
725 return MVT::getIntegerVT(DL
.getPointerSizeInBits(0));
728 EVT
TargetLoweringBase::getShiftAmountTy(EVT LHSTy
, const DataLayout
&DL
,
729 bool LegalTypes
) const {
730 assert(LHSTy
.isInteger() && "Shift amount is not an integer type!");
731 if (LHSTy
.isVector())
733 return LegalTypes
? getScalarShiftAmountTy(DL
, LHSTy
)
737 bool TargetLoweringBase::canOpTrap(unsigned Op
, EVT VT
) const {
738 assert(isTypeLegal(VT
));
750 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive
) {
751 // If the command-line option was specified, ignore this request.
752 if (!JumpIsExpensiveOverride
.getNumOccurrences())
753 JumpIsExpensive
= isExpensive
;
756 TargetLoweringBase::LegalizeKind
757 TargetLoweringBase::getTypeConversion(LLVMContext
&Context
, EVT VT
) const {
758 // If this is a simple type, use the ComputeRegisterProp mechanism.
760 MVT SVT
= VT
.getSimpleVT();
761 assert((unsigned)SVT
.SimpleTy
< array_lengthof(TransformToType
));
762 MVT NVT
= TransformToType
[SVT
.SimpleTy
];
763 LegalizeTypeAction LA
= ValueTypeActions
.getTypeAction(SVT
);
765 assert((LA
== TypeLegal
|| LA
== TypeSoftenFloat
||
766 ValueTypeActions
.getTypeAction(NVT
) != TypePromoteInteger
) &&
767 "Promote may not follow Expand or Promote");
769 if (LA
== TypeSplitVector
)
770 return LegalizeKind(LA
,
771 EVT::getVectorVT(Context
, SVT
.getVectorElementType(),
772 SVT
.getVectorNumElements() / 2));
773 if (LA
== TypeScalarizeVector
)
774 return LegalizeKind(LA
, SVT
.getVectorElementType());
775 return LegalizeKind(LA
, NVT
);
778 // Handle Extended Scalar Types.
779 if (!VT
.isVector()) {
780 assert(VT
.isInteger() && "Float types must be simple");
781 unsigned BitSize
= VT
.getSizeInBits();
782 // First promote to a power-of-two size, then expand if necessary.
783 if (BitSize
< 8 || !isPowerOf2_32(BitSize
)) {
784 EVT NVT
= VT
.getRoundIntegerType(Context
);
785 assert(NVT
!= VT
&& "Unable to round integer VT");
786 LegalizeKind NextStep
= getTypeConversion(Context
, NVT
);
787 // Avoid multi-step promotion.
788 if (NextStep
.first
== TypePromoteInteger
)
790 // Return rounded integer type.
791 return LegalizeKind(TypePromoteInteger
, NVT
);
794 return LegalizeKind(TypeExpandInteger
,
795 EVT::getIntegerVT(Context
, VT
.getSizeInBits() / 2));
798 // Handle vector types.
799 unsigned NumElts
= VT
.getVectorNumElements();
800 EVT EltVT
= VT
.getVectorElementType();
802 // Vectors with only one element are always scalarized.
804 return LegalizeKind(TypeScalarizeVector
, EltVT
);
806 // Try to widen vector elements until the element type is a power of two and
807 // promote it to a legal type later on, for example:
808 // <3 x i8> -> <4 x i8> -> <4 x i32>
809 if (EltVT
.isInteger()) {
810 // Vectors with a number of elements that is not a power of two are always
811 // widened, for example <3 x i8> -> <4 x i8>.
812 if (!VT
.isPow2VectorType()) {
813 NumElts
= (unsigned)NextPowerOf2(NumElts
);
814 EVT NVT
= EVT::getVectorVT(Context
, EltVT
, NumElts
);
815 return LegalizeKind(TypeWidenVector
, NVT
);
818 // Examine the element type.
819 LegalizeKind LK
= getTypeConversion(Context
, EltVT
);
821 // If type is to be expanded, split the vector.
822 // <4 x i140> -> <2 x i140>
823 if (LK
.first
== TypeExpandInteger
)
824 return LegalizeKind(TypeSplitVector
,
825 EVT::getVectorVT(Context
, EltVT
, NumElts
/ 2));
827 // Promote the integer element types until a legal vector type is found
828 // or until the element integer type is too big. If a legal type was not
829 // found, fallback to the usual mechanism of widening/splitting the
831 EVT OldEltVT
= EltVT
;
833 // Increase the bitwidth of the element to the next pow-of-two
834 // (which is greater than 8 bits).
835 EltVT
= EVT::getIntegerVT(Context
, 1 + EltVT
.getSizeInBits())
836 .getRoundIntegerType(Context
);
838 // Stop trying when getting a non-simple element type.
839 // Note that vector elements may be greater than legal vector element
840 // types. Example: X86 XMM registers hold 64bit element on 32bit
842 if (!EltVT
.isSimple())
845 // Build a new vector type and check if it is legal.
846 MVT NVT
= MVT::getVectorVT(EltVT
.getSimpleVT(), NumElts
);
847 // Found a legal promoted vector type.
848 if (NVT
!= MVT() && ValueTypeActions
.getTypeAction(NVT
) == TypeLegal
)
849 return LegalizeKind(TypePromoteInteger
,
850 EVT::getVectorVT(Context
, EltVT
, NumElts
));
853 // Reset the type to the unexpanded type if we did not find a legal vector
854 // type with a promoted vector element type.
858 // Try to widen the vector until a legal type is found.
859 // If there is no wider legal type, split the vector.
861 // Round up to the next power of 2.
862 NumElts
= (unsigned)NextPowerOf2(NumElts
);
864 // If there is no simple vector type with this many elements then there
865 // cannot be a larger legal vector type. Note that this assumes that
866 // there are no skipped intermediate vector types in the simple types.
867 if (!EltVT
.isSimple())
869 MVT LargerVector
= MVT::getVectorVT(EltVT
.getSimpleVT(), NumElts
);
870 if (LargerVector
== MVT())
873 // If this type is legal then widen the vector.
874 if (ValueTypeActions
.getTypeAction(LargerVector
) == TypeLegal
)
875 return LegalizeKind(TypeWidenVector
, LargerVector
);
878 // Widen odd vectors to next power of two.
879 if (!VT
.isPow2VectorType()) {
880 EVT NVT
= VT
.getPow2VectorType(Context
);
881 return LegalizeKind(TypeWidenVector
, NVT
);
884 // Vectors with illegal element types are expanded.
885 EVT NVT
= EVT::getVectorVT(Context
, EltVT
, VT
.getVectorNumElements() / 2);
886 return LegalizeKind(TypeSplitVector
, NVT
);
889 static unsigned getVectorTypeBreakdownMVT(MVT VT
, MVT
&IntermediateVT
,
890 unsigned &NumIntermediates
,
892 TargetLoweringBase
*TLI
) {
893 // Figure out the right, legal destination reg to copy into.
894 unsigned NumElts
= VT
.getVectorNumElements();
895 MVT EltTy
= VT
.getVectorElementType();
897 unsigned NumVectorRegs
= 1;
899 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
900 // could break down into LHS/RHS like LegalizeDAG does.
901 if (!isPowerOf2_32(NumElts
)) {
902 NumVectorRegs
= NumElts
;
906 // Divide the input until we get to a supported size. This will always
907 // end with a scalar if the target doesn't support vectors.
908 while (NumElts
> 1 && !TLI
->isTypeLegal(MVT::getVectorVT(EltTy
, NumElts
))) {
913 NumIntermediates
= NumVectorRegs
;
915 MVT NewVT
= MVT::getVectorVT(EltTy
, NumElts
);
916 if (!TLI
->isTypeLegal(NewVT
))
918 IntermediateVT
= NewVT
;
920 unsigned NewVTSize
= NewVT
.getSizeInBits();
922 // Convert sizes such as i33 to i64.
923 if (!isPowerOf2_32(NewVTSize
))
924 NewVTSize
= NextPowerOf2(NewVTSize
);
926 MVT DestVT
= TLI
->getRegisterType(NewVT
);
928 if (EVT(DestVT
).bitsLT(NewVT
)) // Value is expanded, e.g. i64 -> i16.
929 return NumVectorRegs
*(NewVTSize
/DestVT
.getSizeInBits());
931 // Otherwise, promotion or legal types use the same number of registers as
932 // the vector decimated to the appropriate level.
933 return NumVectorRegs
;
936 /// isLegalRC - Return true if the value types that can be represented by the
937 /// specified register class are all legal.
938 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo
&TRI
,
939 const TargetRegisterClass
&RC
) const {
940 for (auto I
= TRI
.legalclasstypes_begin(RC
); *I
!= MVT::Other
; ++I
)
946 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
947 /// sequence of memory operands that is recognized by PrologEpilogInserter.
949 TargetLoweringBase::emitPatchPoint(MachineInstr
&InitialMI
,
950 MachineBasicBlock
*MBB
) const {
951 MachineInstr
*MI
= &InitialMI
;
952 MachineFunction
&MF
= *MI
->getMF();
953 MachineFrameInfo
&MFI
= MF
.getFrameInfo();
955 // We're handling multiple types of operands here:
956 // PATCHPOINT MetaArgs - live-in, read only, direct
957 // STATEPOINT Deopt Spill - live-through, read only, indirect
958 // STATEPOINT Deopt Alloca - live-through, read only, direct
959 // (We're currently conservative and mark the deopt slots read/write in
961 // STATEPOINT GC Spill - live-through, read/write, indirect
962 // STATEPOINT GC Alloca - live-through, read/write, direct
963 // The live-in vs live-through is handled already (the live through ones are
964 // all stack slots), but we need to handle the different type of stackmap
965 // operands and memory effects here.
967 // MI changes inside this loop as we grow operands.
968 for(unsigned OperIdx
= 0; OperIdx
!= MI
->getNumOperands(); ++OperIdx
) {
969 MachineOperand
&MO
= MI
->getOperand(OperIdx
);
973 // foldMemoryOperand builds a new MI after replacing a single FI operand
974 // with the canonical set of five x86 addressing-mode operands.
975 int FI
= MO
.getIndex();
976 MachineInstrBuilder MIB
= BuildMI(MF
, MI
->getDebugLoc(), MI
->getDesc());
978 // Copy operands before the frame-index.
979 for (unsigned i
= 0; i
< OperIdx
; ++i
)
980 MIB
.add(MI
->getOperand(i
));
981 // Add frame index operands recognized by stackmaps.cpp
982 if (MFI
.isStatepointSpillSlotObjectIndex(FI
)) {
983 // indirect-mem-ref tag, size, #FI, offset.
984 // Used for spills inserted by StatepointLowering. This codepath is not
985 // used for patchpoints/stackmaps at all, for these spilling is done via
986 // foldMemoryOperand callback only.
987 assert(MI
->getOpcode() == TargetOpcode::STATEPOINT
&& "sanity");
988 MIB
.addImm(StackMaps::IndirectMemRefOp
);
989 MIB
.addImm(MFI
.getObjectSize(FI
));
990 MIB
.add(MI
->getOperand(OperIdx
));
993 // direct-mem-ref tag, #FI, offset.
994 // Used by patchpoint, and direct alloca arguments to statepoints
995 MIB
.addImm(StackMaps::DirectMemRefOp
);
996 MIB
.add(MI
->getOperand(OperIdx
));
999 // Copy the operands after the frame index.
1000 for (unsigned i
= OperIdx
+ 1; i
!= MI
->getNumOperands(); ++i
)
1001 MIB
.add(MI
->getOperand(i
));
1003 // Inherit previous memory operands.
1004 MIB
.cloneMemRefs(*MI
);
1005 assert(MIB
->mayLoad() && "Folded a stackmap use to a non-load!");
1007 // Add a new memory operand for this FI.
1008 assert(MFI
.getObjectOffset(FI
) != -1);
1010 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and
1011 // PATCHPOINT should be updated to do the same. (TODO)
1012 if (MI
->getOpcode() != TargetOpcode::STATEPOINT
) {
1013 auto Flags
= MachineMemOperand::MOLoad
;
1014 MachineMemOperand
*MMO
= MF
.getMachineMemOperand(
1015 MachinePointerInfo::getFixedStack(MF
, FI
), Flags
,
1016 MF
.getDataLayout().getPointerSize(), MFI
.getObjectAlignment(FI
));
1017 MIB
->addMemOperand(MF
, MMO
);
1020 // Replace the instruction and update the operand index.
1021 MBB
->insert(MachineBasicBlock::iterator(MI
), MIB
);
1022 OperIdx
+= (MIB
->getNumOperands() - MI
->getNumOperands()) - 1;
1023 MI
->eraseFromParent();
1030 TargetLoweringBase::emitXRayCustomEvent(MachineInstr
&MI
,
1031 MachineBasicBlock
*MBB
) const {
1032 assert(MI
.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL
&&
1033 "Called emitXRayCustomEvent on the wrong MI!");
1034 auto &MF
= *MI
.getMF();
1035 auto MIB
= BuildMI(MF
, MI
.getDebugLoc(), MI
.getDesc());
1036 for (unsigned OpIdx
= 0; OpIdx
!= MI
.getNumOperands(); ++OpIdx
)
1037 MIB
.add(MI
.getOperand(OpIdx
));
1039 MBB
->insert(MachineBasicBlock::iterator(MI
), MIB
);
1040 MI
.eraseFromParent();
1045 TargetLoweringBase::emitXRayTypedEvent(MachineInstr
&MI
,
1046 MachineBasicBlock
*MBB
) const {
1047 assert(MI
.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL
&&
1048 "Called emitXRayTypedEvent on the wrong MI!");
1049 auto &MF
= *MI
.getMF();
1050 auto MIB
= BuildMI(MF
, MI
.getDebugLoc(), MI
.getDesc());
1051 for (unsigned OpIdx
= 0; OpIdx
!= MI
.getNumOperands(); ++OpIdx
)
1052 MIB
.add(MI
.getOperand(OpIdx
));
1054 MBB
->insert(MachineBasicBlock::iterator(MI
), MIB
);
1055 MI
.eraseFromParent();
1059 /// findRepresentativeClass - Return the largest legal super-reg register class
1060 /// of the register class for the specified type and its associated "cost".
1061 // This function is in TargetLowering because it uses RegClassForVT which would
1062 // need to be moved to TargetRegisterInfo and would necessitate moving
1063 // isTypeLegal over as well - a massive change that would just require
1064 // TargetLowering having a TargetRegisterInfo class member that it would use.
1065 std::pair
<const TargetRegisterClass
*, uint8_t>
1066 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo
*TRI
,
1068 const TargetRegisterClass
*RC
= RegClassForVT
[VT
.SimpleTy
];
1070 return std::make_pair(RC
, 0);
1072 // Compute the set of all super-register classes.
1073 BitVector
SuperRegRC(TRI
->getNumRegClasses());
1074 for (SuperRegClassIterator
RCI(RC
, TRI
); RCI
.isValid(); ++RCI
)
1075 SuperRegRC
.setBitsInMask(RCI
.getMask());
1077 // Find the first legal register class with the largest spill size.
1078 const TargetRegisterClass
*BestRC
= RC
;
1079 for (unsigned i
: SuperRegRC
.set_bits()) {
1080 const TargetRegisterClass
*SuperRC
= TRI
->getRegClass(i
);
1081 // We want the largest possible spill size.
1082 if (TRI
->getSpillSize(*SuperRC
) <= TRI
->getSpillSize(*BestRC
))
1084 if (!isLegalRC(*TRI
, *SuperRC
))
1088 return std::make_pair(BestRC
, 1);
1091 /// computeRegisterProperties - Once all of the register classes are added,
1092 /// this allows us to compute derived properties we expose.
1093 void TargetLoweringBase::computeRegisterProperties(
1094 const TargetRegisterInfo
*TRI
) {
1095 static_assert(MVT::LAST_VALUETYPE
<= MVT::MAX_ALLOWED_VALUETYPE
,
1096 "Too many value types for ValueTypeActions to hold!");
1098 // Everything defaults to needing one register.
1099 for (unsigned i
= 0; i
!= MVT::LAST_VALUETYPE
; ++i
) {
1100 NumRegistersForVT
[i
] = 1;
1101 RegisterTypeForVT
[i
] = TransformToType
[i
] = (MVT::SimpleValueType
)i
;
1103 // ...except isVoid, which doesn't need any registers.
1104 NumRegistersForVT
[MVT::isVoid
] = 0;
1106 // Find the largest integer register class.
1107 unsigned LargestIntReg
= MVT::LAST_INTEGER_VALUETYPE
;
1108 for (; RegClassForVT
[LargestIntReg
] == nullptr; --LargestIntReg
)
1109 assert(LargestIntReg
!= MVT::i1
&& "No integer registers defined!");
1111 // Every integer value type larger than this largest register takes twice as
1112 // many registers to represent as the previous ValueType.
1113 for (unsigned ExpandedReg
= LargestIntReg
+ 1;
1114 ExpandedReg
<= MVT::LAST_INTEGER_VALUETYPE
; ++ExpandedReg
) {
1115 NumRegistersForVT
[ExpandedReg
] = 2*NumRegistersForVT
[ExpandedReg
-1];
1116 RegisterTypeForVT
[ExpandedReg
] = (MVT::SimpleValueType
)LargestIntReg
;
1117 TransformToType
[ExpandedReg
] = (MVT::SimpleValueType
)(ExpandedReg
- 1);
1118 ValueTypeActions
.setTypeAction((MVT::SimpleValueType
)ExpandedReg
,
1122 // Inspect all of the ValueType's smaller than the largest integer
1123 // register to see which ones need promotion.
1124 unsigned LegalIntReg
= LargestIntReg
;
1125 for (unsigned IntReg
= LargestIntReg
- 1;
1126 IntReg
>= (unsigned)MVT::i1
; --IntReg
) {
1127 MVT IVT
= (MVT::SimpleValueType
)IntReg
;
1128 if (isTypeLegal(IVT
)) {
1129 LegalIntReg
= IntReg
;
1131 RegisterTypeForVT
[IntReg
] = TransformToType
[IntReg
] =
1132 (MVT::SimpleValueType
)LegalIntReg
;
1133 ValueTypeActions
.setTypeAction(IVT
, TypePromoteInteger
);
1137 // ppcf128 type is really two f64's.
1138 if (!isTypeLegal(MVT::ppcf128
)) {
1139 if (isTypeLegal(MVT::f64
)) {
1140 NumRegistersForVT
[MVT::ppcf128
] = 2*NumRegistersForVT
[MVT::f64
];
1141 RegisterTypeForVT
[MVT::ppcf128
] = MVT::f64
;
1142 TransformToType
[MVT::ppcf128
] = MVT::f64
;
1143 ValueTypeActions
.setTypeAction(MVT::ppcf128
, TypeExpandFloat
);
1145 NumRegistersForVT
[MVT::ppcf128
] = NumRegistersForVT
[MVT::i128
];
1146 RegisterTypeForVT
[MVT::ppcf128
] = RegisterTypeForVT
[MVT::i128
];
1147 TransformToType
[MVT::ppcf128
] = MVT::i128
;
1148 ValueTypeActions
.setTypeAction(MVT::ppcf128
, TypeSoftenFloat
);
1152 // Decide how to handle f128. If the target does not have native f128 support,
1153 // expand it to i128 and we will be generating soft float library calls.
1154 if (!isTypeLegal(MVT::f128
)) {
1155 NumRegistersForVT
[MVT::f128
] = NumRegistersForVT
[MVT::i128
];
1156 RegisterTypeForVT
[MVT::f128
] = RegisterTypeForVT
[MVT::i128
];
1157 TransformToType
[MVT::f128
] = MVT::i128
;
1158 ValueTypeActions
.setTypeAction(MVT::f128
, TypeSoftenFloat
);
1161 // Decide how to handle f64. If the target does not have native f64 support,
1162 // expand it to i64 and we will be generating soft float library calls.
1163 if (!isTypeLegal(MVT::f64
)) {
1164 NumRegistersForVT
[MVT::f64
] = NumRegistersForVT
[MVT::i64
];
1165 RegisterTypeForVT
[MVT::f64
] = RegisterTypeForVT
[MVT::i64
];
1166 TransformToType
[MVT::f64
] = MVT::i64
;
1167 ValueTypeActions
.setTypeAction(MVT::f64
, TypeSoftenFloat
);
1170 // Decide how to handle f32. If the target does not have native f32 support,
1171 // expand it to i32 and we will be generating soft float library calls.
1172 if (!isTypeLegal(MVT::f32
)) {
1173 NumRegistersForVT
[MVT::f32
] = NumRegistersForVT
[MVT::i32
];
1174 RegisterTypeForVT
[MVT::f32
] = RegisterTypeForVT
[MVT::i32
];
1175 TransformToType
[MVT::f32
] = MVT::i32
;
1176 ValueTypeActions
.setTypeAction(MVT::f32
, TypeSoftenFloat
);
1179 // Decide how to handle f16. If the target does not have native f16 support,
1180 // promote it to f32, because there are no f16 library calls (except for
1182 if (!isTypeLegal(MVT::f16
)) {
1183 NumRegistersForVT
[MVT::f16
] = NumRegistersForVT
[MVT::f32
];
1184 RegisterTypeForVT
[MVT::f16
] = RegisterTypeForVT
[MVT::f32
];
1185 TransformToType
[MVT::f16
] = MVT::f32
;
1186 ValueTypeActions
.setTypeAction(MVT::f16
, TypePromoteFloat
);
1189 // Loop over all of the vector value types to see which need transformations.
1190 for (unsigned i
= MVT::FIRST_VECTOR_VALUETYPE
;
1191 i
<= (unsigned)MVT::LAST_VECTOR_VALUETYPE
; ++i
) {
1192 MVT VT
= (MVT::SimpleValueType
) i
;
1193 if (isTypeLegal(VT
))
1196 MVT EltVT
= VT
.getVectorElementType();
1197 unsigned NElts
= VT
.getVectorNumElements();
1198 bool IsLegalWiderType
= false;
1199 LegalizeTypeAction PreferredAction
= getPreferredVectorAction(VT
);
1200 switch (PreferredAction
) {
1201 case TypePromoteInteger
:
1202 // Try to promote the elements of integer vectors. If no legal
1203 // promotion was found, fall through to the widen-vector method.
1204 for (unsigned nVT
= i
+ 1; nVT
<= MVT::LAST_INTEGER_VECTOR_VALUETYPE
; ++nVT
) {
1205 MVT SVT
= (MVT::SimpleValueType
) nVT
;
1206 // Promote vectors of integers to vectors with the same number
1207 // of elements, with a wider element type.
1208 if (SVT
.getScalarSizeInBits() > EltVT
.getSizeInBits() &&
1209 SVT
.getVectorNumElements() == NElts
&& isTypeLegal(SVT
)) {
1210 TransformToType
[i
] = SVT
;
1211 RegisterTypeForVT
[i
] = SVT
;
1212 NumRegistersForVT
[i
] = 1;
1213 ValueTypeActions
.setTypeAction(VT
, TypePromoteInteger
);
1214 IsLegalWiderType
= true;
1218 if (IsLegalWiderType
)
1222 case TypeWidenVector
:
1223 // Try to widen the vector.
1224 for (unsigned nVT
= i
+ 1; nVT
<= MVT::LAST_VECTOR_VALUETYPE
; ++nVT
) {
1225 MVT SVT
= (MVT::SimpleValueType
) nVT
;
1226 if (SVT
.getVectorElementType() == EltVT
1227 && SVT
.getVectorNumElements() > NElts
&& isTypeLegal(SVT
)) {
1228 TransformToType
[i
] = SVT
;
1229 RegisterTypeForVT
[i
] = SVT
;
1230 NumRegistersForVT
[i
] = 1;
1231 ValueTypeActions
.setTypeAction(VT
, TypeWidenVector
);
1232 IsLegalWiderType
= true;
1236 if (IsLegalWiderType
)
1240 case TypeSplitVector
:
1241 case TypeScalarizeVector
: {
1244 unsigned NumIntermediates
;
1245 NumRegistersForVT
[i
] = getVectorTypeBreakdownMVT(VT
, IntermediateVT
,
1246 NumIntermediates
, RegisterVT
, this);
1247 RegisterTypeForVT
[i
] = RegisterVT
;
1249 MVT NVT
= VT
.getPow2VectorType();
1251 // Type is already a power of 2. The default action is to split.
1252 TransformToType
[i
] = MVT::Other
;
1253 if (PreferredAction
== TypeScalarizeVector
)
1254 ValueTypeActions
.setTypeAction(VT
, TypeScalarizeVector
);
1255 else if (PreferredAction
== TypeSplitVector
)
1256 ValueTypeActions
.setTypeAction(VT
, TypeSplitVector
);
1258 // Set type action according to the number of elements.
1259 ValueTypeActions
.setTypeAction(VT
, NElts
== 1 ? TypeScalarizeVector
1262 TransformToType
[i
] = NVT
;
1263 ValueTypeActions
.setTypeAction(VT
, TypeWidenVector
);
1268 llvm_unreachable("Unknown vector legalization action!");
1272 // Determine the 'representative' register class for each value type.
1273 // An representative register class is the largest (meaning one which is
1274 // not a sub-register class / subreg register class) legal register class for
1275 // a group of value types. For example, on i386, i8, i16, and i32
1276 // representative would be GR32; while on x86_64 it's GR64.
1277 for (unsigned i
= 0; i
!= MVT::LAST_VALUETYPE
; ++i
) {
1278 const TargetRegisterClass
* RRC
;
1280 std::tie(RRC
, Cost
) = findRepresentativeClass(TRI
, (MVT::SimpleValueType
)i
);
1281 RepRegClassForVT
[i
] = RRC
;
1282 RepRegClassCostForVT
[i
] = Cost
;
1286 EVT
TargetLoweringBase::getSetCCResultType(const DataLayout
&DL
, LLVMContext
&,
1288 assert(!VT
.isVector() && "No default SetCC type for vectors!");
1289 return getPointerTy(DL
).SimpleTy
;
1292 MVT::SimpleValueType
TargetLoweringBase::getCmpLibcallReturnType() const {
1293 return MVT::i32
; // return the default value
1296 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1297 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1298 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1299 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1301 /// This method returns the number of registers needed, and the VT for each
1302 /// register. It also returns the VT and quantity of the intermediate values
1303 /// before they are promoted/expanded.
1304 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext
&Context
, EVT VT
,
1305 EVT
&IntermediateVT
,
1306 unsigned &NumIntermediates
,
1307 MVT
&RegisterVT
) const {
1308 unsigned NumElts
= VT
.getVectorNumElements();
1310 // If there is a wider vector type with the same element type as this one,
1311 // or a promoted vector type that has the same number of elements which
1312 // are wider, then we should convert to that legal vector type.
1313 // This handles things like <2 x float> -> <4 x float> and
1314 // <4 x i1> -> <4 x i32>.
1315 LegalizeTypeAction TA
= getTypeAction(Context
, VT
);
1316 if (NumElts
!= 1 && (TA
== TypeWidenVector
|| TA
== TypePromoteInteger
)) {
1317 EVT RegisterEVT
= getTypeToTransformTo(Context
, VT
);
1318 if (isTypeLegal(RegisterEVT
)) {
1319 IntermediateVT
= RegisterEVT
;
1320 RegisterVT
= RegisterEVT
.getSimpleVT();
1321 NumIntermediates
= 1;
1326 // Figure out the right, legal destination reg to copy into.
1327 EVT EltTy
= VT
.getVectorElementType();
1329 unsigned NumVectorRegs
= 1;
1331 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1332 // could break down into LHS/RHS like LegalizeDAG does.
1333 if (!isPowerOf2_32(NumElts
)) {
1334 NumVectorRegs
= NumElts
;
1338 // Divide the input until we get to a supported size. This will always
1339 // end with a scalar if the target doesn't support vectors.
1340 while (NumElts
> 1 && !isTypeLegal(
1341 EVT::getVectorVT(Context
, EltTy
, NumElts
))) {
1343 NumVectorRegs
<<= 1;
1346 NumIntermediates
= NumVectorRegs
;
1348 EVT NewVT
= EVT::getVectorVT(Context
, EltTy
, NumElts
);
1349 if (!isTypeLegal(NewVT
))
1351 IntermediateVT
= NewVT
;
1353 MVT DestVT
= getRegisterType(Context
, NewVT
);
1354 RegisterVT
= DestVT
;
1355 unsigned NewVTSize
= NewVT
.getSizeInBits();
1357 // Convert sizes such as i33 to i64.
1358 if (!isPowerOf2_32(NewVTSize
))
1359 NewVTSize
= NextPowerOf2(NewVTSize
);
1361 if (EVT(DestVT
).bitsLT(NewVT
)) // Value is expanded, e.g. i64 -> i16.
1362 return NumVectorRegs
*(NewVTSize
/DestVT
.getSizeInBits());
1364 // Otherwise, promotion or legal types use the same number of registers as
1365 // the vector decimated to the appropriate level.
1366 return NumVectorRegs
;
1369 /// Get the EVTs and ArgFlags collections that represent the legalized return
1370 /// type of the given function. This does not require a DAG or a return value,
1371 /// and is suitable for use before any DAGs for the function are constructed.
1372 /// TODO: Move this out of TargetLowering.cpp.
1373 void llvm::GetReturnInfo(CallingConv::ID CC
, Type
*ReturnType
,
1375 SmallVectorImpl
<ISD::OutputArg
> &Outs
,
1376 const TargetLowering
&TLI
, const DataLayout
&DL
) {
1377 SmallVector
<EVT
, 4> ValueVTs
;
1378 ComputeValueVTs(TLI
, DL
, ReturnType
, ValueVTs
);
1379 unsigned NumValues
= ValueVTs
.size();
1380 if (NumValues
== 0) return;
1382 for (unsigned j
= 0, f
= NumValues
; j
!= f
; ++j
) {
1383 EVT VT
= ValueVTs
[j
];
1384 ISD::NodeType ExtendKind
= ISD::ANY_EXTEND
;
1386 if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::SExt
))
1387 ExtendKind
= ISD::SIGN_EXTEND
;
1388 else if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::ZExt
))
1389 ExtendKind
= ISD::ZERO_EXTEND
;
1391 // FIXME: C calling convention requires the return type to be promoted to
1392 // at least 32-bit. But this is not necessary for non-C calling
1393 // conventions. The frontend should mark functions whose return values
1394 // require promoting with signext or zeroext attributes.
1395 if (ExtendKind
!= ISD::ANY_EXTEND
&& VT
.isInteger()) {
1396 MVT MinVT
= TLI
.getRegisterType(ReturnType
->getContext(), MVT::i32
);
1397 if (VT
.bitsLT(MinVT
))
1402 TLI
.getNumRegistersForCallingConv(ReturnType
->getContext(), CC
, VT
);
1404 TLI
.getRegisterTypeForCallingConv(ReturnType
->getContext(), CC
, VT
);
1406 // 'inreg' on function refers to return value
1407 ISD::ArgFlagsTy Flags
= ISD::ArgFlagsTy();
1408 if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::InReg
))
1411 // Propagate extension type if any
1412 if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::SExt
))
1414 else if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::ZExt
))
1417 for (unsigned i
= 0; i
< NumParts
; ++i
)
1418 Outs
.push_back(ISD::OutputArg(Flags
, PartVT
, VT
, /*isFixed=*/true, 0, 0));
1422 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1423 /// function arguments in the caller parameter area. This is the actual
1424 /// alignment, not its logarithm.
1425 unsigned TargetLoweringBase::getByValTypeAlignment(Type
*Ty
,
1426 const DataLayout
&DL
) const {
1427 return DL
.getABITypeAlignment(Ty
);
1430 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext
&Context
,
1431 const DataLayout
&DL
, EVT VT
,
1435 // Check if the specified alignment is sufficient based on the data layout.
1436 // TODO: While using the data layout works in practice, a better solution
1437 // would be to implement this check directly (make this a virtual function).
1438 // For example, the ABI alignment may change based on software platform while
1439 // this function should only be affected by hardware implementation.
1440 Type
*Ty
= VT
.getTypeForEVT(Context
);
1441 if (Alignment
>= DL
.getABITypeAlignment(Ty
)) {
1442 // Assume that an access that meets the ABI-specified alignment is fast.
1443 if (Fast
!= nullptr)
1448 // This is a misaligned access.
1449 return allowsMisalignedMemoryAccesses(VT
, AddrSpace
, Alignment
, Fast
);
1452 BranchProbability
TargetLoweringBase::getPredictableBranchThreshold() const {
1453 return BranchProbability(MinPercentageForPredictableBranch
, 100);
1456 //===----------------------------------------------------------------------===//
1457 // TargetTransformInfo Helpers
1458 //===----------------------------------------------------------------------===//
1460 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode
) const {
1461 enum InstructionOpcodes
{
1462 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1463 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1464 #include "llvm/IR/Instruction.def"
1466 switch (static_cast<InstructionOpcodes
>(Opcode
)) {
1469 case Switch
: return 0;
1470 case IndirectBr
: return 0;
1471 case Invoke
: return 0;
1472 case CallBr
: return 0;
1473 case Resume
: return 0;
1474 case Unreachable
: return 0;
1475 case CleanupRet
: return 0;
1476 case CatchRet
: return 0;
1477 case CatchPad
: return 0;
1478 case CatchSwitch
: return 0;
1479 case CleanupPad
: return 0;
1480 case FNeg
: return ISD::FNEG
;
1481 case Add
: return ISD::ADD
;
1482 case FAdd
: return ISD::FADD
;
1483 case Sub
: return ISD::SUB
;
1484 case FSub
: return ISD::FSUB
;
1485 case Mul
: return ISD::MUL
;
1486 case FMul
: return ISD::FMUL
;
1487 case UDiv
: return ISD::UDIV
;
1488 case SDiv
: return ISD::SDIV
;
1489 case FDiv
: return ISD::FDIV
;
1490 case URem
: return ISD::UREM
;
1491 case SRem
: return ISD::SREM
;
1492 case FRem
: return ISD::FREM
;
1493 case Shl
: return ISD::SHL
;
1494 case LShr
: return ISD::SRL
;
1495 case AShr
: return ISD::SRA
;
1496 case And
: return ISD::AND
;
1497 case Or
: return ISD::OR
;
1498 case Xor
: return ISD::XOR
;
1499 case Alloca
: return 0;
1500 case Load
: return ISD::LOAD
;
1501 case Store
: return ISD::STORE
;
1502 case GetElementPtr
: return 0;
1503 case Fence
: return 0;
1504 case AtomicCmpXchg
: return 0;
1505 case AtomicRMW
: return 0;
1506 case Trunc
: return ISD::TRUNCATE
;
1507 case ZExt
: return ISD::ZERO_EXTEND
;
1508 case SExt
: return ISD::SIGN_EXTEND
;
1509 case FPToUI
: return ISD::FP_TO_UINT
;
1510 case FPToSI
: return ISD::FP_TO_SINT
;
1511 case UIToFP
: return ISD::UINT_TO_FP
;
1512 case SIToFP
: return ISD::SINT_TO_FP
;
1513 case FPTrunc
: return ISD::FP_ROUND
;
1514 case FPExt
: return ISD::FP_EXTEND
;
1515 case PtrToInt
: return ISD::BITCAST
;
1516 case IntToPtr
: return ISD::BITCAST
;
1517 case BitCast
: return ISD::BITCAST
;
1518 case AddrSpaceCast
: return ISD::ADDRSPACECAST
;
1519 case ICmp
: return ISD::SETCC
;
1520 case FCmp
: return ISD::SETCC
;
1522 case Call
: return 0;
1523 case Select
: return ISD::SELECT
;
1524 case UserOp1
: return 0;
1525 case UserOp2
: return 0;
1526 case VAArg
: return 0;
1527 case ExtractElement
: return ISD::EXTRACT_VECTOR_ELT
;
1528 case InsertElement
: return ISD::INSERT_VECTOR_ELT
;
1529 case ShuffleVector
: return ISD::VECTOR_SHUFFLE
;
1530 case ExtractValue
: return ISD::MERGE_VALUES
;
1531 case InsertValue
: return ISD::MERGE_VALUES
;
1532 case LandingPad
: return 0;
1535 llvm_unreachable("Unknown instruction type encountered!");
1539 TargetLoweringBase::getTypeLegalizationCost(const DataLayout
&DL
,
1541 LLVMContext
&C
= Ty
->getContext();
1542 EVT MTy
= getValueType(DL
, Ty
);
1545 // We keep legalizing the type until we find a legal kind. We assume that
1546 // the only operation that costs anything is the split. After splitting
1547 // we need to handle two types.
1549 LegalizeKind LK
= getTypeConversion(C
, MTy
);
1551 if (LK
.first
== TypeLegal
)
1552 return std::make_pair(Cost
, MTy
.getSimpleVT());
1554 if (LK
.first
== TypeSplitVector
|| LK
.first
== TypeExpandInteger
)
1557 // Do not loop with f128 type.
1558 if (MTy
== LK
.second
)
1559 return std::make_pair(Cost
, MTy
.getSimpleVT());
1561 // Keep legalizing the type.
1566 Value
*TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder
<> &IRB
,
1567 bool UseTLS
) const {
1568 // compiler-rt provides a variable with a magic name. Targets that do not
1569 // link with compiler-rt may also provide such a variable.
1570 Module
*M
= IRB
.GetInsertBlock()->getParent()->getParent();
1571 const char *UnsafeStackPtrVar
= "__safestack_unsafe_stack_ptr";
1572 auto UnsafeStackPtr
=
1573 dyn_cast_or_null
<GlobalVariable
>(M
->getNamedValue(UnsafeStackPtrVar
));
1575 Type
*StackPtrTy
= Type::getInt8PtrTy(M
->getContext());
1577 if (!UnsafeStackPtr
) {
1578 auto TLSModel
= UseTLS
?
1579 GlobalValue::InitialExecTLSModel
:
1580 GlobalValue::NotThreadLocal
;
1581 // The global variable is not defined yet, define it ourselves.
1582 // We use the initial-exec TLS model because we do not support the
1583 // variable living anywhere other than in the main executable.
1584 UnsafeStackPtr
= new GlobalVariable(
1585 *M
, StackPtrTy
, false, GlobalValue::ExternalLinkage
, nullptr,
1586 UnsafeStackPtrVar
, nullptr, TLSModel
);
1588 // The variable exists, check its type and attributes.
1589 if (UnsafeStackPtr
->getValueType() != StackPtrTy
)
1590 report_fatal_error(Twine(UnsafeStackPtrVar
) + " must have void* type");
1591 if (UseTLS
!= UnsafeStackPtr
->isThreadLocal())
1592 report_fatal_error(Twine(UnsafeStackPtrVar
) + " must " +
1593 (UseTLS
? "" : "not ") + "be thread-local");
1595 return UnsafeStackPtr
;
1598 Value
*TargetLoweringBase::getSafeStackPointerLocation(IRBuilder
<> &IRB
) const {
1599 if (!TM
.getTargetTriple().isAndroid())
1600 return getDefaultSafeStackPointerLocation(IRB
, true);
1602 // Android provides a libc function to retrieve the address of the current
1603 // thread's unsafe stack pointer.
1604 Module
*M
= IRB
.GetInsertBlock()->getParent()->getParent();
1605 Type
*StackPtrTy
= Type::getInt8PtrTy(M
->getContext());
1606 FunctionCallee Fn
= M
->getOrInsertFunction("__safestack_pointer_address",
1607 StackPtrTy
->getPointerTo(0));
1608 return IRB
.CreateCall(Fn
);
1611 //===----------------------------------------------------------------------===//
1612 // Loop Strength Reduction hooks
1613 //===----------------------------------------------------------------------===//
1615 /// isLegalAddressingMode - Return true if the addressing mode represented
1616 /// by AM is legal for this target, for a load/store of the specified type.
1617 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout
&DL
,
1618 const AddrMode
&AM
, Type
*Ty
,
1619 unsigned AS
, Instruction
*I
) const {
1620 // The default implementation of this implements a conservative RISCy, r+r and
1623 // Allows a sign-extended 16-bit immediate field.
1624 if (AM
.BaseOffs
<= -(1LL << 16) || AM
.BaseOffs
>= (1LL << 16)-1)
1627 // No global is ever allowed as a base.
1631 // Only support r+r,
1633 case 0: // "r+i" or just "i", depending on HasBaseReg.
1636 if (AM
.HasBaseReg
&& AM
.BaseOffs
) // "r+r+i" is not allowed.
1638 // Otherwise we have r+r or r+i.
1641 if (AM
.HasBaseReg
|| AM
.BaseOffs
) // 2*r+r or 2*r+i is not allowed.
1643 // Allow 2*r as r+r.
1645 default: // Don't allow n * r
1652 //===----------------------------------------------------------------------===//
1654 //===----------------------------------------------------------------------===//
1656 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1657 // so that SelectionDAG handle SSP.
1658 Value
*TargetLoweringBase::getIRStackGuard(IRBuilder
<> &IRB
) const {
1659 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1660 Module
&M
= *IRB
.GetInsertBlock()->getParent()->getParent();
1661 PointerType
*PtrTy
= Type::getInt8PtrTy(M
.getContext());
1662 return M
.getOrInsertGlobal("__guard_local", PtrTy
);
1667 // Currently only support "standard" __stack_chk_guard.
1668 // TODO: add LOAD_STACK_GUARD support.
1669 void TargetLoweringBase::insertSSPDeclarations(Module
&M
) const {
1670 if (!M
.getNamedValue("__stack_chk_guard"))
1671 new GlobalVariable(M
, Type::getInt8PtrTy(M
.getContext()), false,
1672 GlobalVariable::ExternalLinkage
,
1673 nullptr, "__stack_chk_guard");
1676 // Currently only support "standard" __stack_chk_guard.
1677 // TODO: add LOAD_STACK_GUARD support.
1678 Value
*TargetLoweringBase::getSDagStackGuard(const Module
&M
) const {
1679 return M
.getNamedValue("__stack_chk_guard");
1682 Function
*TargetLoweringBase::getSSPStackGuardCheck(const Module
&M
) const {
1686 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1687 return MinimumJumpTableEntries
;
1690 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val
) {
1691 MinimumJumpTableEntries
= Val
;
1694 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize
) const {
1695 return OptForSize
? OptsizeJumpTableDensity
: JumpTableDensity
;
1698 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1699 return MaximumJumpTableSize
;
1702 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val
) {
1703 MaximumJumpTableSize
= Val
;
1706 //===----------------------------------------------------------------------===//
1707 // Reciprocal Estimates
1708 //===----------------------------------------------------------------------===//
1710 /// Get the reciprocal estimate attribute string for a function that will
1711 /// override the target defaults.
1712 static StringRef
getRecipEstimateForFunc(MachineFunction
&MF
) {
1713 const Function
&F
= MF
.getFunction();
1714 return F
.getFnAttribute("reciprocal-estimates").getValueAsString();
1717 /// Construct a string for the given reciprocal operation of the given type.
1718 /// This string should match the corresponding option to the front-end's
1719 /// "-mrecip" flag assuming those strings have been passed through in an
1720 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1721 static std::string
getReciprocalOpName(bool IsSqrt
, EVT VT
) {
1722 std::string Name
= VT
.isVector() ? "vec-" : "";
1724 Name
+= IsSqrt
? "sqrt" : "div";
1726 // TODO: Handle "half" or other float types?
1727 if (VT
.getScalarType() == MVT::f64
) {
1730 assert(VT
.getScalarType() == MVT::f32
&&
1731 "Unexpected FP type for reciprocal estimate");
1738 /// Return the character position and value (a single numeric character) of a
1739 /// customized refinement operation in the input string if it exists. Return
1740 /// false if there is no customized refinement step count.
1741 static bool parseRefinementStep(StringRef In
, size_t &Position
,
1743 const char RefStepToken
= ':';
1744 Position
= In
.find(RefStepToken
);
1745 if (Position
== StringRef::npos
)
1748 StringRef RefStepString
= In
.substr(Position
+ 1);
1749 // Allow exactly one numeric character for the additional refinement
1751 if (RefStepString
.size() == 1) {
1752 char RefStepChar
= RefStepString
[0];
1753 if (RefStepChar
>= '0' && RefStepChar
<= '9') {
1754 Value
= RefStepChar
- '0';
1758 report_fatal_error("Invalid refinement step for -recip.");
1761 /// For the input attribute string, return one of the ReciprocalEstimate enum
1762 /// status values (enabled, disabled, or not specified) for this operation on
1763 /// the specified data type.
1764 static int getOpEnabled(bool IsSqrt
, EVT VT
, StringRef Override
) {
1765 if (Override
.empty())
1766 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1768 SmallVector
<StringRef
, 4> OverrideVector
;
1769 Override
.split(OverrideVector
, ',');
1770 unsigned NumArgs
= OverrideVector
.size();
1772 // Check if "all", "none", or "default" was specified.
1774 // Look for an optional setting of the number of refinement steps needed
1775 // for this type of reciprocal operation.
1778 if (parseRefinementStep(Override
, RefPos
, RefSteps
)) {
1779 // Split the string for further processing.
1780 Override
= Override
.substr(0, RefPos
);
1783 // All reciprocal types are enabled.
1784 if (Override
== "all")
1785 return TargetLoweringBase::ReciprocalEstimate::Enabled
;
1787 // All reciprocal types are disabled.
1788 if (Override
== "none")
1789 return TargetLoweringBase::ReciprocalEstimate::Disabled
;
1791 // Target defaults for enablement are used.
1792 if (Override
== "default")
1793 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1796 // The attribute string may omit the size suffix ('f'/'d').
1797 std::string VTName
= getReciprocalOpName(IsSqrt
, VT
);
1798 std::string VTNameNoSize
= VTName
;
1799 VTNameNoSize
.pop_back();
1800 static const char DisabledPrefix
= '!';
1802 for (StringRef RecipType
: OverrideVector
) {
1805 if (parseRefinementStep(RecipType
, RefPos
, RefSteps
))
1806 RecipType
= RecipType
.substr(0, RefPos
);
1808 // Ignore the disablement token for string matching.
1809 bool IsDisabled
= RecipType
[0] == DisabledPrefix
;
1811 RecipType
= RecipType
.substr(1);
1813 if (RecipType
.equals(VTName
) || RecipType
.equals(VTNameNoSize
))
1814 return IsDisabled
? TargetLoweringBase::ReciprocalEstimate::Disabled
1815 : TargetLoweringBase::ReciprocalEstimate::Enabled
;
1818 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1821 /// For the input attribute string, return the customized refinement step count
1822 /// for this operation on the specified data type. If the step count does not
1823 /// exist, return the ReciprocalEstimate enum value for unspecified.
1824 static int getOpRefinementSteps(bool IsSqrt
, EVT VT
, StringRef Override
) {
1825 if (Override
.empty())
1826 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1828 SmallVector
<StringRef
, 4> OverrideVector
;
1829 Override
.split(OverrideVector
, ',');
1830 unsigned NumArgs
= OverrideVector
.size();
1832 // Check if "all", "default", or "none" was specified.
1834 // Look for an optional setting of the number of refinement steps needed
1835 // for this type of reciprocal operation.
1838 if (!parseRefinementStep(Override
, RefPos
, RefSteps
))
1839 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1841 // Split the string for further processing.
1842 Override
= Override
.substr(0, RefPos
);
1843 assert(Override
!= "none" &&
1844 "Disabled reciprocals, but specifed refinement steps?");
1846 // If this is a general override, return the specified number of steps.
1847 if (Override
== "all" || Override
== "default")
1851 // The attribute string may omit the size suffix ('f'/'d').
1852 std::string VTName
= getReciprocalOpName(IsSqrt
, VT
);
1853 std::string VTNameNoSize
= VTName
;
1854 VTNameNoSize
.pop_back();
1856 for (StringRef RecipType
: OverrideVector
) {
1859 if (!parseRefinementStep(RecipType
, RefPos
, RefSteps
))
1862 RecipType
= RecipType
.substr(0, RefPos
);
1863 if (RecipType
.equals(VTName
) || RecipType
.equals(VTNameNoSize
))
1867 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1870 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT
,
1871 MachineFunction
&MF
) const {
1872 return getOpEnabled(true, VT
, getRecipEstimateForFunc(MF
));
1875 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT
,
1876 MachineFunction
&MF
) const {
1877 return getOpEnabled(false, VT
, getRecipEstimateForFunc(MF
));
1880 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT
,
1881 MachineFunction
&MF
) const {
1882 return getOpRefinementSteps(true, VT
, getRecipEstimateForFunc(MF
));
1885 int TargetLoweringBase::getDivRefinementSteps(EVT VT
,
1886 MachineFunction
&MF
) const {
1887 return getOpRefinementSteps(false, VT
, getRecipEstimateForFunc(MF
));
1890 void TargetLoweringBase::finalizeLowering(MachineFunction
&MF
) const {
1891 MF
.getRegInfo().freezeReservedRegs(MF
);