1 | //===-- X86TargetMachine.cpp - Define TargetMachine for the X86 -----------===// |
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 file defines the X86 specific subclass of TargetMachine. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "X86TargetMachine.h" |
14 | #include "MCTargetDesc/X86MCTargetDesc.h" |
15 | #include "TargetInfo/X86TargetInfo.h" |
16 | #include "X86.h" |
17 | #include "X86MachineFunctionInfo.h" |
18 | #include "X86MacroFusion.h" |
19 | #include "X86Subtarget.h" |
20 | #include "X86TargetObjectFile.h" |
21 | #include "X86TargetTransformInfo.h" |
22 | #include "llvm/ADT/STLExtras.h" |
23 | #include "llvm/ADT/SmallString.h" |
24 | #include "llvm/ADT/StringRef.h" |
25 | #include "llvm/Analysis/TargetTransformInfo.h" |
26 | #include "llvm/CodeGen/ExecutionDomainFix.h" |
27 | #include "llvm/CodeGen/GlobalISel/CSEInfo.h" |
28 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
29 | #include "llvm/CodeGen/GlobalISel/IRTranslator.h" |
30 | #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" |
31 | #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" |
32 | #include "llvm/CodeGen/GlobalISel/Legalizer.h" |
33 | #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" |
34 | #include "llvm/CodeGen/MIRParser/MIParser.h" |
35 | #include "llvm/CodeGen/MIRYamlMapping.h" |
36 | #include "llvm/CodeGen/MachineScheduler.h" |
37 | #include "llvm/CodeGen/Passes.h" |
38 | #include "llvm/CodeGen/RegAllocRegistry.h" |
39 | #include "llvm/CodeGen/TargetPassConfig.h" |
40 | #include "llvm/IR/Attributes.h" |
41 | #include "llvm/IR/DataLayout.h" |
42 | #include "llvm/IR/Function.h" |
43 | #include "llvm/MC/MCAsmInfo.h" |
44 | #include "llvm/MC/TargetRegistry.h" |
45 | #include "llvm/Pass.h" |
46 | #include "llvm/Support/CodeGen.h" |
47 | #include "llvm/Support/CommandLine.h" |
48 | #include "llvm/Support/ErrorHandling.h" |
49 | #include "llvm/Target/TargetLoweringObjectFile.h" |
50 | #include "llvm/Target/TargetOptions.h" |
51 | #include "llvm/TargetParser/Triple.h" |
52 | #include "llvm/Transforms/CFGuard.h" |
53 | #include <memory> |
54 | #include <optional> |
55 | #include <string> |
56 | |
57 | using namespace llvm; |
58 | |
59 | static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner" , |
60 | cl::desc("Enable the machine combiner pass" ), |
61 | cl::init(Val: true), cl::Hidden); |
62 | |
63 | static cl::opt<bool> |
64 | EnableTileRAPass("x86-tile-ra" , |
65 | cl::desc("Enable the tile register allocation pass" ), |
66 | cl::init(Val: true), cl::Hidden); |
67 | |
68 | extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeX86Target() { |
69 | // Register the target. |
70 | RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); |
71 | RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); |
72 | |
73 | PassRegistry &PR = *PassRegistry::getPassRegistry(); |
74 | initializeX86LowerAMXIntrinsicsLegacyPassPass(PR); |
75 | initializeX86LowerAMXTypeLegacyPassPass(PR); |
76 | initializeX86PreTileConfigPass(PR); |
77 | initializeGlobalISel(PR); |
78 | initializeWinEHStatePassPass(PR); |
79 | initializeFixupBWInstPassPass(PR); |
80 | initializeCompressEVEXPassPass(PR); |
81 | initializeFixupLEAPassPass(PR); |
82 | initializeFPSPass(PR); |
83 | initializeX86FixupSetCCPassPass(PR); |
84 | initializeX86CallFrameOptimizationPass(PR); |
85 | initializeX86CmovConverterPassPass(PR); |
86 | initializeX86TileConfigPass(PR); |
87 | initializeX86FastPreTileConfigPass(PR); |
88 | initializeX86FastTileConfigPass(PR); |
89 | initializeKCFIPass(PR); |
90 | initializeX86LowerTileCopyPass(PR); |
91 | initializeX86ExpandPseudoPass(PR); |
92 | initializeX86ExecutionDomainFixPass(PR); |
93 | initializeX86DomainReassignmentPass(PR); |
94 | initializeX86AvoidSFBPassPass(PR); |
95 | initializeX86AvoidTrailingCallPassPass(PR); |
96 | initializeX86SpeculativeLoadHardeningPassPass(PR); |
97 | initializeX86SpeculativeExecutionSideEffectSuppressionPass(PR); |
98 | initializeX86FlagsCopyLoweringPassPass(PR); |
99 | initializeX86LoadValueInjectionLoadHardeningPassPass(PR); |
100 | initializeX86LoadValueInjectionRetHardeningPassPass(PR); |
101 | initializeX86OptimizeLEAPassPass(PR); |
102 | initializeX86PartialReductionPass(PR); |
103 | initializePseudoProbeInserterPass(PR); |
104 | initializeX86ReturnThunksPass(PR); |
105 | initializeX86DAGToDAGISelLegacyPass(PR); |
106 | initializeX86ArgumentStackSlotPassPass(PR); |
107 | initializeX86FixupInstTuningPassPass(PR); |
108 | initializeX86FixupVectorConstantsPassPass(PR); |
109 | } |
110 | |
111 | static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { |
112 | if (TT.isOSBinFormatMachO()) { |
113 | if (TT.getArch() == Triple::x86_64) |
114 | return std::make_unique<X86_64MachoTargetObjectFile>(); |
115 | return std::make_unique<TargetLoweringObjectFileMachO>(); |
116 | } |
117 | |
118 | if (TT.isOSBinFormatCOFF()) |
119 | return std::make_unique<TargetLoweringObjectFileCOFF>(); |
120 | |
121 | if (TT.getArch() == Triple::x86_64) |
122 | return std::make_unique<X86_64ELFTargetObjectFile>(); |
123 | return std::make_unique<X86ELFTargetObjectFile>(); |
124 | } |
125 | |
126 | static std::string computeDataLayout(const Triple &TT) { |
127 | // X86 is little endian |
128 | std::string Ret = "e" ; |
129 | |
130 | Ret += DataLayout::getManglingComponent(T: TT); |
131 | // X86 and x32 have 32 bit pointers. |
132 | if (!TT.isArch64Bit() || TT.isX32() || TT.isOSNaCl()) |
133 | Ret += "-p:32:32" ; |
134 | |
135 | // Address spaces for 32 bit signed, 32 bit unsigned, and 64 bit pointers. |
136 | Ret += "-p270:32:32-p271:32:32-p272:64:64" ; |
137 | |
138 | // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. |
139 | // 128 bit integers are not specified in the 32-bit ABIs but are used |
140 | // internally for lowering f128, so we match the alignment to that. |
141 | if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) |
142 | Ret += "-i64:64-i128:128" ; |
143 | else if (TT.isOSIAMCU()) |
144 | Ret += "-i64:32-f64:32" ; |
145 | else |
146 | Ret += "-i128:128-f64:32:64" ; |
147 | |
148 | // Some ABIs align long double to 128 bits, others to 32. |
149 | if (TT.isOSNaCl() || TT.isOSIAMCU()) |
150 | ; // No f80 |
151 | else if (TT.isArch64Bit() || TT.isOSDarwin() || TT.isWindowsMSVCEnvironment()) |
152 | Ret += "-f80:128" ; |
153 | else |
154 | Ret += "-f80:32" ; |
155 | |
156 | if (TT.isOSIAMCU()) |
157 | Ret += "-f128:32" ; |
158 | |
159 | // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. |
160 | if (TT.isArch64Bit()) |
161 | Ret += "-n8:16:32:64" ; |
162 | else |
163 | Ret += "-n8:16:32" ; |
164 | |
165 | // The stack is aligned to 32 bits on some ABIs and 128 bits on others. |
166 | if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) |
167 | Ret += "-a:0:32-S32" ; |
168 | else |
169 | Ret += "-S128" ; |
170 | |
171 | return Ret; |
172 | } |
173 | |
174 | static Reloc::Model getEffectiveRelocModel(const Triple &TT, bool JIT, |
175 | std::optional<Reloc::Model> RM) { |
176 | bool is64Bit = TT.getArch() == Triple::x86_64; |
177 | if (!RM) { |
178 | // JIT codegen should use static relocations by default, since it's |
179 | // typically executed in process and not relocatable. |
180 | if (JIT) |
181 | return Reloc::Static; |
182 | |
183 | // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. |
184 | // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we |
185 | // use static relocation model by default. |
186 | if (TT.isOSDarwin()) { |
187 | if (is64Bit) |
188 | return Reloc::PIC_; |
189 | return Reloc::DynamicNoPIC; |
190 | } |
191 | if (TT.isOSWindows() && is64Bit) |
192 | return Reloc::PIC_; |
193 | return Reloc::Static; |
194 | } |
195 | |
196 | // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC |
197 | // is defined as a model for code which may be used in static or dynamic |
198 | // executables but not necessarily a shared library. On X86-32 we just |
199 | // compile in -static mode, in x86-64 we use PIC. |
200 | if (*RM == Reloc::DynamicNoPIC) { |
201 | if (is64Bit) |
202 | return Reloc::PIC_; |
203 | if (!TT.isOSDarwin()) |
204 | return Reloc::Static; |
205 | } |
206 | |
207 | // If we are on Darwin, disallow static relocation model in X86-64 mode, since |
208 | // the Mach-O file format doesn't support it. |
209 | if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) |
210 | return Reloc::PIC_; |
211 | |
212 | return *RM; |
213 | } |
214 | |
215 | static CodeModel::Model |
216 | getEffectiveX86CodeModel(const Triple &TT, std::optional<CodeModel::Model> CM, |
217 | bool JIT) { |
218 | bool Is64Bit = TT.getArch() == Triple::x86_64; |
219 | if (CM) { |
220 | if (*CM == CodeModel::Tiny) |
221 | report_fatal_error(reason: "Target does not support the tiny CodeModel" , gen_crash_diag: false); |
222 | return *CM; |
223 | } |
224 | if (JIT) |
225 | return Is64Bit ? CodeModel::Large : CodeModel::Small; |
226 | return CodeModel::Small; |
227 | } |
228 | |
229 | /// Create an X86 target. |
230 | /// |
231 | X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, |
232 | StringRef CPU, StringRef FS, |
233 | const TargetOptions &Options, |
234 | std::optional<Reloc::Model> RM, |
235 | std::optional<CodeModel::Model> CM, |
236 | CodeGenOptLevel OL, bool JIT) |
237 | : LLVMTargetMachine( |
238 | T, computeDataLayout(TT), TT, CPU, FS, Options, |
239 | getEffectiveRelocModel(TT, JIT, RM), |
240 | getEffectiveX86CodeModel(TT, CM, JIT), |
241 | OL), |
242 | TLOF(createTLOF(TT: getTargetTriple())), IsJIT(JIT) { |
243 | // On PS4/PS5, the "return address" of a 'noreturn' call must still be within |
244 | // the calling function, and TrapUnreachable is an easy way to get that. |
245 | if (TT.isPS() || TT.isOSBinFormatMachO()) { |
246 | this->Options.TrapUnreachable = true; |
247 | this->Options.NoTrapAfterNoreturn = TT.isOSBinFormatMachO(); |
248 | } |
249 | |
250 | setMachineOutliner(true); |
251 | |
252 | // x86 supports the debug entry values. |
253 | setSupportsDebugEntryValues(true); |
254 | |
255 | initAsmInfo(); |
256 | } |
257 | |
258 | X86TargetMachine::~X86TargetMachine() = default; |
259 | |
260 | const X86Subtarget * |
261 | X86TargetMachine::getSubtargetImpl(const Function &F) const { |
262 | Attribute CPUAttr = F.getFnAttribute(Kind: "target-cpu" ); |
263 | Attribute TuneAttr = F.getFnAttribute(Kind: "tune-cpu" ); |
264 | Attribute FSAttr = F.getFnAttribute(Kind: "target-features" ); |
265 | |
266 | StringRef CPU = |
267 | CPUAttr.isValid() ? CPUAttr.getValueAsString() : (StringRef)TargetCPU; |
268 | // "x86-64" is a default target setting for many front ends. In these cases, |
269 | // they actually request for "generic" tuning unless the "tune-cpu" was |
270 | // specified. |
271 | StringRef TuneCPU = TuneAttr.isValid() ? TuneAttr.getValueAsString() |
272 | : CPU == "x86-64" ? "generic" |
273 | : (StringRef)CPU; |
274 | StringRef FS = |
275 | FSAttr.isValid() ? FSAttr.getValueAsString() : (StringRef)TargetFS; |
276 | |
277 | SmallString<512> Key; |
278 | // The additions here are ordered so that the definitely short strings are |
279 | // added first so we won't exceed the small size. We append the |
280 | // much longer FS string at the end so that we only heap allocate at most |
281 | // one time. |
282 | |
283 | // Extract prefer-vector-width attribute. |
284 | unsigned PreferVectorWidthOverride = 0; |
285 | Attribute PreferVecWidthAttr = F.getFnAttribute(Kind: "prefer-vector-width" ); |
286 | if (PreferVecWidthAttr.isValid()) { |
287 | StringRef Val = PreferVecWidthAttr.getValueAsString(); |
288 | unsigned Width; |
289 | if (!Val.getAsInteger(Radix: 0, Result&: Width)) { |
290 | Key += 'p'; |
291 | Key += Val; |
292 | PreferVectorWidthOverride = Width; |
293 | } |
294 | } |
295 | |
296 | // Extract min-legal-vector-width attribute. |
297 | unsigned RequiredVectorWidth = UINT32_MAX; |
298 | Attribute MinLegalVecWidthAttr = F.getFnAttribute(Kind: "min-legal-vector-width" ); |
299 | if (MinLegalVecWidthAttr.isValid()) { |
300 | StringRef Val = MinLegalVecWidthAttr.getValueAsString(); |
301 | unsigned Width; |
302 | if (!Val.getAsInteger(Radix: 0, Result&: Width)) { |
303 | Key += 'm'; |
304 | Key += Val; |
305 | RequiredVectorWidth = Width; |
306 | } |
307 | } |
308 | |
309 | // Add CPU to the Key. |
310 | Key += CPU; |
311 | |
312 | // Add tune CPU to the Key. |
313 | Key += TuneCPU; |
314 | |
315 | // Keep track of the start of the feature portion of the string. |
316 | unsigned FSStart = Key.size(); |
317 | |
318 | // FIXME: This is related to the code below to reset the target options, |
319 | // we need to know whether or not the soft float flag is set on the |
320 | // function before we can generate a subtarget. We also need to use |
321 | // it as a key for the subtarget since that can be the only difference |
322 | // between two functions. |
323 | bool SoftFloat = F.getFnAttribute(Kind: "use-soft-float" ).getValueAsBool(); |
324 | // If the soft float attribute is set on the function turn on the soft float |
325 | // subtarget feature. |
326 | if (SoftFloat) |
327 | Key += FS.empty() ? "+soft-float" : "+soft-float," ; |
328 | |
329 | Key += FS; |
330 | |
331 | // We may have added +soft-float to the features so move the StringRef to |
332 | // point to the full string in the Key. |
333 | FS = Key.substr(Start: FSStart); |
334 | |
335 | auto &I = SubtargetMap[Key]; |
336 | if (!I) { |
337 | // This needs to be done before we create a new subtarget since any |
338 | // creation will depend on the TM and the code generation flags on the |
339 | // function that reside in TargetOptions. |
340 | resetTargetOptions(F); |
341 | I = std::make_unique<X86Subtarget>( |
342 | args: TargetTriple, args&: CPU, args&: TuneCPU, args&: FS, args: *this, |
343 | args: MaybeAlign(F.getParent()->getOverrideStackAlignment()), |
344 | args&: PreferVectorWidthOverride, args&: RequiredVectorWidth); |
345 | } |
346 | return I.get(); |
347 | } |
348 | |
349 | yaml::MachineFunctionInfo *X86TargetMachine::createDefaultFuncInfoYAML() const { |
350 | return new yaml::X86MachineFunctionInfo(); |
351 | } |
352 | |
353 | yaml::MachineFunctionInfo * |
354 | X86TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { |
355 | const auto *MFI = MF.getInfo<X86MachineFunctionInfo>(); |
356 | return new yaml::X86MachineFunctionInfo(*MFI); |
357 | } |
358 | |
359 | bool X86TargetMachine::parseMachineFunctionInfo( |
360 | const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS, |
361 | SMDiagnostic &Error, SMRange &SourceRange) const { |
362 | const auto &YamlMFI = static_cast<const yaml::X86MachineFunctionInfo &>(MFI); |
363 | PFS.MF.getInfo<X86MachineFunctionInfo>()->initializeBaseYamlFields(YamlMFI); |
364 | return false; |
365 | } |
366 | |
367 | bool X86TargetMachine::isNoopAddrSpaceCast(unsigned SrcAS, |
368 | unsigned DestAS) const { |
369 | assert(SrcAS != DestAS && "Expected different address spaces!" ); |
370 | if (getPointerSize(AS: SrcAS) != getPointerSize(AS: DestAS)) |
371 | return false; |
372 | return SrcAS < 256 && DestAS < 256; |
373 | } |
374 | |
375 | //===----------------------------------------------------------------------===// |
376 | // X86 TTI query. |
377 | //===----------------------------------------------------------------------===// |
378 | |
379 | TargetTransformInfo |
380 | X86TargetMachine::getTargetTransformInfo(const Function &F) const { |
381 | return TargetTransformInfo(X86TTIImpl(this, F)); |
382 | } |
383 | |
384 | //===----------------------------------------------------------------------===// |
385 | // Pass Pipeline Configuration |
386 | //===----------------------------------------------------------------------===// |
387 | |
388 | namespace { |
389 | |
390 | /// X86 Code Generator Pass Configuration Options. |
391 | class X86PassConfig : public TargetPassConfig { |
392 | public: |
393 | X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM) |
394 | : TargetPassConfig(TM, PM) {} |
395 | |
396 | X86TargetMachine &getX86TargetMachine() const { |
397 | return getTM<X86TargetMachine>(); |
398 | } |
399 | |
400 | ScheduleDAGInstrs * |
401 | createMachineScheduler(MachineSchedContext *C) const override { |
402 | ScheduleDAGMILive *DAG = createGenericSchedLive(C); |
403 | DAG->addMutation(Mutation: createX86MacroFusionDAGMutation()); |
404 | return DAG; |
405 | } |
406 | |
407 | ScheduleDAGInstrs * |
408 | createPostMachineScheduler(MachineSchedContext *C) const override { |
409 | ScheduleDAGMI *DAG = createGenericSchedPostRA(C); |
410 | DAG->addMutation(Mutation: createX86MacroFusionDAGMutation()); |
411 | return DAG; |
412 | } |
413 | |
414 | void addIRPasses() override; |
415 | bool addInstSelector() override; |
416 | bool addIRTranslator() override; |
417 | bool addLegalizeMachineIR() override; |
418 | bool addRegBankSelect() override; |
419 | bool addGlobalInstructionSelect() override; |
420 | bool addILPOpts() override; |
421 | bool addPreISel() override; |
422 | void addMachineSSAOptimization() override; |
423 | void addPreRegAlloc() override; |
424 | bool addPostFastRegAllocRewrite() override; |
425 | void addPostRegAlloc() override; |
426 | void addPreEmitPass() override; |
427 | void addPreEmitPass2() override; |
428 | void addPreSched2() override; |
429 | bool addRegAssignAndRewriteOptimized() override; |
430 | |
431 | std::unique_ptr<CSEConfigBase> getCSEConfig() const override; |
432 | }; |
433 | |
434 | class X86ExecutionDomainFix : public ExecutionDomainFix { |
435 | public: |
436 | static char ID; |
437 | X86ExecutionDomainFix() : ExecutionDomainFix(ID, X86::VR128XRegClass) {} |
438 | StringRef getPassName() const override { |
439 | return "X86 Execution Dependency Fix" ; |
440 | } |
441 | }; |
442 | char X86ExecutionDomainFix::ID; |
443 | |
444 | } // end anonymous namespace |
445 | |
446 | INITIALIZE_PASS_BEGIN(X86ExecutionDomainFix, "x86-execution-domain-fix" , |
447 | "X86 Execution Domain Fix" , false, false) |
448 | INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis) |
449 | INITIALIZE_PASS_END(X86ExecutionDomainFix, "x86-execution-domain-fix" , |
450 | "X86 Execution Domain Fix" , false, false) |
451 | |
452 | TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { |
453 | return new X86PassConfig(*this, PM); |
454 | } |
455 | |
456 | MachineFunctionInfo *X86TargetMachine::createMachineFunctionInfo( |
457 | BumpPtrAllocator &Allocator, const Function &F, |
458 | const TargetSubtargetInfo *STI) const { |
459 | return X86MachineFunctionInfo::create<X86MachineFunctionInfo>(Allocator, F, |
460 | STI); |
461 | } |
462 | |
463 | void X86PassConfig::addIRPasses() { |
464 | addPass(P: createAtomicExpandLegacyPass()); |
465 | |
466 | // We add both pass anyway and when these two passes run, we skip the pass |
467 | // based on the option level and option attribute. |
468 | addPass(P: createX86LowerAMXIntrinsicsPass()); |
469 | addPass(P: createX86LowerAMXTypePass()); |
470 | |
471 | TargetPassConfig::addIRPasses(); |
472 | |
473 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
474 | addPass(P: createInterleavedAccessPass()); |
475 | addPass(P: createX86PartialReductionPass()); |
476 | } |
477 | |
478 | // Add passes that handle indirect branch removal and insertion of a retpoline |
479 | // thunk. These will be a no-op unless a function subtarget has the retpoline |
480 | // feature enabled. |
481 | addPass(P: createIndirectBrExpandPass()); |
482 | |
483 | // Add Control Flow Guard checks. |
484 | const Triple &TT = TM->getTargetTriple(); |
485 | if (TT.isOSWindows()) { |
486 | if (TT.getArch() == Triple::x86_64) { |
487 | addPass(P: createCFGuardDispatchPass()); |
488 | } else { |
489 | addPass(P: createCFGuardCheckPass()); |
490 | } |
491 | } |
492 | |
493 | if (TM->Options.JMCInstrument) |
494 | addPass(P: createJMCInstrumenterPass()); |
495 | } |
496 | |
497 | bool X86PassConfig::addInstSelector() { |
498 | // Install an instruction selector. |
499 | addPass(P: createX86ISelDag(TM&: getX86TargetMachine(), OptLevel: getOptLevel())); |
500 | |
501 | // For ELF, cleanup any local-dynamic TLS accesses. |
502 | if (TM->getTargetTriple().isOSBinFormatELF() && |
503 | getOptLevel() != CodeGenOptLevel::None) |
504 | addPass(P: createCleanupLocalDynamicTLSPass()); |
505 | |
506 | addPass(P: createX86GlobalBaseRegPass()); |
507 | addPass(P: createX86ArgumentStackSlotPass()); |
508 | return false; |
509 | } |
510 | |
511 | bool X86PassConfig::addIRTranslator() { |
512 | addPass(P: new IRTranslator(getOptLevel())); |
513 | return false; |
514 | } |
515 | |
516 | bool X86PassConfig::addLegalizeMachineIR() { |
517 | addPass(P: new Legalizer()); |
518 | return false; |
519 | } |
520 | |
521 | bool X86PassConfig::addRegBankSelect() { |
522 | addPass(P: new RegBankSelect()); |
523 | return false; |
524 | } |
525 | |
526 | bool X86PassConfig::addGlobalInstructionSelect() { |
527 | addPass(P: new InstructionSelect(getOptLevel())); |
528 | // Add GlobalBaseReg in case there is no SelectionDAG passes afterwards |
529 | if (isGlobalISelAbortEnabled()) |
530 | addPass(P: createX86GlobalBaseRegPass()); |
531 | return false; |
532 | } |
533 | |
534 | bool X86PassConfig::addILPOpts() { |
535 | addPass(PassID: &EarlyIfConverterID); |
536 | if (EnableMachineCombinerPass) |
537 | addPass(PassID: &MachineCombinerID); |
538 | addPass(P: createX86CmovConverterPass()); |
539 | return true; |
540 | } |
541 | |
542 | bool X86PassConfig::addPreISel() { |
543 | // Only add this pass for 32-bit x86 Windows. |
544 | const Triple &TT = TM->getTargetTriple(); |
545 | if (TT.isOSWindows() && TT.getArch() == Triple::x86) |
546 | addPass(P: createX86WinEHStatePass()); |
547 | return true; |
548 | } |
549 | |
550 | void X86PassConfig::addPreRegAlloc() { |
551 | if (getOptLevel() != CodeGenOptLevel::None) { |
552 | addPass(PassID: &LiveRangeShrinkID); |
553 | addPass(P: createX86WinFixupBufferSecurityCheckPass()); |
554 | addPass(P: createX86FixupSetCC()); |
555 | addPass(P: createX86OptimizeLEAs()); |
556 | addPass(P: createX86CallFrameOptimization()); |
557 | addPass(P: createX86AvoidStoreForwardingBlocks()); |
558 | } |
559 | |
560 | addPass(P: createX86SpeculativeLoadHardeningPass()); |
561 | addPass(P: createX86FlagsCopyLoweringPass()); |
562 | addPass(P: createX86DynAllocaExpander()); |
563 | |
564 | if (getOptLevel() != CodeGenOptLevel::None) |
565 | addPass(P: createX86PreTileConfigPass()); |
566 | else |
567 | addPass(P: createX86FastPreTileConfigPass()); |
568 | } |
569 | |
570 | void X86PassConfig::addMachineSSAOptimization() { |
571 | addPass(P: createX86DomainReassignmentPass()); |
572 | TargetPassConfig::addMachineSSAOptimization(); |
573 | } |
574 | |
575 | void X86PassConfig::addPostRegAlloc() { |
576 | addPass(P: createX86LowerTileCopyPass()); |
577 | addPass(P: createX86FloatingPointStackifierPass()); |
578 | // When -O0 is enabled, the Load Value Injection Hardening pass will fall back |
579 | // to using the Speculative Execution Side Effect Suppression pass for |
580 | // mitigation. This is to prevent slow downs due to |
581 | // analyses needed by the LVIHardening pass when compiling at -O0. |
582 | if (getOptLevel() != CodeGenOptLevel::None) |
583 | addPass(P: createX86LoadValueInjectionLoadHardeningPass()); |
584 | } |
585 | |
586 | void X86PassConfig::addPreSched2() { |
587 | addPass(P: createX86ExpandPseudoPass()); |
588 | addPass(P: createKCFIPass()); |
589 | } |
590 | |
591 | void X86PassConfig::addPreEmitPass() { |
592 | if (getOptLevel() != CodeGenOptLevel::None) { |
593 | addPass(P: new X86ExecutionDomainFix()); |
594 | addPass(P: createBreakFalseDeps()); |
595 | } |
596 | |
597 | addPass(P: createX86IndirectBranchTrackingPass()); |
598 | |
599 | addPass(P: createX86IssueVZeroUpperPass()); |
600 | |
601 | if (getOptLevel() != CodeGenOptLevel::None) { |
602 | addPass(P: createX86FixupBWInsts()); |
603 | addPass(P: createX86PadShortFunctions()); |
604 | addPass(P: createX86FixupLEAs()); |
605 | addPass(P: createX86FixupInstTuning()); |
606 | addPass(P: createX86FixupVectorConstants()); |
607 | } |
608 | addPass(P: createX86CompressEVEXPass()); |
609 | addPass(P: createX86DiscriminateMemOpsPass()); |
610 | addPass(P: createX86InsertPrefetchPass()); |
611 | addPass(P: createX86InsertX87waitPass()); |
612 | } |
613 | |
614 | void X86PassConfig::addPreEmitPass2() { |
615 | const Triple &TT = TM->getTargetTriple(); |
616 | const MCAsmInfo *MAI = TM->getMCAsmInfo(); |
617 | |
618 | // The X86 Speculative Execution Pass must run after all control |
619 | // flow graph modifying passes. As a result it was listed to run right before |
620 | // the X86 Retpoline Thunks pass. The reason it must run after control flow |
621 | // graph modifications is that the model of LFENCE in LLVM has to be updated |
622 | // (FIXME: https://bugs.llvm.org/show_bug.cgi?id=45167). Currently the |
623 | // placement of this pass was hand checked to ensure that the subsequent |
624 | // passes don't move the code around the LFENCEs in a way that will hurt the |
625 | // correctness of this pass. This placement has been shown to work based on |
626 | // hand inspection of the codegen output. |
627 | addPass(P: createX86SpeculativeExecutionSideEffectSuppression()); |
628 | addPass(P: createX86IndirectThunksPass()); |
629 | addPass(P: createX86ReturnThunksPass()); |
630 | |
631 | // Insert extra int3 instructions after trailing call instructions to avoid |
632 | // issues in the unwinder. |
633 | if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) |
634 | addPass(P: createX86AvoidTrailingCallPass()); |
635 | |
636 | // Verify basic block incoming and outgoing cfa offset and register values and |
637 | // correct CFA calculation rule where needed by inserting appropriate CFI |
638 | // instructions. |
639 | if (!TT.isOSDarwin() && |
640 | (!TT.isOSWindows() || |
641 | MAI->getExceptionHandlingType() == ExceptionHandling::DwarfCFI)) |
642 | addPass(P: createCFIInstrInserter()); |
643 | |
644 | if (TT.isOSWindows()) { |
645 | // Identify valid longjmp targets for Windows Control Flow Guard. |
646 | addPass(P: createCFGuardLongjmpPass()); |
647 | // Identify valid eh continuation targets for Windows EHCont Guard. |
648 | addPass(P: createEHContGuardCatchretPass()); |
649 | } |
650 | addPass(P: createX86LoadValueInjectionRetHardeningPass()); |
651 | |
652 | // Insert pseudo probe annotation for callsite profiling |
653 | addPass(P: createPseudoProbeInserter()); |
654 | |
655 | // KCFI indirect call checks are lowered to a bundle, and on Darwin platforms, |
656 | // also CALL_RVMARKER. |
657 | addPass(P: createUnpackMachineBundles(Ftor: [&TT](const MachineFunction &MF) { |
658 | // Only run bundle expansion if the module uses kcfi, or there are relevant |
659 | // ObjC runtime functions present in the module. |
660 | const Function &F = MF.getFunction(); |
661 | const Module *M = F.getParent(); |
662 | return M->getModuleFlag(Key: "kcfi" ) || |
663 | (TT.isOSDarwin() && |
664 | (M->getFunction(Name: "objc_retainAutoreleasedReturnValue" ) || |
665 | M->getFunction(Name: "objc_unsafeClaimAutoreleasedReturnValue" ))); |
666 | })); |
667 | } |
668 | |
669 | bool X86PassConfig::addPostFastRegAllocRewrite() { |
670 | addPass(P: createX86FastTileConfigPass()); |
671 | return true; |
672 | } |
673 | |
674 | std::unique_ptr<CSEConfigBase> X86PassConfig::getCSEConfig() const { |
675 | return getStandardCSEConfigForOpt(Level: TM->getOptLevel()); |
676 | } |
677 | |
678 | static bool onlyAllocateTileRegisters(const TargetRegisterInfo &TRI, |
679 | const MachineRegisterInfo &MRI, |
680 | const Register Reg) { |
681 | const TargetRegisterClass *RC = MRI.getRegClass(Reg); |
682 | return static_cast<const X86RegisterInfo &>(TRI).isTileRegisterClass(RC); |
683 | } |
684 | |
685 | bool X86PassConfig::addRegAssignAndRewriteOptimized() { |
686 | // Don't support tile RA when RA is specified by command line "-regalloc". |
687 | if (!isCustomizedRegAlloc() && EnableTileRAPass) { |
688 | // Allocate tile register first. |
689 | addPass(P: createGreedyRegisterAllocator(F: onlyAllocateTileRegisters)); |
690 | addPass(P: createX86TileConfigPass()); |
691 | } |
692 | return TargetPassConfig::addRegAssignAndRewriteOptimized(); |
693 | } |
694 | |