lto.cpp revision 0701a2f70df66134ead84d4fe86b20b8f28c4fc3
1//===-lto.cpp - LLVM Link Time Optimizer ----------------------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by Devang Patel and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implementes link time optimization library. This library is
11// intended to be used by linker to optimize code at link time.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/Module.h"
16#include "llvm/PassManager.h"
17#include "llvm/Linker.h"
18#include "llvm/Constants.h"
19#include "llvm/DerivedTypes.h"
20#include "llvm/SymbolTable.h"
21#include "llvm/Bytecode/Reader.h"
22#include "llvm/Bytecode/Writer.h"
23#include "llvm/Support/CommandLine.h"
24#include "llvm/Support/FileUtilities.h"
25#include "llvm/Support/SystemUtils.h"
26#include "llvm/Support/Mangler.h"
27#include "llvm/System/Program.h"
28#include "llvm/System/Signals.h"
29#include "llvm/Analysis/Passes.h"
30#include "llvm/Analysis/Verifier.h"
31#include "llvm/Target/SubtargetFeature.h"
32#include "llvm/Target/TargetData.h"
33#include "llvm/Target/TargetMachine.h"
34#include "llvm/Target/TargetMachineRegistry.h"
35#include "llvm/Transforms/IPO.h"
36#include "llvm/Transforms/Scalar.h"
37#include "llvm/Analysis/LoadValueNumbering.h"
38#include "llvm/LinkTimeOptimizer.h"
39#include <fstream>
40#include <iostream>
41
42using namespace llvm;
43
44extern "C"
45llvm::LinkTimeOptimizer *createLLVMOptimizer()
46{
47  llvm::LinkTimeOptimizer *l = new llvm::LinkTimeOptimizer();
48  return l;
49}
50
51
52
53/// If symbol is not used then make it internal and let optimizer takes
54/// care of it.
55void LLVMSymbol::mayBeNotUsed() {
56  gv->setLinkage(GlobalValue::InternalLinkage);
57}
58
59// Helper routine
60// FIXME : Take advantage of GlobalPrefix from AsmPrinter
61static const char *addUnderscore(const char *name) {
62  size_t namelen = strlen(name);
63  char *symName = (char*)malloc(namelen+2);
64  symName[0] = '_';
65  strcpy(&symName[1], name);
66  return symName;
67}
68
69// Map LLVM LinkageType to LTO LinakgeType
70static LTOLinkageTypes
71getLTOLinkageType(GlobalValue *v)
72{
73  LTOLinkageTypes lt;
74  if (v->hasExternalLinkage())
75    lt = LTOExternalLinkage;
76  else if (v->hasLinkOnceLinkage())
77    lt = LTOLinkOnceLinkage;
78  else if (v->hasWeakLinkage())
79    lt = LTOWeakLinkage;
80  else
81    // Otherwise it is internal linkage for link time optimizer
82    lt = LTOInternalLinkage;
83  return lt;
84}
85
86// Find exeternal symbols referenced by VALUE. This is a recursive function.
87static void
88findExternalRefs(Value *value, std::set<std::string> &references,
89                 Mangler &mangler) {
90
91  if (GlobalValue *gv = dyn_cast<GlobalValue>(value)) {
92    LTOLinkageTypes lt = getLTOLinkageType(gv);
93    if (lt != LTOInternalLinkage && strncmp (gv->getName().c_str(), "llvm.", 5))
94      references.insert(mangler.getValueName(gv));
95  }
96  else if (Constant *c = dyn_cast<Constant>(value))
97    // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
98    for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
99      findExternalRefs(c->getOperand(i), references, mangler);
100}
101
102/// InputFilename is a LLVM bytecode file. If Module with InputFilename is
103/// available then return it. Otherwise parseInputFilename.
104Module *
105LinkTimeOptimizer::getModule(const std::string &InputFilename)
106{
107  Module *m = NULL;
108
109  NameToModuleMap::iterator pos = allModules.find(InputFilename.c_str());
110  if (pos != allModules.end())
111    m = allModules[InputFilename.c_str()];
112  else {
113    m = ParseBytecodeFile(InputFilename);
114    allModules[InputFilename.c_str()] = m;
115  }
116  return m;
117}
118
119/// InputFilename is a LLVM bytecode file. Read it using bytecode reader.
120/// Collect global functions and symbol names in symbols vector.
121/// Collect external references in references vector.
122/// Return LTO_READ_SUCCESS if there is no error.
123enum LTOStatus
124LinkTimeOptimizer::readLLVMObjectFile(const std::string &InputFilename,
125                                      NameToSymbolMap &symbols,
126                                      std::set<std::string> &references)
127{
128  Module *m = getModule(InputFilename);
129  if (!m)
130    return LTO_READ_FAILURE;
131
132  // Use mangler to add GlobalPrefix to names to match linker names.
133  // FIXME : Instead of hard coding "-" use GlobalPrefix.
134  Mangler mangler(*m, "_");
135
136  modules.push_back(m);
137
138  for (Module::iterator f = m->begin(), e = m->end(); f != e; ++f) {
139
140    LTOLinkageTypes lt = getLTOLinkageType(f);
141
142    if (!f->isExternal() && lt != LTOInternalLinkage
143        && strncmp (f->getName().c_str(), "llvm.", 5)) {
144      LLVMSymbol *newSymbol = new LLVMSymbol(lt, f, f->getName(),
145                                             mangler.getValueName(f));
146      symbols[newSymbol->getMangledName()] = newSymbol;
147      allSymbols[newSymbol->getMangledName()] = newSymbol;
148    }
149
150    // Collect external symbols referenced by this function.
151    for (Function::iterator b = f->begin(), fe = f->end(); b != fe; ++b)
152      for (BasicBlock::iterator i = b->begin(), be = b->end();
153           i != be; ++i)
154        for (unsigned count = 0, total = i->getNumOperands();
155             count != total; ++count)
156          findExternalRefs(i->getOperand(count), references, mangler);
157  }
158
159  for (Module::global_iterator v = m->global_begin(), e = m->global_end();
160       v !=  e; ++v) {
161    LTOLinkageTypes lt = getLTOLinkageType(v);
162    if (!v->isExternal() && lt != LTOInternalLinkage
163        && strncmp (v->getName().c_str(), "llvm.", 5)) {
164      LLVMSymbol *newSymbol = new LLVMSymbol(lt, v, v->getName(),
165                                             mangler.getValueName(v));
166      symbols[newSymbol->getMangledName()] = newSymbol;
167      allSymbols[newSymbol->getMangledName()] = newSymbol;
168
169      for (unsigned count = 0, total = v->getNumOperands();
170           count != total; ++count)
171        findExternalRefs(v->getOperand(count), references, mangler);
172
173    }
174  }
175
176  return LTO_READ_SUCCESS;
177}
178
179/// Optimize module M using various IPO passes. Use exportList to
180/// internalize selected symbols. Target platform is selected
181/// based on information available to module M. No new target
182/// features are selected.
183static enum LTOStatus lto_optimize(Module *M, std::ostream &Out,
184                                   std::vector<const char *> &exportList)
185{
186  // Instantiate the pass manager to organize the passes.
187  PassManager Passes;
188
189  // Collect Target info
190  std::string Err;
191  const TargetMachineRegistry::Entry* March =
192    TargetMachineRegistry::getClosestStaticTargetForModule(*M, Err);
193
194  if (March == 0)
195    return LTO_NO_TARGET;
196
197  // Create target
198  std::string Features;
199  std::auto_ptr<TargetMachine> target(March->CtorFn(*M, Features));
200  if (!target.get())
201    return LTO_NO_TARGET;
202
203  TargetMachine &Target = *target.get();
204
205  // Start off with a verification pass.
206  Passes.add(createVerifierPass());
207
208  // Add an appropriate TargetData instance for this module...
209  Passes.add(new TargetData(*Target.getTargetData()));
210
211  // Often if the programmer does not specify proper prototypes for the
212  // functions they are calling, they end up calling a vararg version of the
213  // function that does not get a body filled in (the real function has typed
214  // arguments).  This pass merges the two functions.
215  Passes.add(createFunctionResolvingPass());
216
217  // Internalize symbols if export list is nonemty
218  if (!exportList.empty())
219    Passes.add(createInternalizePass(exportList));
220
221  // Now that we internalized some globals, see if we can hack on them!
222  Passes.add(createGlobalOptimizerPass());
223
224  // Linking modules together can lead to duplicated global constants, only
225  // keep one copy of each constant...
226  Passes.add(createConstantMergePass());
227
228  // If the -s command line option was specified, strip the symbols out of the
229  // resulting program to make it smaller.  -s is a GLD option that we are
230  // supporting.
231  Passes.add(createStripSymbolsPass());
232
233  // Propagate constants at call sites into the functions they call.
234  Passes.add(createIPConstantPropagationPass());
235
236  // Remove unused arguments from functions...
237  Passes.add(createDeadArgEliminationPass());
238
239  Passes.add(createFunctionInliningPass()); // Inline small functions
240
241  Passes.add(createPruneEHPass());            // Remove dead EH info
242
243  Passes.add(createGlobalDCEPass());          // Remove dead functions
244
245  // If we didn't decide to inline a function, check to see if we can
246  // transform it to pass arguments by value instead of by reference.
247  Passes.add(createArgumentPromotionPass());
248
249  // The IPO passes may leave cruft around.  Clean up after them.
250  Passes.add(createInstructionCombiningPass());
251
252  Passes.add(createScalarReplAggregatesPass()); // Break up allocas
253
254  // Run a few AA driven optimizations here and now, to cleanup the code.
255  Passes.add(createGlobalsModRefPass());      // IP alias analysis
256
257  Passes.add(createLICMPass());               // Hoist loop invariants
258  Passes.add(createLoadValueNumberingPass()); // GVN for load instrs
259  Passes.add(createGCSEPass());               // Remove common subexprs
260  Passes.add(createDeadStoreEliminationPass()); // Nuke dead stores
261
262  // Cleanup and simplify the code after the scalar optimizations.
263  Passes.add(createInstructionCombiningPass());
264
265  // Delete basic blocks, which optimization passes may have killed...
266  Passes.add(createCFGSimplificationPass());
267
268  // Now that we have optimized the program, discard unreachable functions...
269  Passes.add(createGlobalDCEPass());
270
271  // Make sure everything is still good.
272  Passes.add(createVerifierPass());
273
274  Target.addPassesToEmitFile(Passes, Out, TargetMachine::AssemblyFile, true);
275
276  // Run our queue of passes all at once now, efficiently.
277  Passes.run(*M);
278
279  return LTO_OPT_SUCCESS;
280}
281
282///Link all modules together and optimize them using IPO. Generate
283/// native object file using OutputFilename
284/// Return appropriate LTOStatus.
285enum LTOStatus
286LinkTimeOptimizer::optimizeModules(const std::string &OutputFilename,
287                                   std::vector<const char *> &exportList,
288                                   std::string &targetTriple)
289{
290  if (modules.empty())
291    return LTO_NO_WORK;
292
293  std::ios::openmode io_mode =
294    std::ios::out | std::ios::trunc | std::ios::binary;
295  std::string *errMsg = NULL;
296  Module *bigOne = modules[0];
297  Linker theLinker("LinkTimeOptimizer", bigOne, false);
298  for (unsigned i = 1, e = modules.size(); i != e; ++i)
299    if (theLinker.LinkModules(bigOne, modules[i], errMsg))
300      return LTO_MODULE_MERGE_FAILURE;
301
302#if 0
303  // Enable this when -save-temps is used
304  std::ofstream Out("big.bc", io_mode);
305  WriteBytecodeToFile(bigOne, Out, true);
306#endif
307
308  // Strip leading underscore because it was added to match names
309  // seen by linker.
310  for (unsigned i = 0, e = exportList.size(); i != e; ++i) {
311    const char *name = exportList[i];
312    NameToSymbolMap::iterator itr = allSymbols.find(name);
313    if (itr != allSymbols.end())
314      exportList[i] = allSymbols[name]->getName();
315  }
316
317  sys::Path tmpAsmFilePath("/tmp/");
318  std::string ErrMsg;
319  if (tmpAsmFilePath.createTemporaryFileOnDisk(&ErrMsg)) {
320    std::cerr << "lto: " << ErrMsg << "\n";
321    return LTO_WRITE_FAILURE;
322  }
323  sys::RemoveFileOnSignal(tmpAsmFilePath);
324
325  std::ofstream asmFile(tmpAsmFilePath.c_str(), io_mode);
326  if (!asmFile.is_open() || asmFile.bad()) {
327    if (tmpAsmFilePath.exists())
328      tmpAsmFilePath.eraseFromDisk();
329    return LTO_WRITE_FAILURE;
330  }
331
332  enum LTOStatus status = lto_optimize(bigOne, asmFile, exportList);
333  asmFile.close();
334  if (status != LTO_OPT_SUCCESS) {
335    tmpAsmFilePath.eraseFromDisk();
336    return status;
337  }
338
339  targetTriple = bigOne->getTargetTriple();
340
341  // Run GCC to assemble and link the program into native code.
342  //
343  // Note:
344  //  We can't just assemble and link the file with the system assembler
345  //  and linker because we don't know where to put the _start symbol.
346  //  GCC mysteriously knows how to do it.
347  const sys::Path gcc = FindExecutable("gcc", "/");
348  if (gcc.isEmpty()) {
349    tmpAsmFilePath.eraseFromDisk();
350    return LTO_ASM_FAILURE;
351  }
352
353  std::vector<const char*> args;
354  args.push_back(gcc.c_str());
355  args.push_back("-c");
356  args.push_back("-x");
357  args.push_back("assembler");
358  args.push_back("-o");
359  args.push_back(OutputFilename.c_str());
360  args.push_back(tmpAsmFilePath.c_str());
361  args.push_back(0);
362
363  sys::Program::ExecuteAndWait(gcc, &args[0], 0, 0, 1);
364
365  tmpAsmFilePath.eraseFromDisk();
366
367  return LTO_OPT_SUCCESS;
368}
369