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/SmallString.h" |
23 | #include "llvm/ADT/StringRef.h" |
24 | #include "llvm/Analysis/TargetTransformInfo.h" |
25 | #include "llvm/CodeGen/ExecutionDomainFix.h" |
26 | #include "llvm/CodeGen/GlobalISel/CSEInfo.h" |
27 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
28 | #include "llvm/CodeGen/GlobalISel/IRTranslator.h" |
29 | #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" |
30 | #include "llvm/CodeGen/GlobalISel/Legalizer.h" |
31 | #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" |
32 | #include "llvm/CodeGen/MIRParser/MIParser.h" |
33 | #include "llvm/CodeGen/MIRYamlMapping.h" |
34 | #include "llvm/CodeGen/MachineScheduler.h" |
35 | #include "llvm/CodeGen/Passes.h" |
36 | #include "llvm/CodeGen/TargetPassConfig.h" |
37 | #include "llvm/IR/Attributes.h" |
38 | #include "llvm/IR/DataLayout.h" |
39 | #include "llvm/IR/Function.h" |
40 | #include "llvm/MC/MCAsmInfo.h" |
41 | #include "llvm/MC/TargetRegistry.h" |
42 | #include "llvm/Pass.h" |
43 | #include "llvm/Support/CodeGen.h" |
44 | #include "llvm/Support/CommandLine.h" |
45 | #include "llvm/Support/ErrorHandling.h" |
46 | #include "llvm/Target/TargetLoweringObjectFile.h" |
47 | #include "llvm/Target/TargetOptions.h" |
48 | #include "llvm/TargetParser/Triple.h" |
49 | #include "llvm/Transforms/CFGuard.h" |
50 | #include <memory> |
51 | #include <optional> |
52 | #include <string> |
53 | |
54 | using namespace llvm; |
55 | |
56 | static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner" , |
57 | cl::desc("Enable the machine combiner pass" ), |
58 | cl::init(Val: true), cl::Hidden); |
59 | |
60 | static cl::opt<bool> |
61 | EnableTileRAPass("x86-tile-ra" , |
62 | cl::desc("Enable the tile register allocation pass" ), |
63 | cl::init(Val: true), cl::Hidden); |
64 | |
65 | extern "C" LLVM_C_ABI void LLVMInitializeX86Target() { |
66 | // Register the target. |
67 | RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); |
68 | RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); |
69 | |
70 | PassRegistry &PR = *PassRegistry::getPassRegistry(); |
71 | initializeX86LowerAMXIntrinsicsLegacyPassPass(PR); |
72 | initializeX86LowerAMXTypeLegacyPassPass(PR); |
73 | initializeX86PreTileConfigPass(PR); |
74 | initializeGlobalISel(PR); |
75 | initializeWinEHStatePassPass(PR); |
76 | initializeFixupBWInstPassPass(PR); |
77 | initializeCompressEVEXPassPass(PR); |
78 | initializeFixupLEAPassPass(PR); |
79 | initializeFPSPass(PR); |
80 | initializeX86FixupSetCCPassPass(PR); |
81 | initializeX86CallFrameOptimizationPass(PR); |
82 | initializeX86CmovConverterPassPass(PR); |
83 | initializeX86TileConfigPass(PR); |
84 | initializeX86FastPreTileConfigPass(PR); |
85 | initializeX86FastTileConfigPass(PR); |
86 | initializeKCFIPass(PR); |
87 | initializeX86LowerTileCopyPass(PR); |
88 | initializeX86ExpandPseudoPass(PR); |
89 | initializeX86ExecutionDomainFixPass(PR); |
90 | initializeX86DomainReassignmentPass(PR); |
91 | initializeX86AvoidSFBPassPass(PR); |
92 | initializeX86AvoidTrailingCallPassPass(PR); |
93 | initializeX86SpeculativeLoadHardeningPassPass(PR); |
94 | initializeX86SpeculativeExecutionSideEffectSuppressionPass(PR); |
95 | initializeX86FlagsCopyLoweringPassPass(PR); |
96 | initializeX86LoadValueInjectionLoadHardeningPassPass(PR); |
97 | initializeX86LoadValueInjectionRetHardeningPassPass(PR); |
98 | initializeX86OptimizeLEAPassPass(PR); |
99 | initializeX86PartialReductionPass(PR); |
100 | initializePseudoProbeInserterPass(PR); |
101 | initializeX86ReturnThunksPass(PR); |
102 | initializeX86DAGToDAGISelLegacyPass(PR); |
103 | initializeX86ArgumentStackSlotPassPass(PR); |
104 | initializeX86AsmPrinterPass(PR); |
105 | initializeX86FixupInstTuningPassPass(PR); |
106 | initializeX86FixupVectorConstantsPassPass(PR); |
107 | initializeX86DynAllocaExpanderPass(PR); |
108 | initializeX86SuppressAPXForRelocationPassPass(PR); |
109 | initializeX86WinEHUnwindV2Pass(PR); |
110 | } |
111 | |
112 | static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { |
113 | if (TT.isOSBinFormatMachO()) { |
114 | if (TT.getArch() == Triple::x86_64) |
115 | return std::make_unique<X86_64MachoTargetObjectFile>(); |
116 | return std::make_unique<TargetLoweringObjectFileMachO>(); |
117 | } |
118 | |
119 | if (TT.isOSBinFormatCOFF()) |
120 | return std::make_unique<TargetLoweringObjectFileCOFF>(); |
121 | |
122 | if (TT.getArch() == Triple::x86_64) |
123 | return std::make_unique<X86_64ELFTargetObjectFile>(); |
124 | return std::make_unique<X86ELFTargetObjectFile>(); |
125 | } |
126 | |
127 | static std::string computeDataLayout(const Triple &TT) { |
128 | // X86 is little endian |
129 | std::string Ret = "e" ; |
130 | |
131 | Ret += DataLayout::getManglingComponent(T: TT); |
132 | // X86 and x32 have 32 bit pointers. |
133 | if (!TT.isArch64Bit() || TT.isX32() || TT.isOSNaCl()) |
134 | Ret += "-p:32:32" ; |
135 | |
136 | // Address spaces for 32 bit signed, 32 bit unsigned, and 64 bit pointers. |
137 | Ret += "-p270:32:32-p271:32:32-p272:64:64" ; |
138 | |
139 | // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. |
140 | // 128 bit integers are not specified in the 32-bit ABIs but are used |
141 | // internally for lowering f128, so we match the alignment to that. |
142 | if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) |
143 | Ret += "-i64:64-i128:128" ; |
144 | else if (TT.isOSIAMCU()) |
145 | Ret += "-i64:32-f64:32" ; |
146 | else |
147 | Ret += "-i128:128-f64:32:64" ; |
148 | |
149 | // Some ABIs align long double to 128 bits, others to 32. |
150 | if (TT.isOSNaCl() || TT.isOSIAMCU()) |
151 | ; // No f80 |
152 | else if (TT.isArch64Bit() || TT.isOSDarwin() || TT.isWindowsMSVCEnvironment()) |
153 | Ret += "-f80:128" ; |
154 | else |
155 | Ret += "-f80:32" ; |
156 | |
157 | if (TT.isOSIAMCU()) |
158 | Ret += "-f128:32" ; |
159 | |
160 | // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. |
161 | if (TT.isArch64Bit()) |
162 | Ret += "-n8:16:32:64" ; |
163 | else |
164 | Ret += "-n8:16:32" ; |
165 | |
166 | // The stack is aligned to 32 bits on some ABIs and 128 bits on others. |
167 | if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) |
168 | Ret += "-a:0:32-S32" ; |
169 | else |
170 | Ret += "-S128" ; |
171 | |
172 | return Ret; |
173 | } |
174 | |
175 | static Reloc::Model getEffectiveRelocModel(const Triple &TT, bool JIT, |
176 | std::optional<Reloc::Model> RM) { |
177 | bool is64Bit = TT.getArch() == Triple::x86_64; |
178 | if (!RM) { |
179 | // JIT codegen should use static relocations by default, since it's |
180 | // typically executed in process and not relocatable. |
181 | if (JIT) |
182 | return Reloc::Static; |
183 | |
184 | // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. |
185 | // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we |
186 | // use static relocation model by default. |
187 | if (TT.isOSDarwin()) { |
188 | if (is64Bit) |
189 | return Reloc::PIC_; |
190 | return Reloc::DynamicNoPIC; |
191 | } |
192 | if (TT.isOSWindows() && is64Bit) |
193 | return Reloc::PIC_; |
194 | return Reloc::Static; |
195 | } |
196 | |
197 | // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC |
198 | // is defined as a model for code which may be used in static or dynamic |
199 | // executables but not necessarily a shared library. On X86-32 we just |
200 | // compile in -static mode, in x86-64 we use PIC. |
201 | if (*RM == Reloc::DynamicNoPIC) { |
202 | if (is64Bit) |
203 | return Reloc::PIC_; |
204 | if (!TT.isOSDarwin()) |
205 | return Reloc::Static; |
206 | } |
207 | |
208 | // If we are on Darwin, disallow static relocation model in X86-64 mode, since |
209 | // the Mach-O file format doesn't support it. |
210 | if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) |
211 | return Reloc::PIC_; |
212 | |
213 | return *RM; |
214 | } |
215 | |
216 | static CodeModel::Model |
217 | getEffectiveX86CodeModel(const Triple &TT, std::optional<CodeModel::Model> CM, |
218 | bool JIT) { |
219 | bool Is64Bit = TT.getArch() == Triple::x86_64; |
220 | if (CM) { |
221 | if (*CM == CodeModel::Tiny) |
222 | reportFatalUsageError(reason: "target does not support the tiny CodeModel" ); |
223 | return *CM; |
224 | } |
225 | if (JIT) |
226 | return Is64Bit ? CodeModel::Large : CodeModel::Small; |
227 | return CodeModel::Small; |
228 | } |
229 | |
230 | /// Create an X86 target. |
231 | /// |
232 | X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, |
233 | StringRef CPU, StringRef FS, |
234 | const TargetOptions &Options, |
235 | std::optional<Reloc::Model> RM, |
236 | std::optional<CodeModel::Model> CM, |
237 | CodeGenOptLevel OL, bool JIT) |
238 | : CodeGenTargetMachineImpl(T, computeDataLayout(TT), TT, CPU, FS, Options, |
239 | getEffectiveRelocModel(TT, JIT, RM), |
240 | getEffectiveX86CodeModel(TT, CM, JIT), OL), |
241 | TLOF(createTLOF(TT: getTargetTriple())), IsJIT(JIT) { |
242 | // On PS4/PS5, the "return address" of a 'noreturn' call must still be within |
243 | // the calling function. Note that this also includes __stack_chk_fail, |
244 | // so there was some target-specific logic in the instruction selectors |
245 | // to handle that. That code has since been generalized, so the only thing |
246 | // needed is to set TrapUnreachable here. |
247 | if (TT.isPS() || TT.isOSBinFormatMachO()) { |
248 | this->Options.TrapUnreachable = true; |
249 | this->Options.NoTrapAfterNoreturn = TT.isOSBinFormatMachO(); |
250 | } |
251 | |
252 | setMachineOutliner(true); |
253 | |
254 | // x86 supports the debug entry values. |
255 | setSupportsDebugEntryValues(true); |
256 | |
257 | initAsmInfo(); |
258 | } |
259 | |
260 | X86TargetMachine::~X86TargetMachine() = default; |
261 | |
262 | const X86Subtarget * |
263 | X86TargetMachine::getSubtargetImpl(const Function &F) const { |
264 | Attribute CPUAttr = F.getFnAttribute(Kind: "target-cpu" ); |
265 | Attribute TuneAttr = F.getFnAttribute(Kind: "tune-cpu" ); |
266 | Attribute FSAttr = F.getFnAttribute(Kind: "target-features" ); |
267 | |
268 | StringRef CPU = |
269 | CPUAttr.isValid() ? CPUAttr.getValueAsString() : (StringRef)TargetCPU; |
270 | // "x86-64" is a default target setting for many front ends. In these cases, |
271 | // they actually request for "generic" tuning unless the "tune-cpu" was |
272 | // specified. |
273 | StringRef TuneCPU = TuneAttr.isValid() ? TuneAttr.getValueAsString() |
274 | : CPU == "x86-64" ? "generic" |
275 | : (StringRef)CPU; |
276 | StringRef FS = |
277 | FSAttr.isValid() ? FSAttr.getValueAsString() : (StringRef)TargetFS; |
278 | |
279 | SmallString<512> Key; |
280 | // The additions here are ordered so that the definitely short strings are |
281 | // added first so we won't exceed the small size. We append the |
282 | // much longer FS string at the end so that we only heap allocate at most |
283 | // one time. |
284 | |
285 | // Extract prefer-vector-width attribute. |
286 | unsigned PreferVectorWidthOverride = 0; |
287 | Attribute PreferVecWidthAttr = F.getFnAttribute(Kind: "prefer-vector-width" ); |
288 | if (PreferVecWidthAttr.isValid()) { |
289 | StringRef Val = PreferVecWidthAttr.getValueAsString(); |
290 | unsigned Width; |
291 | if (!Val.getAsInteger(Radix: 0, Result&: Width)) { |
292 | Key += 'p'; |
293 | Key += Val; |
294 | PreferVectorWidthOverride = Width; |
295 | } |
296 | } |
297 | |
298 | // Extract min-legal-vector-width attribute. |
299 | unsigned RequiredVectorWidth = UINT32_MAX; |
300 | Attribute MinLegalVecWidthAttr = F.getFnAttribute(Kind: "min-legal-vector-width" ); |
301 | if (MinLegalVecWidthAttr.isValid()) { |
302 | StringRef Val = MinLegalVecWidthAttr.getValueAsString(); |
303 | unsigned Width; |
304 | if (!Val.getAsInteger(Radix: 0, Result&: Width)) { |
305 | Key += 'm'; |
306 | Key += Val; |
307 | RequiredVectorWidth = Width; |
308 | } |
309 | } |
310 | |
311 | // Add CPU to the Key. |
312 | Key += CPU; |
313 | |
314 | // Add tune CPU to the Key. |
315 | Key += TuneCPU; |
316 | |
317 | // Keep track of the start of the feature portion of the string. |
318 | unsigned FSStart = Key.size(); |
319 | |
320 | // FIXME: This is related to the code below to reset the target options, |
321 | // we need to know whether or not the soft float flag is set on the |
322 | // function before we can generate a subtarget. We also need to use |
323 | // it as a key for the subtarget since that can be the only difference |
324 | // between two functions. |
325 | bool SoftFloat = F.getFnAttribute(Kind: "use-soft-float" ).getValueAsBool(); |
326 | // If the soft float attribute is set on the function turn on the soft float |
327 | // subtarget feature. |
328 | if (SoftFloat) |
329 | Key += FS.empty() ? "+soft-float" : "+soft-float," ; |
330 | |
331 | Key += FS; |
332 | |
333 | // We may have added +soft-float to the features so move the StringRef to |
334 | // point to the full string in the Key. |
335 | FS = Key.substr(Start: FSStart); |
336 | |
337 | auto &I = SubtargetMap[Key]; |
338 | if (!I) { |
339 | // This needs to be done before we create a new subtarget since any |
340 | // creation will depend on the TM and the code generation flags on the |
341 | // function that reside in TargetOptions. |
342 | resetTargetOptions(F); |
343 | I = std::make_unique<X86Subtarget>( |
344 | args: TargetTriple, args&: CPU, args&: TuneCPU, args&: FS, args: *this, |
345 | args: MaybeAlign(F.getParent()->getOverrideStackAlignment()), |
346 | args&: PreferVectorWidthOverride, args&: RequiredVectorWidth); |
347 | } |
348 | return I.get(); |
349 | } |
350 | |
351 | yaml::MachineFunctionInfo *X86TargetMachine::createDefaultFuncInfoYAML() const { |
352 | return new yaml::X86MachineFunctionInfo(); |
353 | } |
354 | |
355 | yaml::MachineFunctionInfo * |
356 | X86TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { |
357 | const auto *MFI = MF.getInfo<X86MachineFunctionInfo>(); |
358 | return new yaml::X86MachineFunctionInfo(*MFI); |
359 | } |
360 | |
361 | bool X86TargetMachine::parseMachineFunctionInfo( |
362 | const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS, |
363 | SMDiagnostic &Error, SMRange &SourceRange) const { |
364 | const auto &YamlMFI = static_cast<const yaml::X86MachineFunctionInfo &>(MFI); |
365 | PFS.MF.getInfo<X86MachineFunctionInfo>()->initializeBaseYamlFields(YamlMFI); |
366 | return false; |
367 | } |
368 | |
369 | bool X86TargetMachine::isNoopAddrSpaceCast(unsigned SrcAS, |
370 | unsigned DestAS) const { |
371 | assert(SrcAS != DestAS && "Expected different address spaces!" ); |
372 | if (getPointerSize(AS: SrcAS) != getPointerSize(AS: DestAS)) |
373 | return false; |
374 | return SrcAS < 256 && DestAS < 256; |
375 | } |
376 | |
377 | void X86TargetMachine::reset() { SubtargetMap.clear(); } |
378 | |
379 | ScheduleDAGInstrs * |
380 | X86TargetMachine::createMachineScheduler(MachineSchedContext *C) const { |
381 | ScheduleDAGMILive *DAG = createSchedLive(C); |
382 | DAG->addMutation(Mutation: createX86MacroFusionDAGMutation()); |
383 | return DAG; |
384 | } |
385 | |
386 | ScheduleDAGInstrs * |
387 | X86TargetMachine::createPostMachineScheduler(MachineSchedContext *C) const { |
388 | ScheduleDAGMI *DAG = createSchedPostRA(C); |
389 | DAG->addMutation(Mutation: createX86MacroFusionDAGMutation()); |
390 | return DAG; |
391 | } |
392 | |
393 | //===----------------------------------------------------------------------===// |
394 | // X86 TTI query. |
395 | //===----------------------------------------------------------------------===// |
396 | |
397 | TargetTransformInfo |
398 | X86TargetMachine::getTargetTransformInfo(const Function &F) const { |
399 | return TargetTransformInfo(std::make_unique<X86TTIImpl>(args: this, args: F)); |
400 | } |
401 | |
402 | //===----------------------------------------------------------------------===// |
403 | // Pass Pipeline Configuration |
404 | //===----------------------------------------------------------------------===// |
405 | |
406 | namespace { |
407 | |
408 | /// X86 Code Generator Pass Configuration Options. |
409 | class X86PassConfig : public TargetPassConfig { |
410 | public: |
411 | X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM) |
412 | : TargetPassConfig(TM, PM) {} |
413 | |
414 | X86TargetMachine &getX86TargetMachine() const { |
415 | return getTM<X86TargetMachine>(); |
416 | } |
417 | |
418 | void addIRPasses() override; |
419 | bool addInstSelector() override; |
420 | bool addIRTranslator() override; |
421 | bool addLegalizeMachineIR() override; |
422 | bool addRegBankSelect() override; |
423 | bool addGlobalInstructionSelect() override; |
424 | bool addILPOpts() override; |
425 | bool addPreISel() override; |
426 | void addMachineSSAOptimization() override; |
427 | void addPreRegAlloc() override; |
428 | bool addPostFastRegAllocRewrite() override; |
429 | void addPostRegAlloc() override; |
430 | void addPreEmitPass() override; |
431 | void addPreEmitPass2() override; |
432 | void addPreSched2() override; |
433 | bool addRegAssignAndRewriteOptimized() override; |
434 | |
435 | std::unique_ptr<CSEConfigBase> getCSEConfig() const override; |
436 | }; |
437 | |
438 | class X86ExecutionDomainFix : public ExecutionDomainFix { |
439 | public: |
440 | static char ID; |
441 | X86ExecutionDomainFix() : ExecutionDomainFix(ID, X86::VR128XRegClass) {} |
442 | StringRef getPassName() const override { |
443 | return "X86 Execution Dependency Fix" ; |
444 | } |
445 | }; |
446 | char X86ExecutionDomainFix::ID; |
447 | |
448 | } // end anonymous namespace |
449 | |
450 | INITIALIZE_PASS_BEGIN(X86ExecutionDomainFix, "x86-execution-domain-fix" , |
451 | "X86 Execution Domain Fix" , false, false) |
452 | INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis) |
453 | INITIALIZE_PASS_END(X86ExecutionDomainFix, "x86-execution-domain-fix" , |
454 | "X86 Execution Domain Fix" , false, false) |
455 | |
456 | TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { |
457 | return new X86PassConfig(*this, PM); |
458 | } |
459 | |
460 | MachineFunctionInfo *X86TargetMachine::createMachineFunctionInfo( |
461 | BumpPtrAllocator &Allocator, const Function &F, |
462 | const TargetSubtargetInfo *STI) const { |
463 | return X86MachineFunctionInfo::create<X86MachineFunctionInfo>(Allocator, F, |
464 | STI); |
465 | } |
466 | |
467 | void X86PassConfig::addIRPasses() { |
468 | addPass(P: createAtomicExpandLegacyPass()); |
469 | |
470 | // We add both pass anyway and when these two passes run, we skip the pass |
471 | // based on the option level and option attribute. |
472 | addPass(P: createX86LowerAMXIntrinsicsPass()); |
473 | addPass(P: createX86LowerAMXTypePass()); |
474 | |
475 | TargetPassConfig::addIRPasses(); |
476 | |
477 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
478 | addPass(P: createInterleavedAccessPass()); |
479 | addPass(P: createX86PartialReductionPass()); |
480 | } |
481 | |
482 | // Add passes that handle indirect branch removal and insertion of a retpoline |
483 | // thunk. These will be a no-op unless a function subtarget has the retpoline |
484 | // feature enabled. |
485 | addPass(P: createIndirectBrExpandPass()); |
486 | |
487 | // Add Control Flow Guard checks. |
488 | const Triple &TT = TM->getTargetTriple(); |
489 | if (TT.isOSWindows()) { |
490 | if (TT.getArch() == Triple::x86_64) { |
491 | addPass(P: createCFGuardDispatchPass()); |
492 | } else { |
493 | addPass(P: createCFGuardCheckPass()); |
494 | } |
495 | } |
496 | |
497 | if (TM->Options.JMCInstrument) |
498 | addPass(P: createJMCInstrumenterPass()); |
499 | } |
500 | |
501 | bool X86PassConfig::addInstSelector() { |
502 | // Install an instruction selector. |
503 | addPass(P: createX86ISelDag(TM&: getX86TargetMachine(), OptLevel: getOptLevel())); |
504 | |
505 | // For ELF, cleanup any local-dynamic TLS accesses. |
506 | if (TM->getTargetTriple().isOSBinFormatELF() && |
507 | getOptLevel() != CodeGenOptLevel::None) |
508 | addPass(P: createCleanupLocalDynamicTLSPass()); |
509 | |
510 | addPass(P: createX86GlobalBaseRegPass()); |
511 | addPass(P: createX86ArgumentStackSlotPass()); |
512 | return false; |
513 | } |
514 | |
515 | bool X86PassConfig::addIRTranslator() { |
516 | addPass(P: new IRTranslator(getOptLevel())); |
517 | return false; |
518 | } |
519 | |
520 | bool X86PassConfig::addLegalizeMachineIR() { |
521 | addPass(P: new Legalizer()); |
522 | return false; |
523 | } |
524 | |
525 | bool X86PassConfig::addRegBankSelect() { |
526 | addPass(P: new RegBankSelect()); |
527 | return false; |
528 | } |
529 | |
530 | bool X86PassConfig::addGlobalInstructionSelect() { |
531 | addPass(P: new InstructionSelect(getOptLevel())); |
532 | // Add GlobalBaseReg in case there is no SelectionDAG passes afterwards |
533 | if (isGlobalISelAbortEnabled()) |
534 | addPass(P: createX86GlobalBaseRegPass()); |
535 | return false; |
536 | } |
537 | |
538 | bool X86PassConfig::addILPOpts() { |
539 | addPass(PassID: &EarlyIfConverterLegacyID); |
540 | if (EnableMachineCombinerPass) |
541 | addPass(PassID: &MachineCombinerID); |
542 | addPass(P: createX86CmovConverterPass()); |
543 | return true; |
544 | } |
545 | |
546 | bool X86PassConfig::addPreISel() { |
547 | // Only add this pass for 32-bit x86 Windows. |
548 | const Triple &TT = TM->getTargetTriple(); |
549 | if (TT.isOSWindows() && TT.getArch() == Triple::x86) |
550 | addPass(P: createX86WinEHStatePass()); |
551 | return true; |
552 | } |
553 | |
554 | void X86PassConfig::addPreRegAlloc() { |
555 | if (getOptLevel() != CodeGenOptLevel::None) { |
556 | addPass(PassID: &LiveRangeShrinkID); |
557 | addPass(P: createX86FixupSetCC()); |
558 | addPass(P: createX86OptimizeLEAs()); |
559 | addPass(P: createX86CallFrameOptimization()); |
560 | addPass(P: createX86AvoidStoreForwardingBlocks()); |
561 | } |
562 | |
563 | addPass(P: createX86SuppressAPXForRelocationPass()); |
564 | |
565 | addPass(P: createX86SpeculativeLoadHardeningPass()); |
566 | addPass(P: createX86FlagsCopyLoweringPass()); |
567 | addPass(P: createX86DynAllocaExpander()); |
568 | |
569 | if (getOptLevel() != CodeGenOptLevel::None) |
570 | addPass(P: createX86PreTileConfigPass()); |
571 | else |
572 | addPass(P: createX86FastPreTileConfigPass()); |
573 | } |
574 | |
575 | void X86PassConfig::addMachineSSAOptimization() { |
576 | addPass(P: createX86DomainReassignmentPass()); |
577 | TargetPassConfig::addMachineSSAOptimization(); |
578 | } |
579 | |
580 | void X86PassConfig::addPostRegAlloc() { |
581 | addPass(P: createX86LowerTileCopyPass()); |
582 | addPass(P: createX86FloatingPointStackifierPass()); |
583 | // When -O0 is enabled, the Load Value Injection Hardening pass will fall back |
584 | // to using the Speculative Execution Side Effect Suppression pass for |
585 | // mitigation. This is to prevent slow downs due to |
586 | // analyses needed by the LVIHardening pass when compiling at -O0. |
587 | if (getOptLevel() != CodeGenOptLevel::None) |
588 | addPass(P: createX86LoadValueInjectionLoadHardeningPass()); |
589 | } |
590 | |
591 | void X86PassConfig::addPreSched2() { |
592 | addPass(P: createX86ExpandPseudoPass()); |
593 | addPass(P: createKCFIPass()); |
594 | } |
595 | |
596 | void X86PassConfig::addPreEmitPass() { |
597 | if (getOptLevel() != CodeGenOptLevel::None) { |
598 | addPass(P: new X86ExecutionDomainFix()); |
599 | addPass(P: createBreakFalseDeps()); |
600 | } |
601 | |
602 | addPass(P: createX86IndirectBranchTrackingPass()); |
603 | |
604 | addPass(P: createX86IssueVZeroUpperPass()); |
605 | |
606 | if (getOptLevel() != CodeGenOptLevel::None) { |
607 | addPass(P: createX86FixupBWInsts()); |
608 | addPass(P: createX86PadShortFunctions()); |
609 | addPass(P: createX86FixupLEAs()); |
610 | addPass(P: createX86FixupInstTuning()); |
611 | addPass(P: createX86FixupVectorConstants()); |
612 | } |
613 | addPass(P: createX86CompressEVEXPass()); |
614 | addPass(P: createX86DiscriminateMemOpsPass()); |
615 | addPass(P: createX86InsertPrefetchPass()); |
616 | addPass(P: createX86InsertX87waitPass()); |
617 | } |
618 | |
619 | void X86PassConfig::addPreEmitPass2() { |
620 | const Triple &TT = TM->getTargetTriple(); |
621 | const MCAsmInfo *MAI = TM->getMCAsmInfo(); |
622 | |
623 | // The X86 Speculative Execution Pass must run after all control |
624 | // flow graph modifying passes. As a result it was listed to run right before |
625 | // the X86 Retpoline Thunks pass. The reason it must run after control flow |
626 | // graph modifications is that the model of LFENCE in LLVM has to be updated |
627 | // (FIXME: https://bugs.llvm.org/show_bug.cgi?id=45167). Currently the |
628 | // placement of this pass was hand checked to ensure that the subsequent |
629 | // passes don't move the code around the LFENCEs in a way that will hurt the |
630 | // correctness of this pass. This placement has been shown to work based on |
631 | // hand inspection of the codegen output. |
632 | addPass(P: createX86SpeculativeExecutionSideEffectSuppression()); |
633 | addPass(P: createX86IndirectThunksPass()); |
634 | addPass(P: createX86ReturnThunksPass()); |
635 | |
636 | // Insert extra int3 instructions after trailing call instructions to avoid |
637 | // issues in the unwinder. |
638 | if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) |
639 | addPass(P: createX86AvoidTrailingCallPass()); |
640 | |
641 | // Verify basic block incoming and outgoing cfa offset and register values and |
642 | // correct CFA calculation rule where needed by inserting appropriate CFI |
643 | // instructions. |
644 | if (!TT.isOSDarwin() && |
645 | (!TT.isOSWindows() || |
646 | MAI->getExceptionHandlingType() == ExceptionHandling::DwarfCFI)) |
647 | addPass(P: createCFIInstrInserter()); |
648 | |
649 | if (TT.isOSWindows()) { |
650 | // Identify valid longjmp targets for Windows Control Flow Guard. |
651 | addPass(P: createCFGuardLongjmpPass()); |
652 | // Identify valid eh continuation targets for Windows EHCont Guard. |
653 | addPass(P: createEHContGuardTargetsPass()); |
654 | } |
655 | addPass(P: createX86LoadValueInjectionRetHardeningPass()); |
656 | |
657 | // Insert pseudo probe annotation for callsite profiling |
658 | addPass(P: createPseudoProbeInserter()); |
659 | |
660 | // KCFI indirect call checks are lowered to a bundle, and on Darwin platforms, |
661 | // also CALL_RVMARKER. |
662 | addPass(P: createUnpackMachineBundles(Ftor: [&TT](const MachineFunction &MF) { |
663 | // Only run bundle expansion if the module uses kcfi, or there are relevant |
664 | // ObjC runtime functions present in the module. |
665 | const Function &F = MF.getFunction(); |
666 | const Module *M = F.getParent(); |
667 | return M->getModuleFlag(Key: "kcfi" ) || |
668 | (TT.isOSDarwin() && |
669 | (M->getFunction(Name: "objc_retainAutoreleasedReturnValue" ) || |
670 | M->getFunction(Name: "objc_unsafeClaimAutoreleasedReturnValue" ))); |
671 | })); |
672 | |
673 | // Analyzes and emits pseudos to support Win x64 Unwind V2. This pass must run |
674 | // after all real instructions have been added to the epilog. |
675 | if (TT.isOSWindows() && (TT.getArch() == Triple::x86_64)) |
676 | addPass(P: createX86WinEHUnwindV2Pass()); |
677 | } |
678 | |
679 | bool X86PassConfig::addPostFastRegAllocRewrite() { |
680 | addPass(P: createX86FastTileConfigPass()); |
681 | return true; |
682 | } |
683 | |
684 | std::unique_ptr<CSEConfigBase> X86PassConfig::getCSEConfig() const { |
685 | return getStandardCSEConfigForOpt(Level: TM->getOptLevel()); |
686 | } |
687 | |
688 | static bool onlyAllocateTileRegisters(const TargetRegisterInfo &TRI, |
689 | const MachineRegisterInfo &MRI, |
690 | const Register Reg) { |
691 | const TargetRegisterClass *RC = MRI.getRegClass(Reg); |
692 | return static_cast<const X86RegisterInfo &>(TRI).isTileRegisterClass(RC); |
693 | } |
694 | |
695 | bool X86PassConfig::addRegAssignAndRewriteOptimized() { |
696 | // Don't support tile RA when RA is specified by command line "-regalloc". |
697 | if (!isCustomizedRegAlloc() && EnableTileRAPass) { |
698 | // Allocate tile register first. |
699 | addPass(P: createGreedyRegisterAllocator(F: onlyAllocateTileRegisters)); |
700 | addPass(P: createX86TileConfigPass()); |
701 | } |
702 | return TargetPassConfig::addRegAssignAndRewriteOptimized(); |
703 | } |
704 | |