xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86TargetMachine.cpp (revision 700637cbb5e582861067a11aaca4d053546871d2)
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-c/Visibility.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/Legalizer.h"
32 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
33 #include "llvm/CodeGen/MIRParser/MIParser.h"
34 #include "llvm/CodeGen/MIRYamlMapping.h"
35 #include "llvm/CodeGen/MachineScheduler.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/CodeGen/TargetPassConfig.h"
38 #include "llvm/IR/Attributes.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/MC/MCAsmInfo.h"
42 #include "llvm/MC/TargetRegistry.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/CodeGen.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Target/TargetLoweringObjectFile.h"
48 #include "llvm/Target/TargetOptions.h"
49 #include "llvm/TargetParser/Triple.h"
50 #include "llvm/Transforms/CFGuard.h"
51 #include <memory>
52 #include <optional>
53 #include <string>
54 
55 using namespace llvm;
56 
57 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner",
58                                cl::desc("Enable the machine combiner pass"),
59                                cl::init(true), cl::Hidden);
60 
61 static cl::opt<bool>
62     EnableTileRAPass("x86-tile-ra",
63                      cl::desc("Enable the tile register allocation pass"),
64                      cl::init(true), cl::Hidden);
65 
LLVMInitializeX86Target()66 extern "C" LLVM_C_ABI void LLVMInitializeX86Target() {
67   // Register the target.
68   RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target());
69   RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target());
70 
71   PassRegistry &PR = *PassRegistry::getPassRegistry();
72   initializeX86LowerAMXIntrinsicsLegacyPassPass(PR);
73   initializeX86LowerAMXTypeLegacyPassPass(PR);
74   initializeX86PreTileConfigPass(PR);
75   initializeGlobalISel(PR);
76   initializeWinEHStatePassPass(PR);
77   initializeFixupBWInstPassPass(PR);
78   initializeCompressEVEXPassPass(PR);
79   initializeFixupLEAPassPass(PR);
80   initializeFPSPass(PR);
81   initializeX86FixupSetCCPassPass(PR);
82   initializeX86CallFrameOptimizationPass(PR);
83   initializeX86CmovConverterPassPass(PR);
84   initializeX86TileConfigPass(PR);
85   initializeX86FastPreTileConfigPass(PR);
86   initializeX86FastTileConfigPass(PR);
87   initializeKCFIPass(PR);
88   initializeX86LowerTileCopyPass(PR);
89   initializeX86ExpandPseudoPass(PR);
90   initializeX86ExecutionDomainFixPass(PR);
91   initializeX86DomainReassignmentPass(PR);
92   initializeX86AvoidSFBPassPass(PR);
93   initializeX86AvoidTrailingCallPassPass(PR);
94   initializeX86SpeculativeLoadHardeningPassPass(PR);
95   initializeX86SpeculativeExecutionSideEffectSuppressionPass(PR);
96   initializeX86FlagsCopyLoweringPassPass(PR);
97   initializeX86LoadValueInjectionLoadHardeningPassPass(PR);
98   initializeX86LoadValueInjectionRetHardeningPassPass(PR);
99   initializeX86OptimizeLEAPassPass(PR);
100   initializeX86PartialReductionPass(PR);
101   initializePseudoProbeInserterPass(PR);
102   initializeX86ReturnThunksPass(PR);
103   initializeX86DAGToDAGISelLegacyPass(PR);
104   initializeX86ArgumentStackSlotPassPass(PR);
105   initializeX86AsmPrinterPass(PR);
106   initializeX86FixupInstTuningPassPass(PR);
107   initializeX86FixupVectorConstantsPassPass(PR);
108   initializeX86DynAllocaExpanderPass(PR);
109   initializeX86SuppressAPXForRelocationPassPass(PR);
110   initializeX86WinEHUnwindV2Pass(PR);
111 }
112 
createTLOF(const Triple & TT)113 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
114   if (TT.isOSBinFormatMachO()) {
115     if (TT.getArch() == Triple::x86_64)
116       return std::make_unique<X86_64MachoTargetObjectFile>();
117     return std::make_unique<TargetLoweringObjectFileMachO>();
118   }
119 
120   if (TT.isOSBinFormatCOFF())
121     return std::make_unique<TargetLoweringObjectFileCOFF>();
122 
123   if (TT.getArch() == Triple::x86_64)
124     return std::make_unique<X86_64ELFTargetObjectFile>();
125   return std::make_unique<X86ELFTargetObjectFile>();
126 }
127 
computeDataLayout(const Triple & TT)128 static std::string computeDataLayout(const Triple &TT) {
129   // X86 is little endian
130   std::string Ret = "e";
131 
132   Ret += DataLayout::getManglingComponent(TT);
133   // X86 and x32 have 32 bit pointers.
134   if (!TT.isArch64Bit() || TT.isX32() || TT.isOSNaCl())
135     Ret += "-p:32:32";
136 
137   // Address spaces for 32 bit signed, 32 bit unsigned, and 64 bit pointers.
138   Ret += "-p270:32:32-p271:32:32-p272:64:64";
139 
140   // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32.
141   // 128 bit integers are not specified in the 32-bit ABIs but are used
142   // internally for lowering f128, so we match the alignment to that.
143   if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl())
144     Ret += "-i64:64-i128:128";
145   else if (TT.isOSIAMCU())
146     Ret += "-i64:32-f64:32";
147   else
148     Ret += "-i128:128-f64:32:64";
149 
150   // Some ABIs align long double to 128 bits, others to 32.
151   if (TT.isOSNaCl() || TT.isOSIAMCU())
152     ; // No f80
153   else if (TT.isArch64Bit() || TT.isOSDarwin() || TT.isWindowsMSVCEnvironment())
154     Ret += "-f80:128";
155   else
156     Ret += "-f80:32";
157 
158   if (TT.isOSIAMCU())
159     Ret += "-f128:32";
160 
161   // The registers can hold 8, 16, 32 or, in x86-64, 64 bits.
162   if (TT.isArch64Bit())
163     Ret += "-n8:16:32:64";
164   else
165     Ret += "-n8:16:32";
166 
167   // The stack is aligned to 32 bits on some ABIs and 128 bits on others.
168   if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU())
169     Ret += "-a:0:32-S32";
170   else
171     Ret += "-S128";
172 
173   return Ret;
174 }
175 
getEffectiveRelocModel(const Triple & TT,bool JIT,std::optional<Reloc::Model> RM)176 static Reloc::Model getEffectiveRelocModel(const Triple &TT, bool JIT,
177                                            std::optional<Reloc::Model> RM) {
178   bool is64Bit = TT.getArch() == Triple::x86_64;
179   if (!RM) {
180     // JIT codegen should use static relocations by default, since it's
181     // typically executed in process and not relocatable.
182     if (JIT)
183       return Reloc::Static;
184 
185     // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode.
186     // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we
187     // use static relocation model by default.
188     if (TT.isOSDarwin()) {
189       if (is64Bit)
190         return Reloc::PIC_;
191       return Reloc::DynamicNoPIC;
192     }
193     if (TT.isOSWindows() && is64Bit)
194       return Reloc::PIC_;
195     return Reloc::Static;
196   }
197 
198   // ELF and X86-64 don't have a distinct DynamicNoPIC model.  DynamicNoPIC
199   // is defined as a model for code which may be used in static or dynamic
200   // executables but not necessarily a shared library. On X86-32 we just
201   // compile in -static mode, in x86-64 we use PIC.
202   if (*RM == Reloc::DynamicNoPIC) {
203     if (is64Bit)
204       return Reloc::PIC_;
205     if (!TT.isOSDarwin())
206       return Reloc::Static;
207   }
208 
209   // If we are on Darwin, disallow static relocation model in X86-64 mode, since
210   // the Mach-O file format doesn't support it.
211   if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit)
212     return Reloc::PIC_;
213 
214   return *RM;
215 }
216 
217 static CodeModel::Model
getEffectiveX86CodeModel(const Triple & TT,std::optional<CodeModel::Model> CM,bool JIT)218 getEffectiveX86CodeModel(const Triple &TT, std::optional<CodeModel::Model> CM,
219                          bool JIT) {
220   bool Is64Bit = TT.getArch() == Triple::x86_64;
221   if (CM) {
222     if (*CM == CodeModel::Tiny)
223       reportFatalUsageError("target does not support the tiny CodeModel");
224     return *CM;
225   }
226   if (JIT)
227     return Is64Bit ? CodeModel::Large : CodeModel::Small;
228   return CodeModel::Small;
229 }
230 
231 /// Create an X86 target.
232 ///
X86TargetMachine(const Target & T,const Triple & TT,StringRef CPU,StringRef FS,const TargetOptions & Options,std::optional<Reloc::Model> RM,std::optional<CodeModel::Model> CM,CodeGenOptLevel OL,bool JIT)233 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT,
234                                    StringRef CPU, StringRef FS,
235                                    const TargetOptions &Options,
236                                    std::optional<Reloc::Model> RM,
237                                    std::optional<CodeModel::Model> CM,
238                                    CodeGenOptLevel OL, bool JIT)
239     : CodeGenTargetMachineImpl(T, computeDataLayout(TT), TT, CPU, FS, Options,
240                                getEffectiveRelocModel(TT, JIT, RM),
241                                getEffectiveX86CodeModel(TT, CM, JIT), OL),
242       TLOF(createTLOF(getTargetTriple())), IsJIT(JIT) {
243   // On PS4/PS5, the "return address" of a 'noreturn' call must still be within
244   // the calling function. Note that this also includes __stack_chk_fail,
245   // so there was some target-specific logic in the instruction selectors
246   // to handle that. That code has since been generalized, so the only thing
247   // needed is to set TrapUnreachable here.
248   if (TT.isPS() || TT.isOSBinFormatMachO()) {
249     this->Options.TrapUnreachable = true;
250     this->Options.NoTrapAfterNoreturn = TT.isOSBinFormatMachO();
251   }
252 
253   setMachineOutliner(true);
254 
255   // x86 supports the debug entry values.
256   setSupportsDebugEntryValues(true);
257 
258   initAsmInfo();
259 }
260 
261 X86TargetMachine::~X86TargetMachine() = default;
262 
263 const X86Subtarget *
getSubtargetImpl(const Function & F) const264 X86TargetMachine::getSubtargetImpl(const Function &F) const {
265   Attribute CPUAttr = F.getFnAttribute("target-cpu");
266   Attribute TuneAttr = F.getFnAttribute("tune-cpu");
267   Attribute FSAttr = F.getFnAttribute("target-features");
268 
269   StringRef CPU =
270       CPUAttr.isValid() ? CPUAttr.getValueAsString() : (StringRef)TargetCPU;
271   // "x86-64" is a default target setting for many front ends. In these cases,
272   // they actually request for "generic" tuning unless the "tune-cpu" was
273   // specified.
274   StringRef TuneCPU = TuneAttr.isValid() ? TuneAttr.getValueAsString()
275                       : CPU == "x86-64"  ? "generic"
276                                          : (StringRef)CPU;
277   StringRef FS =
278       FSAttr.isValid() ? FSAttr.getValueAsString() : (StringRef)TargetFS;
279 
280   SmallString<512> Key;
281   // The additions here are ordered so that the definitely short strings are
282   // added first so we won't exceed the small size. We append the
283   // much longer FS string at the end so that we only heap allocate at most
284   // one time.
285 
286   // Extract prefer-vector-width attribute.
287   unsigned PreferVectorWidthOverride = 0;
288   Attribute PreferVecWidthAttr = F.getFnAttribute("prefer-vector-width");
289   if (PreferVecWidthAttr.isValid()) {
290     StringRef Val = PreferVecWidthAttr.getValueAsString();
291     unsigned Width;
292     if (!Val.getAsInteger(0, Width)) {
293       Key += 'p';
294       Key += Val;
295       PreferVectorWidthOverride = Width;
296     }
297   }
298 
299   // Extract min-legal-vector-width attribute.
300   unsigned RequiredVectorWidth = UINT32_MAX;
301   Attribute MinLegalVecWidthAttr = F.getFnAttribute("min-legal-vector-width");
302   if (MinLegalVecWidthAttr.isValid()) {
303     StringRef Val = MinLegalVecWidthAttr.getValueAsString();
304     unsigned Width;
305     if (!Val.getAsInteger(0, Width)) {
306       Key += 'm';
307       Key += Val;
308       RequiredVectorWidth = Width;
309     }
310   }
311 
312   // Add CPU to the Key.
313   Key += CPU;
314 
315   // Add tune CPU to the Key.
316   Key += TuneCPU;
317 
318   // Keep track of the start of the feature portion of the string.
319   unsigned FSStart = Key.size();
320 
321   // FIXME: This is related to the code below to reset the target options,
322   // we need to know whether or not the soft float flag is set on the
323   // function before we can generate a subtarget. We also need to use
324   // it as a key for the subtarget since that can be the only difference
325   // between two functions.
326   bool SoftFloat = F.getFnAttribute("use-soft-float").getValueAsBool();
327   // If the soft float attribute is set on the function turn on the soft float
328   // subtarget feature.
329   if (SoftFloat)
330     Key += FS.empty() ? "+soft-float" : "+soft-float,";
331 
332   Key += FS;
333 
334   // We may have added +soft-float to the features so move the StringRef to
335   // point to the full string in the Key.
336   FS = Key.substr(FSStart);
337 
338   auto &I = SubtargetMap[Key];
339   if (!I) {
340     // This needs to be done before we create a new subtarget since any
341     // creation will depend on the TM and the code generation flags on the
342     // function that reside in TargetOptions.
343     resetTargetOptions(F);
344     I = std::make_unique<X86Subtarget>(
345         TargetTriple, CPU, TuneCPU, FS, *this,
346         MaybeAlign(F.getParent()->getOverrideStackAlignment()),
347         PreferVectorWidthOverride, RequiredVectorWidth);
348   }
349   return I.get();
350 }
351 
createDefaultFuncInfoYAML() const352 yaml::MachineFunctionInfo *X86TargetMachine::createDefaultFuncInfoYAML() const {
353   return new yaml::X86MachineFunctionInfo();
354 }
355 
356 yaml::MachineFunctionInfo *
convertFuncInfoToYAML(const MachineFunction & MF) const357 X86TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
358   const auto *MFI = MF.getInfo<X86MachineFunctionInfo>();
359   return new yaml::X86MachineFunctionInfo(*MFI);
360 }
361 
parseMachineFunctionInfo(const yaml::MachineFunctionInfo & MFI,PerFunctionMIParsingState & PFS,SMDiagnostic & Error,SMRange & SourceRange) const362 bool X86TargetMachine::parseMachineFunctionInfo(
363     const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS,
364     SMDiagnostic &Error, SMRange &SourceRange) const {
365   const auto &YamlMFI = static_cast<const yaml::X86MachineFunctionInfo &>(MFI);
366   PFS.MF.getInfo<X86MachineFunctionInfo>()->initializeBaseYamlFields(YamlMFI);
367   return false;
368 }
369 
isNoopAddrSpaceCast(unsigned SrcAS,unsigned DestAS) const370 bool X86TargetMachine::isNoopAddrSpaceCast(unsigned SrcAS,
371                                            unsigned DestAS) const {
372   assert(SrcAS != DestAS && "Expected different address spaces!");
373   if (getPointerSize(SrcAS) != getPointerSize(DestAS))
374     return false;
375   return SrcAS < 256 && DestAS < 256;
376 }
377 
reset()378 void X86TargetMachine::reset() { SubtargetMap.clear(); }
379 
380 ScheduleDAGInstrs *
createMachineScheduler(MachineSchedContext * C) const381 X86TargetMachine::createMachineScheduler(MachineSchedContext *C) const {
382   ScheduleDAGMILive *DAG = createSchedLive(C);
383   DAG->addMutation(createX86MacroFusionDAGMutation());
384   return DAG;
385 }
386 
387 ScheduleDAGInstrs *
createPostMachineScheduler(MachineSchedContext * C) const388 X86TargetMachine::createPostMachineScheduler(MachineSchedContext *C) const {
389   ScheduleDAGMI *DAG = createSchedPostRA(C);
390   DAG->addMutation(createX86MacroFusionDAGMutation());
391   return DAG;
392 }
393 
394 //===----------------------------------------------------------------------===//
395 // X86 TTI query.
396 //===----------------------------------------------------------------------===//
397 
398 TargetTransformInfo
getTargetTransformInfo(const Function & F) const399 X86TargetMachine::getTargetTransformInfo(const Function &F) const {
400   return TargetTransformInfo(std::make_unique<X86TTIImpl>(this, F));
401 }
402 
403 //===----------------------------------------------------------------------===//
404 // Pass Pipeline Configuration
405 //===----------------------------------------------------------------------===//
406 
407 namespace {
408 
409 /// X86 Code Generator Pass Configuration Options.
410 class X86PassConfig : public TargetPassConfig {
411 public:
X86PassConfig(X86TargetMachine & TM,PassManagerBase & PM)412   X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM)
413     : TargetPassConfig(TM, PM) {}
414 
getX86TargetMachine() const415   X86TargetMachine &getX86TargetMachine() const {
416     return getTM<X86TargetMachine>();
417   }
418 
419   void addIRPasses() override;
420   bool addInstSelector() override;
421   bool addIRTranslator() override;
422   bool addLegalizeMachineIR() override;
423   bool addRegBankSelect() override;
424   bool addGlobalInstructionSelect() override;
425   bool addILPOpts() override;
426   bool addPreISel() override;
427   void addMachineSSAOptimization() override;
428   void addPreRegAlloc() override;
429   bool addPostFastRegAllocRewrite() override;
430   void addPostRegAlloc() override;
431   void addPreEmitPass() override;
432   void addPreEmitPass2() override;
433   void addPreSched2() override;
434   bool addRegAssignAndRewriteOptimized() override;
435 
436   std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
437 };
438 
439 class X86ExecutionDomainFix : public ExecutionDomainFix {
440 public:
441   static char ID;
X86ExecutionDomainFix()442   X86ExecutionDomainFix() : ExecutionDomainFix(ID, X86::VR128XRegClass) {}
getPassName() const443   StringRef getPassName() const override {
444     return "X86 Execution Dependency Fix";
445   }
446 };
447 char X86ExecutionDomainFix::ID;
448 
449 } // end anonymous namespace
450 
451 INITIALIZE_PASS_BEGIN(X86ExecutionDomainFix, "x86-execution-domain-fix",
452   "X86 Execution Domain Fix", false, false)
INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis)453 INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis)
454 INITIALIZE_PASS_END(X86ExecutionDomainFix, "x86-execution-domain-fix",
455   "X86 Execution Domain Fix", false, false)
456 
457 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) {
458   return new X86PassConfig(*this, PM);
459 }
460 
createMachineFunctionInfo(BumpPtrAllocator & Allocator,const Function & F,const TargetSubtargetInfo * STI) const461 MachineFunctionInfo *X86TargetMachine::createMachineFunctionInfo(
462     BumpPtrAllocator &Allocator, const Function &F,
463     const TargetSubtargetInfo *STI) const {
464   return X86MachineFunctionInfo::create<X86MachineFunctionInfo>(Allocator, F,
465                                                                 STI);
466 }
467 
addIRPasses()468 void X86PassConfig::addIRPasses() {
469   addPass(createAtomicExpandLegacyPass());
470 
471   // We add both pass anyway and when these two passes run, we skip the pass
472   // based on the option level and option attribute.
473   addPass(createX86LowerAMXIntrinsicsPass());
474   addPass(createX86LowerAMXTypePass());
475 
476   TargetPassConfig::addIRPasses();
477 
478   if (TM->getOptLevel() != CodeGenOptLevel::None) {
479     addPass(createInterleavedAccessPass());
480     addPass(createX86PartialReductionPass());
481   }
482 
483   // Add passes that handle indirect branch removal and insertion of a retpoline
484   // thunk. These will be a no-op unless a function subtarget has the retpoline
485   // feature enabled.
486   addPass(createIndirectBrExpandPass());
487 
488   // Add Control Flow Guard checks.
489   const Triple &TT = TM->getTargetTriple();
490   if (TT.isOSWindows()) {
491     if (TT.getArch() == Triple::x86_64) {
492       addPass(createCFGuardDispatchPass());
493     } else {
494       addPass(createCFGuardCheckPass());
495     }
496   }
497 
498   if (TM->Options.JMCInstrument)
499     addPass(createJMCInstrumenterPass());
500 }
501 
addInstSelector()502 bool X86PassConfig::addInstSelector() {
503   // Install an instruction selector.
504   addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel()));
505 
506   // For ELF, cleanup any local-dynamic TLS accesses.
507   if (TM->getTargetTriple().isOSBinFormatELF() &&
508       getOptLevel() != CodeGenOptLevel::None)
509     addPass(createCleanupLocalDynamicTLSPass());
510 
511   addPass(createX86GlobalBaseRegPass());
512   addPass(createX86ArgumentStackSlotPass());
513   return false;
514 }
515 
addIRTranslator()516 bool X86PassConfig::addIRTranslator() {
517   addPass(new IRTranslator(getOptLevel()));
518   return false;
519 }
520 
addLegalizeMachineIR()521 bool X86PassConfig::addLegalizeMachineIR() {
522   addPass(new Legalizer());
523   return false;
524 }
525 
addRegBankSelect()526 bool X86PassConfig::addRegBankSelect() {
527   addPass(new RegBankSelect());
528   return false;
529 }
530 
addGlobalInstructionSelect()531 bool X86PassConfig::addGlobalInstructionSelect() {
532   addPass(new InstructionSelect(getOptLevel()));
533   // Add GlobalBaseReg in case there is no SelectionDAG passes afterwards
534   if (isGlobalISelAbortEnabled())
535     addPass(createX86GlobalBaseRegPass());
536   return false;
537 }
538 
addILPOpts()539 bool X86PassConfig::addILPOpts() {
540   addPass(&EarlyIfConverterLegacyID);
541   if (EnableMachineCombinerPass)
542     addPass(&MachineCombinerID);
543   addPass(createX86CmovConverterPass());
544   return true;
545 }
546 
addPreISel()547 bool X86PassConfig::addPreISel() {
548   // Only add this pass for 32-bit x86 Windows.
549   const Triple &TT = TM->getTargetTriple();
550   if (TT.isOSWindows() && TT.getArch() == Triple::x86)
551     addPass(createX86WinEHStatePass());
552   return true;
553 }
554 
addPreRegAlloc()555 void X86PassConfig::addPreRegAlloc() {
556   if (getOptLevel() != CodeGenOptLevel::None) {
557     addPass(&LiveRangeShrinkID);
558     addPass(createX86FixupSetCC());
559     addPass(createX86OptimizeLEAs());
560     addPass(createX86CallFrameOptimization());
561     addPass(createX86AvoidStoreForwardingBlocks());
562   }
563 
564   addPass(createX86SuppressAPXForRelocationPass());
565 
566   addPass(createX86SpeculativeLoadHardeningPass());
567   addPass(createX86FlagsCopyLoweringPass());
568   addPass(createX86DynAllocaExpander());
569 
570   if (getOptLevel() != CodeGenOptLevel::None)
571     addPass(createX86PreTileConfigPass());
572   else
573     addPass(createX86FastPreTileConfigPass());
574 }
575 
addMachineSSAOptimization()576 void X86PassConfig::addMachineSSAOptimization() {
577   addPass(createX86DomainReassignmentPass());
578   TargetPassConfig::addMachineSSAOptimization();
579 }
580 
addPostRegAlloc()581 void X86PassConfig::addPostRegAlloc() {
582   addPass(createX86LowerTileCopyPass());
583   addPass(createX86FloatingPointStackifierPass());
584   // When -O0 is enabled, the Load Value Injection Hardening pass will fall back
585   // to using the Speculative Execution Side Effect Suppression pass for
586   // mitigation. This is to prevent slow downs due to
587   // analyses needed by the LVIHardening pass when compiling at -O0.
588   if (getOptLevel() != CodeGenOptLevel::None)
589     addPass(createX86LoadValueInjectionLoadHardeningPass());
590 }
591 
addPreSched2()592 void X86PassConfig::addPreSched2() {
593   addPass(createX86ExpandPseudoPass());
594   addPass(createKCFIPass());
595 }
596 
addPreEmitPass()597 void X86PassConfig::addPreEmitPass() {
598   if (getOptLevel() != CodeGenOptLevel::None) {
599     addPass(new X86ExecutionDomainFix());
600     addPass(createBreakFalseDeps());
601   }
602 
603   addPass(createX86IndirectBranchTrackingPass());
604 
605   addPass(createX86IssueVZeroUpperPass());
606 
607   if (getOptLevel() != CodeGenOptLevel::None) {
608     addPass(createX86FixupBWInsts());
609     addPass(createX86PadShortFunctions());
610     addPass(createX86FixupLEAs());
611     addPass(createX86FixupInstTuning());
612     addPass(createX86FixupVectorConstants());
613   }
614   addPass(createX86CompressEVEXPass());
615   addPass(createX86DiscriminateMemOpsPass());
616   addPass(createX86InsertPrefetchPass());
617   addPass(createX86InsertX87waitPass());
618 }
619 
addPreEmitPass2()620 void X86PassConfig::addPreEmitPass2() {
621   const Triple &TT = TM->getTargetTriple();
622   const MCAsmInfo *MAI = TM->getMCAsmInfo();
623 
624   // The X86 Speculative Execution Pass must run after all control
625   // flow graph modifying passes. As a result it was listed to run right before
626   // the X86 Retpoline Thunks pass. The reason it must run after control flow
627   // graph modifications is that the model of LFENCE in LLVM has to be updated
628   // (FIXME: https://bugs.llvm.org/show_bug.cgi?id=45167). Currently the
629   // placement of this pass was hand checked to ensure that the subsequent
630   // passes don't move the code around the LFENCEs in a way that will hurt the
631   // correctness of this pass. This placement has been shown to work based on
632   // hand inspection of the codegen output.
633   addPass(createX86SpeculativeExecutionSideEffectSuppression());
634   addPass(createX86IndirectThunksPass());
635   addPass(createX86ReturnThunksPass());
636 
637   // Insert extra int3 instructions after trailing call instructions to avoid
638   // issues in the unwinder.
639   if (TT.isOSWindows() && TT.getArch() == Triple::x86_64)
640     addPass(createX86AvoidTrailingCallPass());
641 
642   // Verify basic block incoming and outgoing cfa offset and register values and
643   // correct CFA calculation rule where needed by inserting appropriate CFI
644   // instructions.
645   if (!TT.isOSDarwin() &&
646       (!TT.isOSWindows() ||
647        MAI->getExceptionHandlingType() == ExceptionHandling::DwarfCFI))
648     addPass(createCFIInstrInserter());
649 
650   if (TT.isOSWindows()) {
651     // Identify valid longjmp targets for Windows Control Flow Guard.
652     addPass(createCFGuardLongjmpPass());
653     // Identify valid eh continuation targets for Windows EHCont Guard.
654     addPass(createEHContGuardTargetsPass());
655   }
656   addPass(createX86LoadValueInjectionRetHardeningPass());
657 
658   // Insert pseudo probe annotation for callsite profiling
659   addPass(createPseudoProbeInserter());
660 
661   // KCFI indirect call checks are lowered to a bundle, and on Darwin platforms,
662   // also CALL_RVMARKER.
663   addPass(createUnpackMachineBundles([&TT](const MachineFunction &MF) {
664     // Only run bundle expansion if the module uses kcfi, or there are relevant
665     // ObjC runtime functions present in the module.
666     const Function &F = MF.getFunction();
667     const Module *M = F.getParent();
668     return M->getModuleFlag("kcfi") ||
669            (TT.isOSDarwin() &&
670             (M->getFunction("objc_retainAutoreleasedReturnValue") ||
671              M->getFunction("objc_unsafeClaimAutoreleasedReturnValue")));
672   }));
673 
674   // Analyzes and emits pseudos to support Win x64 Unwind V2. This pass must run
675   // after all real instructions have been added to the epilog.
676   if (TT.isOSWindows() && (TT.getArch() == Triple::x86_64))
677     addPass(createX86WinEHUnwindV2Pass());
678 }
679 
addPostFastRegAllocRewrite()680 bool X86PassConfig::addPostFastRegAllocRewrite() {
681   addPass(createX86FastTileConfigPass());
682   return true;
683 }
684 
getCSEConfig() const685 std::unique_ptr<CSEConfigBase> X86PassConfig::getCSEConfig() const {
686   return getStandardCSEConfigForOpt(TM->getOptLevel());
687 }
688 
onlyAllocateTileRegisters(const TargetRegisterInfo & TRI,const MachineRegisterInfo & MRI,const Register Reg)689 static bool onlyAllocateTileRegisters(const TargetRegisterInfo &TRI,
690                                       const MachineRegisterInfo &MRI,
691                                       const Register Reg) {
692   const TargetRegisterClass *RC = MRI.getRegClass(Reg);
693   return static_cast<const X86RegisterInfo &>(TRI).isTileRegisterClass(RC);
694 }
695 
addRegAssignAndRewriteOptimized()696 bool X86PassConfig::addRegAssignAndRewriteOptimized() {
697   // Don't support tile RA when RA is specified by command line "-regalloc".
698   if (!isCustomizedRegAlloc() && EnableTileRAPass) {
699     // Allocate tile register first.
700     addPass(createGreedyRegisterAllocator(onlyAllocateTileRegisters));
701     addPass(createX86TileConfigPass());
702   }
703   return TargetPassConfig::addRegAssignAndRewriteOptimized();
704 }
705