1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the TargetLoweringBase class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/ADT/BitVector.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/StringExtras.h"
18 #include "llvm/ADT/StringRef.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/ADT/Twine.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/RuntimeLibcalls.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/IR/Attributes.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/Support/BranchProbability.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/MachineValueType.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Target/TargetMachine.h"
67 static cl::opt
<bool> JumpIsExpensiveOverride(
68 "jump-is-expensive", cl::init(false),
69 cl::desc("Do not create extra branches to split comparison logic."),
72 static cl::opt
<unsigned> MinimumJumpTableEntries
73 ("min-jump-table-entries", cl::init(4), cl::Hidden
,
74 cl::desc("Set minimum number of entries to use a jump table."));
76 static cl::opt
<unsigned> MaximumJumpTableSize
77 ("max-jump-table-size", cl::init(0), cl::Hidden
,
78 cl::desc("Set maximum size of jump tables; zero for no limit."));
80 /// Minimum jump table density for normal functions.
81 static cl::opt
<unsigned>
82 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden
,
83 cl::desc("Minimum density for building a jump table in "
84 "a normal function"));
86 /// Minimum jump table density for -Os or -Oz functions.
87 static cl::opt
<unsigned> OptsizeJumpTableDensity(
88 "optsize-jump-table-density", cl::init(40), cl::Hidden
,
89 cl::desc("Minimum density for building a jump table in "
90 "an optsize function"));
92 static bool darwinHasSinCos(const Triple
&TT
) {
93 assert(TT
.isOSDarwin() && "should be called with darwin triple");
94 // Don't bother with 32 bit x86.
95 if (TT
.getArch() == Triple::x86
)
97 // Macos < 10.9 has no sincos_stret.
99 return !TT
.isMacOSXVersionLT(10, 9) && TT
.isArch64Bit();
100 // iOS < 7.0 has no sincos_stret.
102 return !TT
.isOSVersionLT(7, 0);
103 // Any other darwin such as WatchOS/TvOS is new enough.
107 // Although this default value is arbitrary, it is not random. It is assumed
108 // that a condition that evaluates the same way by a higher percentage than this
109 // is best represented as control flow. Therefore, the default value N should be
110 // set such that the win from N% correct executions is greater than the loss
111 // from (100 - N)% mispredicted executions for the majority of intended targets.
112 static cl::opt
<int> MinPercentageForPredictableBranch(
113 "min-predictable-branch", cl::init(99),
114 cl::desc("Minimum percentage (0-100) that a condition must be either true "
115 "or false to assume that the condition is predictable"),
118 void TargetLoweringBase::InitLibcalls(const Triple
&TT
) {
119 #define HANDLE_LIBCALL(code, name) \
120 setLibcallName(RTLIB::code, name);
121 #include "llvm/IR/RuntimeLibcalls.def"
122 #undef HANDLE_LIBCALL
123 // Initialize calling conventions to their default.
124 for (int LC
= 0; LC
< RTLIB::UNKNOWN_LIBCALL
; ++LC
)
125 setLibcallCallingConv((RTLIB::Libcall
)LC
, CallingConv::C
);
127 // A few names are different on particular architectures or environments.
128 if (TT
.isOSDarwin()) {
129 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
130 // of the gnueabi-style __gnu_*_ieee.
131 // FIXME: What about other targets?
132 setLibcallName(RTLIB::FPEXT_F16_F32
, "__extendhfsf2");
133 setLibcallName(RTLIB::FPROUND_F32_F16
, "__truncsfhf2");
135 // Some darwins have an optimized __bzero/bzero function.
136 switch (TT
.getArch()) {
139 if (TT
.isMacOSX() && !TT
.isMacOSXVersionLT(10, 6))
140 setLibcallName(RTLIB::BZERO
, "__bzero");
142 case Triple::aarch64
:
143 setLibcallName(RTLIB::BZERO
, "bzero");
149 if (darwinHasSinCos(TT
)) {
150 setLibcallName(RTLIB::SINCOS_STRET_F32
, "__sincosf_stret");
151 setLibcallName(RTLIB::SINCOS_STRET_F64
, "__sincos_stret");
152 if (TT
.isWatchABI()) {
153 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32
,
154 CallingConv::ARM_AAPCS_VFP
);
155 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64
,
156 CallingConv::ARM_AAPCS_VFP
);
160 setLibcallName(RTLIB::FPEXT_F16_F32
, "__gnu_h2f_ieee");
161 setLibcallName(RTLIB::FPROUND_F32_F16
, "__gnu_f2h_ieee");
164 if (TT
.isGNUEnvironment() || TT
.isOSFuchsia() ||
165 (TT
.isAndroid() && !TT
.isAndroidVersionLT(9))) {
166 setLibcallName(RTLIB::SINCOS_F32
, "sincosf");
167 setLibcallName(RTLIB::SINCOS_F64
, "sincos");
168 setLibcallName(RTLIB::SINCOS_F80
, "sincosl");
169 setLibcallName(RTLIB::SINCOS_F128
, "sincosl");
170 setLibcallName(RTLIB::SINCOS_PPCF128
, "sincosl");
173 if (TT
.isOSOpenBSD()) {
174 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL
, nullptr);
178 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
179 /// UNKNOWN_LIBCALL if there is none.
180 RTLIB::Libcall
RTLIB::getFPEXT(EVT OpVT
, EVT RetVT
) {
181 if (OpVT
== MVT::f16
) {
182 if (RetVT
== MVT::f32
)
183 return FPEXT_F16_F32
;
184 } else if (OpVT
== MVT::f32
) {
185 if (RetVT
== MVT::f64
)
186 return FPEXT_F32_F64
;
187 if (RetVT
== MVT::f128
)
188 return FPEXT_F32_F128
;
189 if (RetVT
== MVT::ppcf128
)
190 return FPEXT_F32_PPCF128
;
191 } else if (OpVT
== MVT::f64
) {
192 if (RetVT
== MVT::f128
)
193 return FPEXT_F64_F128
;
194 else if (RetVT
== MVT::ppcf128
)
195 return FPEXT_F64_PPCF128
;
196 } else if (OpVT
== MVT::f80
) {
197 if (RetVT
== MVT::f128
)
198 return FPEXT_F80_F128
;
201 return UNKNOWN_LIBCALL
;
204 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
205 /// UNKNOWN_LIBCALL if there is none.
206 RTLIB::Libcall
RTLIB::getFPROUND(EVT OpVT
, EVT RetVT
) {
207 if (RetVT
== MVT::f16
) {
208 if (OpVT
== MVT::f32
)
209 return FPROUND_F32_F16
;
210 if (OpVT
== MVT::f64
)
211 return FPROUND_F64_F16
;
212 if (OpVT
== MVT::f80
)
213 return FPROUND_F80_F16
;
214 if (OpVT
== MVT::f128
)
215 return FPROUND_F128_F16
;
216 if (OpVT
== MVT::ppcf128
)
217 return FPROUND_PPCF128_F16
;
218 } else if (RetVT
== MVT::f32
) {
219 if (OpVT
== MVT::f64
)
220 return FPROUND_F64_F32
;
221 if (OpVT
== MVT::f80
)
222 return FPROUND_F80_F32
;
223 if (OpVT
== MVT::f128
)
224 return FPROUND_F128_F32
;
225 if (OpVT
== MVT::ppcf128
)
226 return FPROUND_PPCF128_F32
;
227 } else if (RetVT
== MVT::f64
) {
228 if (OpVT
== MVT::f80
)
229 return FPROUND_F80_F64
;
230 if (OpVT
== MVT::f128
)
231 return FPROUND_F128_F64
;
232 if (OpVT
== MVT::ppcf128
)
233 return FPROUND_PPCF128_F64
;
234 } else if (RetVT
== MVT::f80
) {
235 if (OpVT
== MVT::f128
)
236 return FPROUND_F128_F80
;
239 return UNKNOWN_LIBCALL
;
242 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
243 /// UNKNOWN_LIBCALL if there is none.
244 RTLIB::Libcall
RTLIB::getFPTOSINT(EVT OpVT
, EVT RetVT
) {
245 if (OpVT
== MVT::f32
) {
246 if (RetVT
== MVT::i32
)
247 return FPTOSINT_F32_I32
;
248 if (RetVT
== MVT::i64
)
249 return FPTOSINT_F32_I64
;
250 if (RetVT
== MVT::i128
)
251 return FPTOSINT_F32_I128
;
252 } else if (OpVT
== MVT::f64
) {
253 if (RetVT
== MVT::i32
)
254 return FPTOSINT_F64_I32
;
255 if (RetVT
== MVT::i64
)
256 return FPTOSINT_F64_I64
;
257 if (RetVT
== MVT::i128
)
258 return FPTOSINT_F64_I128
;
259 } else if (OpVT
== MVT::f80
) {
260 if (RetVT
== MVT::i32
)
261 return FPTOSINT_F80_I32
;
262 if (RetVT
== MVT::i64
)
263 return FPTOSINT_F80_I64
;
264 if (RetVT
== MVT::i128
)
265 return FPTOSINT_F80_I128
;
266 } else if (OpVT
== MVT::f128
) {
267 if (RetVT
== MVT::i32
)
268 return FPTOSINT_F128_I32
;
269 if (RetVT
== MVT::i64
)
270 return FPTOSINT_F128_I64
;
271 if (RetVT
== MVT::i128
)
272 return FPTOSINT_F128_I128
;
273 } else if (OpVT
== MVT::ppcf128
) {
274 if (RetVT
== MVT::i32
)
275 return FPTOSINT_PPCF128_I32
;
276 if (RetVT
== MVT::i64
)
277 return FPTOSINT_PPCF128_I64
;
278 if (RetVT
== MVT::i128
)
279 return FPTOSINT_PPCF128_I128
;
281 return UNKNOWN_LIBCALL
;
284 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
285 /// UNKNOWN_LIBCALL if there is none.
286 RTLIB::Libcall
RTLIB::getFPTOUINT(EVT OpVT
, EVT RetVT
) {
287 if (OpVT
== MVT::f32
) {
288 if (RetVT
== MVT::i32
)
289 return FPTOUINT_F32_I32
;
290 if (RetVT
== MVT::i64
)
291 return FPTOUINT_F32_I64
;
292 if (RetVT
== MVT::i128
)
293 return FPTOUINT_F32_I128
;
294 } else if (OpVT
== MVT::f64
) {
295 if (RetVT
== MVT::i32
)
296 return FPTOUINT_F64_I32
;
297 if (RetVT
== MVT::i64
)
298 return FPTOUINT_F64_I64
;
299 if (RetVT
== MVT::i128
)
300 return FPTOUINT_F64_I128
;
301 } else if (OpVT
== MVT::f80
) {
302 if (RetVT
== MVT::i32
)
303 return FPTOUINT_F80_I32
;
304 if (RetVT
== MVT::i64
)
305 return FPTOUINT_F80_I64
;
306 if (RetVT
== MVT::i128
)
307 return FPTOUINT_F80_I128
;
308 } else if (OpVT
== MVT::f128
) {
309 if (RetVT
== MVT::i32
)
310 return FPTOUINT_F128_I32
;
311 if (RetVT
== MVT::i64
)
312 return FPTOUINT_F128_I64
;
313 if (RetVT
== MVT::i128
)
314 return FPTOUINT_F128_I128
;
315 } else if (OpVT
== MVT::ppcf128
) {
316 if (RetVT
== MVT::i32
)
317 return FPTOUINT_PPCF128_I32
;
318 if (RetVT
== MVT::i64
)
319 return FPTOUINT_PPCF128_I64
;
320 if (RetVT
== MVT::i128
)
321 return FPTOUINT_PPCF128_I128
;
323 return UNKNOWN_LIBCALL
;
326 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
327 /// UNKNOWN_LIBCALL if there is none.
328 RTLIB::Libcall
RTLIB::getSINTTOFP(EVT OpVT
, EVT RetVT
) {
329 if (OpVT
== MVT::i32
) {
330 if (RetVT
== MVT::f32
)
331 return SINTTOFP_I32_F32
;
332 if (RetVT
== MVT::f64
)
333 return SINTTOFP_I32_F64
;
334 if (RetVT
== MVT::f80
)
335 return SINTTOFP_I32_F80
;
336 if (RetVT
== MVT::f128
)
337 return SINTTOFP_I32_F128
;
338 if (RetVT
== MVT::ppcf128
)
339 return SINTTOFP_I32_PPCF128
;
340 } else if (OpVT
== MVT::i64
) {
341 if (RetVT
== MVT::f32
)
342 return SINTTOFP_I64_F32
;
343 if (RetVT
== MVT::f64
)
344 return SINTTOFP_I64_F64
;
345 if (RetVT
== MVT::f80
)
346 return SINTTOFP_I64_F80
;
347 if (RetVT
== MVT::f128
)
348 return SINTTOFP_I64_F128
;
349 if (RetVT
== MVT::ppcf128
)
350 return SINTTOFP_I64_PPCF128
;
351 } else if (OpVT
== MVT::i128
) {
352 if (RetVT
== MVT::f32
)
353 return SINTTOFP_I128_F32
;
354 if (RetVT
== MVT::f64
)
355 return SINTTOFP_I128_F64
;
356 if (RetVT
== MVT::f80
)
357 return SINTTOFP_I128_F80
;
358 if (RetVT
== MVT::f128
)
359 return SINTTOFP_I128_F128
;
360 if (RetVT
== MVT::ppcf128
)
361 return SINTTOFP_I128_PPCF128
;
363 return UNKNOWN_LIBCALL
;
366 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
367 /// UNKNOWN_LIBCALL if there is none.
368 RTLIB::Libcall
RTLIB::getUINTTOFP(EVT OpVT
, EVT RetVT
) {
369 if (OpVT
== MVT::i32
) {
370 if (RetVT
== MVT::f32
)
371 return UINTTOFP_I32_F32
;
372 if (RetVT
== MVT::f64
)
373 return UINTTOFP_I32_F64
;
374 if (RetVT
== MVT::f80
)
375 return UINTTOFP_I32_F80
;
376 if (RetVT
== MVT::f128
)
377 return UINTTOFP_I32_F128
;
378 if (RetVT
== MVT::ppcf128
)
379 return UINTTOFP_I32_PPCF128
;
380 } else if (OpVT
== MVT::i64
) {
381 if (RetVT
== MVT::f32
)
382 return UINTTOFP_I64_F32
;
383 if (RetVT
== MVT::f64
)
384 return UINTTOFP_I64_F64
;
385 if (RetVT
== MVT::f80
)
386 return UINTTOFP_I64_F80
;
387 if (RetVT
== MVT::f128
)
388 return UINTTOFP_I64_F128
;
389 if (RetVT
== MVT::ppcf128
)
390 return UINTTOFP_I64_PPCF128
;
391 } else if (OpVT
== MVT::i128
) {
392 if (RetVT
== MVT::f32
)
393 return UINTTOFP_I128_F32
;
394 if (RetVT
== MVT::f64
)
395 return UINTTOFP_I128_F64
;
396 if (RetVT
== MVT::f80
)
397 return UINTTOFP_I128_F80
;
398 if (RetVT
== MVT::f128
)
399 return UINTTOFP_I128_F128
;
400 if (RetVT
== MVT::ppcf128
)
401 return UINTTOFP_I128_PPCF128
;
403 return UNKNOWN_LIBCALL
;
406 RTLIB::Libcall
RTLIB::getSYNC(unsigned Opc
, MVT VT
) {
407 #define OP_TO_LIBCALL(Name, Enum) \
409 switch (VT.SimpleTy) { \
411 return UNKNOWN_LIBCALL; \
425 OP_TO_LIBCALL(ISD::ATOMIC_SWAP
, SYNC_LOCK_TEST_AND_SET
)
426 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP
, SYNC_VAL_COMPARE_AND_SWAP
)
427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD
, SYNC_FETCH_AND_ADD
)
428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB
, SYNC_FETCH_AND_SUB
)
429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND
, SYNC_FETCH_AND_AND
)
430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR
, SYNC_FETCH_AND_OR
)
431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR
, SYNC_FETCH_AND_XOR
)
432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND
, SYNC_FETCH_AND_NAND
)
433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX
, SYNC_FETCH_AND_MAX
)
434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX
, SYNC_FETCH_AND_UMAX
)
435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN
, SYNC_FETCH_AND_MIN
)
436 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN
, SYNC_FETCH_AND_UMIN
)
441 return UNKNOWN_LIBCALL
;
444 RTLIB::Libcall
RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize
) {
445 switch (ElementSize
) {
447 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1
;
449 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2
;
451 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4
;
453 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8
;
455 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16
;
457 return UNKNOWN_LIBCALL
;
461 RTLIB::Libcall
RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize
) {
462 switch (ElementSize
) {
464 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1
;
466 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2
;
468 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4
;
470 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8
;
472 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16
;
474 return UNKNOWN_LIBCALL
;
478 RTLIB::Libcall
RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize
) {
479 switch (ElementSize
) {
481 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1
;
483 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2
;
485 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4
;
487 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8
;
489 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16
;
491 return UNKNOWN_LIBCALL
;
495 /// InitCmpLibcallCCs - Set default comparison libcall CC.
496 static void InitCmpLibcallCCs(ISD::CondCode
*CCs
) {
497 memset(CCs
, ISD::SETCC_INVALID
, sizeof(ISD::CondCode
)*RTLIB::UNKNOWN_LIBCALL
);
498 CCs
[RTLIB::OEQ_F32
] = ISD::SETEQ
;
499 CCs
[RTLIB::OEQ_F64
] = ISD::SETEQ
;
500 CCs
[RTLIB::OEQ_F128
] = ISD::SETEQ
;
501 CCs
[RTLIB::OEQ_PPCF128
] = ISD::SETEQ
;
502 CCs
[RTLIB::UNE_F32
] = ISD::SETNE
;
503 CCs
[RTLIB::UNE_F64
] = ISD::SETNE
;
504 CCs
[RTLIB::UNE_F128
] = ISD::SETNE
;
505 CCs
[RTLIB::UNE_PPCF128
] = ISD::SETNE
;
506 CCs
[RTLIB::OGE_F32
] = ISD::SETGE
;
507 CCs
[RTLIB::OGE_F64
] = ISD::SETGE
;
508 CCs
[RTLIB::OGE_F128
] = ISD::SETGE
;
509 CCs
[RTLIB::OGE_PPCF128
] = ISD::SETGE
;
510 CCs
[RTLIB::OLT_F32
] = ISD::SETLT
;
511 CCs
[RTLIB::OLT_F64
] = ISD::SETLT
;
512 CCs
[RTLIB::OLT_F128
] = ISD::SETLT
;
513 CCs
[RTLIB::OLT_PPCF128
] = ISD::SETLT
;
514 CCs
[RTLIB::OLE_F32
] = ISD::SETLE
;
515 CCs
[RTLIB::OLE_F64
] = ISD::SETLE
;
516 CCs
[RTLIB::OLE_F128
] = ISD::SETLE
;
517 CCs
[RTLIB::OLE_PPCF128
] = ISD::SETLE
;
518 CCs
[RTLIB::OGT_F32
] = ISD::SETGT
;
519 CCs
[RTLIB::OGT_F64
] = ISD::SETGT
;
520 CCs
[RTLIB::OGT_F128
] = ISD::SETGT
;
521 CCs
[RTLIB::OGT_PPCF128
] = ISD::SETGT
;
522 CCs
[RTLIB::UO_F32
] = ISD::SETNE
;
523 CCs
[RTLIB::UO_F64
] = ISD::SETNE
;
524 CCs
[RTLIB::UO_F128
] = ISD::SETNE
;
525 CCs
[RTLIB::UO_PPCF128
] = ISD::SETNE
;
526 CCs
[RTLIB::O_F32
] = ISD::SETEQ
;
527 CCs
[RTLIB::O_F64
] = ISD::SETEQ
;
528 CCs
[RTLIB::O_F128
] = ISD::SETEQ
;
529 CCs
[RTLIB::O_PPCF128
] = ISD::SETEQ
;
532 /// NOTE: The TargetMachine owns TLOF.
533 TargetLoweringBase::TargetLoweringBase(const TargetMachine
&tm
) : TM(tm
) {
536 // Perform these initializations only once.
537 MaxStoresPerMemset
= MaxStoresPerMemcpy
= MaxStoresPerMemmove
=
538 MaxLoadsPerMemcmp
= 8;
539 MaxGluedStoresPerMemcpy
= 0;
540 MaxStoresPerMemsetOptSize
= MaxStoresPerMemcpyOptSize
=
541 MaxStoresPerMemmoveOptSize
= MaxLoadsPerMemcmpOptSize
= 4;
542 UseUnderscoreSetJmp
= false;
543 UseUnderscoreLongJmp
= false;
544 HasMultipleConditionRegisters
= false;
545 HasExtractBitsInsn
= false;
546 JumpIsExpensive
= JumpIsExpensiveOverride
;
547 PredictableSelectIsExpensive
= false;
548 EnableExtLdPromotion
= false;
549 HasFloatingPointExceptions
= true;
550 StackPointerRegisterToSaveRestore
= 0;
551 BooleanContents
= UndefinedBooleanContent
;
552 BooleanFloatContents
= UndefinedBooleanContent
;
553 BooleanVectorContents
= UndefinedBooleanContent
;
554 SchedPreferenceInfo
= Sched::ILP
;
556 JumpBufAlignment
= 0;
557 MinFunctionAlignment
= 0;
558 PrefFunctionAlignment
= 0;
559 PrefLoopAlignment
= 0;
560 GatherAllAliasesMaxDepth
= 18;
561 MinStackArgumentAlignment
= 1;
562 // TODO: the default will be switched to 0 in the next commit, along
563 // with the Target-specific changes necessary.
564 MaxAtomicSizeInBitsSupported
= 1024;
566 MinCmpXchgSizeInBits
= 0;
567 SupportsUnalignedAtomics
= false;
569 std::fill(std::begin(LibcallRoutineNames
), std::end(LibcallRoutineNames
), nullptr);
571 InitLibcalls(TM
.getTargetTriple());
572 InitCmpLibcallCCs(CmpLibcallCCs
);
575 void TargetLoweringBase::initActions() {
576 // All operations default to being supported.
577 memset(OpActions
, 0, sizeof(OpActions
));
578 memset(LoadExtActions
, 0, sizeof(LoadExtActions
));
579 memset(TruncStoreActions
, 0, sizeof(TruncStoreActions
));
580 memset(IndexedModeActions
, 0, sizeof(IndexedModeActions
));
581 memset(CondCodeActions
, 0, sizeof(CondCodeActions
));
582 std::fill(std::begin(RegClassForVT
), std::end(RegClassForVT
), nullptr);
583 std::fill(std::begin(TargetDAGCombineArray
),
584 std::end(TargetDAGCombineArray
), 0);
586 // Set default actions for various operations.
587 for (MVT VT
: MVT::all_valuetypes()) {
588 // Default all indexed load / store to expand.
589 for (unsigned IM
= (unsigned)ISD::PRE_INC
;
590 IM
!= (unsigned)ISD::LAST_INDEXED_MODE
; ++IM
) {
591 setIndexedLoadAction(IM
, VT
, Expand
);
592 setIndexedStoreAction(IM
, VT
, Expand
);
595 // Most backends expect to see the node which just returns the value loaded.
596 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS
, VT
, Expand
);
598 // These operations default to expand.
599 setOperationAction(ISD::FGETSIGN
, VT
, Expand
);
600 setOperationAction(ISD::CONCAT_VECTORS
, VT
, Expand
);
601 setOperationAction(ISD::FMINNUM
, VT
, Expand
);
602 setOperationAction(ISD::FMAXNUM
, VT
, Expand
);
603 setOperationAction(ISD::FMINNAN
, VT
, Expand
);
604 setOperationAction(ISD::FMAXNAN
, VT
, Expand
);
605 setOperationAction(ISD::FMAD
, VT
, Expand
);
606 setOperationAction(ISD::SMIN
, VT
, Expand
);
607 setOperationAction(ISD::SMAX
, VT
, Expand
);
608 setOperationAction(ISD::UMIN
, VT
, Expand
);
609 setOperationAction(ISD::UMAX
, VT
, Expand
);
610 setOperationAction(ISD::ABS
, VT
, Expand
);
612 // Overflow operations default to expand
613 setOperationAction(ISD::SADDO
, VT
, Expand
);
614 setOperationAction(ISD::SSUBO
, VT
, Expand
);
615 setOperationAction(ISD::UADDO
, VT
, Expand
);
616 setOperationAction(ISD::USUBO
, VT
, Expand
);
617 setOperationAction(ISD::SMULO
, VT
, Expand
);
618 setOperationAction(ISD::UMULO
, VT
, Expand
);
620 // ADDCARRY operations default to expand
621 setOperationAction(ISD::ADDCARRY
, VT
, Expand
);
622 setOperationAction(ISD::SUBCARRY
, VT
, Expand
);
623 setOperationAction(ISD::SETCCCARRY
, VT
, Expand
);
625 // ADDC/ADDE/SUBC/SUBE default to expand.
626 setOperationAction(ISD::ADDC
, VT
, Expand
);
627 setOperationAction(ISD::ADDE
, VT
, Expand
);
628 setOperationAction(ISD::SUBC
, VT
, Expand
);
629 setOperationAction(ISD::SUBE
, VT
, Expand
);
631 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
632 setOperationAction(ISD::CTLZ_ZERO_UNDEF
, VT
, Expand
);
633 setOperationAction(ISD::CTTZ_ZERO_UNDEF
, VT
, Expand
);
635 setOperationAction(ISD::BITREVERSE
, VT
, Expand
);
637 // These library functions default to expand.
638 setOperationAction(ISD::FROUND
, VT
, Expand
);
639 setOperationAction(ISD::FPOWI
, VT
, Expand
);
641 // These operations default to expand for vector types.
643 setOperationAction(ISD::FCOPYSIGN
, VT
, Expand
);
644 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG
, VT
, Expand
);
645 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG
, VT
, Expand
);
646 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG
, VT
, Expand
);
649 // For most targets @llvm.get.dynamic.area.offset just returns 0.
650 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET
, VT
, Expand
);
653 // Most targets ignore the @llvm.prefetch intrinsic.
654 setOperationAction(ISD::PREFETCH
, MVT::Other
, Expand
);
656 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
657 setOperationAction(ISD::READCYCLECOUNTER
, MVT::i64
, Expand
);
659 // ConstantFP nodes default to expand. Targets can either change this to
660 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
661 // to optimize expansions for certain constants.
662 setOperationAction(ISD::ConstantFP
, MVT::f16
, Expand
);
663 setOperationAction(ISD::ConstantFP
, MVT::f32
, Expand
);
664 setOperationAction(ISD::ConstantFP
, MVT::f64
, Expand
);
665 setOperationAction(ISD::ConstantFP
, MVT::f80
, Expand
);
666 setOperationAction(ISD::ConstantFP
, MVT::f128
, Expand
);
668 // These library functions default to expand.
669 for (MVT VT
: {MVT::f32
, MVT::f64
, MVT::f128
}) {
670 setOperationAction(ISD::FCBRT
, VT
, Expand
);
671 setOperationAction(ISD::FLOG
, VT
, Expand
);
672 setOperationAction(ISD::FLOG2
, VT
, Expand
);
673 setOperationAction(ISD::FLOG10
, VT
, Expand
);
674 setOperationAction(ISD::FEXP
, VT
, Expand
);
675 setOperationAction(ISD::FEXP2
, VT
, Expand
);
676 setOperationAction(ISD::FFLOOR
, VT
, Expand
);
677 setOperationAction(ISD::FNEARBYINT
, VT
, Expand
);
678 setOperationAction(ISD::FCEIL
, VT
, Expand
);
679 setOperationAction(ISD::FRINT
, VT
, Expand
);
680 setOperationAction(ISD::FTRUNC
, VT
, Expand
);
681 setOperationAction(ISD::FROUND
, VT
, Expand
);
684 // Default ISD::TRAP to expand (which turns it into abort).
685 setOperationAction(ISD::TRAP
, MVT::Other
, Expand
);
687 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
688 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
689 setOperationAction(ISD::DEBUGTRAP
, MVT::Other
, Expand
);
692 MVT
TargetLoweringBase::getScalarShiftAmountTy(const DataLayout
&DL
,
694 return MVT::getIntegerVT(8 * DL
.getPointerSize(0));
697 EVT
TargetLoweringBase::getShiftAmountTy(EVT LHSTy
, const DataLayout
&DL
,
698 bool LegalTypes
) const {
699 assert(LHSTy
.isInteger() && "Shift amount is not an integer type!");
700 if (LHSTy
.isVector())
702 return LegalTypes
? getScalarShiftAmountTy(DL
, LHSTy
)
706 bool TargetLoweringBase::canOpTrap(unsigned Op
, EVT VT
) const {
707 assert(isTypeLegal(VT
));
719 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive
) {
720 // If the command-line option was specified, ignore this request.
721 if (!JumpIsExpensiveOverride
.getNumOccurrences())
722 JumpIsExpensive
= isExpensive
;
725 TargetLoweringBase::LegalizeKind
726 TargetLoweringBase::getTypeConversion(LLVMContext
&Context
, EVT VT
) const {
727 // If this is a simple type, use the ComputeRegisterProp mechanism.
729 MVT SVT
= VT
.getSimpleVT();
730 assert((unsigned)SVT
.SimpleTy
< array_lengthof(TransformToType
));
731 MVT NVT
= TransformToType
[SVT
.SimpleTy
];
732 LegalizeTypeAction LA
= ValueTypeActions
.getTypeAction(SVT
);
734 assert((LA
== TypeLegal
|| LA
== TypeSoftenFloat
||
735 ValueTypeActions
.getTypeAction(NVT
) != TypePromoteInteger
) &&
736 "Promote may not follow Expand or Promote");
738 if (LA
== TypeSplitVector
)
739 return LegalizeKind(LA
,
740 EVT::getVectorVT(Context
, SVT
.getVectorElementType(),
741 SVT
.getVectorNumElements() / 2));
742 if (LA
== TypeScalarizeVector
)
743 return LegalizeKind(LA
, SVT
.getVectorElementType());
744 return LegalizeKind(LA
, NVT
);
747 // Handle Extended Scalar Types.
748 if (!VT
.isVector()) {
749 assert(VT
.isInteger() && "Float types must be simple");
750 unsigned BitSize
= VT
.getSizeInBits();
751 // First promote to a power-of-two size, then expand if necessary.
752 if (BitSize
< 8 || !isPowerOf2_32(BitSize
)) {
753 EVT NVT
= VT
.getRoundIntegerType(Context
);
754 assert(NVT
!= VT
&& "Unable to round integer VT");
755 LegalizeKind NextStep
= getTypeConversion(Context
, NVT
);
756 // Avoid multi-step promotion.
757 if (NextStep
.first
== TypePromoteInteger
)
759 // Return rounded integer type.
760 return LegalizeKind(TypePromoteInteger
, NVT
);
763 return LegalizeKind(TypeExpandInteger
,
764 EVT::getIntegerVT(Context
, VT
.getSizeInBits() / 2));
767 // Handle vector types.
768 unsigned NumElts
= VT
.getVectorNumElements();
769 EVT EltVT
= VT
.getVectorElementType();
771 // Vectors with only one element are always scalarized.
773 return LegalizeKind(TypeScalarizeVector
, EltVT
);
775 // Try to widen vector elements until the element type is a power of two and
776 // promote it to a legal type later on, for example:
777 // <3 x i8> -> <4 x i8> -> <4 x i32>
778 if (EltVT
.isInteger()) {
779 // Vectors with a number of elements that is not a power of two are always
780 // widened, for example <3 x i8> -> <4 x i8>.
781 if (!VT
.isPow2VectorType()) {
782 NumElts
= (unsigned)NextPowerOf2(NumElts
);
783 EVT NVT
= EVT::getVectorVT(Context
, EltVT
, NumElts
);
784 return LegalizeKind(TypeWidenVector
, NVT
);
787 // Examine the element type.
788 LegalizeKind LK
= getTypeConversion(Context
, EltVT
);
790 // If type is to be expanded, split the vector.
791 // <4 x i140> -> <2 x i140>
792 if (LK
.first
== TypeExpandInteger
)
793 return LegalizeKind(TypeSplitVector
,
794 EVT::getVectorVT(Context
, EltVT
, NumElts
/ 2));
796 // Promote the integer element types until a legal vector type is found
797 // or until the element integer type is too big. If a legal type was not
798 // found, fallback to the usual mechanism of widening/splitting the
800 EVT OldEltVT
= EltVT
;
802 // Increase the bitwidth of the element to the next pow-of-two
803 // (which is greater than 8 bits).
804 EltVT
= EVT::getIntegerVT(Context
, 1 + EltVT
.getSizeInBits())
805 .getRoundIntegerType(Context
);
807 // Stop trying when getting a non-simple element type.
808 // Note that vector elements may be greater than legal vector element
809 // types. Example: X86 XMM registers hold 64bit element on 32bit
811 if (!EltVT
.isSimple())
814 // Build a new vector type and check if it is legal.
815 MVT NVT
= MVT::getVectorVT(EltVT
.getSimpleVT(), NumElts
);
816 // Found a legal promoted vector type.
817 if (NVT
!= MVT() && ValueTypeActions
.getTypeAction(NVT
) == TypeLegal
)
818 return LegalizeKind(TypePromoteInteger
,
819 EVT::getVectorVT(Context
, EltVT
, NumElts
));
822 // Reset the type to the unexpanded type if we did not find a legal vector
823 // type with a promoted vector element type.
827 // Try to widen the vector until a legal type is found.
828 // If there is no wider legal type, split the vector.
830 // Round up to the next power of 2.
831 NumElts
= (unsigned)NextPowerOf2(NumElts
);
833 // If there is no simple vector type with this many elements then there
834 // cannot be a larger legal vector type. Note that this assumes that
835 // there are no skipped intermediate vector types in the simple types.
836 if (!EltVT
.isSimple())
838 MVT LargerVector
= MVT::getVectorVT(EltVT
.getSimpleVT(), NumElts
);
839 if (LargerVector
== MVT())
842 // If this type is legal then widen the vector.
843 if (ValueTypeActions
.getTypeAction(LargerVector
) == TypeLegal
)
844 return LegalizeKind(TypeWidenVector
, LargerVector
);
847 // Widen odd vectors to next power of two.
848 if (!VT
.isPow2VectorType()) {
849 EVT NVT
= VT
.getPow2VectorType(Context
);
850 return LegalizeKind(TypeWidenVector
, NVT
);
853 // Vectors with illegal element types are expanded.
854 EVT NVT
= EVT::getVectorVT(Context
, EltVT
, VT
.getVectorNumElements() / 2);
855 return LegalizeKind(TypeSplitVector
, NVT
);
858 static unsigned getVectorTypeBreakdownMVT(MVT VT
, MVT
&IntermediateVT
,
859 unsigned &NumIntermediates
,
861 TargetLoweringBase
*TLI
) {
862 // Figure out the right, legal destination reg to copy into.
863 unsigned NumElts
= VT
.getVectorNumElements();
864 MVT EltTy
= VT
.getVectorElementType();
866 unsigned NumVectorRegs
= 1;
868 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
869 // could break down into LHS/RHS like LegalizeDAG does.
870 if (!isPowerOf2_32(NumElts
)) {
871 NumVectorRegs
= NumElts
;
875 // Divide the input until we get to a supported size. This will always
876 // end with a scalar if the target doesn't support vectors.
877 while (NumElts
> 1 && !TLI
->isTypeLegal(MVT::getVectorVT(EltTy
, NumElts
))) {
882 NumIntermediates
= NumVectorRegs
;
884 MVT NewVT
= MVT::getVectorVT(EltTy
, NumElts
);
885 if (!TLI
->isTypeLegal(NewVT
))
887 IntermediateVT
= NewVT
;
889 unsigned NewVTSize
= NewVT
.getSizeInBits();
891 // Convert sizes such as i33 to i64.
892 if (!isPowerOf2_32(NewVTSize
))
893 NewVTSize
= NextPowerOf2(NewVTSize
);
895 MVT DestVT
= TLI
->getRegisterType(NewVT
);
897 if (EVT(DestVT
).bitsLT(NewVT
)) // Value is expanded, e.g. i64 -> i16.
898 return NumVectorRegs
*(NewVTSize
/DestVT
.getSizeInBits());
900 // Otherwise, promotion or legal types use the same number of registers as
901 // the vector decimated to the appropriate level.
902 return NumVectorRegs
;
905 /// isLegalRC - Return true if the value types that can be represented by the
906 /// specified register class are all legal.
907 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo
&TRI
,
908 const TargetRegisterClass
&RC
) const {
909 for (auto I
= TRI
.legalclasstypes_begin(RC
); *I
!= MVT::Other
; ++I
)
915 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
916 /// sequence of memory operands that is recognized by PrologEpilogInserter.
918 TargetLoweringBase::emitPatchPoint(MachineInstr
&InitialMI
,
919 MachineBasicBlock
*MBB
) const {
920 MachineInstr
*MI
= &InitialMI
;
921 MachineFunction
&MF
= *MI
->getMF();
922 MachineFrameInfo
&MFI
= MF
.getFrameInfo();
924 // We're handling multiple types of operands here:
925 // PATCHPOINT MetaArgs - live-in, read only, direct
926 // STATEPOINT Deopt Spill - live-through, read only, indirect
927 // STATEPOINT Deopt Alloca - live-through, read only, direct
928 // (We're currently conservative and mark the deopt slots read/write in
930 // STATEPOINT GC Spill - live-through, read/write, indirect
931 // STATEPOINT GC Alloca - live-through, read/write, direct
932 // The live-in vs live-through is handled already (the live through ones are
933 // all stack slots), but we need to handle the different type of stackmap
934 // operands and memory effects here.
936 // MI changes inside this loop as we grow operands.
937 for(unsigned OperIdx
= 0; OperIdx
!= MI
->getNumOperands(); ++OperIdx
) {
938 MachineOperand
&MO
= MI
->getOperand(OperIdx
);
942 // foldMemoryOperand builds a new MI after replacing a single FI operand
943 // with the canonical set of five x86 addressing-mode operands.
944 int FI
= MO
.getIndex();
945 MachineInstrBuilder MIB
= BuildMI(MF
, MI
->getDebugLoc(), MI
->getDesc());
947 // Copy operands before the frame-index.
948 for (unsigned i
= 0; i
< OperIdx
; ++i
)
949 MIB
.add(MI
->getOperand(i
));
950 // Add frame index operands recognized by stackmaps.cpp
951 if (MFI
.isStatepointSpillSlotObjectIndex(FI
)) {
952 // indirect-mem-ref tag, size, #FI, offset.
953 // Used for spills inserted by StatepointLowering. This codepath is not
954 // used for patchpoints/stackmaps at all, for these spilling is done via
955 // foldMemoryOperand callback only.
956 assert(MI
->getOpcode() == TargetOpcode::STATEPOINT
&& "sanity");
957 MIB
.addImm(StackMaps::IndirectMemRefOp
);
958 MIB
.addImm(MFI
.getObjectSize(FI
));
959 MIB
.add(MI
->getOperand(OperIdx
));
962 // direct-mem-ref tag, #FI, offset.
963 // Used by patchpoint, and direct alloca arguments to statepoints
964 MIB
.addImm(StackMaps::DirectMemRefOp
);
965 MIB
.add(MI
->getOperand(OperIdx
));
968 // Copy the operands after the frame index.
969 for (unsigned i
= OperIdx
+ 1; i
!= MI
->getNumOperands(); ++i
)
970 MIB
.add(MI
->getOperand(i
));
972 // Inherit previous memory operands.
973 MIB
.cloneMemRefs(*MI
);
974 assert(MIB
->mayLoad() && "Folded a stackmap use to a non-load!");
976 // Add a new memory operand for this FI.
977 assert(MFI
.getObjectOffset(FI
) != -1);
979 auto Flags
= MachineMemOperand::MOLoad
;
980 if (MI
->getOpcode() == TargetOpcode::STATEPOINT
) {
981 Flags
|= MachineMemOperand::MOStore
;
982 Flags
|= MachineMemOperand::MOVolatile
;
984 MachineMemOperand
*MMO
= MF
.getMachineMemOperand(
985 MachinePointerInfo::getFixedStack(MF
, FI
), Flags
,
986 MF
.getDataLayout().getPointerSize(), MFI
.getObjectAlignment(FI
));
987 MIB
->addMemOperand(MF
, MMO
);
989 // Replace the instruction and update the operand index.
990 MBB
->insert(MachineBasicBlock::iterator(MI
), MIB
);
991 OperIdx
+= (MIB
->getNumOperands() - MI
->getNumOperands()) - 1;
992 MI
->eraseFromParent();
999 TargetLoweringBase::emitXRayCustomEvent(MachineInstr
&MI
,
1000 MachineBasicBlock
*MBB
) const {
1001 assert(MI
.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL
&&
1002 "Called emitXRayCustomEvent on the wrong MI!");
1003 auto &MF
= *MI
.getMF();
1004 auto MIB
= BuildMI(MF
, MI
.getDebugLoc(), MI
.getDesc());
1005 for (unsigned OpIdx
= 0; OpIdx
!= MI
.getNumOperands(); ++OpIdx
)
1006 MIB
.add(MI
.getOperand(OpIdx
));
1008 MBB
->insert(MachineBasicBlock::iterator(MI
), MIB
);
1009 MI
.eraseFromParent();
1014 TargetLoweringBase::emitXRayTypedEvent(MachineInstr
&MI
,
1015 MachineBasicBlock
*MBB
) const {
1016 assert(MI
.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL
&&
1017 "Called emitXRayTypedEvent on the wrong MI!");
1018 auto &MF
= *MI
.getMF();
1019 auto MIB
= BuildMI(MF
, MI
.getDebugLoc(), MI
.getDesc());
1020 for (unsigned OpIdx
= 0; OpIdx
!= MI
.getNumOperands(); ++OpIdx
)
1021 MIB
.add(MI
.getOperand(OpIdx
));
1023 MBB
->insert(MachineBasicBlock::iterator(MI
), MIB
);
1024 MI
.eraseFromParent();
1028 /// findRepresentativeClass - Return the largest legal super-reg register class
1029 /// of the register class for the specified type and its associated "cost".
1030 // This function is in TargetLowering because it uses RegClassForVT which would
1031 // need to be moved to TargetRegisterInfo and would necessitate moving
1032 // isTypeLegal over as well - a massive change that would just require
1033 // TargetLowering having a TargetRegisterInfo class member that it would use.
1034 std::pair
<const TargetRegisterClass
*, uint8_t>
1035 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo
*TRI
,
1037 const TargetRegisterClass
*RC
= RegClassForVT
[VT
.SimpleTy
];
1039 return std::make_pair(RC
, 0);
1041 // Compute the set of all super-register classes.
1042 BitVector
SuperRegRC(TRI
->getNumRegClasses());
1043 for (SuperRegClassIterator
RCI(RC
, TRI
); RCI
.isValid(); ++RCI
)
1044 SuperRegRC
.setBitsInMask(RCI
.getMask());
1046 // Find the first legal register class with the largest spill size.
1047 const TargetRegisterClass
*BestRC
= RC
;
1048 for (unsigned i
: SuperRegRC
.set_bits()) {
1049 const TargetRegisterClass
*SuperRC
= TRI
->getRegClass(i
);
1050 // We want the largest possible spill size.
1051 if (TRI
->getSpillSize(*SuperRC
) <= TRI
->getSpillSize(*BestRC
))
1053 if (!isLegalRC(*TRI
, *SuperRC
))
1057 return std::make_pair(BestRC
, 1);
1060 /// computeRegisterProperties - Once all of the register classes are added,
1061 /// this allows us to compute derived properties we expose.
1062 void TargetLoweringBase::computeRegisterProperties(
1063 const TargetRegisterInfo
*TRI
) {
1064 static_assert(MVT::LAST_VALUETYPE
<= MVT::MAX_ALLOWED_VALUETYPE
,
1065 "Too many value types for ValueTypeActions to hold!");
1067 // Everything defaults to needing one register.
1068 for (unsigned i
= 0; i
!= MVT::LAST_VALUETYPE
; ++i
) {
1069 NumRegistersForVT
[i
] = 1;
1070 RegisterTypeForVT
[i
] = TransformToType
[i
] = (MVT::SimpleValueType
)i
;
1072 // ...except isVoid, which doesn't need any registers.
1073 NumRegistersForVT
[MVT::isVoid
] = 0;
1075 // Find the largest integer register class.
1076 unsigned LargestIntReg
= MVT::LAST_INTEGER_VALUETYPE
;
1077 for (; RegClassForVT
[LargestIntReg
] == nullptr; --LargestIntReg
)
1078 assert(LargestIntReg
!= MVT::i1
&& "No integer registers defined!");
1080 // Every integer value type larger than this largest register takes twice as
1081 // many registers to represent as the previous ValueType.
1082 for (unsigned ExpandedReg
= LargestIntReg
+ 1;
1083 ExpandedReg
<= MVT::LAST_INTEGER_VALUETYPE
; ++ExpandedReg
) {
1084 NumRegistersForVT
[ExpandedReg
] = 2*NumRegistersForVT
[ExpandedReg
-1];
1085 RegisterTypeForVT
[ExpandedReg
] = (MVT::SimpleValueType
)LargestIntReg
;
1086 TransformToType
[ExpandedReg
] = (MVT::SimpleValueType
)(ExpandedReg
- 1);
1087 ValueTypeActions
.setTypeAction((MVT::SimpleValueType
)ExpandedReg
,
1091 // Inspect all of the ValueType's smaller than the largest integer
1092 // register to see which ones need promotion.
1093 unsigned LegalIntReg
= LargestIntReg
;
1094 for (unsigned IntReg
= LargestIntReg
- 1;
1095 IntReg
>= (unsigned)MVT::i1
; --IntReg
) {
1096 MVT IVT
= (MVT::SimpleValueType
)IntReg
;
1097 if (isTypeLegal(IVT
)) {
1098 LegalIntReg
= IntReg
;
1100 RegisterTypeForVT
[IntReg
] = TransformToType
[IntReg
] =
1101 (const MVT::SimpleValueType
)LegalIntReg
;
1102 ValueTypeActions
.setTypeAction(IVT
, TypePromoteInteger
);
1106 // ppcf128 type is really two f64's.
1107 if (!isTypeLegal(MVT::ppcf128
)) {
1108 if (isTypeLegal(MVT::f64
)) {
1109 NumRegistersForVT
[MVT::ppcf128
] = 2*NumRegistersForVT
[MVT::f64
];
1110 RegisterTypeForVT
[MVT::ppcf128
] = MVT::f64
;
1111 TransformToType
[MVT::ppcf128
] = MVT::f64
;
1112 ValueTypeActions
.setTypeAction(MVT::ppcf128
, TypeExpandFloat
);
1114 NumRegistersForVT
[MVT::ppcf128
] = NumRegistersForVT
[MVT::i128
];
1115 RegisterTypeForVT
[MVT::ppcf128
] = RegisterTypeForVT
[MVT::i128
];
1116 TransformToType
[MVT::ppcf128
] = MVT::i128
;
1117 ValueTypeActions
.setTypeAction(MVT::ppcf128
, TypeSoftenFloat
);
1121 // Decide how to handle f128. If the target does not have native f128 support,
1122 // expand it to i128 and we will be generating soft float library calls.
1123 if (!isTypeLegal(MVT::f128
)) {
1124 NumRegistersForVT
[MVT::f128
] = NumRegistersForVT
[MVT::i128
];
1125 RegisterTypeForVT
[MVT::f128
] = RegisterTypeForVT
[MVT::i128
];
1126 TransformToType
[MVT::f128
] = MVT::i128
;
1127 ValueTypeActions
.setTypeAction(MVT::f128
, TypeSoftenFloat
);
1130 // Decide how to handle f64. If the target does not have native f64 support,
1131 // expand it to i64 and we will be generating soft float library calls.
1132 if (!isTypeLegal(MVT::f64
)) {
1133 NumRegistersForVT
[MVT::f64
] = NumRegistersForVT
[MVT::i64
];
1134 RegisterTypeForVT
[MVT::f64
] = RegisterTypeForVT
[MVT::i64
];
1135 TransformToType
[MVT::f64
] = MVT::i64
;
1136 ValueTypeActions
.setTypeAction(MVT::f64
, TypeSoftenFloat
);
1139 // Decide how to handle f32. If the target does not have native f32 support,
1140 // expand it to i32 and we will be generating soft float library calls.
1141 if (!isTypeLegal(MVT::f32
)) {
1142 NumRegistersForVT
[MVT::f32
] = NumRegistersForVT
[MVT::i32
];
1143 RegisterTypeForVT
[MVT::f32
] = RegisterTypeForVT
[MVT::i32
];
1144 TransformToType
[MVT::f32
] = MVT::i32
;
1145 ValueTypeActions
.setTypeAction(MVT::f32
, TypeSoftenFloat
);
1148 // Decide how to handle f16. If the target does not have native f16 support,
1149 // promote it to f32, because there are no f16 library calls (except for
1151 if (!isTypeLegal(MVT::f16
)) {
1152 NumRegistersForVT
[MVT::f16
] = NumRegistersForVT
[MVT::f32
];
1153 RegisterTypeForVT
[MVT::f16
] = RegisterTypeForVT
[MVT::f32
];
1154 TransformToType
[MVT::f16
] = MVT::f32
;
1155 ValueTypeActions
.setTypeAction(MVT::f16
, TypePromoteFloat
);
1158 // Loop over all of the vector value types to see which need transformations.
1159 for (unsigned i
= MVT::FIRST_VECTOR_VALUETYPE
;
1160 i
<= (unsigned)MVT::LAST_VECTOR_VALUETYPE
; ++i
) {
1161 MVT VT
= (MVT::SimpleValueType
) i
;
1162 if (isTypeLegal(VT
))
1165 MVT EltVT
= VT
.getVectorElementType();
1166 unsigned NElts
= VT
.getVectorNumElements();
1167 bool IsLegalWiderType
= false;
1168 LegalizeTypeAction PreferredAction
= getPreferredVectorAction(VT
);
1169 switch (PreferredAction
) {
1170 case TypePromoteInteger
:
1171 // Try to promote the elements of integer vectors. If no legal
1172 // promotion was found, fall through to the widen-vector method.
1173 for (unsigned nVT
= i
+ 1; nVT
<= MVT::LAST_INTEGER_VECTOR_VALUETYPE
; ++nVT
) {
1174 MVT SVT
= (MVT::SimpleValueType
) nVT
;
1175 // Promote vectors of integers to vectors with the same number
1176 // of elements, with a wider element type.
1177 if (SVT
.getScalarSizeInBits() > EltVT
.getSizeInBits() &&
1178 SVT
.getVectorNumElements() == NElts
&& isTypeLegal(SVT
)) {
1179 TransformToType
[i
] = SVT
;
1180 RegisterTypeForVT
[i
] = SVT
;
1181 NumRegistersForVT
[i
] = 1;
1182 ValueTypeActions
.setTypeAction(VT
, TypePromoteInteger
);
1183 IsLegalWiderType
= true;
1187 if (IsLegalWiderType
)
1191 case TypeWidenVector
:
1192 // Try to widen the vector.
1193 for (unsigned nVT
= i
+ 1; nVT
<= MVT::LAST_VECTOR_VALUETYPE
; ++nVT
) {
1194 MVT SVT
= (MVT::SimpleValueType
) nVT
;
1195 if (SVT
.getVectorElementType() == EltVT
1196 && SVT
.getVectorNumElements() > NElts
&& isTypeLegal(SVT
)) {
1197 TransformToType
[i
] = SVT
;
1198 RegisterTypeForVT
[i
] = SVT
;
1199 NumRegistersForVT
[i
] = 1;
1200 ValueTypeActions
.setTypeAction(VT
, TypeWidenVector
);
1201 IsLegalWiderType
= true;
1205 if (IsLegalWiderType
)
1209 case TypeSplitVector
:
1210 case TypeScalarizeVector
: {
1213 unsigned NumIntermediates
;
1214 NumRegistersForVT
[i
] = getVectorTypeBreakdownMVT(VT
, IntermediateVT
,
1215 NumIntermediates
, RegisterVT
, this);
1216 RegisterTypeForVT
[i
] = RegisterVT
;
1218 MVT NVT
= VT
.getPow2VectorType();
1220 // Type is already a power of 2. The default action is to split.
1221 TransformToType
[i
] = MVT::Other
;
1222 if (PreferredAction
== TypeScalarizeVector
)
1223 ValueTypeActions
.setTypeAction(VT
, TypeScalarizeVector
);
1224 else if (PreferredAction
== TypeSplitVector
)
1225 ValueTypeActions
.setTypeAction(VT
, TypeSplitVector
);
1227 // Set type action according to the number of elements.
1228 ValueTypeActions
.setTypeAction(VT
, NElts
== 1 ? TypeScalarizeVector
1231 TransformToType
[i
] = NVT
;
1232 ValueTypeActions
.setTypeAction(VT
, TypeWidenVector
);
1237 llvm_unreachable("Unknown vector legalization action!");
1241 // Determine the 'representative' register class for each value type.
1242 // An representative register class is the largest (meaning one which is
1243 // not a sub-register class / subreg register class) legal register class for
1244 // a group of value types. For example, on i386, i8, i16, and i32
1245 // representative would be GR32; while on x86_64 it's GR64.
1246 for (unsigned i
= 0; i
!= MVT::LAST_VALUETYPE
; ++i
) {
1247 const TargetRegisterClass
* RRC
;
1249 std::tie(RRC
, Cost
) = findRepresentativeClass(TRI
, (MVT::SimpleValueType
)i
);
1250 RepRegClassForVT
[i
] = RRC
;
1251 RepRegClassCostForVT
[i
] = Cost
;
1255 EVT
TargetLoweringBase::getSetCCResultType(const DataLayout
&DL
, LLVMContext
&,
1257 assert(!VT
.isVector() && "No default SetCC type for vectors!");
1258 return getPointerTy(DL
).SimpleTy
;
1261 MVT::SimpleValueType
TargetLoweringBase::getCmpLibcallReturnType() const {
1262 return MVT::i32
; // return the default value
1265 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1266 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1267 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1268 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1270 /// This method returns the number of registers needed, and the VT for each
1271 /// register. It also returns the VT and quantity of the intermediate values
1272 /// before they are promoted/expanded.
1273 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext
&Context
, EVT VT
,
1274 EVT
&IntermediateVT
,
1275 unsigned &NumIntermediates
,
1276 MVT
&RegisterVT
) const {
1277 unsigned NumElts
= VT
.getVectorNumElements();
1279 // If there is a wider vector type with the same element type as this one,
1280 // or a promoted vector type that has the same number of elements which
1281 // are wider, then we should convert to that legal vector type.
1282 // This handles things like <2 x float> -> <4 x float> and
1283 // <4 x i1> -> <4 x i32>.
1284 LegalizeTypeAction TA
= getTypeAction(Context
, VT
);
1285 if (NumElts
!= 1 && (TA
== TypeWidenVector
|| TA
== TypePromoteInteger
)) {
1286 EVT RegisterEVT
= getTypeToTransformTo(Context
, VT
);
1287 if (isTypeLegal(RegisterEVT
)) {
1288 IntermediateVT
= RegisterEVT
;
1289 RegisterVT
= RegisterEVT
.getSimpleVT();
1290 NumIntermediates
= 1;
1295 // Figure out the right, legal destination reg to copy into.
1296 EVT EltTy
= VT
.getVectorElementType();
1298 unsigned NumVectorRegs
= 1;
1300 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1301 // could break down into LHS/RHS like LegalizeDAG does.
1302 if (!isPowerOf2_32(NumElts
)) {
1303 NumVectorRegs
= NumElts
;
1307 // Divide the input until we get to a supported size. This will always
1308 // end with a scalar if the target doesn't support vectors.
1309 while (NumElts
> 1 && !isTypeLegal(
1310 EVT::getVectorVT(Context
, EltTy
, NumElts
))) {
1312 NumVectorRegs
<<= 1;
1315 NumIntermediates
= NumVectorRegs
;
1317 EVT NewVT
= EVT::getVectorVT(Context
, EltTy
, NumElts
);
1318 if (!isTypeLegal(NewVT
))
1320 IntermediateVT
= NewVT
;
1322 MVT DestVT
= getRegisterType(Context
, NewVT
);
1323 RegisterVT
= DestVT
;
1324 unsigned NewVTSize
= NewVT
.getSizeInBits();
1326 // Convert sizes such as i33 to i64.
1327 if (!isPowerOf2_32(NewVTSize
))
1328 NewVTSize
= NextPowerOf2(NewVTSize
);
1330 if (EVT(DestVT
).bitsLT(NewVT
)) // Value is expanded, e.g. i64 -> i16.
1331 return NumVectorRegs
*(NewVTSize
/DestVT
.getSizeInBits());
1333 // Otherwise, promotion or legal types use the same number of registers as
1334 // the vector decimated to the appropriate level.
1335 return NumVectorRegs
;
1338 /// Get the EVTs and ArgFlags collections that represent the legalized return
1339 /// type of the given function. This does not require a DAG or a return value,
1340 /// and is suitable for use before any DAGs for the function are constructed.
1341 /// TODO: Move this out of TargetLowering.cpp.
1342 void llvm::GetReturnInfo(CallingConv::ID CC
, Type
*ReturnType
,
1344 SmallVectorImpl
<ISD::OutputArg
> &Outs
,
1345 const TargetLowering
&TLI
, const DataLayout
&DL
) {
1346 SmallVector
<EVT
, 4> ValueVTs
;
1347 ComputeValueVTs(TLI
, DL
, ReturnType
, ValueVTs
);
1348 unsigned NumValues
= ValueVTs
.size();
1349 if (NumValues
== 0) return;
1351 for (unsigned j
= 0, f
= NumValues
; j
!= f
; ++j
) {
1352 EVT VT
= ValueVTs
[j
];
1353 ISD::NodeType ExtendKind
= ISD::ANY_EXTEND
;
1355 if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::SExt
))
1356 ExtendKind
= ISD::SIGN_EXTEND
;
1357 else if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::ZExt
))
1358 ExtendKind
= ISD::ZERO_EXTEND
;
1360 // FIXME: C calling convention requires the return type to be promoted to
1361 // at least 32-bit. But this is not necessary for non-C calling
1362 // conventions. The frontend should mark functions whose return values
1363 // require promoting with signext or zeroext attributes.
1364 if (ExtendKind
!= ISD::ANY_EXTEND
&& VT
.isInteger()) {
1365 MVT MinVT
= TLI
.getRegisterType(ReturnType
->getContext(), MVT::i32
);
1366 if (VT
.bitsLT(MinVT
))
1371 TLI
.getNumRegistersForCallingConv(ReturnType
->getContext(), CC
, VT
);
1373 TLI
.getRegisterTypeForCallingConv(ReturnType
->getContext(), CC
, VT
);
1375 // 'inreg' on function refers to return value
1376 ISD::ArgFlagsTy Flags
= ISD::ArgFlagsTy();
1377 if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::InReg
))
1380 // Propagate extension type if any
1381 if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::SExt
))
1383 else if (attr
.hasAttribute(AttributeList::ReturnIndex
, Attribute::ZExt
))
1386 for (unsigned i
= 0; i
< NumParts
; ++i
)
1387 Outs
.push_back(ISD::OutputArg(Flags
, PartVT
, VT
, /*isFixed=*/true, 0, 0));
1391 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1392 /// function arguments in the caller parameter area. This is the actual
1393 /// alignment, not its logarithm.
1394 unsigned TargetLoweringBase::getByValTypeAlignment(Type
*Ty
,
1395 const DataLayout
&DL
) const {
1396 return DL
.getABITypeAlignment(Ty
);
1399 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext
&Context
,
1400 const DataLayout
&DL
, EVT VT
,
1404 // Check if the specified alignment is sufficient based on the data layout.
1405 // TODO: While using the data layout works in practice, a better solution
1406 // would be to implement this check directly (make this a virtual function).
1407 // For example, the ABI alignment may change based on software platform while
1408 // this function should only be affected by hardware implementation.
1409 Type
*Ty
= VT
.getTypeForEVT(Context
);
1410 if (Alignment
>= DL
.getABITypeAlignment(Ty
)) {
1411 // Assume that an access that meets the ABI-specified alignment is fast.
1412 if (Fast
!= nullptr)
1417 // This is a misaligned access.
1418 return allowsMisalignedMemoryAccesses(VT
, AddrSpace
, Alignment
, Fast
);
1421 BranchProbability
TargetLoweringBase::getPredictableBranchThreshold() const {
1422 return BranchProbability(MinPercentageForPredictableBranch
, 100);
1425 //===----------------------------------------------------------------------===//
1426 // TargetTransformInfo Helpers
1427 //===----------------------------------------------------------------------===//
1429 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode
) const {
1430 enum InstructionOpcodes
{
1431 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1432 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1433 #include "llvm/IR/Instruction.def"
1435 switch (static_cast<InstructionOpcodes
>(Opcode
)) {
1438 case Switch
: return 0;
1439 case IndirectBr
: return 0;
1440 case Invoke
: return 0;
1441 case Resume
: return 0;
1442 case Unreachable
: return 0;
1443 case CleanupRet
: return 0;
1444 case CatchRet
: return 0;
1445 case CatchPad
: return 0;
1446 case CatchSwitch
: return 0;
1447 case CleanupPad
: return 0;
1448 case Add
: return ISD::ADD
;
1449 case FAdd
: return ISD::FADD
;
1450 case Sub
: return ISD::SUB
;
1451 case FSub
: return ISD::FSUB
;
1452 case Mul
: return ISD::MUL
;
1453 case FMul
: return ISD::FMUL
;
1454 case UDiv
: return ISD::UDIV
;
1455 case SDiv
: return ISD::SDIV
;
1456 case FDiv
: return ISD::FDIV
;
1457 case URem
: return ISD::UREM
;
1458 case SRem
: return ISD::SREM
;
1459 case FRem
: return ISD::FREM
;
1460 case Shl
: return ISD::SHL
;
1461 case LShr
: return ISD::SRL
;
1462 case AShr
: return ISD::SRA
;
1463 case And
: return ISD::AND
;
1464 case Or
: return ISD::OR
;
1465 case Xor
: return ISD::XOR
;
1466 case Alloca
: return 0;
1467 case Load
: return ISD::LOAD
;
1468 case Store
: return ISD::STORE
;
1469 case GetElementPtr
: return 0;
1470 case Fence
: return 0;
1471 case AtomicCmpXchg
: return 0;
1472 case AtomicRMW
: return 0;
1473 case Trunc
: return ISD::TRUNCATE
;
1474 case ZExt
: return ISD::ZERO_EXTEND
;
1475 case SExt
: return ISD::SIGN_EXTEND
;
1476 case FPToUI
: return ISD::FP_TO_UINT
;
1477 case FPToSI
: return ISD::FP_TO_SINT
;
1478 case UIToFP
: return ISD::UINT_TO_FP
;
1479 case SIToFP
: return ISD::SINT_TO_FP
;
1480 case FPTrunc
: return ISD::FP_ROUND
;
1481 case FPExt
: return ISD::FP_EXTEND
;
1482 case PtrToInt
: return ISD::BITCAST
;
1483 case IntToPtr
: return ISD::BITCAST
;
1484 case BitCast
: return ISD::BITCAST
;
1485 case AddrSpaceCast
: return ISD::ADDRSPACECAST
;
1486 case ICmp
: return ISD::SETCC
;
1487 case FCmp
: return ISD::SETCC
;
1489 case Call
: return 0;
1490 case Select
: return ISD::SELECT
;
1491 case UserOp1
: return 0;
1492 case UserOp2
: return 0;
1493 case VAArg
: return 0;
1494 case ExtractElement
: return ISD::EXTRACT_VECTOR_ELT
;
1495 case InsertElement
: return ISD::INSERT_VECTOR_ELT
;
1496 case ShuffleVector
: return ISD::VECTOR_SHUFFLE
;
1497 case ExtractValue
: return ISD::MERGE_VALUES
;
1498 case InsertValue
: return ISD::MERGE_VALUES
;
1499 case LandingPad
: return 0;
1502 llvm_unreachable("Unknown instruction type encountered!");
1506 TargetLoweringBase::getTypeLegalizationCost(const DataLayout
&DL
,
1508 LLVMContext
&C
= Ty
->getContext();
1509 EVT MTy
= getValueType(DL
, Ty
);
1512 // We keep legalizing the type until we find a legal kind. We assume that
1513 // the only operation that costs anything is the split. After splitting
1514 // we need to handle two types.
1516 LegalizeKind LK
= getTypeConversion(C
, MTy
);
1518 if (LK
.first
== TypeLegal
)
1519 return std::make_pair(Cost
, MTy
.getSimpleVT());
1521 if (LK
.first
== TypeSplitVector
|| LK
.first
== TypeExpandInteger
)
1524 // Do not loop with f128 type.
1525 if (MTy
== LK
.second
)
1526 return std::make_pair(Cost
, MTy
.getSimpleVT());
1528 // Keep legalizing the type.
1533 Value
*TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder
<> &IRB
,
1534 bool UseTLS
) const {
1535 // compiler-rt provides a variable with a magic name. Targets that do not
1536 // link with compiler-rt may also provide such a variable.
1537 Module
*M
= IRB
.GetInsertBlock()->getParent()->getParent();
1538 const char *UnsafeStackPtrVar
= "__safestack_unsafe_stack_ptr";
1539 auto UnsafeStackPtr
=
1540 dyn_cast_or_null
<GlobalVariable
>(M
->getNamedValue(UnsafeStackPtrVar
));
1542 Type
*StackPtrTy
= Type::getInt8PtrTy(M
->getContext());
1544 if (!UnsafeStackPtr
) {
1545 auto TLSModel
= UseTLS
?
1546 GlobalValue::InitialExecTLSModel
:
1547 GlobalValue::NotThreadLocal
;
1548 // The global variable is not defined yet, define it ourselves.
1549 // We use the initial-exec TLS model because we do not support the
1550 // variable living anywhere other than in the main executable.
1551 UnsafeStackPtr
= new GlobalVariable(
1552 *M
, StackPtrTy
, false, GlobalValue::ExternalLinkage
, nullptr,
1553 UnsafeStackPtrVar
, nullptr, TLSModel
);
1555 // The variable exists, check its type and attributes.
1556 if (UnsafeStackPtr
->getValueType() != StackPtrTy
)
1557 report_fatal_error(Twine(UnsafeStackPtrVar
) + " must have void* type");
1558 if (UseTLS
!= UnsafeStackPtr
->isThreadLocal())
1559 report_fatal_error(Twine(UnsafeStackPtrVar
) + " must " +
1560 (UseTLS
? "" : "not ") + "be thread-local");
1562 return UnsafeStackPtr
;
1565 Value
*TargetLoweringBase::getSafeStackPointerLocation(IRBuilder
<> &IRB
) const {
1566 if (!TM
.getTargetTriple().isAndroid())
1567 return getDefaultSafeStackPointerLocation(IRB
, true);
1569 // Android provides a libc function to retrieve the address of the current
1570 // thread's unsafe stack pointer.
1571 Module
*M
= IRB
.GetInsertBlock()->getParent()->getParent();
1572 Type
*StackPtrTy
= Type::getInt8PtrTy(M
->getContext());
1573 Value
*Fn
= M
->getOrInsertFunction("__safestack_pointer_address",
1574 StackPtrTy
->getPointerTo(0));
1575 return IRB
.CreateCall(Fn
);
1578 //===----------------------------------------------------------------------===//
1579 // Loop Strength Reduction hooks
1580 //===----------------------------------------------------------------------===//
1582 /// isLegalAddressingMode - Return true if the addressing mode represented
1583 /// by AM is legal for this target, for a load/store of the specified type.
1584 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout
&DL
,
1585 const AddrMode
&AM
, Type
*Ty
,
1586 unsigned AS
, Instruction
*I
) const {
1587 // The default implementation of this implements a conservative RISCy, r+r and
1590 // Allows a sign-extended 16-bit immediate field.
1591 if (AM
.BaseOffs
<= -(1LL << 16) || AM
.BaseOffs
>= (1LL << 16)-1)
1594 // No global is ever allowed as a base.
1598 // Only support r+r,
1600 case 0: // "r+i" or just "i", depending on HasBaseReg.
1603 if (AM
.HasBaseReg
&& AM
.BaseOffs
) // "r+r+i" is not allowed.
1605 // Otherwise we have r+r or r+i.
1608 if (AM
.HasBaseReg
|| AM
.BaseOffs
) // 2*r+r or 2*r+i is not allowed.
1610 // Allow 2*r as r+r.
1612 default: // Don't allow n * r
1619 //===----------------------------------------------------------------------===//
1621 //===----------------------------------------------------------------------===//
1623 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1624 // so that SelectionDAG handle SSP.
1625 Value
*TargetLoweringBase::getIRStackGuard(IRBuilder
<> &IRB
) const {
1626 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1627 Module
&M
= *IRB
.GetInsertBlock()->getParent()->getParent();
1628 PointerType
*PtrTy
= Type::getInt8PtrTy(M
.getContext());
1629 return M
.getOrInsertGlobal("__guard_local", PtrTy
);
1634 // Currently only support "standard" __stack_chk_guard.
1635 // TODO: add LOAD_STACK_GUARD support.
1636 void TargetLoweringBase::insertSSPDeclarations(Module
&M
) const {
1637 if (!M
.getNamedValue("__stack_chk_guard"))
1638 new GlobalVariable(M
, Type::getInt8PtrTy(M
.getContext()), false,
1639 GlobalVariable::ExternalLinkage
,
1640 nullptr, "__stack_chk_guard");
1643 // Currently only support "standard" __stack_chk_guard.
1644 // TODO: add LOAD_STACK_GUARD support.
1645 Value
*TargetLoweringBase::getSDagStackGuard(const Module
&M
) const {
1646 return M
.getNamedValue("__stack_chk_guard");
1649 Value
*TargetLoweringBase::getSSPStackGuardCheck(const Module
&M
) const {
1653 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1654 return MinimumJumpTableEntries
;
1657 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val
) {
1658 MinimumJumpTableEntries
= Val
;
1661 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize
) const {
1662 return OptForSize
? OptsizeJumpTableDensity
: JumpTableDensity
;
1665 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1666 return MaximumJumpTableSize
;
1669 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val
) {
1670 MaximumJumpTableSize
= Val
;
1673 //===----------------------------------------------------------------------===//
1674 // Reciprocal Estimates
1675 //===----------------------------------------------------------------------===//
1677 /// Get the reciprocal estimate attribute string for a function that will
1678 /// override the target defaults.
1679 static StringRef
getRecipEstimateForFunc(MachineFunction
&MF
) {
1680 const Function
&F
= MF
.getFunction();
1681 return F
.getFnAttribute("reciprocal-estimates").getValueAsString();
1684 /// Construct a string for the given reciprocal operation of the given type.
1685 /// This string should match the corresponding option to the front-end's
1686 /// "-mrecip" flag assuming those strings have been passed through in an
1687 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1688 static std::string
getReciprocalOpName(bool IsSqrt
, EVT VT
) {
1689 std::string Name
= VT
.isVector() ? "vec-" : "";
1691 Name
+= IsSqrt
? "sqrt" : "div";
1693 // TODO: Handle "half" or other float types?
1694 if (VT
.getScalarType() == MVT::f64
) {
1697 assert(VT
.getScalarType() == MVT::f32
&&
1698 "Unexpected FP type for reciprocal estimate");
1705 /// Return the character position and value (a single numeric character) of a
1706 /// customized refinement operation in the input string if it exists. Return
1707 /// false if there is no customized refinement step count.
1708 static bool parseRefinementStep(StringRef In
, size_t &Position
,
1710 const char RefStepToken
= ':';
1711 Position
= In
.find(RefStepToken
);
1712 if (Position
== StringRef::npos
)
1715 StringRef RefStepString
= In
.substr(Position
+ 1);
1716 // Allow exactly one numeric character for the additional refinement
1718 if (RefStepString
.size() == 1) {
1719 char RefStepChar
= RefStepString
[0];
1720 if (RefStepChar
>= '0' && RefStepChar
<= '9') {
1721 Value
= RefStepChar
- '0';
1725 report_fatal_error("Invalid refinement step for -recip.");
1728 /// For the input attribute string, return one of the ReciprocalEstimate enum
1729 /// status values (enabled, disabled, or not specified) for this operation on
1730 /// the specified data type.
1731 static int getOpEnabled(bool IsSqrt
, EVT VT
, StringRef Override
) {
1732 if (Override
.empty())
1733 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1735 SmallVector
<StringRef
, 4> OverrideVector
;
1736 Override
.split(OverrideVector
, ',');
1737 unsigned NumArgs
= OverrideVector
.size();
1739 // Check if "all", "none", or "default" was specified.
1741 // Look for an optional setting of the number of refinement steps needed
1742 // for this type of reciprocal operation.
1745 if (parseRefinementStep(Override
, RefPos
, RefSteps
)) {
1746 // Split the string for further processing.
1747 Override
= Override
.substr(0, RefPos
);
1750 // All reciprocal types are enabled.
1751 if (Override
== "all")
1752 return TargetLoweringBase::ReciprocalEstimate::Enabled
;
1754 // All reciprocal types are disabled.
1755 if (Override
== "none")
1756 return TargetLoweringBase::ReciprocalEstimate::Disabled
;
1758 // Target defaults for enablement are used.
1759 if (Override
== "default")
1760 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1763 // The attribute string may omit the size suffix ('f'/'d').
1764 std::string VTName
= getReciprocalOpName(IsSqrt
, VT
);
1765 std::string VTNameNoSize
= VTName
;
1766 VTNameNoSize
.pop_back();
1767 static const char DisabledPrefix
= '!';
1769 for (StringRef RecipType
: OverrideVector
) {
1772 if (parseRefinementStep(RecipType
, RefPos
, RefSteps
))
1773 RecipType
= RecipType
.substr(0, RefPos
);
1775 // Ignore the disablement token for string matching.
1776 bool IsDisabled
= RecipType
[0] == DisabledPrefix
;
1778 RecipType
= RecipType
.substr(1);
1780 if (RecipType
.equals(VTName
) || RecipType
.equals(VTNameNoSize
))
1781 return IsDisabled
? TargetLoweringBase::ReciprocalEstimate::Disabled
1782 : TargetLoweringBase::ReciprocalEstimate::Enabled
;
1785 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1788 /// For the input attribute string, return the customized refinement step count
1789 /// for this operation on the specified data type. If the step count does not
1790 /// exist, return the ReciprocalEstimate enum value for unspecified.
1791 static int getOpRefinementSteps(bool IsSqrt
, EVT VT
, StringRef Override
) {
1792 if (Override
.empty())
1793 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1795 SmallVector
<StringRef
, 4> OverrideVector
;
1796 Override
.split(OverrideVector
, ',');
1797 unsigned NumArgs
= OverrideVector
.size();
1799 // Check if "all", "default", or "none" was specified.
1801 // Look for an optional setting of the number of refinement steps needed
1802 // for this type of reciprocal operation.
1805 if (!parseRefinementStep(Override
, RefPos
, RefSteps
))
1806 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1808 // Split the string for further processing.
1809 Override
= Override
.substr(0, RefPos
);
1810 assert(Override
!= "none" &&
1811 "Disabled reciprocals, but specifed refinement steps?");
1813 // If this is a general override, return the specified number of steps.
1814 if (Override
== "all" || Override
== "default")
1818 // The attribute string may omit the size suffix ('f'/'d').
1819 std::string VTName
= getReciprocalOpName(IsSqrt
, VT
);
1820 std::string VTNameNoSize
= VTName
;
1821 VTNameNoSize
.pop_back();
1823 for (StringRef RecipType
: OverrideVector
) {
1826 if (!parseRefinementStep(RecipType
, RefPos
, RefSteps
))
1829 RecipType
= RecipType
.substr(0, RefPos
);
1830 if (RecipType
.equals(VTName
) || RecipType
.equals(VTNameNoSize
))
1834 return TargetLoweringBase::ReciprocalEstimate::Unspecified
;
1837 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT
,
1838 MachineFunction
&MF
) const {
1839 return getOpEnabled(true, VT
, getRecipEstimateForFunc(MF
));
1842 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT
,
1843 MachineFunction
&MF
) const {
1844 return getOpEnabled(false, VT
, getRecipEstimateForFunc(MF
));
1847 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT
,
1848 MachineFunction
&MF
) const {
1849 return getOpRefinementSteps(true, VT
, getRecipEstimateForFunc(MF
));
1852 int TargetLoweringBase::getDivRefinementSteps(EVT VT
,
1853 MachineFunction
&MF
) const {
1854 return getOpRefinementSteps(false, VT
, getRecipEstimateForFunc(MF
));
1857 void TargetLoweringBase::finalizeLowering(MachineFunction
&MF
) const {
1858 MF
.getRegInfo().freezeReservedRegs(MF
);