?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
[Alignment][NFC] Use Align with TargetLowering::setMinFunctionAlignment
[llvm-core.git] / lib / CodeGen / TargetLoweringBase.cpp
blob320f931845a38f5aee75473ea9e0bf4fb081a904
1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements the TargetLoweringBase class.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/ADT/BitVector.h"
14 #include "llvm/ADT/STLExtras.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/StringExtras.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/ADT/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"
54 #include <algorithm>
55 #include <cassert>
56 #include <cstddef>
57 #include <cstdint>
58 #include <cstring>
59 #include <iterator>
60 #include <string>
61 #include <tuple>
62 #include <utility>
63
64 using namespace llvm;
65
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."),
69 cl::Hidden);
70
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."));
74
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."));
78
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"));
84
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"));
90
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)
95 return false;
96 // Macos < 10.9 has no sincos_stret.
97 if (TT.isMacOSX())
98 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit();
99 // iOS < 7.0 has no sincos_stret.
100 if (TT.isiOS())
101 return !TT.isOSVersionLT(7, 0);
102 // Any other darwin such as WatchOS/TvOS is new enough.
103 return true;
104 }
105
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"),
115 cl::Hidden);
116
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);
125
126 // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf".
127 if (TT.getArch() == Triple::ppc || TT.isPPC64()) {
128 setLibcallName(RTLIB::ADD_F128, "__addkf3");
129 setLibcallName(RTLIB::SUB_F128, "__subkf3");
130 setLibcallName(RTLIB::MUL_F128, "__mulkf3");
131 setLibcallName(RTLIB::DIV_F128, "__divkf3");
132 setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2");
133 setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2");
134 setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2");
135 setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2");
136 setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi");
137 setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi");
138 setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi");
139 setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi");
140 setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf");
141 setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf");
142 setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf");
143 setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf");
144 setLibcallName(RTLIB::OEQ_F128, "__eqkf2");
145 setLibcallName(RTLIB::UNE_F128, "__nekf2");
146 setLibcallName(RTLIB::OGE_F128, "__gekf2");
147 setLibcallName(RTLIB::OLT_F128, "__ltkf2");
148 setLibcallName(RTLIB::OLE_F128, "__lekf2");
149 setLibcallName(RTLIB::OGT_F128, "__gtkf2");
150 setLibcallName(RTLIB::UO_F128, "__unordkf2");
151 setLibcallName(RTLIB::O_F128, "__unordkf2");
152 }
153
154 // A few names are different on particular architectures or environments.
155 if (TT.isOSDarwin()) {
156 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
157 // of the gnueabi-style __gnu_*_ieee.
158 // FIXME: What about other targets?
159 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
160 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
161
162 // Some darwins have an optimized __bzero/bzero function.
163 switch (TT.getArch()) {
164 case Triple::x86:
165 case Triple::x86_64:
166 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6))
167 setLibcallName(RTLIB::BZERO, "__bzero");
168 break;
169 case Triple::aarch64:
170 setLibcallName(RTLIB::BZERO, "bzero");
171 break;
172 default:
173 break;
174 }
175
176 if (darwinHasSinCos(TT)) {
177 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret");
178 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret");
179 if (TT.isWatchABI()) {
180 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32,
181 CallingConv::ARM_AAPCS_VFP);
182 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64,
183 CallingConv::ARM_AAPCS_VFP);
184 }
185 }
186 } else {
187 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee");
188 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee");
189 }
190
191 if (TT.isGNUEnvironment() || TT.isOSFuchsia() ||
192 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) {
193 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
194 setLibcallName(RTLIB::SINCOS_F64, "sincos");
195 setLibcallName(RTLIB::SINCOS_F80, "sincosl");
196 setLibcallName(RTLIB::SINCOS_F128, "sincosl");
197 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl");
198 }
199
200 if (TT.isPS4CPU()) {
201 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
202 setLibcallName(RTLIB::SINCOS_F64, "sincos");
203 }
204
205 if (TT.isOSOpenBSD()) {
206 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr);
207 }
208 }
209
210 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
211 /// UNKNOWN_LIBCALL if there is none.
212 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
213 if (OpVT == MVT::f16) {
214 if (RetVT == MVT::f32)
215 return FPEXT_F16_F32;
216 } else if (OpVT == MVT::f32) {
217 if (RetVT == MVT::f64)
218 return FPEXT_F32_F64;
219 if (RetVT == MVT::f128)
220 return FPEXT_F32_F128;
221 if (RetVT == MVT::ppcf128)
222 return FPEXT_F32_PPCF128;
223 } else if (OpVT == MVT::f64) {
224 if (RetVT == MVT::f128)
225 return FPEXT_F64_F128;
226 else if (RetVT == MVT::ppcf128)
227 return FPEXT_F64_PPCF128;
228 } else if (OpVT == MVT::f80) {
229 if (RetVT == MVT::f128)
230 return FPEXT_F80_F128;
231 }
232
233 return UNKNOWN_LIBCALL;
234 }
235
236 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
237 /// UNKNOWN_LIBCALL if there is none.
238 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
239 if (RetVT == MVT::f16) {
240 if (OpVT == MVT::f32)
241 return FPROUND_F32_F16;
242 if (OpVT == MVT::f64)
243 return FPROUND_F64_F16;
244 if (OpVT == MVT::f80)
245 return FPROUND_F80_F16;
246 if (OpVT == MVT::f128)
247 return FPROUND_F128_F16;
248 if (OpVT == MVT::ppcf128)
249 return FPROUND_PPCF128_F16;
250 } else if (RetVT == MVT::f32) {
251 if (OpVT == MVT::f64)
252 return FPROUND_F64_F32;
253 if (OpVT == MVT::f80)
254 return FPROUND_F80_F32;
255 if (OpVT == MVT::f128)
256 return FPROUND_F128_F32;
257 if (OpVT == MVT::ppcf128)
258 return FPROUND_PPCF128_F32;
259 } else if (RetVT == MVT::f64) {
260 if (OpVT == MVT::f80)
261 return FPROUND_F80_F64;
262 if (OpVT == MVT::f128)
263 return FPROUND_F128_F64;
264 if (OpVT == MVT::ppcf128)
265 return FPROUND_PPCF128_F64;
266 } else if (RetVT == MVT::f80) {
267 if (OpVT == MVT::f128)
268 return FPROUND_F128_F80;
269 }
270
271 return UNKNOWN_LIBCALL;
272 }
273
274 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
275 /// UNKNOWN_LIBCALL if there is none.
276 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
277 if (OpVT == MVT::f32) {
278 if (RetVT == MVT::i32)
279 return FPTOSINT_F32_I32;
280 if (RetVT == MVT::i64)
281 return FPTOSINT_F32_I64;
282 if (RetVT == MVT::i128)
283 return FPTOSINT_F32_I128;
284 } else if (OpVT == MVT::f64) {
285 if (RetVT == MVT::i32)
286 return FPTOSINT_F64_I32;
287 if (RetVT == MVT::i64)
288 return FPTOSINT_F64_I64;
289 if (RetVT == MVT::i128)
290 return FPTOSINT_F64_I128;
291 } else if (OpVT == MVT::f80) {
292 if (RetVT == MVT::i32)
293 return FPTOSINT_F80_I32;
294 if (RetVT == MVT::i64)
295 return FPTOSINT_F80_I64;
296 if (RetVT == MVT::i128)
297 return FPTOSINT_F80_I128;
298 } else if (OpVT == MVT::f128) {
299 if (RetVT == MVT::i32)
300 return FPTOSINT_F128_I32;
301 if (RetVT == MVT::i64)
302 return FPTOSINT_F128_I64;
303 if (RetVT == MVT::i128)
304 return FPTOSINT_F128_I128;
305 } else if (OpVT == MVT::ppcf128) {
306 if (RetVT == MVT::i32)
307 return FPTOSINT_PPCF128_I32;
308 if (RetVT == MVT::i64)
309 return FPTOSINT_PPCF128_I64;
310 if (RetVT == MVT::i128)
311 return FPTOSINT_PPCF128_I128;
312 }
313 return UNKNOWN_LIBCALL;
314 }
315
316 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
317 /// UNKNOWN_LIBCALL if there is none.
318 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
319 if (OpVT == MVT::f32) {
320 if (RetVT == MVT::i32)
321 return FPTOUINT_F32_I32;
322 if (RetVT == MVT::i64)
323 return FPTOUINT_F32_I64;
324 if (RetVT == MVT::i128)
325 return FPTOUINT_F32_I128;
326 } else if (OpVT == MVT::f64) {
327 if (RetVT == MVT::i32)
328 return FPTOUINT_F64_I32;
329 if (RetVT == MVT::i64)
330 return FPTOUINT_F64_I64;
331 if (RetVT == MVT::i128)
332 return FPTOUINT_F64_I128;
333 } else if (OpVT == MVT::f80) {
334 if (RetVT == MVT::i32)
335 return FPTOUINT_F80_I32;
336 if (RetVT == MVT::i64)
337 return FPTOUINT_F80_I64;
338 if (RetVT == MVT::i128)
339 return FPTOUINT_F80_I128;
340 } else if (OpVT == MVT::f128) {
341 if (RetVT == MVT::i32)
342 return FPTOUINT_F128_I32;
343 if (RetVT == MVT::i64)
344 return FPTOUINT_F128_I64;
345 if (RetVT == MVT::i128)
346 return FPTOUINT_F128_I128;
347 } else if (OpVT == MVT::ppcf128) {
348 if (RetVT == MVT::i32)
349 return FPTOUINT_PPCF128_I32;
350 if (RetVT == MVT::i64)
351 return FPTOUINT_PPCF128_I64;
352 if (RetVT == MVT::i128)
353 return FPTOUINT_PPCF128_I128;
354 }
355 return UNKNOWN_LIBCALL;
356 }
357
358 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
359 /// UNKNOWN_LIBCALL if there is none.
360 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
361 if (OpVT == MVT::i32) {
362 if (RetVT == MVT::f32)
363 return SINTTOFP_I32_F32;
364 if (RetVT == MVT::f64)
365 return SINTTOFP_I32_F64;
366 if (RetVT == MVT::f80)
367 return SINTTOFP_I32_F80;
368 if (RetVT == MVT::f128)
369 return SINTTOFP_I32_F128;
370 if (RetVT == MVT::ppcf128)
371 return SINTTOFP_I32_PPCF128;
372 } else if (OpVT == MVT::i64) {
373 if (RetVT == MVT::f32)
374 return SINTTOFP_I64_F32;
375 if (RetVT == MVT::f64)
376 return SINTTOFP_I64_F64;
377 if (RetVT == MVT::f80)
378 return SINTTOFP_I64_F80;
379 if (RetVT == MVT::f128)
380 return SINTTOFP_I64_F128;
381 if (RetVT == MVT::ppcf128)
382 return SINTTOFP_I64_PPCF128;
383 } else if (OpVT == MVT::i128) {
384 if (RetVT == MVT::f32)
385 return SINTTOFP_I128_F32;
386 if (RetVT == MVT::f64)
387 return SINTTOFP_I128_F64;
388 if (RetVT == MVT::f80)
389 return SINTTOFP_I128_F80;
390 if (RetVT == MVT::f128)
391 return SINTTOFP_I128_F128;
392 if (RetVT == MVT::ppcf128)
393 return SINTTOFP_I128_PPCF128;
394 }
395 return UNKNOWN_LIBCALL;
396 }
397
398 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
399 /// UNKNOWN_LIBCALL if there is none.
400 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
401 if (OpVT == MVT::i32) {
402 if (RetVT == MVT::f32)
403 return UINTTOFP_I32_F32;
404 if (RetVT == MVT::f64)
405 return UINTTOFP_I32_F64;
406 if (RetVT == MVT::f80)
407 return UINTTOFP_I32_F80;
408 if (RetVT == MVT::f128)
409 return UINTTOFP_I32_F128;
410 if (RetVT == MVT::ppcf128)
411 return UINTTOFP_I32_PPCF128;
412 } else if (OpVT == MVT::i64) {
413 if (RetVT == MVT::f32)
414 return UINTTOFP_I64_F32;
415 if (RetVT == MVT::f64)
416 return UINTTOFP_I64_F64;
417 if (RetVT == MVT::f80)
418 return UINTTOFP_I64_F80;
419 if (RetVT == MVT::f128)
420 return UINTTOFP_I64_F128;
421 if (RetVT == MVT::ppcf128)
422 return UINTTOFP_I64_PPCF128;
423 } else if (OpVT == MVT::i128) {
424 if (RetVT == MVT::f32)
425 return UINTTOFP_I128_F32;
426 if (RetVT == MVT::f64)
427 return UINTTOFP_I128_F64;
428 if (RetVT == MVT::f80)
429 return UINTTOFP_I128_F80;
430 if (RetVT == MVT::f128)
431 return UINTTOFP_I128_F128;
432 if (RetVT == MVT::ppcf128)
433 return UINTTOFP_I128_PPCF128;
434 }
435 return UNKNOWN_LIBCALL;
436 }
437
438 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
439 #define OP_TO_LIBCALL(Name, Enum) \
440 case Name: \
441 switch (VT.SimpleTy) { \
442 default: \
443 return UNKNOWN_LIBCALL; \
444 case MVT::i8: \
445 return Enum##_1; \
446 case MVT::i16: \
447 return Enum##_2; \
448 case MVT::i32: \
449 return Enum##_4; \
450 case MVT::i64: \
451 return Enum##_8; \
452 case MVT::i128: \
453 return Enum##_16; \
454 }
455
456 switch (Opc) {
457 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
458 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
459 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
460 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
461 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
462 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
463 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
464 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
465 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
466 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
467 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
468 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
469 }
470
471 #undef OP_TO_LIBCALL
472
473 return UNKNOWN_LIBCALL;
474 }
475
476 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
477 switch (ElementSize) {
478 case 1:
479 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
480 case 2:
481 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
482 case 4:
483 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
484 case 8:
485 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
486 case 16:
487 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
488 default:
489 return UNKNOWN_LIBCALL;
490 }
491 }
492
493 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
494 switch (ElementSize) {
495 case 1:
496 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
497 case 2:
498 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
499 case 4:
500 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
501 case 8:
502 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
503 case 16:
504 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
505 default:
506 return UNKNOWN_LIBCALL;
507 }
508 }
509
510 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
511 switch (ElementSize) {
512 case 1:
513 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
514 case 2:
515 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
516 case 4:
517 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
518 case 8:
519 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
520 case 16:
521 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
522 default:
523 return UNKNOWN_LIBCALL;
524 }
525 }
526
527 /// InitCmpLibcallCCs - Set default comparison libcall CC.
528 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
529 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
530 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
531 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
532 CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
533 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ;
534 CCs[RTLIB::UNE_F32] = ISD::SETNE;
535 CCs[RTLIB::UNE_F64] = ISD::SETNE;
536 CCs[RTLIB::UNE_F128] = ISD::SETNE;
537 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE;
538 CCs[RTLIB::OGE_F32] = ISD::SETGE;
539 CCs[RTLIB::OGE_F64] = ISD::SETGE;
540 CCs[RTLIB::OGE_F128] = ISD::SETGE;
541 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE;
542 CCs[RTLIB::OLT_F32] = ISD::SETLT;
543 CCs[RTLIB::OLT_F64] = ISD::SETLT;
544 CCs[RTLIB::OLT_F128] = ISD::SETLT;
545 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT;
546 CCs[RTLIB::OLE_F32] = ISD::SETLE;
547 CCs[RTLIB::OLE_F64] = ISD::SETLE;
548 CCs[RTLIB::OLE_F128] = ISD::SETLE;
549 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE;
550 CCs[RTLIB::OGT_F32] = ISD::SETGT;
551 CCs[RTLIB::OGT_F64] = ISD::SETGT;
552 CCs[RTLIB::OGT_F128] = ISD::SETGT;
553 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT;
554 CCs[RTLIB::UO_F32] = ISD::SETNE;
555 CCs[RTLIB::UO_F64] = ISD::SETNE;
556 CCs[RTLIB::UO_F128] = ISD::SETNE;
557 CCs[RTLIB::UO_PPCF128] = ISD::SETNE;
558 CCs[RTLIB::O_F32] = ISD::SETEQ;
559 CCs[RTLIB::O_F64] = ISD::SETEQ;
560 CCs[RTLIB::O_F128] = ISD::SETEQ;
561 CCs[RTLIB::O_PPCF128] = ISD::SETEQ;
562 }
563
564 /// NOTE: The TargetMachine owns TLOF.
565 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
566 initActions();
567
568 // Perform these initializations only once.
569 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
570 MaxLoadsPerMemcmp = 8;
571 MaxGluedStoresPerMemcpy = 0;
572 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
573 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
574 UseUnderscoreSetJmp = false;
575 UseUnderscoreLongJmp = false;
576 HasMultipleConditionRegisters = false;
577 HasExtractBitsInsn = false;
578 JumpIsExpensive = JumpIsExpensiveOverride;
579 PredictableSelectIsExpensive = false;
580 EnableExtLdPromotion = false;
581 StackPointerRegisterToSaveRestore = 0;
582 BooleanContents = UndefinedBooleanContent;
583 BooleanFloatContents = UndefinedBooleanContent;
584 BooleanVectorContents = UndefinedBooleanContent;
585 SchedPreferenceInfo = Sched::ILP;
586 GatherAllAliasesMaxDepth = 18;
587 // TODO: the default will be switched to 0 in the next commit, along
588 // with the Target-specific changes necessary.
589 MaxAtomicSizeInBitsSupported = 1024;
590
591 MinCmpXchgSizeInBits = 0;
592 SupportsUnalignedAtomics = false;
593
594 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr);
595
596 InitLibcalls(TM.getTargetTriple());
597 InitCmpLibcallCCs(CmpLibcallCCs);
598 }
599
600 void TargetLoweringBase::initActions() {
601 // All operations default to being supported.
602 memset(OpActions, 0, sizeof(OpActions));
603 memset(LoadExtActions, 0, sizeof(LoadExtActions));
604 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
605 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
606 memset(CondCodeActions, 0, sizeof(CondCodeActions));
607 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr);
608 std::fill(std::begin(TargetDAGCombineArray),
609 std::end(TargetDAGCombineArray), 0);
610
611 for (MVT VT : MVT::fp_valuetypes()) {
612 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits());
613 if (IntVT.isValid()) {
614 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote);
615 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT);
616 }
617 }
618
619 // Set default actions for various operations.
620 for (MVT VT : MVT::all_valuetypes()) {
621 // Default all indexed load / store to expand.
622 for (unsigned IM = (unsigned)ISD::PRE_INC;
623 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
624 setIndexedLoadAction(IM, VT, Expand);
625 setIndexedStoreAction(IM, VT, Expand);
626 }
627
628 // Most backends expect to see the node which just returns the value loaded.
629 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
630
631 // These operations default to expand.
632 setOperationAction(ISD::FGETSIGN, VT, Expand);
633 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand);
634 setOperationAction(ISD::FMINNUM, VT, Expand);
635 setOperationAction(ISD::FMAXNUM, VT, Expand);
636 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand);
637 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand);
638 setOperationAction(ISD::FMINIMUM, VT, Expand);
639 setOperationAction(ISD::FMAXIMUM, VT, Expand);
640 setOperationAction(ISD::FMAD, VT, Expand);
641 setOperationAction(ISD::SMIN, VT, Expand);
642 setOperationAction(ISD::SMAX, VT, Expand);
643 setOperationAction(ISD::UMIN, VT, Expand);
644 setOperationAction(ISD::UMAX, VT, Expand);
645 setOperationAction(ISD::ABS, VT, Expand);
646 setOperationAction(ISD::FSHL, VT, Expand);
647 setOperationAction(ISD::FSHR, VT, Expand);
648 setOperationAction(ISD::SADDSAT, VT, Expand);
649 setOperationAction(ISD::UADDSAT, VT, Expand);
650 setOperationAction(ISD::SSUBSAT, VT, Expand);
651 setOperationAction(ISD::USUBSAT, VT, Expand);
652 setOperationAction(ISD::SMULFIX, VT, Expand);
653 setOperationAction(ISD::SMULFIXSAT, VT, Expand);
654 setOperationAction(ISD::UMULFIX, VT, Expand);
655
656 // Overflow operations default to expand
657 setOperationAction(ISD::SADDO, VT, Expand);
658 setOperationAction(ISD::SSUBO, VT, Expand);
659 setOperationAction(ISD::UADDO, VT, Expand);
660 setOperationAction(ISD::USUBO, VT, Expand);
661 setOperationAction(ISD::SMULO, VT, Expand);
662 setOperationAction(ISD::UMULO, VT, Expand);
663
664 // ADDCARRY operations default to expand
665 setOperationAction(ISD::ADDCARRY, VT, Expand);
666 setOperationAction(ISD::SUBCARRY, VT, Expand);
667 setOperationAction(ISD::SETCCCARRY, VT, Expand);
668
669 // ADDC/ADDE/SUBC/SUBE default to expand.
670 setOperationAction(ISD::ADDC, VT, Expand);
671 setOperationAction(ISD::ADDE, VT, Expand);
672 setOperationAction(ISD::SUBC, VT, Expand);
673 setOperationAction(ISD::SUBE, VT, Expand);
674
675 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
676 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
677 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
678
679 setOperationAction(ISD::BITREVERSE, VT, Expand);
680
681 // These library functions default to expand.
682 setOperationAction(ISD::FROUND, VT, Expand);
683 setOperationAction(ISD::FPOWI, VT, Expand);
684
685 // These operations default to expand for vector types.
686 if (VT.isVector()) {
687 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
688 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
689 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
690 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
691 }
692
693 // Constrained floating-point operations default to expand.
694 setOperationAction(ISD::STRICT_FADD, VT, Expand);
695 setOperationAction(ISD::STRICT_FSUB, VT, Expand);
696 setOperationAction(ISD::STRICT_FMUL, VT, Expand);
697 setOperationAction(ISD::STRICT_FDIV, VT, Expand);
698 setOperationAction(ISD::STRICT_FREM, VT, Expand);
699 setOperationAction(ISD::STRICT_FMA, VT, Expand);
700 setOperationAction(ISD::STRICT_FSQRT, VT, Expand);
701 setOperationAction(ISD::STRICT_FPOW, VT, Expand);
702 setOperationAction(ISD::STRICT_FPOWI, VT, Expand);
703 setOperationAction(ISD::STRICT_FSIN, VT, Expand);
704 setOperationAction(ISD::STRICT_FCOS, VT, Expand);
705 setOperationAction(ISD::STRICT_FEXP, VT, Expand);
706 setOperationAction(ISD::STRICT_FEXP2, VT, Expand);
707 setOperationAction(ISD::STRICT_FLOG, VT, Expand);
708 setOperationAction(ISD::STRICT_FLOG10, VT, Expand);
709 setOperationAction(ISD::STRICT_FLOG2, VT, Expand);
710 setOperationAction(ISD::STRICT_FRINT, VT, Expand);
711 setOperationAction(ISD::STRICT_FNEARBYINT, VT, Expand);
712 setOperationAction(ISD::STRICT_FCEIL, VT, Expand);
713 setOperationAction(ISD::STRICT_FFLOOR, VT, Expand);
714 setOperationAction(ISD::STRICT_FROUND, VT, Expand);
715 setOperationAction(ISD::STRICT_FTRUNC, VT, Expand);
716 setOperationAction(ISD::STRICT_FMAXNUM, VT, Expand);
717 setOperationAction(ISD::STRICT_FMINNUM, VT, Expand);
718 setOperationAction(ISD::STRICT_FP_ROUND, VT, Expand);
719 setOperationAction(ISD::STRICT_FP_EXTEND, VT, Expand);
720 setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Expand);
721 setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Expand);
722
723 // For most targets @llvm.get.dynamic.area.offset just returns 0.
724 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
725
726 // Vector reduction default to expand.
727 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand);
728 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand);
729 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand);
730 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand);
731 setOperationAction(ISD::VECREDUCE_AND, VT, Expand);
732 setOperationAction(ISD::VECREDUCE_OR, VT, Expand);
733 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand);
734 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand);
735 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand);
736 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand);
737 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand);
738 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand);
739 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand);
740 }
741
742 // Most targets ignore the @llvm.prefetch intrinsic.
743 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
744
745 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
746 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
747
748 // ConstantFP nodes default to expand. Targets can either change this to
749 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
750 // to optimize expansions for certain constants.
751 setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
752 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
753 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
754 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
755 setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
756
757 // These library functions default to expand.
758 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
759 setOperationAction(ISD::FCBRT, VT, Expand);
760 setOperationAction(ISD::FLOG , VT, Expand);
761 setOperationAction(ISD::FLOG2, VT, Expand);
762 setOperationAction(ISD::FLOG10, VT, Expand);
763 setOperationAction(ISD::FEXP , VT, Expand);
764 setOperationAction(ISD::FEXP2, VT, Expand);
765 setOperationAction(ISD::FFLOOR, VT, Expand);
766 setOperationAction(ISD::FNEARBYINT, VT, Expand);
767 setOperationAction(ISD::FCEIL, VT, Expand);
768 setOperationAction(ISD::FRINT, VT, Expand);
769 setOperationAction(ISD::FTRUNC, VT, Expand);
770 setOperationAction(ISD::FROUND, VT, Expand);
771 setOperationAction(ISD::LROUND, VT, Expand);
772 setOperationAction(ISD::LLROUND, VT, Expand);
773 setOperationAction(ISD::LRINT, VT, Expand);
774 setOperationAction(ISD::LLRINT, VT, Expand);
775 }
776
777 // Default ISD::TRAP to expand (which turns it into abort).
778 setOperationAction(ISD::TRAP, MVT::Other, Expand);
779
780 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
781 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
782 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
783 }
784
785 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
786 EVT) const {
787 return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
788 }
789
790 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
791 bool LegalTypes) const {
792 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
793 if (LHSTy.isVector())
794 return LHSTy;
795 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
796 : getPointerTy(DL);
797 }
798
799 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
800 assert(isTypeLegal(VT));
801 switch (Op) {
802 default:
803 return false;
804 case ISD::SDIV:
805 case ISD::UDIV:
806 case ISD::SREM:
807 case ISD::UREM:
808 return true;
809 }
810 }
811
812 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
813 // If the command-line option was specified, ignore this request.
814 if (!JumpIsExpensiveOverride.getNumOccurrences())
815 JumpIsExpensive = isExpensive;
816 }
817
818 TargetLoweringBase::LegalizeKind
819 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
820 // If this is a simple type, use the ComputeRegisterProp mechanism.
821 if (VT.isSimple()) {
822 MVT SVT = VT.getSimpleVT();
823 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
824 MVT NVT = TransformToType[SVT.SimpleTy];
825 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
826
827 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
828 (NVT.isVector() ||
829 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
830 "Promote may not follow Expand or Promote");
831
832 if (LA == TypeSplitVector)
833 return LegalizeKind(LA,
834 EVT::getVectorVT(Context, SVT.getVectorElementType(),
835 SVT.getVectorNumElements() / 2));
836 if (LA == TypeScalarizeVector)
837 return LegalizeKind(LA, SVT.getVectorElementType());
838 return LegalizeKind(LA, NVT);
839 }
840
841 // Handle Extended Scalar Types.
842 if (!VT.isVector()) {
843 assert(VT.isInteger() && "Float types must be simple");
844 unsigned BitSize = VT.getSizeInBits();
845 // First promote to a power-of-two size, then expand if necessary.
846 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
847 EVT NVT = VT.getRoundIntegerType(Context);
848 assert(NVT != VT && "Unable to round integer VT");
849 LegalizeKind NextStep = getTypeConversion(Context, NVT);
850 // Avoid multi-step promotion.
851 if (NextStep.first == TypePromoteInteger)
852 return NextStep;
853 // Return rounded integer type.
854 return LegalizeKind(TypePromoteInteger, NVT);
855 }
856
857 return LegalizeKind(TypeExpandInteger,
858 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
859 }
860
861 // Handle vector types.
862 unsigned NumElts = VT.getVectorNumElements();
863 EVT EltVT = VT.getVectorElementType();
864
865 // Vectors with only one element are always scalarized.
866 if (NumElts == 1)
867 return LegalizeKind(TypeScalarizeVector, EltVT);
868
869 // Try to widen vector elements until the element type is a power of two and
870 // promote it to a legal type later on, for example:
871 // <3 x i8> -> <4 x i8> -> <4 x i32>
872 if (EltVT.isInteger()) {
873 // Vectors with a number of elements that is not a power of two are always
874 // widened, for example <3 x i8> -> <4 x i8>.
875 if (!VT.isPow2VectorType()) {
876 NumElts = (unsigned)NextPowerOf2(NumElts);
877 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
878 return LegalizeKind(TypeWidenVector, NVT);
879 }
880
881 // Examine the element type.
882 LegalizeKind LK = getTypeConversion(Context, EltVT);
883
884 // If type is to be expanded, split the vector.
885 // <4 x i140> -> <2 x i140>
886 if (LK.first == TypeExpandInteger)
887 return LegalizeKind(TypeSplitVector,
888 EVT::getVectorVT(Context, EltVT, NumElts / 2));
889
890 // Promote the integer element types until a legal vector type is found
891 // or until the element integer type is too big. If a legal type was not
892 // found, fallback to the usual mechanism of widening/splitting the
893 // vector.
894 EVT OldEltVT = EltVT;
895 while (true) {
896 // Increase the bitwidth of the element to the next pow-of-two
897 // (which is greater than 8 bits).
898 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
899 .getRoundIntegerType(Context);
900
901 // Stop trying when getting a non-simple element type.
902 // Note that vector elements may be greater than legal vector element
903 // types. Example: X86 XMM registers hold 64bit element on 32bit
904 // systems.
905 if (!EltVT.isSimple())
906 break;
907
908 // Build a new vector type and check if it is legal.
909 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
910 // Found a legal promoted vector type.
911 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
912 return LegalizeKind(TypePromoteInteger,
913 EVT::getVectorVT(Context, EltVT, NumElts));
914 }
915
916 // Reset the type to the unexpanded type if we did not find a legal vector
917 // type with a promoted vector element type.
918 EltVT = OldEltVT;
919 }
920
921 // Try to widen the vector until a legal type is found.
922 // If there is no wider legal type, split the vector.
923 while (true) {
924 // Round up to the next power of 2.
925 NumElts = (unsigned)NextPowerOf2(NumElts);
926
927 // If there is no simple vector type with this many elements then there
928 // cannot be a larger legal vector type. Note that this assumes that
929 // there are no skipped intermediate vector types in the simple types.
930 if (!EltVT.isSimple())
931 break;
932 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
933 if (LargerVector == MVT())
934 break;
935
936 // If this type is legal then widen the vector.
937 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
938 return LegalizeKind(TypeWidenVector, LargerVector);
939 }
940
941 // Widen odd vectors to next power of two.
942 if (!VT.isPow2VectorType()) {
943 EVT NVT = VT.getPow2VectorType(Context);
944 return LegalizeKind(TypeWidenVector, NVT);
945 }
946
947 // Vectors with illegal element types are expanded.
948 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
949 return LegalizeKind(TypeSplitVector, NVT);
950 }
951
952 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
953 unsigned &NumIntermediates,
954 MVT &RegisterVT,
955 TargetLoweringBase *TLI) {
956 // Figure out the right, legal destination reg to copy into.
957 unsigned NumElts = VT.getVectorNumElements();
958 MVT EltTy = VT.getVectorElementType();
959
960 unsigned NumVectorRegs = 1;
961
962 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
963 // could break down into LHS/RHS like LegalizeDAG does.
964 if (!isPowerOf2_32(NumElts)) {
965 NumVectorRegs = NumElts;
966 NumElts = 1;
967 }
968
969 // Divide the input until we get to a supported size. This will always
970 // end with a scalar if the target doesn't support vectors.
971 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
972 NumElts >>= 1;
973 NumVectorRegs <<= 1;
974 }
975
976 NumIntermediates = NumVectorRegs;
977
978 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
979 if (!TLI->isTypeLegal(NewVT))
980 NewVT = EltTy;
981 IntermediateVT = NewVT;
982
983 unsigned NewVTSize = NewVT.getSizeInBits();
984
985 // Convert sizes such as i33 to i64.
986 if (!isPowerOf2_32(NewVTSize))
987 NewVTSize = NextPowerOf2(NewVTSize);
988
989 MVT DestVT = TLI->getRegisterType(NewVT);
990 RegisterVT = DestVT;
991 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
992 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
993
994 // Otherwise, promotion or legal types use the same number of registers as
995 // the vector decimated to the appropriate level.
996 return NumVectorRegs;
997 }
998
999 /// isLegalRC - Return true if the value types that can be represented by the
1000 /// specified register class are all legal.
1001 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
1002 const TargetRegisterClass &RC) const {
1003 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1004 if (isTypeLegal(*I))
1005 return true;
1006 return false;
1007 }
1008
1009 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1010 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1011 MachineBasicBlock *
1012 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
1013 MachineBasicBlock *MBB) const {
1014 MachineInstr *MI = &InitialMI;
1015 MachineFunction &MF = *MI->getMF();
1016 MachineFrameInfo &MFI = MF.getFrameInfo();
1017
1018 // We're handling multiple types of operands here:
1019 // PATCHPOINT MetaArgs - live-in, read only, direct
1020 // STATEPOINT Deopt Spill - live-through, read only, indirect
1021 // STATEPOINT Deopt Alloca - live-through, read only, direct
1022 // (We're currently conservative and mark the deopt slots read/write in
1023 // practice.)
1024 // STATEPOINT GC Spill - live-through, read/write, indirect
1025 // STATEPOINT GC Alloca - live-through, read/write, direct
1026 // The live-in vs live-through is handled already (the live through ones are
1027 // all stack slots), but we need to handle the different type of stackmap
1028 // operands and memory effects here.
1029
1030 // MI changes inside this loop as we grow operands.
1031 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) {
1032 MachineOperand &MO = MI->getOperand(OperIdx);
1033 if (!MO.isFI())
1034 continue;
1035
1036 // foldMemoryOperand builds a new MI after replacing a single FI operand
1037 // with the canonical set of five x86 addressing-mode operands.
1038 int FI = MO.getIndex();
1039 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1040
1041 // Copy operands before the frame-index.
1042 for (unsigned i = 0; i < OperIdx; ++i)
1043 MIB.add(MI->getOperand(i));
1044 // Add frame index operands recognized by stackmaps.cpp
1045 if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
1046 // indirect-mem-ref tag, size, #FI, offset.
1047 // Used for spills inserted by StatepointLowering. This codepath is not
1048 // used for patchpoints/stackmaps at all, for these spilling is done via
1049 // foldMemoryOperand callback only.
1050 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1051 MIB.addImm(StackMaps::IndirectMemRefOp);
1052 MIB.addImm(MFI.getObjectSize(FI));
1053 MIB.add(MI->getOperand(OperIdx));
1054 MIB.addImm(0);
1055 } else {
1056 // direct-mem-ref tag, #FI, offset.
1057 // Used by patchpoint, and direct alloca arguments to statepoints
1058 MIB.addImm(StackMaps::DirectMemRefOp);
1059 MIB.add(MI->getOperand(OperIdx));
1060 MIB.addImm(0);
1061 }
1062 // Copy the operands after the frame index.
1063 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i)
1064 MIB.add(MI->getOperand(i));
1065
1066 // Inherit previous memory operands.
1067 MIB.cloneMemRefs(*MI);
1068 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1069
1070 // Add a new memory operand for this FI.
1071 assert(MFI.getObjectOffset(FI) != -1);
1072
1073 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and
1074 // PATCHPOINT should be updated to do the same. (TODO)
1075 if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1076 auto Flags = MachineMemOperand::MOLoad;
1077 MachineMemOperand *MMO = MF.getMachineMemOperand(
1078 MachinePointerInfo::getFixedStack(MF, FI), Flags,
1079 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI));
1080 MIB->addMemOperand(MF, MMO);
1081 }
1082
1083 // Replace the instruction and update the operand index.
1084 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1085 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1;
1086 MI->eraseFromParent();
1087 MI = MIB;
1088 }
1089 return MBB;
1090 }
1091
1092 MachineBasicBlock *
1093 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI,
1094 MachineBasicBlock *MBB) const {
1095 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL &&
1096 "Called emitXRayCustomEvent on the wrong MI!");
1097 auto &MF = *MI.getMF();
1098 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1099 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1100 MIB.add(MI.getOperand(OpIdx));
1101
1102 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1103 MI.eraseFromParent();
1104 return MBB;
1105 }
1106
1107 MachineBasicBlock *
1108 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI,
1109 MachineBasicBlock *MBB) const {
1110 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL &&
1111 "Called emitXRayTypedEvent on the wrong MI!");
1112 auto &MF = *MI.getMF();
1113 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1114 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1115 MIB.add(MI.getOperand(OpIdx));
1116
1117 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1118 MI.eraseFromParent();
1119 return MBB;
1120 }
1121
1122 /// findRepresentativeClass - Return the largest legal super-reg register class
1123 /// of the register class for the specified type and its associated "cost".
1124 // This function is in TargetLowering because it uses RegClassForVT which would
1125 // need to be moved to TargetRegisterInfo and would necessitate moving
1126 // isTypeLegal over as well - a massive change that would just require
1127 // TargetLowering having a TargetRegisterInfo class member that it would use.
1128 std::pair<const TargetRegisterClass *, uint8_t>
1129 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1130 MVT VT) const {
1131 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1132 if (!RC)
1133 return std::make_pair(RC, 0);
1134
1135 // Compute the set of all super-register classes.
1136 BitVector SuperRegRC(TRI->getNumRegClasses());
1137 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1138 SuperRegRC.setBitsInMask(RCI.getMask());
1139
1140 // Find the first legal register class with the largest spill size.
1141 const TargetRegisterClass *BestRC = RC;
1142 for (unsigned i : SuperRegRC.set_bits()) {
1143 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1144 // We want the largest possible spill size.
1145 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1146 continue;
1147 if (!isLegalRC(*TRI, *SuperRC))
1148 continue;
1149 BestRC = SuperRC;
1150 }
1151 return std::make_pair(BestRC, 1);
1152 }
1153
1154 /// computeRegisterProperties - Once all of the register classes are added,
1155 /// this allows us to compute derived properties we expose.
1156 void TargetLoweringBase::computeRegisterProperties(
1157 const TargetRegisterInfo *TRI) {
1158 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1159 "Too many value types for ValueTypeActions to hold!");
1160
1161 // Everything defaults to needing one register.
1162 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1163 NumRegistersForVT[i] = 1;
1164 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1165 }
1166 // ...except isVoid, which doesn't need any registers.
1167 NumRegistersForVT[MVT::isVoid] = 0;
1168
1169 // Find the largest integer register class.
1170 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1171 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1172 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1173
1174 // Every integer value type larger than this largest register takes twice as
1175 // many registers to represent as the previous ValueType.
1176 for (unsigned ExpandedReg = LargestIntReg + 1;
1177 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1178 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1179 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1180 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1181 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1182 TypeExpandInteger);
1183 }
1184
1185 // Inspect all of the ValueType's smaller than the largest integer
1186 // register to see which ones need promotion.
1187 unsigned LegalIntReg = LargestIntReg;
1188 for (unsigned IntReg = LargestIntReg - 1;
1189 IntReg >= (unsigned)MVT::i1; --IntReg) {
1190 MVT IVT = (MVT::SimpleValueType)IntReg;
1191 if (isTypeLegal(IVT)) {
1192 LegalIntReg = IntReg;
1193 } else {
1194 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1195 (MVT::SimpleValueType)LegalIntReg;
1196 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1197 }
1198 }
1199
1200 // ppcf128 type is really two f64's.
1201 if (!isTypeLegal(MVT::ppcf128)) {
1202 if (isTypeLegal(MVT::f64)) {
1203 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1204 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1205 TransformToType[MVT::ppcf128] = MVT::f64;
1206 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1207 } else {
1208 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1209 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1210 TransformToType[MVT::ppcf128] = MVT::i128;
1211 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1212 }
1213 }
1214
1215 // Decide how to handle f128. If the target does not have native f128 support,
1216 // expand it to i128 and we will be generating soft float library calls.
1217 if (!isTypeLegal(MVT::f128)) {
1218 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1219 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1220 TransformToType[MVT::f128] = MVT::i128;
1221 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1222 }
1223
1224 // Decide how to handle f64. If the target does not have native f64 support,
1225 // expand it to i64 and we will be generating soft float library calls.
1226 if (!isTypeLegal(MVT::f64)) {
1227 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1228 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1229 TransformToType[MVT::f64] = MVT::i64;
1230 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1231 }
1232
1233 // Decide how to handle f32. If the target does not have native f32 support,
1234 // expand it to i32 and we will be generating soft float library calls.
1235 if (!isTypeLegal(MVT::f32)) {
1236 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1237 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1238 TransformToType[MVT::f32] = MVT::i32;
1239 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1240 }
1241
1242 // Decide how to handle f16. If the target does not have native f16 support,
1243 // promote it to f32, because there are no f16 library calls (except for
1244 // conversions).
1245 if (!isTypeLegal(MVT::f16)) {
1246 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1247 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1248 TransformToType[MVT::f16] = MVT::f32;
1249 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1250 }
1251
1252 // Loop over all of the vector value types to see which need transformations.
1253 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1254 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1255 MVT VT = (MVT::SimpleValueType) i;
1256 if (isTypeLegal(VT))
1257 continue;
1258
1259 MVT EltVT = VT.getVectorElementType();
1260 unsigned NElts = VT.getVectorNumElements();
1261 bool IsLegalWiderType = false;
1262 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1263 switch (PreferredAction) {
1264 case TypePromoteInteger:
1265 // Try to promote the elements of integer vectors. If no legal
1266 // promotion was found, fall through to the widen-vector method.
1267 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) {
1268 MVT SVT = (MVT::SimpleValueType) nVT;
1269 // Promote vectors of integers to vectors with the same number
1270 // of elements, with a wider element type.
1271 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() &&
1272 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) {
1273 TransformToType[i] = SVT;
1274 RegisterTypeForVT[i] = SVT;
1275 NumRegistersForVT[i] = 1;
1276 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1277 IsLegalWiderType = true;
1278 break;
1279 }
1280 }
1281 if (IsLegalWiderType)
1282 break;
1283 LLVM_FALLTHROUGH;
1284
1285 case TypeWidenVector:
1286 if (isPowerOf2_32(NElts)) {
1287 // Try to widen the vector.
1288 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1289 MVT SVT = (MVT::SimpleValueType) nVT;
1290 if (SVT.getVectorElementType() == EltVT
1291 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) {
1292 TransformToType[i] = SVT;
1293 RegisterTypeForVT[i] = SVT;
1294 NumRegistersForVT[i] = 1;
1295 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1296 IsLegalWiderType = true;
1297 break;
1298 }
1299 }
1300 if (IsLegalWiderType)
1301 break;
1302 } else {
1303 // Only widen to the next power of 2 to keep consistency with EVT.
1304 MVT NVT = VT.getPow2VectorType();
1305 if (isTypeLegal(NVT)) {
1306 TransformToType[i] = NVT;
1307 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1308 RegisterTypeForVT[i] = NVT;
1309 NumRegistersForVT[i] = 1;
1310 break;
1311 }
1312 }
1313 LLVM_FALLTHROUGH;
1314
1315 case TypeSplitVector:
1316 case TypeScalarizeVector: {
1317 MVT IntermediateVT;
1318 MVT RegisterVT;
1319 unsigned NumIntermediates;
1320 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1321 NumIntermediates, RegisterVT, this);
1322 RegisterTypeForVT[i] = RegisterVT;
1323
1324 MVT NVT = VT.getPow2VectorType();
1325 if (NVT == VT) {
1326 // Type is already a power of 2. The default action is to split.
1327 TransformToType[i] = MVT::Other;
1328 if (PreferredAction == TypeScalarizeVector)
1329 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1330 else if (PreferredAction == TypeSplitVector)
1331 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1332 else
1333 // Set type action according to the number of elements.
1334 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector
1335 : TypeSplitVector);
1336 } else {
1337 TransformToType[i] = NVT;
1338 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1339 }
1340 break;
1341 }
1342 default:
1343 llvm_unreachable("Unknown vector legalization action!");
1344 }
1345 }
1346
1347 // Determine the 'representative' register class for each value type.
1348 // An representative register class is the largest (meaning one which is
1349 // not a sub-register class / subreg register class) legal register class for
1350 // a group of value types. For example, on i386, i8, i16, and i32
1351 // representative would be GR32; while on x86_64 it's GR64.
1352 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1353 const TargetRegisterClass* RRC;
1354 uint8_t Cost;
1355 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1356 RepRegClassForVT[i] = RRC;
1357 RepRegClassCostForVT[i] = Cost;
1358 }
1359 }
1360
1361 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1362 EVT VT) const {
1363 assert(!VT.isVector() && "No default SetCC type for vectors!");
1364 return getPointerTy(DL).SimpleTy;
1365 }
1366
1367 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1368 return MVT::i32; // return the default value
1369 }
1370
1371 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1372 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1373 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1374 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1375 ///
1376 /// This method returns the number of registers needed, and the VT for each
1377 /// register. It also returns the VT and quantity of the intermediate values
1378 /// before they are promoted/expanded.
1379 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1380 EVT &IntermediateVT,
1381 unsigned &NumIntermediates,
1382 MVT &RegisterVT) const {
1383 unsigned NumElts = VT.getVectorNumElements();
1384
1385 // If there is a wider vector type with the same element type as this one,
1386 // or a promoted vector type that has the same number of elements which
1387 // are wider, then we should convert to that legal vector type.
1388 // This handles things like <2 x float> -> <4 x float> and
1389 // <4 x i1> -> <4 x i32>.
1390 LegalizeTypeAction TA = getTypeAction(Context, VT);
1391 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1392 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1393 if (isTypeLegal(RegisterEVT)) {
1394 IntermediateVT = RegisterEVT;
1395 RegisterVT = RegisterEVT.getSimpleVT();
1396 NumIntermediates = 1;
1397 return 1;
1398 }
1399 }
1400
1401 // Figure out the right, legal destination reg to copy into.
1402 EVT EltTy = VT.getVectorElementType();
1403
1404 unsigned NumVectorRegs = 1;
1405
1406 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1407 // could break down into LHS/RHS like LegalizeDAG does.
1408 if (!isPowerOf2_32(NumElts)) {
1409 NumVectorRegs = NumElts;
1410 NumElts = 1;
1411 }
1412
1413 // Divide the input until we get to a supported size. This will always
1414 // end with a scalar if the target doesn't support vectors.
1415 while (NumElts > 1 && !isTypeLegal(
1416 EVT::getVectorVT(Context, EltTy, NumElts))) {
1417 NumElts >>= 1;
1418 NumVectorRegs <<= 1;
1419 }
1420
1421 NumIntermediates = NumVectorRegs;
1422
1423 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
1424 if (!isTypeLegal(NewVT))
1425 NewVT = EltTy;
1426 IntermediateVT = NewVT;
1427
1428 MVT DestVT = getRegisterType(Context, NewVT);
1429 RegisterVT = DestVT;
1430 unsigned NewVTSize = NewVT.getSizeInBits();
1431
1432 // Convert sizes such as i33 to i64.
1433 if (!isPowerOf2_32(NewVTSize))
1434 NewVTSize = NextPowerOf2(NewVTSize);
1435
1436 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1437 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1438
1439 // Otherwise, promotion or legal types use the same number of registers as
1440 // the vector decimated to the appropriate level.
1441 return NumVectorRegs;
1442 }
1443
1444 /// Get the EVTs and ArgFlags collections that represent the legalized return
1445 /// type of the given function. This does not require a DAG or a return value,
1446 /// and is suitable for use before any DAGs for the function are constructed.
1447 /// TODO: Move this out of TargetLowering.cpp.
1448 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1449 AttributeList attr,
1450 SmallVectorImpl<ISD::OutputArg> &Outs,
1451 const TargetLowering &TLI, const DataLayout &DL) {
1452 SmallVector<EVT, 4> ValueVTs;
1453 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1454 unsigned NumValues = ValueVTs.size();
1455 if (NumValues == 0) return;
1456
1457 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1458 EVT VT = ValueVTs[j];
1459 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1460
1461 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1462 ExtendKind = ISD::SIGN_EXTEND;
1463 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1464 ExtendKind = ISD::ZERO_EXTEND;
1465
1466 // FIXME: C calling convention requires the return type to be promoted to
1467 // at least 32-bit. But this is not necessary for non-C calling
1468 // conventions. The frontend should mark functions whose return values
1469 // require promoting with signext or zeroext attributes.
1470 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1471 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1472 if (VT.bitsLT(MinVT))
1473 VT = MinVT;
1474 }
1475
1476 unsigned NumParts =
1477 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
1478 MVT PartVT =
1479 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
1480
1481 // 'inreg' on function refers to return value
1482 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1483 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
1484 Flags.setInReg();
1485
1486 // Propagate extension type if any
1487 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1488 Flags.setSExt();
1489 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1490 Flags.setZExt();
1491
1492 for (unsigned i = 0; i < NumParts; ++i)
1493 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0));
1494 }
1495 }
1496
1497 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1498 /// function arguments in the caller parameter area. This is the actual
1499 /// alignment, not its logarithm.
1500 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1501 const DataLayout &DL) const {
1502 return DL.getABITypeAlignment(Ty);
1503 }
1504
1505 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1506 const DataLayout &DL, EVT VT,
1507 unsigned AddrSpace,
1508 unsigned Alignment,
1509 MachineMemOperand::Flags Flags,
1510 bool *Fast) const {
1511 // Check if the specified alignment is sufficient based on the data layout.
1512 // TODO: While using the data layout works in practice, a better solution
1513 // would be to implement this check directly (make this a virtual function).
1514 // For example, the ABI alignment may change based on software platform while
1515 // this function should only be affected by hardware implementation.
1516 Type *Ty = VT.getTypeForEVT(Context);
1517 if (Alignment >= DL.getABITypeAlignment(Ty)) {
1518 // Assume that an access that meets the ABI-specified alignment is fast.
1519 if (Fast != nullptr)
1520 *Fast = true;
1521 return true;
1522 }
1523
1524 // This is a misaligned access.
1525 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast);
1526 }
1527
1528 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1529 const DataLayout &DL, EVT VT,
1530 const MachineMemOperand &MMO,
1531 bool *Fast) const {
1532 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(),
1533 MMO.getAlignment(), MMO.getFlags(), Fast);
1534 }
1535
1536 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const {
1537 return BranchProbability(MinPercentageForPredictableBranch, 100);
1538 }
1539
1540 //===----------------------------------------------------------------------===//
1541 // TargetTransformInfo Helpers
1542 //===----------------------------------------------------------------------===//
1543
1544 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1545 enum InstructionOpcodes {
1546 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1547 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1548 #include "llvm/IR/Instruction.def"
1549 };
1550 switch (static_cast<InstructionOpcodes>(Opcode)) {
1551 case Ret: return 0;
1552 case Br: return 0;
1553 case Switch: return 0;
1554 case IndirectBr: return 0;
1555 case Invoke: return 0;
1556 case CallBr: return 0;
1557 case Resume: return 0;
1558 case Unreachable: return 0;
1559 case CleanupRet: return 0;
1560 case CatchRet: return 0;
1561 case CatchPad: return 0;
1562 case CatchSwitch: return 0;
1563 case CleanupPad: return 0;
1564 case FNeg: return ISD::FNEG;
1565 case Add: return ISD::ADD;
1566 case FAdd: return ISD::FADD;
1567 case Sub: return ISD::SUB;
1568 case FSub: return ISD::FSUB;
1569 case Mul: return ISD::MUL;
1570 case FMul: return ISD::FMUL;
1571 case UDiv: return ISD::UDIV;
1572 case SDiv: return ISD::SDIV;
1573 case FDiv: return ISD::FDIV;
1574 case URem: return ISD::UREM;
1575 case SRem: return ISD::SREM;
1576 case FRem: return ISD::FREM;
1577 case Shl: return ISD::SHL;
1578 case LShr: return ISD::SRL;
1579 case AShr: return ISD::SRA;
1580 case And: return ISD::AND;
1581 case Or: return ISD::OR;
1582 case Xor: return ISD::XOR;
1583 case Alloca: return 0;
1584 case Load: return ISD::LOAD;
1585 case Store: return ISD::STORE;
1586 case GetElementPtr: return 0;
1587 case Fence: return 0;
1588 case AtomicCmpXchg: return 0;
1589 case AtomicRMW: return 0;
1590 case Trunc: return ISD::TRUNCATE;
1591 case ZExt: return ISD::ZERO_EXTEND;
1592 case SExt: return ISD::SIGN_EXTEND;
1593 case FPToUI: return ISD::FP_TO_UINT;
1594 case FPToSI: return ISD::FP_TO_SINT;
1595 case UIToFP: return ISD::UINT_TO_FP;
1596 case SIToFP: return ISD::SINT_TO_FP;
1597 case FPTrunc: return ISD::FP_ROUND;
1598 case FPExt: return ISD::FP_EXTEND;
1599 case PtrToInt: return ISD::BITCAST;
1600 case IntToPtr: return ISD::BITCAST;
1601 case BitCast: return ISD::BITCAST;
1602 case AddrSpaceCast: return ISD::ADDRSPACECAST;
1603 case ICmp: return ISD::SETCC;
1604 case FCmp: return ISD::SETCC;
1605 case PHI: return 0;
1606 case Call: return 0;
1607 case Select: return ISD::SELECT;
1608 case UserOp1: return 0;
1609 case UserOp2: return 0;
1610 case VAArg: return 0;
1611 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1612 case InsertElement: return ISD::INSERT_VECTOR_ELT;
1613 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
1614 case ExtractValue: return ISD::MERGE_VALUES;
1615 case InsertValue: return ISD::MERGE_VALUES;
1616 case LandingPad: return 0;
1617 }
1618
1619 llvm_unreachable("Unknown instruction type encountered!");
1620 }
1621
1622 std::pair<int, MVT>
1623 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1624 Type *Ty) const {
1625 LLVMContext &C = Ty->getContext();
1626 EVT MTy = getValueType(DL, Ty);
1627
1628 int Cost = 1;
1629 // We keep legalizing the type until we find a legal kind. We assume that
1630 // the only operation that costs anything is the split. After splitting
1631 // we need to handle two types.
1632 while (true) {
1633 LegalizeKind LK = getTypeConversion(C, MTy);
1634
1635 if (LK.first == TypeLegal)
1636 return std::make_pair(Cost, MTy.getSimpleVT());
1637
1638 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1639 Cost *= 2;
1640
1641 // Do not loop with f128 type.
1642 if (MTy == LK.second)
1643 return std::make_pair(Cost, MTy.getSimpleVT());
1644
1645 // Keep legalizing the type.
1646 MTy = LK.second;
1647 }
1648 }
1649
1650 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB,
1651 bool UseTLS) const {
1652 // compiler-rt provides a variable with a magic name. Targets that do not
1653 // link with compiler-rt may also provide such a variable.
1654 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1655 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1656 auto UnsafeStackPtr =
1657 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1658
1659 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1660
1661 if (!UnsafeStackPtr) {
1662 auto TLSModel = UseTLS ?
1663 GlobalValue::InitialExecTLSModel :
1664 GlobalValue::NotThreadLocal;
1665 // The global variable is not defined yet, define it ourselves.
1666 // We use the initial-exec TLS model because we do not support the
1667 // variable living anywhere other than in the main executable.
1668 UnsafeStackPtr = new GlobalVariable(
1669 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1670 UnsafeStackPtrVar, nullptr, TLSModel);
1671 } else {
1672 // The variable exists, check its type and attributes.
1673 if (UnsafeStackPtr->getValueType() != StackPtrTy)
1674 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1675 if (UseTLS != UnsafeStackPtr->isThreadLocal())
1676 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1677 (UseTLS ? "" : "not ") + "be thread-local");
1678 }
1679 return UnsafeStackPtr;
1680 }
1681
1682 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
1683 if (!TM.getTargetTriple().isAndroid())
1684 return getDefaultSafeStackPointerLocation(IRB, true);
1685
1686 // Android provides a libc function to retrieve the address of the current
1687 // thread's unsafe stack pointer.
1688 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1689 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1690 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address",
1691 StackPtrTy->getPointerTo(0));
1692 return IRB.CreateCall(Fn);
1693 }
1694
1695 //===----------------------------------------------------------------------===//
1696 // Loop Strength Reduction hooks
1697 //===----------------------------------------------------------------------===//
1698
1699 /// isLegalAddressingMode - Return true if the addressing mode represented
1700 /// by AM is legal for this target, for a load/store of the specified type.
1701 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1702 const AddrMode &AM, Type *Ty,
1703 unsigned AS, Instruction *I) const {
1704 // The default implementation of this implements a conservative RISCy, r+r and
1705 // r+i addr mode.
1706
1707 // Allows a sign-extended 16-bit immediate field.
1708 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1709 return false;
1710
1711 // No global is ever allowed as a base.
1712 if (AM.BaseGV)
1713 return false;
1714
1715 // Only support r+r,
1716 switch (AM.Scale) {
1717 case 0: // "r+i" or just "i", depending on HasBaseReg.
1718 break;
1719 case 1:
1720 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1721 return false;
1722 // Otherwise we have r+r or r+i.
1723 break;
1724 case 2:
1725 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1726 return false;
1727 // Allow 2*r as r+r.
1728 break;
1729 default: // Don't allow n * r
1730 return false;
1731 }
1732
1733 return true;
1734 }
1735
1736 //===----------------------------------------------------------------------===//
1737 // Stack Protector
1738 //===----------------------------------------------------------------------===//
1739
1740 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1741 // so that SelectionDAG handle SSP.
1742 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const {
1743 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1744 Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1745 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1746 return M.getOrInsertGlobal("__guard_local", PtrTy);
1747 }
1748 return nullptr;
1749 }
1750
1751 // Currently only support "standard" __stack_chk_guard.
1752 // TODO: add LOAD_STACK_GUARD support.
1753 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1754 if (!M.getNamedValue("__stack_chk_guard"))
1755 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1756 GlobalVariable::ExternalLinkage,
1757 nullptr, "__stack_chk_guard");
1758 }
1759
1760 // Currently only support "standard" __stack_chk_guard.
1761 // TODO: add LOAD_STACK_GUARD support.
1762 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
1763 return M.getNamedValue("__stack_chk_guard");
1764 }
1765
1766 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
1767 return nullptr;
1768 }
1769
1770 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1771 return MinimumJumpTableEntries;
1772 }
1773
1774 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
1775 MinimumJumpTableEntries = Val;
1776 }
1777
1778 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
1779 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
1780 }
1781
1782 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1783 return MaximumJumpTableSize;
1784 }
1785
1786 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
1787 MaximumJumpTableSize = Val;
1788 }
1789
1790 //===----------------------------------------------------------------------===//
1791 // Reciprocal Estimates
1792 //===----------------------------------------------------------------------===//
1793
1794 /// Get the reciprocal estimate attribute string for a function that will
1795 /// override the target defaults.
1796 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
1797 const Function &F = MF.getFunction();
1798 return F.getFnAttribute("reciprocal-estimates").getValueAsString();
1799 }
1800
1801 /// Construct a string for the given reciprocal operation of the given type.
1802 /// This string should match the corresponding option to the front-end's
1803 /// "-mrecip" flag assuming those strings have been passed through in an
1804 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1805 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
1806 std::string Name = VT.isVector() ? "vec-" : "";
1807
1808 Name += IsSqrt ? "sqrt" : "div";
1809
1810 // TODO: Handle "half" or other float types?
1811 if (VT.getScalarType() == MVT::f64) {
1812 Name += "d";
1813 } else {
1814 assert(VT.getScalarType() == MVT::f32 &&
1815 "Unexpected FP type for reciprocal estimate");
1816 Name += "f";
1817 }
1818
1819 return Name;
1820 }
1821
1822 /// Return the character position and value (a single numeric character) of a
1823 /// customized refinement operation in the input string if it exists. Return
1824 /// false if there is no customized refinement step count.
1825 static bool parseRefinementStep(StringRef In, size_t &Position,
1826 uint8_t &Value) {
1827 const char RefStepToken = ':';
1828 Position = In.find(RefStepToken);
1829 if (Position == StringRef::npos)
1830 return false;
1831
1832 StringRef RefStepString = In.substr(Position + 1);
1833 // Allow exactly one numeric character for the additional refinement
1834 // step parameter.
1835 if (RefStepString.size() == 1) {
1836 char RefStepChar = RefStepString[0];
1837 if (RefStepChar >= '0' && RefStepChar <= '9') {
1838 Value = RefStepChar - '0';
1839 return true;
1840 }
1841 }
1842 report_fatal_error("Invalid refinement step for -recip.");
1843 }
1844
1845 /// For the input attribute string, return one of the ReciprocalEstimate enum
1846 /// status values (enabled, disabled, or not specified) for this operation on
1847 /// the specified data type.
1848 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
1849 if (Override.empty())
1850 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1851
1852 SmallVector<StringRef, 4> OverrideVector;
1853 Override.split(OverrideVector, ',');
1854 unsigned NumArgs = OverrideVector.size();
1855
1856 // Check if "all", "none", or "default" was specified.
1857 if (NumArgs == 1) {
1858 // Look for an optional setting of the number of refinement steps needed
1859 // for this type of reciprocal operation.
1860 size_t RefPos;
1861 uint8_t RefSteps;
1862 if (parseRefinementStep(Override, RefPos, RefSteps)) {
1863 // Split the string for further processing.
1864 Override = Override.substr(0, RefPos);
1865 }
1866
1867 // All reciprocal types are enabled.
1868 if (Override == "all")
1869 return TargetLoweringBase::ReciprocalEstimate::Enabled;
1870
1871 // All reciprocal types are disabled.
1872 if (Override == "none")
1873 return TargetLoweringBase::ReciprocalEstimate::Disabled;
1874
1875 // Target defaults for enablement are used.
1876 if (Override == "default")
1877 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1878 }
1879
1880 // The attribute string may omit the size suffix ('f'/'d').
1881 std::string VTName = getReciprocalOpName(IsSqrt, VT);
1882 std::string VTNameNoSize = VTName;
1883 VTNameNoSize.pop_back();
1884 static const char DisabledPrefix = '!';
1885
1886 for (StringRef RecipType : OverrideVector) {
1887 size_t RefPos;
1888 uint8_t RefSteps;
1889 if (parseRefinementStep(RecipType, RefPos, RefSteps))
1890 RecipType = RecipType.substr(0, RefPos);
1891
1892 // Ignore the disablement token for string matching.
1893 bool IsDisabled = RecipType[0] == DisabledPrefix;
1894 if (IsDisabled)
1895 RecipType = RecipType.substr(1);
1896
1897 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1898 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
1899 : TargetLoweringBase::ReciprocalEstimate::Enabled;
1900 }
1901
1902 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1903 }
1904
1905 /// For the input attribute string, return the customized refinement step count
1906 /// for this operation on the specified data type. If the step count does not
1907 /// exist, return the ReciprocalEstimate enum value for unspecified.
1908 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
1909 if (Override.empty())
1910 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1911
1912 SmallVector<StringRef, 4> OverrideVector;
1913 Override.split(OverrideVector, ',');
1914 unsigned NumArgs = OverrideVector.size();
1915
1916 // Check if "all", "default", or "none" was specified.
1917 if (NumArgs == 1) {
1918 // Look for an optional setting of the number of refinement steps needed
1919 // for this type of reciprocal operation.
1920 size_t RefPos;
1921 uint8_t RefSteps;
1922 if (!parseRefinementStep(Override, RefPos, RefSteps))
1923 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1924
1925 // Split the string for further processing.
1926 Override = Override.substr(0, RefPos);
1927 assert(Override != "none" &&
1928 "Disabled reciprocals, but specifed refinement steps?");
1929
1930 // If this is a general override, return the specified number of steps.
1931 if (Override == "all" || Override == "default")
1932 return RefSteps;
1933 }
1934
1935 // The attribute string may omit the size suffix ('f'/'d').
1936 std::string VTName = getReciprocalOpName(IsSqrt, VT);
1937 std::string VTNameNoSize = VTName;
1938 VTNameNoSize.pop_back();
1939
1940 for (StringRef RecipType : OverrideVector) {
1941 size_t RefPos;
1942 uint8_t RefSteps;
1943 if (!parseRefinementStep(RecipType, RefPos, RefSteps))
1944 continue;
1945
1946 RecipType = RecipType.substr(0, RefPos);
1947 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1948 return RefSteps;
1949 }
1950
1951 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1952 }
1953
1954 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
1955 MachineFunction &MF) const {
1956 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
1957 }
1958
1959 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
1960 MachineFunction &MF) const {
1961 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
1962 }
1963
1964 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
1965 MachineFunction &MF) const {
1966 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
1967 }
1968
1969 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
1970 MachineFunction &MF) const {
1971 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
1972 }
1973
1974 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
1975 MF.getRegInfo().freezeReservedRegs(MF);
1976 }
The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.