1 //===-- LLVMTargetMachine.cpp - Implement the LLVMTargetMachine class -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the LLVMTargetMachine class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetMachine.h"
15 #include "llvm/PassManager.h"
16 #include "llvm/Analysis/Verifier.h"
17 #include "llvm/Assembly/PrintModulePass.h"
18 #include "llvm/CodeGen/AsmPrinter.h"
19 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
20 #include "llvm/CodeGen/MachineModuleInfo.h"
21 #include "llvm/CodeGen/GCStrategy.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/Target/TargetOptions.h"
24 #include "llvm/MC/MCAsmInfo.h"
25 #include "llvm/MC/MCStreamer.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetRegistry.h"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/ADT/OwningPtr.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/FormattedStream.h"
39 static cl::opt<bool> DisablePostRA("disable-post-ra", cl::Hidden,
40 cl::desc("Disable Post Regalloc"));
41 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden,
42 cl::desc("Disable branch folding"));
43 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden,
44 cl::desc("Disable tail duplication"));
45 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden,
46 cl::desc("Disable pre-register allocation tail duplication"));
47 static cl::opt<bool> DisableCodePlace("disable-code-place", cl::Hidden,
48 cl::desc("Disable code placement"));
49 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden,
50 cl::desc("Disable Stack Slot Coloring"));
51 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden,
52 cl::desc("Disable Machine LICM"));
53 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
55 cl::desc("Disable Machine LICM"));
56 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden,
57 cl::desc("Disable Machine Sinking"));
58 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden,
59 cl::desc("Disable Loop Strength Reduction Pass"));
60 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden,
61 cl::desc("Disable Codegen Prepare"));
62 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden,
63 cl::desc("Print LLVM IR produced by the loop-reduce pass"));
64 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden,
65 cl::desc("Print LLVM IR input to isel pass"));
66 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden,
67 cl::desc("Dump garbage collector data"));
68 static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden,
69 cl::desc("Verify generated machine code"),
70 cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL));
72 static cl::opt<cl::boolOrDefault>
73 AsmVerbose("asm-verbose", cl::desc("Add comments to directives."),
74 cl::init(cl::BOU_UNSET));
76 static bool getVerboseAsm() {
79 case cl::BOU_UNSET: return TargetMachine::getAsmVerbosityDefault();
80 case cl::BOU_TRUE: return true;
81 case cl::BOU_FALSE: return false;
85 // Enable or disable FastISel. Both options are needed, because
86 // FastISel is enabled by default with -fast, and we wish to be
87 // able to enable or disable fast-isel independently from -O0.
88 static cl::opt<cl::boolOrDefault>
89 EnableFastISelOption("fast-isel", cl::Hidden,
90 cl::desc("Enable the \"fast\" instruction selector"));
92 // Enable or disable an experimental optimization to split GEPs
93 // and run a special GVN pass which does not examine loads, in
94 // an effort to factor out redundancy implicit in complex GEPs.
95 static cl::opt<bool> EnableSplitGEPGVN("split-gep-gvn", cl::Hidden,
96 cl::desc("Split GEPs and run no-load GVN"));
98 LLVMTargetMachine::LLVMTargetMachine(const Target &T,
99 const std::string &Triple)
100 : TargetMachine(T), TargetTriple(Triple) {
101 AsmInfo = T.createAsmInfo(TargetTriple);
104 // Set the default code model for the JIT for a generic target.
105 // FIXME: Is small right here? or .is64Bit() ? Large : Small?
106 void LLVMTargetMachine::setCodeModelForJIT() {
107 setCodeModel(CodeModel::Small);
110 // Set the default code model for static compilation for a generic target.
111 void LLVMTargetMachine::setCodeModelForStatic() {
112 setCodeModel(CodeModel::Small);
115 bool LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
116 formatted_raw_ostream &Out,
117 CodeGenFileType FileType,
118 CodeGenOpt::Level OptLevel,
119 bool DisableVerify) {
120 // Add common CodeGen passes.
121 MCContext *Context = 0;
122 if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify, Context))
124 assert(Context != 0 && "Failed to get MCContext");
126 const MCAsmInfo &MAI = *getMCAsmInfo();
127 OwningPtr<MCStreamer> AsmStreamer;
130 default: return true;
131 case CGFT_AssemblyFile: {
132 MCInstPrinter *InstPrinter =
133 getTarget().createMCInstPrinter(MAI.getAssemblerDialect(), MAI);
134 AsmStreamer.reset(createAsmStreamer(*Context, Out,
135 getTargetData()->isLittleEndian(),
136 getVerboseAsm(), InstPrinter,
140 case CGFT_ObjectFile: {
141 // Create the code emitter for the target if it exists. If not, .o file
143 MCCodeEmitter *MCE = getTarget().createCodeEmitter(*this, *Context);
144 TargetAsmBackend *TAB = getTarget().createAsmBackend(TargetTriple);
145 if (MCE == 0 || TAB == 0)
148 AsmStreamer.reset(createMachOStreamer(*Context, *TAB, Out, MCE));
152 // The Null output is intended for use for performance analysis and testing,
154 AsmStreamer.reset(createNullStreamer(*Context));
158 // Create the AsmPrinter, which takes ownership of AsmStreamer if successful.
159 FunctionPass *Printer = getTarget().createAsmPrinter(*this, *AsmStreamer);
163 // If successful, createAsmPrinter took ownership of AsmStreamer.
168 // Make sure the code model is set.
169 setCodeModelForStatic();
170 PM.add(createGCInfoDeleter());
174 /// addPassesToEmitMachineCode - Add passes to the specified pass manager to
175 /// get machine code emitted. This uses a JITCodeEmitter object to handle
176 /// actually outputting the machine code and resolving things like the address
177 /// of functions. This method should returns true if machine code emission is
180 bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
182 CodeGenOpt::Level OptLevel,
183 bool DisableVerify) {
184 // Make sure the code model is set.
185 setCodeModelForJIT();
187 // Add common CodeGen passes.
189 if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify, Ctx))
192 addCodeEmitter(PM, OptLevel, JCE);
193 PM.add(createGCInfoDeleter());
195 return false; // success!
198 static void printNoVerify(PassManagerBase &PM, const char *Banner) {
199 if (PrintMachineCode)
200 PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
203 static void printAndVerify(PassManagerBase &PM,
205 bool allowDoubleDefs = false) {
206 if (PrintMachineCode)
207 PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
209 if (VerifyMachineCode)
210 PM.add(createMachineVerifierPass(allowDoubleDefs));
213 /// addCommonCodeGenPasses - Add standard LLVM codegen passes used for both
214 /// emitting to assembly files or machine code output.
216 bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
217 CodeGenOpt::Level OptLevel,
219 MCContext *&OutContext) {
220 // Standard LLVM-Level Passes.
222 // Before running any passes, run the verifier to determine if the input
223 // coming from the front-end and/or optimizer is valid.
225 PM.add(createVerifierPass());
227 // Optionally, tun split-GEPs and no-load GVN.
228 if (EnableSplitGEPGVN) {
229 PM.add(createGEPSplitterPass());
230 PM.add(createGVNPass(/*NoLoads=*/true));
233 // Run loop strength reduction before anything else.
234 if (OptLevel != CodeGenOpt::None && !DisableLSR) {
235 PM.add(createLoopStrengthReducePass(getTargetLowering()));
237 PM.add(createPrintFunctionPass("\n\n*** Code after LSR ***\n", &dbgs()));
240 // Turn exception handling constructs into something the code generators can
242 switch (getMCAsmInfo()->getExceptionHandlingType()) {
243 case ExceptionHandling::SjLj:
244 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both
245 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise,
246 // catch info can get misplaced when a selector ends up more than one block
247 // removed from the parent invoke(s). This could happen when a landing
248 // pad is shared by multiple invokes and is also a target of a normal
249 // edge from elsewhere.
250 PM.add(createSjLjEHPass(getTargetLowering()));
251 PM.add(createDwarfEHPass(this, OptLevel==CodeGenOpt::None));
253 case ExceptionHandling::Dwarf:
254 PM.add(createDwarfEHPass(this, OptLevel==CodeGenOpt::None));
256 case ExceptionHandling::None:
257 PM.add(createLowerInvokePass(getTargetLowering()));
261 PM.add(createGCLoweringPass());
263 // Make sure that no unreachable blocks are instruction selected.
264 PM.add(createUnreachableBlockEliminationPass());
266 if (OptLevel != CodeGenOpt::None && !DisableCGP)
267 PM.add(createCodeGenPreparePass(getTargetLowering()));
269 PM.add(createStackProtectorPass(getTargetLowering()));
272 PM.add(createPrintFunctionPass("\n\n"
273 "*** Final LLVM Code input to ISel ***\n",
276 // All passes which modify the LLVM IR are now complete; run the verifier
277 // to ensure that the IR is valid.
279 PM.add(createVerifierPass());
281 // Standard Lower-Level Passes.
283 // Install a MachineModuleInfo class, which is an immutable pass that holds
284 // all the per-module stuff we're generating, including MCContext.
285 MachineModuleInfo *MMI = new MachineModuleInfo(*getMCAsmInfo());
287 OutContext = &MMI->getContext(); // Return the MCContext specifically by-ref.
290 // Set up a MachineFunction for the rest of CodeGen to work on.
291 PM.add(new MachineFunctionAnalysis(*this, OptLevel));
293 // Enable FastISel with -fast, but allow that to be overridden.
294 if (EnableFastISelOption == cl::BOU_TRUE ||
295 (OptLevel == CodeGenOpt::None && EnableFastISelOption != cl::BOU_FALSE))
296 EnableFastISel = true;
298 // Ask the target for an isel.
299 if (addInstSelector(PM, OptLevel))
302 // Print the instruction selected machine code...
303 printAndVerify(PM, "After Instruction Selection",
304 /* allowDoubleDefs= */ true);
306 // Optimize PHIs before DCE: removing dead PHI cycles may make more
307 // instructions dead.
308 if (OptLevel != CodeGenOpt::None)
309 PM.add(createOptimizePHIsPass());
311 // Delete dead machine instructions regardless of optimization level.
312 PM.add(createDeadMachineInstructionElimPass());
313 printAndVerify(PM, "After codegen DCE pass",
314 /* allowDoubleDefs= */ true);
316 if (OptLevel != CodeGenOpt::None) {
317 PM.add(createOptimizeExtsPass());
318 if (!DisableMachineLICM)
319 PM.add(createMachineLICMPass());
320 PM.add(createMachineCSEPass());
321 if (!DisableMachineSink)
322 PM.add(createMachineSinkingPass());
323 printAndVerify(PM, "After Machine LICM, CSE and Sinking passes",
324 /* allowDoubleDefs= */ true);
327 // Pre-ra tail duplication.
328 if (OptLevel != CodeGenOpt::None && !DisableEarlyTailDup) {
329 PM.add(createTailDuplicatePass(true));
330 printAndVerify(PM, "After Pre-RegAlloc TailDuplicate",
331 /* allowDoubleDefs= */ true);
334 // Run pre-ra passes.
335 if (addPreRegAlloc(PM, OptLevel))
336 printAndVerify(PM, "After PreRegAlloc passes",
337 /* allowDoubleDefs= */ true);
339 // Perform register allocation.
340 PM.add(createRegisterAllocator());
341 printAndVerify(PM, "After Register Allocation");
343 // Perform stack slot coloring and post-ra machine LICM.
344 if (OptLevel != CodeGenOpt::None) {
345 // FIXME: Re-enable coloring with register when it's capable of adding
348 PM.add(createStackSlotColoringPass(false));
350 // Run post-ra machine LICM to hoist reloads / remats.
351 if (!DisablePostRAMachineLICM)
352 PM.add(createMachineLICMPass(false));
354 printAndVerify(PM, "After StackSlotColoring and postra Machine LICM");
357 // Run post-ra passes.
358 if (addPostRegAlloc(PM, OptLevel))
359 printAndVerify(PM, "After PostRegAlloc passes");
361 PM.add(createLowerSubregsPass());
362 printAndVerify(PM, "After LowerSubregs");
364 // Insert prolog/epilog code. Eliminate abstract frame index references...
365 PM.add(createPrologEpilogCodeInserter());
366 printAndVerify(PM, "After PrologEpilogCodeInserter");
368 // Run pre-sched2 passes.
369 if (addPreSched2(PM, OptLevel))
370 printAndVerify(PM, "After PreSched2 passes");
372 // Second pass scheduler.
373 if (OptLevel != CodeGenOpt::None && !DisablePostRA) {
374 PM.add(createPostRAScheduler(OptLevel));
375 printAndVerify(PM, "After PostRAScheduler");
378 // Branch folding must be run after regalloc and prolog/epilog insertion.
379 if (OptLevel != CodeGenOpt::None && !DisableBranchFold) {
380 PM.add(createBranchFoldingPass(getEnableTailMergeDefault()));
381 printNoVerify(PM, "After BranchFolding");
385 if (OptLevel != CodeGenOpt::None && !DisableTailDuplicate) {
386 PM.add(createTailDuplicatePass(false));
387 printNoVerify(PM, "After TailDuplicate");
390 PM.add(createGCMachineCodeAnalysisPass());
393 PM.add(createGCInfoPrinter(dbgs()));
395 if (OptLevel != CodeGenOpt::None && !DisableCodePlace) {
396 PM.add(createCodePlacementOptPass());
397 printNoVerify(PM, "After CodePlacementOpt");
400 if (addPreEmitPass(PM, OptLevel))
401 printNoVerify(PM, "After PreEmit passes");