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 "MCTargetDesc/X86MCTargetDesc.h"
15 #include "X86.h"
16 #include "X86CallLowering.h"
17 #include "X86LegalizerInfo.h"
18 #include "X86MacroFusion.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.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/ExecutionDepsFix.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 namespace llvm {
58 
59 void initializeWinEHStatePassPass(PassRegistry &);
60 void initializeFixupLEAPassPass(PassRegistry &);
61 void initializeX86CallFrameOptimizationPass(PassRegistry &);
62 void initializeX86CmovConverterPassPass(PassRegistry &);
63 void initializeX86ExecutionDepsFixPass(PassRegistry &);
64 void initializeX86DomainReassignmentPass(PassRegistry &);
65 
66 } // end namespace llvm
67 
68 extern "C" 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   initializeGlobalISel(PR);
75   initializeWinEHStatePassPass(PR);
76   initializeFixupBWInstPassPass(PR);
77   initializeEvexToVexInstPassPass(PR);
78   initializeFixupLEAPassPass(PR);
79   initializeX86CallFrameOptimizationPass(PR);
80   initializeX86CmovConverterPassPass(PR);
81   initializeX86ExecutionDepsFixPass(PR);
82   initializeX86DomainReassignmentPass(PR);
83 }
84 
85 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
86   if (TT.isOSBinFormatMachO()) {
87     if (TT.getArch() == Triple::x86_64)
88       return llvm::make_unique<X86_64MachoTargetObjectFile>();
89     return llvm::make_unique<TargetLoweringObjectFileMachO>();
90   }
91 
92   if (TT.isOSFreeBSD())
93     return llvm::make_unique<X86FreeBSDTargetObjectFile>();
94   if (TT.isOSLinux() || TT.isOSNaCl() || TT.isOSIAMCU())
95     return llvm::make_unique<X86LinuxNaClTargetObjectFile>();
96   if (TT.isOSSolaris())
97     return llvm::make_unique<X86SolarisTargetObjectFile>();
98   if (TT.isOSFuchsia())
99     return llvm::make_unique<X86FuchsiaTargetObjectFile>();
100   if (TT.isOSBinFormatELF())
101     return llvm::make_unique<X86ELFTargetObjectFile>();
102   if (TT.isKnownWindowsMSVCEnvironment() || TT.isWindowsCoreCLREnvironment())
103     return llvm::make_unique<X86WindowsTargetObjectFile>();
104   if (TT.isOSBinFormatCOFF())
105     return llvm::make_unique<TargetLoweringObjectFileCOFF>();
106   llvm_unreachable("unknown subtarget type");
107 }
108 
109 static std::string computeDataLayout(const Triple &TT) {
110   // X86 is little endian
111   std::string Ret = "e";
112 
113   Ret += DataLayout::getManglingComponent(TT);
114   // X86 and x32 have 32 bit pointers.
115   if ((TT.isArch64Bit() &&
116        (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) ||
117       !TT.isArch64Bit())
118     Ret += "-p:32:32";
119 
120   // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32.
121   if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl())
122     Ret += "-i64:64";
123   else if (TT.isOSIAMCU())
124     Ret += "-i64:32-f64:32";
125   else
126     Ret += "-f64:32:64";
127 
128   // Some ABIs align long double to 128 bits, others to 32.
129   if (TT.isOSNaCl() || TT.isOSIAMCU())
130     ; // No f80
131   else if (TT.isArch64Bit() || TT.isOSDarwin())
132     Ret += "-f80:128";
133   else
134     Ret += "-f80:32";
135 
136   if (TT.isOSIAMCU())
137     Ret += "-f128:32";
138 
139   // The registers can hold 8, 16, 32 or, in x86-64, 64 bits.
140   if (TT.isArch64Bit())
141     Ret += "-n8:16:32:64";
142   else
143     Ret += "-n8:16:32";
144 
145   // The stack is aligned to 32 bits on some ABIs and 128 bits on others.
146   if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU())
147     Ret += "-a:0:32-S32";
148   else
149     Ret += "-S128";
150 
151   return Ret;
152 }
153 
154 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
155                                            Optional<Reloc::Model> RM) {
156   bool is64Bit = TT.getArch() == Triple::x86_64;
157   if (!RM.hasValue()) {
158     // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode.
159     // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we
160     // use static relocation model by default.
161     if (TT.isOSDarwin()) {
162       if (is64Bit)
163         return Reloc::PIC_;
164       return Reloc::DynamicNoPIC;
165     }
166     if (TT.isOSWindows() && is64Bit)
167       return Reloc::PIC_;
168     return Reloc::Static;
169   }
170 
171   // ELF and X86-64 don't have a distinct DynamicNoPIC model.  DynamicNoPIC
172   // is defined as a model for code which may be used in static or dynamic
173   // executables but not necessarily a shared library. On X86-32 we just
174   // compile in -static mode, in x86-64 we use PIC.
175   if (*RM == Reloc::DynamicNoPIC) {
176     if (is64Bit)
177       return Reloc::PIC_;
178     if (!TT.isOSDarwin())
179       return Reloc::Static;
180   }
181 
182   // If we are on Darwin, disallow static relocation model in X86-64 mode, since
183   // the Mach-O file format doesn't support it.
184   if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit)
185     return Reloc::PIC_;
186 
187   return *RM;
188 }
189 
190 static CodeModel::Model getEffectiveCodeModel(Optional<CodeModel::Model> CM,
191                                               bool JIT, bool Is64Bit) {
192   if (CM)
193     return *CM;
194   if (JIT)
195     return Is64Bit ? CodeModel::Large : CodeModel::Small;
196   return CodeModel::Small;
197 }
198 
199 /// Create an X86 target.
200 ///
201 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT,
202                                    StringRef CPU, StringRef FS,
203                                    const TargetOptions &Options,
204                                    Optional<Reloc::Model> RM,
205                                    Optional<CodeModel::Model> CM,
206                                    CodeGenOpt::Level OL, bool JIT)
207     : LLVMTargetMachine(
208           T, computeDataLayout(TT), TT, CPU, FS, Options,
209           getEffectiveRelocModel(TT, RM),
210           getEffectiveCodeModel(CM, JIT, TT.getArch() == Triple::x86_64), OL),
211       TLOF(createTLOF(getTargetTriple())) {
212   // Windows stack unwinder gets confused when execution flow "falls through"
213   // after a call to 'noreturn' function.
214   // To prevent that, we emit a trap for 'unreachable' IR instructions.
215   // (which on X86, happens to be the 'ud2' instruction)
216   // On PS4, the "return address" of a 'noreturn' call must still be within
217   // the calling function, and TrapUnreachable is an easy way to get that.
218   // The check here for 64-bit windows is a bit icky, but as we're unlikely
219   // to ever want to mix 32 and 64-bit windows code in a single module
220   // this should be fine.
221   if ((TT.isOSWindows() && TT.getArch() == Triple::x86_64) || TT.isPS4())
222     this->Options.TrapUnreachable = true;
223 
224   initAsmInfo();
225 }
226 
227 X86TargetMachine::~X86TargetMachine() = default;
228 
229 const X86Subtarget *
230 X86TargetMachine::getSubtargetImpl(const Function &F) const {
231   Attribute CPUAttr = F.getFnAttribute("target-cpu");
232   Attribute FSAttr = F.getFnAttribute("target-features");
233 
234   StringRef CPU = !CPUAttr.hasAttribute(Attribute::None)
235                       ? CPUAttr.getValueAsString()
236                       : (StringRef)TargetCPU;
237   StringRef FS = !FSAttr.hasAttribute(Attribute::None)
238                      ? FSAttr.getValueAsString()
239                      : (StringRef)TargetFS;
240 
241   SmallString<512> Key;
242   Key.reserve(CPU.size() + FS.size());
243   Key += CPU;
244   Key += FS;
245 
246   // FIXME: This is related to the code below to reset the target options,
247   // we need to know whether or not the soft float flag is set on the
248   // function before we can generate a subtarget. We also need to use
249   // it as a key for the subtarget since that can be the only difference
250   // between two functions.
251   bool SoftFloat =
252       F.getFnAttribute("use-soft-float").getValueAsString() == "true";
253   // If the soft float attribute is set on the function turn on the soft float
254   // subtarget feature.
255   if (SoftFloat)
256     Key += FS.empty() ? "+soft-float" : ",+soft-float";
257 
258   FS = Key.substr(CPU.size());
259 
260   auto &I = SubtargetMap[Key];
261   if (!I) {
262     // This needs to be done before we create a new subtarget since any
263     // creation will depend on the TM and the code generation flags on the
264     // function that reside in TargetOptions.
265     resetTargetOptions(F);
266     I = llvm::make_unique<X86Subtarget>(TargetTriple, CPU, FS, *this,
267                                         Options.StackAlignmentOverride);
268   }
269   return I.get();
270 }
271 
272 //===----------------------------------------------------------------------===//
273 // Command line options for x86
274 //===----------------------------------------------------------------------===//
275 static cl::opt<bool>
276 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden,
277   cl::desc("Minimize AVX to SSE transition penalty"),
278   cl::init(true));
279 
280 //===----------------------------------------------------------------------===//
281 // X86 TTI query.
282 //===----------------------------------------------------------------------===//
283 
284 TargetIRAnalysis X86TargetMachine::getTargetIRAnalysis() {
285   return TargetIRAnalysis([this](const Function &F) {
286     return TargetTransformInfo(X86TTIImpl(this, F));
287   });
288 }
289 
290 //===----------------------------------------------------------------------===//
291 // Pass Pipeline Configuration
292 //===----------------------------------------------------------------------===//
293 
294 namespace {
295 
296 /// X86 Code Generator Pass Configuration Options.
297 class X86PassConfig : public TargetPassConfig {
298 public:
299   X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM)
300     : TargetPassConfig(TM, PM) {}
301 
302   X86TargetMachine &getX86TargetMachine() const {
303     return getTM<X86TargetMachine>();
304   }
305 
306   ScheduleDAGInstrs *
307   createMachineScheduler(MachineSchedContext *C) const override {
308     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
309     DAG->addMutation(createX86MacroFusionDAGMutation());
310     return DAG;
311   }
312 
313   void addIRPasses() override;
314   bool addInstSelector() override;
315   bool addIRTranslator() override;
316   bool addLegalizeMachineIR() override;
317   bool addRegBankSelect() override;
318   bool addGlobalInstructionSelect() override;
319   bool addILPOpts() override;
320   bool addPreISel() override;
321   void addMachineSSAOptimization() override;
322   void addPreRegAlloc() override;
323   void addPostRegAlloc() override;
324   void addPreEmitPass() override;
325   void addPreSched2() override;
326 };
327 
328 class X86ExecutionDepsFix : public ExecutionDepsFix {
329 public:
330   static char ID;
331   X86ExecutionDepsFix() : ExecutionDepsFix(ID, X86::VR128XRegClass) {}
332   StringRef getPassName() const override {
333     return "X86 Execution Dependency Fix";
334   }
335 };
336 char X86ExecutionDepsFix::ID;
337 
338 } // end anonymous namespace
339 
340 INITIALIZE_PASS(X86ExecutionDepsFix, "x86-execution-deps-fix",
341                 "X86 Execution Dependency Fix", false, false)
342 
343 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) {
344   return new X86PassConfig(*this, PM);
345 }
346 
347 void X86PassConfig::addIRPasses() {
348   addPass(createAtomicExpandPass());
349 
350   TargetPassConfig::addIRPasses();
351 
352   if (TM->getOptLevel() != CodeGenOpt::None)
353     addPass(createInterleavedAccessPass());
354 }
355 
356 bool X86PassConfig::addInstSelector() {
357   // Install an instruction selector.
358   addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel()));
359 
360   // For ELF, cleanup any local-dynamic TLS accesses.
361   if (TM->getTargetTriple().isOSBinFormatELF() &&
362       getOptLevel() != CodeGenOpt::None)
363     addPass(createCleanupLocalDynamicTLSPass());
364 
365   addPass(createX86GlobalBaseRegPass());
366   return false;
367 }
368 
369 bool X86PassConfig::addIRTranslator() {
370   addPass(new IRTranslator());
371   return false;
372 }
373 
374 bool X86PassConfig::addLegalizeMachineIR() {
375   addPass(new Legalizer());
376   return false;
377 }
378 
379 bool X86PassConfig::addRegBankSelect() {
380   addPass(new RegBankSelect());
381   return false;
382 }
383 
384 bool X86PassConfig::addGlobalInstructionSelect() {
385   addPass(new InstructionSelect());
386   return false;
387 }
388 
389 bool X86PassConfig::addILPOpts() {
390   addPass(&EarlyIfConverterID);
391   if (EnableMachineCombinerPass)
392     addPass(&MachineCombinerID);
393   addPass(createX86CmovConverterPass());
394   return true;
395 }
396 
397 bool X86PassConfig::addPreISel() {
398   // Only add this pass for 32-bit x86 Windows.
399   const Triple &TT = TM->getTargetTriple();
400   if (TT.isOSWindows() && TT.getArch() == Triple::x86)
401     addPass(createX86WinEHStatePass());
402   return true;
403 }
404 
405 void X86PassConfig::addPreRegAlloc() {
406   if (getOptLevel() != CodeGenOpt::None) {
407     addPass(&LiveRangeShrinkID);
408     addPass(createX86FixupSetCC());
409     addPass(createX86OptimizeLEAs());
410     addPass(createX86CallFrameOptimization());
411   }
412 
413   addPass(createX86WinAllocaExpander());
414 }
415 void X86PassConfig::addMachineSSAOptimization() {
416   addPass(createX86DomainReassignmentPass());
417   TargetPassConfig::addMachineSSAOptimization();
418 }
419 
420 void X86PassConfig::addPostRegAlloc() {
421   addPass(createX86FloatingPointStackifierPass());
422 }
423 
424 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); }
425 
426 void X86PassConfig::addPreEmitPass() {
427   if (getOptLevel() != CodeGenOpt::None)
428     addPass(new X86ExecutionDepsFix());
429 
430   if (UseVZeroUpper)
431     addPass(createX86IssueVZeroUpperPass());
432 
433   if (getOptLevel() != CodeGenOpt::None) {
434     addPass(createX86FixupBWInsts());
435     addPass(createX86PadShortFunctions());
436     addPass(createX86FixupLEAs());
437     addPass(createX86EvexToVexInsts());
438   }
439 }
440