1 //===-- X86TargetMachine.cpp - Define TargetMachine for the X86 -----------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the X86 specific subclass of TargetMachine.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "X86TargetMachine.h"
15 #include "MCTargetDesc/X86MCTargetDesc.h"
16 #include "X86.h"
17 #include "X86CallLowering.h"
18 #include "X86LegalizerInfo.h"
19 #include "X86MacroFusion.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetObjectFile.h"
22 #include "X86TargetTransformInfo.h"
23 #include "llvm/ADT/Optional.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SmallString.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/ADT/Triple.h"
28 #include "llvm/Analysis/TargetTransformInfo.h"
29 #include "llvm/CodeGen/ExecutionDomainFix.h"
30 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
31 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
32 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
33 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
34 #include "llvm/CodeGen/GlobalISel/RegBankSelect.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/Pass.h"
42 #include "llvm/Support/CodeGen.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/TargetRegistry.h"
46 #include "llvm/Target/TargetLoweringObjectFile.h"
47 #include "llvm/Target/TargetOptions.h"
48 #include <memory>
49 #include <string>
50 
51 using namespace llvm;
52 
53 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner",
54                                cl::desc("Enable the machine combiner pass"),
55                                cl::init(true), cl::Hidden);
56 
57 static cl::opt<bool> EnableCondBrFoldingPass("x86-condbr-folding",
58                                cl::desc("Enable the conditional branch "
59                                         "folding pass"),
60                                cl::init(true), cl::Hidden);
61 
62 namespace llvm {
63 
64 void initializeWinEHStatePassPass(PassRegistry &);
65 void initializeFixupLEAPassPass(PassRegistry &);
66 void initializeShadowCallStackPass(PassRegistry &);
67 void initializeX86CallFrameOptimizationPass(PassRegistry &);
68 void initializeX86CmovConverterPassPass(PassRegistry &);
69 void initializeX86ExecutionDomainFixPass(PassRegistry &);
70 void initializeX86DomainReassignmentPass(PassRegistry &);
71 void initializeX86AvoidSFBPassPass(PassRegistry &);
72 void initializeX86SpeculativeLoadHardeningPassPass(PassRegistry &);
73 void initializeX86FlagsCopyLoweringPassPass(PassRegistry &);
74 
75 } // end namespace llvm
76 
77 extern "C" void LLVMInitializeX86Target() {
78   // Register the target.
79   RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target());
80   RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target());
81 
82   PassRegistry &PR = *PassRegistry::getPassRegistry();
83   initializeGlobalISel(PR);
84   initializeWinEHStatePassPass(PR);
85   initializeFixupBWInstPassPass(PR);
86   initializeEvexToVexInstPassPass(PR);
87   initializeFixupLEAPassPass(PR);
88   initializeShadowCallStackPass(PR);
89   initializeX86CallFrameOptimizationPass(PR);
90   initializeX86CmovConverterPassPass(PR);
91   initializeX86ExecutionDomainFixPass(PR);
92   initializeX86DomainReassignmentPass(PR);
93   initializeX86AvoidSFBPassPass(PR);
94   initializeX86SpeculativeLoadHardeningPassPass(PR);
95   initializeX86FlagsCopyLoweringPassPass(PR);
96 }
97 
98 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
99   if (TT.isOSBinFormatMachO()) {
100     if (TT.getArch() == Triple::x86_64)
101       return llvm::make_unique<X86_64MachoTargetObjectFile>();
102     return llvm::make_unique<TargetLoweringObjectFileMachO>();
103   }
104 
105   if (TT.isOSFreeBSD())
106     return llvm::make_unique<X86FreeBSDTargetObjectFile>();
107   if (TT.isOSLinux() || TT.isOSNaCl() || TT.isOSIAMCU())
108     return llvm::make_unique<X86LinuxNaClTargetObjectFile>();
109   if (TT.isOSSolaris())
110     return llvm::make_unique<X86SolarisTargetObjectFile>();
111   if (TT.isOSFuchsia())
112     return llvm::make_unique<X86FuchsiaTargetObjectFile>();
113   if (TT.isOSBinFormatELF())
114     return llvm::make_unique<X86ELFTargetObjectFile>();
115   if (TT.isOSBinFormatCOFF())
116     return llvm::make_unique<TargetLoweringObjectFileCOFF>();
117   llvm_unreachable("unknown subtarget type");
118 }
119 
120 static std::string computeDataLayout(const Triple &TT) {
121   // X86 is little endian
122   std::string Ret = "e";
123 
124   Ret += DataLayout::getManglingComponent(TT);
125   // X86 and x32 have 32 bit pointers.
126   if ((TT.isArch64Bit() &&
127        (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) ||
128       !TT.isArch64Bit())
129     Ret += "-p:32:32";
130 
131   // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32.
132   if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl())
133     Ret += "-i64:64";
134   else if (TT.isOSIAMCU())
135     Ret += "-i64:32-f64:32";
136   else
137     Ret += "-f64:32:64";
138 
139   // Some ABIs align long double to 128 bits, others to 32.
140   if (TT.isOSNaCl() || TT.isOSIAMCU())
141     ; // No f80
142   else if (TT.isArch64Bit() || TT.isOSDarwin())
143     Ret += "-f80:128";
144   else
145     Ret += "-f80:32";
146 
147   if (TT.isOSIAMCU())
148     Ret += "-f128:32";
149 
150   // The registers can hold 8, 16, 32 or, in x86-64, 64 bits.
151   if (TT.isArch64Bit())
152     Ret += "-n8:16:32:64";
153   else
154     Ret += "-n8:16:32";
155 
156   // The stack is aligned to 32 bits on some ABIs and 128 bits on others.
157   if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU())
158     Ret += "-a:0:32-S32";
159   else
160     Ret += "-S128";
161 
162   return Ret;
163 }
164 
165 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
166                                            bool JIT,
167                                            Optional<Reloc::Model> RM) {
168   bool is64Bit = TT.getArch() == Triple::x86_64;
169   if (!RM.hasValue()) {
170     // JIT codegen should use static relocations by default, since it's
171     // typically executed in process and not relocatable.
172     if (JIT)
173       return Reloc::Static;
174 
175     // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode.
176     // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we
177     // use static relocation model by default.
178     if (TT.isOSDarwin()) {
179       if (is64Bit)
180         return Reloc::PIC_;
181       return Reloc::DynamicNoPIC;
182     }
183     if (TT.isOSWindows() && is64Bit)
184       return Reloc::PIC_;
185     return Reloc::Static;
186   }
187 
188   // ELF and X86-64 don't have a distinct DynamicNoPIC model.  DynamicNoPIC
189   // is defined as a model for code which may be used in static or dynamic
190   // executables but not necessarily a shared library. On X86-32 we just
191   // compile in -static mode, in x86-64 we use PIC.
192   if (*RM == Reloc::DynamicNoPIC) {
193     if (is64Bit)
194       return Reloc::PIC_;
195     if (!TT.isOSDarwin())
196       return Reloc::Static;
197   }
198 
199   // If we are on Darwin, disallow static relocation model in X86-64 mode, since
200   // the Mach-O file format doesn't support it.
201   if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit)
202     return Reloc::PIC_;
203 
204   return *RM;
205 }
206 
207 static CodeModel::Model getEffectiveCodeModel(Optional<CodeModel::Model> CM,
208                                               bool JIT, bool Is64Bit) {
209   if (CM)
210     return *CM;
211   if (JIT)
212     return Is64Bit ? CodeModel::Large : CodeModel::Small;
213   return CodeModel::Small;
214 }
215 
216 /// Create an X86 target.
217 ///
218 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT,
219                                    StringRef CPU, StringRef FS,
220                                    const TargetOptions &Options,
221                                    Optional<Reloc::Model> RM,
222                                    Optional<CodeModel::Model> CM,
223                                    CodeGenOpt::Level OL, bool JIT)
224     : LLVMTargetMachine(
225           T, computeDataLayout(TT), TT, CPU, FS, Options,
226           getEffectiveRelocModel(TT, JIT, RM),
227           getEffectiveCodeModel(CM, JIT, TT.getArch() == Triple::x86_64), OL),
228       TLOF(createTLOF(getTargetTriple())) {
229   // Windows stack unwinder gets confused when execution flow "falls through"
230   // after a call to 'noreturn' function.
231   // To prevent that, we emit a trap for 'unreachable' IR instructions.
232   // (which on X86, happens to be the 'ud2' instruction)
233   // On PS4, the "return address" of a 'noreturn' call must still be within
234   // the calling function, and TrapUnreachable is an easy way to get that.
235   // The check here for 64-bit windows is a bit icky, but as we're unlikely
236   // to ever want to mix 32 and 64-bit windows code in a single module
237   // this should be fine.
238   if ((TT.isOSWindows() && TT.getArch() == Triple::x86_64) || TT.isPS4() ||
239       TT.isOSBinFormatMachO()) {
240     this->Options.TrapUnreachable = true;
241     this->Options.NoTrapAfterNoreturn = TT.isOSBinFormatMachO();
242   }
243 
244   // Outlining is available for x86-64.
245   if (TT.getArch() == Triple::x86_64)
246     setMachineOutliner(true);
247 
248   initAsmInfo();
249 }
250 
251 X86TargetMachine::~X86TargetMachine() = default;
252 
253 const X86Subtarget *
254 X86TargetMachine::getSubtargetImpl(const Function &F) const {
255   Attribute CPUAttr = F.getFnAttribute("target-cpu");
256   Attribute FSAttr = F.getFnAttribute("target-features");
257 
258   StringRef CPU = !CPUAttr.hasAttribute(Attribute::None)
259                       ? CPUAttr.getValueAsString()
260                       : (StringRef)TargetCPU;
261   StringRef FS = !FSAttr.hasAttribute(Attribute::None)
262                      ? FSAttr.getValueAsString()
263                      : (StringRef)TargetFS;
264 
265   SmallString<512> Key;
266   Key.reserve(CPU.size() + FS.size());
267   Key += CPU;
268   Key += FS;
269 
270   // FIXME: This is related to the code below to reset the target options,
271   // we need to know whether or not the soft float flag is set on the
272   // function before we can generate a subtarget. We also need to use
273   // it as a key for the subtarget since that can be the only difference
274   // between two functions.
275   bool SoftFloat =
276       F.getFnAttribute("use-soft-float").getValueAsString() == "true";
277   // If the soft float attribute is set on the function turn on the soft float
278   // subtarget feature.
279   if (SoftFloat)
280     Key += FS.empty() ? "+soft-float" : ",+soft-float";
281 
282   // Keep track of the key width after all features are added so we can extract
283   // the feature string out later.
284   unsigned CPUFSWidth = Key.size();
285 
286   // Extract prefer-vector-width attribute.
287   unsigned PreferVectorWidthOverride = 0;
288   if (F.hasFnAttribute("prefer-vector-width")) {
289     StringRef Val = F.getFnAttribute("prefer-vector-width").getValueAsString();
290     unsigned Width;
291     if (!Val.getAsInteger(0, Width)) {
292       Key += ",prefer-vector-width=";
293       Key += Val;
294       PreferVectorWidthOverride = Width;
295     }
296   }
297 
298   // Extract required-vector-width attribute.
299   unsigned RequiredVectorWidth = UINT32_MAX;
300   if (F.hasFnAttribute("required-vector-width")) {
301     StringRef Val = F.getFnAttribute("required-vector-width").getValueAsString();
302     unsigned Width;
303     if (!Val.getAsInteger(0, Width)) {
304       Key += ",required-vector-width=";
305       Key += Val;
306       RequiredVectorWidth = Width;
307     }
308   }
309 
310   // Extracted here so that we make sure there is backing for the StringRef. If
311   // we assigned earlier, its possible the SmallString reallocated leaving a
312   // dangling StringRef.
313   FS = Key.slice(CPU.size(), CPUFSWidth);
314 
315   auto &I = SubtargetMap[Key];
316   if (!I) {
317     // This needs to be done before we create a new subtarget since any
318     // creation will depend on the TM and the code generation flags on the
319     // function that reside in TargetOptions.
320     resetTargetOptions(F);
321     I = llvm::make_unique<X86Subtarget>(TargetTriple, CPU, FS, *this,
322                                         Options.StackAlignmentOverride,
323                                         PreferVectorWidthOverride,
324                                         RequiredVectorWidth);
325   }
326   return I.get();
327 }
328 
329 //===----------------------------------------------------------------------===//
330 // Command line options for x86
331 //===----------------------------------------------------------------------===//
332 static cl::opt<bool>
333 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden,
334   cl::desc("Minimize AVX to SSE transition penalty"),
335   cl::init(true));
336 
337 //===----------------------------------------------------------------------===//
338 // X86 TTI query.
339 //===----------------------------------------------------------------------===//
340 
341 TargetTransformInfo
342 X86TargetMachine::getTargetTransformInfo(const Function &F) {
343   return TargetTransformInfo(X86TTIImpl(this, F));
344 }
345 
346 //===----------------------------------------------------------------------===//
347 // Pass Pipeline Configuration
348 //===----------------------------------------------------------------------===//
349 
350 namespace {
351 
352 /// X86 Code Generator Pass Configuration Options.
353 class X86PassConfig : public TargetPassConfig {
354 public:
355   X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM)
356     : TargetPassConfig(TM, PM) {}
357 
358   X86TargetMachine &getX86TargetMachine() const {
359     return getTM<X86TargetMachine>();
360   }
361 
362   ScheduleDAGInstrs *
363   createMachineScheduler(MachineSchedContext *C) const override {
364     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
365     DAG->addMutation(createX86MacroFusionDAGMutation());
366     return DAG;
367   }
368 
369   void addIRPasses() override;
370   bool addInstSelector() override;
371   bool addIRTranslator() override;
372   bool addLegalizeMachineIR() override;
373   bool addRegBankSelect() override;
374   bool addGlobalInstructionSelect() override;
375   bool addILPOpts() override;
376   bool addPreISel() override;
377   void addMachineSSAOptimization() override;
378   void addPreRegAlloc() override;
379   void addPostRegAlloc() override;
380   void addPreEmitPass() override;
381   void addPreEmitPass2() override;
382   void addPreSched2() override;
383 };
384 
385 class X86ExecutionDomainFix : public ExecutionDomainFix {
386 public:
387   static char ID;
388   X86ExecutionDomainFix() : ExecutionDomainFix(ID, X86::VR128XRegClass) {}
389   StringRef getPassName() const override {
390     return "X86 Execution Dependency Fix";
391   }
392 };
393 char X86ExecutionDomainFix::ID;
394 
395 } // end anonymous namespace
396 
397 INITIALIZE_PASS_BEGIN(X86ExecutionDomainFix, "x86-execution-domain-fix",
398   "X86 Execution Domain Fix", false, false)
399 INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis)
400 INITIALIZE_PASS_END(X86ExecutionDomainFix, "x86-execution-domain-fix",
401   "X86 Execution Domain Fix", false, false)
402 
403 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) {
404   return new X86PassConfig(*this, PM);
405 }
406 
407 void X86PassConfig::addIRPasses() {
408   addPass(createAtomicExpandPass());
409 
410   TargetPassConfig::addIRPasses();
411 
412   if (TM->getOptLevel() != CodeGenOpt::None)
413     addPass(createInterleavedAccessPass());
414 
415   // Add passes that handle indirect branch removal and insertion of a retpoline
416   // thunk. These will be a no-op unless a function subtarget has the retpoline
417   // feature enabled.
418   addPass(createIndirectBrExpandPass());
419 }
420 
421 bool X86PassConfig::addInstSelector() {
422   // Install an instruction selector.
423   addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel()));
424 
425   // For ELF, cleanup any local-dynamic TLS accesses.
426   if (TM->getTargetTriple().isOSBinFormatELF() &&
427       getOptLevel() != CodeGenOpt::None)
428     addPass(createCleanupLocalDynamicTLSPass());
429 
430   addPass(createX86GlobalBaseRegPass());
431   return false;
432 }
433 
434 bool X86PassConfig::addIRTranslator() {
435   addPass(new IRTranslator());
436   return false;
437 }
438 
439 bool X86PassConfig::addLegalizeMachineIR() {
440   addPass(new Legalizer());
441   return false;
442 }
443 
444 bool X86PassConfig::addRegBankSelect() {
445   addPass(new RegBankSelect());
446   return false;
447 }
448 
449 bool X86PassConfig::addGlobalInstructionSelect() {
450   addPass(new InstructionSelect());
451   return false;
452 }
453 
454 bool X86PassConfig::addILPOpts() {
455   if (EnableCondBrFoldingPass)
456     addPass(createX86CondBrFolding());
457   addPass(&EarlyIfConverterID);
458   if (EnableMachineCombinerPass)
459     addPass(&MachineCombinerID);
460   addPass(createX86CmovConverterPass());
461   return true;
462 }
463 
464 bool X86PassConfig::addPreISel() {
465   // Only add this pass for 32-bit x86 Windows.
466   const Triple &TT = TM->getTargetTriple();
467   if (TT.isOSWindows() && TT.getArch() == Triple::x86)
468     addPass(createX86WinEHStatePass());
469   return true;
470 }
471 
472 void X86PassConfig::addPreRegAlloc() {
473   if (getOptLevel() != CodeGenOpt::None) {
474     addPass(&LiveRangeShrinkID);
475     addPass(createX86FixupSetCC());
476     addPass(createX86OptimizeLEAs());
477     addPass(createX86CallFrameOptimization());
478     addPass(createX86AvoidStoreForwardingBlocks());
479   }
480 
481   addPass(createX86SpeculativeLoadHardeningPass());
482   addPass(createX86FlagsCopyLoweringPass());
483   addPass(createX86WinAllocaExpander());
484 }
485 void X86PassConfig::addMachineSSAOptimization() {
486   addPass(createX86DomainReassignmentPass());
487   TargetPassConfig::addMachineSSAOptimization();
488 }
489 
490 void X86PassConfig::addPostRegAlloc() {
491   addPass(createX86FloatingPointStackifierPass());
492 }
493 
494 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); }
495 
496 void X86PassConfig::addPreEmitPass() {
497   if (getOptLevel() != CodeGenOpt::None) {
498     addPass(new X86ExecutionDomainFix());
499     addPass(createBreakFalseDeps());
500   }
501 
502   addPass(createShadowCallStackPass());
503   addPass(createX86IndirectBranchTrackingPass());
504 
505   if (UseVZeroUpper)
506     addPass(createX86IssueVZeroUpperPass());
507 
508   if (getOptLevel() != CodeGenOpt::None) {
509     addPass(createX86FixupBWInsts());
510     addPass(createX86PadShortFunctions());
511     addPass(createX86FixupLEAs());
512     addPass(createX86EvexToVexInsts());
513   }
514 }
515 
516 void X86PassConfig::addPreEmitPass2() {
517   addPass(createX86RetpolineThunksPass());
518   // Verify basic block incoming and outgoing cfa offset and register values and
519   // correct CFA calculation rule where needed by inserting appropriate CFI
520   // instructions.
521   const Triple &TT = TM->getTargetTriple();
522   if (!TT.isOSDarwin() && !TT.isOSWindows())
523     addPass(createCFIInstrInserter());
524 }
525