Compiler.cpp revision 8e9089377848628813a697b972773e969b942c3b 
1/*
2 * Copyright 2010-2012, The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *     http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#include "bcc/Compiler.h"
18
19#include <llvm/Analysis/Passes.h>
20#include <llvm/Analysis/TargetTransformInfo.h>
21#include <llvm/CodeGen/RegAllocRegistry.h>
22#include <llvm/IR/LegacyPassManager.h>
23#include <llvm/IR/Module.h>
24#include <llvm/Support/TargetRegistry.h>
25#include <llvm/Support/raw_ostream.h>
26#include <llvm/IR/DataLayout.h>
27#include <llvm/Target/TargetSubtargetInfo.h>
28#include <llvm/Target/TargetMachine.h>
29#include <llvm/Transforms/IPO.h>
30#include <llvm/Transforms/IPO/PassManagerBuilder.h>
31#include <llvm/Transforms/Scalar.h>
32#include <llvm/Transforms/Vectorize.h>
33
34#include "bcc/Assert.h"
35#include "bcc/Renderscript/RSScript.h"
36#include "bcc/Renderscript/RSTransforms.h"
37#include "bcc/Script.h"
38#include "bcc/Source.h"
39#include "bcc/Support/CompilerConfig.h"
40#include "bcc/Support/Log.h"
41#include "bcc/Support/OutputFile.h"
42#include "bcinfo/MetadataExtractor.h"
43#include "rsDefines.h"
44
45#include <string>
46
47using namespace bcc;
48
49const char *Compiler::GetErrorString(enum ErrorCode pErrCode) {
50  switch (pErrCode) {
51  case kSuccess:
52    return "Successfully compiled.";
53  case kInvalidConfigNoTarget:
54    return "Invalid compiler config supplied (getTarget() returns nullptr.) "
55           "(missing call to CompilerConfig::initialize()?)";
56  case kErrCreateTargetMachine:
57    return "Failed to create llvm::TargetMachine.";
58  case kErrSwitchTargetMachine:
59    return  "Failed to switch llvm::TargetMachine.";
60  case kErrNoTargetMachine:
61    return "Failed to compile the script since there's no available "
62           "TargetMachine. (missing call to Compiler::config()?)";
63  case kErrMaterialization:
64    return "Failed to materialize the module.";
65  case kErrInvalidOutputFileState:
66    return "Supplied output file was invalid (in the error state.)";
67  case kErrPrepareOutput:
68    return "Failed to prepare file for output.";
69  case kPrepareCodeGenPass:
70    return "Failed to construct pass list for code-generation.";
71  case kErrCustomPasses:
72    return "Error occurred while adding custom passes.";
73  case kErrInvalidSource:
74    return "Error loading input bitcode";
75  case kIllegalGlobalFunction:
76    return "Use of undefined external function";
77  }
78
79  // This assert should never be reached as the compiler verifies that the
80  // above switch coveres all enum values.
81  bccAssert(false && "Unknown error code encountered");
82  return  "";
83}
84
85//===----------------------------------------------------------------------===//
86// Instance Methods
87//===----------------------------------------------------------------------===//
88Compiler::Compiler() : mTarget(nullptr), mEnableOpt(true) {
89  return;
90}
91
92Compiler::Compiler(const CompilerConfig &pConfig) : mTarget(nullptr),
93                                                    mEnableOpt(true) {
94  const std::string &triple = pConfig.getTriple();
95
96  enum ErrorCode err = config(pConfig);
97  if (err != kSuccess) {
98    ALOGE("%s (%s, features: %s)", GetErrorString(err),
99          triple.c_str(), pConfig.getFeatureString().c_str());
100    return;
101  }
102
103  return;
104}
105
106enum Compiler::ErrorCode Compiler::config(const CompilerConfig &pConfig) {
107  if (pConfig.getTarget() == nullptr) {
108    return kInvalidConfigNoTarget;
109  }
110
111  llvm::TargetMachine *new_target =
112      (pConfig.getTarget())->createTargetMachine(pConfig.getTriple(),
113                                                 pConfig.getCPU(),
114                                                 pConfig.getFeatureString(),
115                                                 pConfig.getTargetOptions(),
116                                                 pConfig.getRelocationModel(),
117                                                 pConfig.getCodeModel(),
118                                                 pConfig.getOptimizationLevel());
119
120  if (new_target == nullptr) {
121    return ((mTarget != nullptr) ? kErrSwitchTargetMachine :
122                                   kErrCreateTargetMachine);
123  }
124
125  // Replace the old TargetMachine.
126  delete mTarget;
127  mTarget = new_target;
128
129  // Adjust register allocation policy according to the optimization level.
130  //  createFastRegisterAllocator: fast but bad quality
131  //  createLinearScanRegisterAllocator: not so fast but good quality
132  if ((pConfig.getOptimizationLevel() == llvm::CodeGenOpt::None)) {
133    llvm::RegisterRegAlloc::setDefault(llvm::createFastRegisterAllocator);
134  } else {
135    llvm::RegisterRegAlloc::setDefault(llvm::createGreedyRegisterAllocator);
136  }
137
138  return kSuccess;
139}
140
141Compiler::~Compiler() {
142  delete mTarget;
143}
144
145
146// This function has complete responsibility for creating and executing the
147// exact list of compiler passes.
148enum Compiler::ErrorCode Compiler::runPasses(Script &pScript,
149                                             llvm::raw_pwrite_stream &pResult) {
150  // Pass manager for link-time optimization
151  llvm::legacy::PassManager transformPasses;
152
153  // Empty MCContext.
154  llvm::MCContext *mc_context = nullptr;
155
156  transformPasses.add(
157      createTargetTransformInfoWrapperPass(mTarget->getTargetIRAnalysis()));
158
159  // Add some initial custom passes.
160  addInvokeHelperPass(transformPasses);
161  addExpandKernelPass(transformPasses);
162  addDebugInfoPass(pScript, transformPasses);
163  addInvariantPass(transformPasses);
164  if (mTarget->getOptLevel() != llvm::CodeGenOpt::None) {
165    if (!addInternalizeSymbolsPass(pScript, transformPasses))
166      return kErrCustomPasses;
167  }
168  addGlobalInfoPass(pScript, transformPasses);
169
170  if (mTarget->getOptLevel() == llvm::CodeGenOpt::None) {
171    transformPasses.add(llvm::createGlobalOptimizerPass());
172    transformPasses.add(llvm::createConstantMergePass());
173
174  } else {
175    // FIXME: Figure out which passes should be executed.
176    llvm::PassManagerBuilder Builder;
177    Builder.Inliner = llvm::createFunctionInliningPass();
178    Builder.populateLTOPassManager(transformPasses);
179
180    /* FIXME: Reenable autovectorization after rebase.
181       bug 19324423
182    // Add vectorization passes after LTO passes are in
183    // additional flag: -unroll-runtime
184    transformPasses.add(llvm::createLoopUnrollPass(-1, 16, 0, 1));
185    // Need to pass appropriate flags here: -scalarize-load-store
186    transformPasses.add(llvm::createScalarizerPass());
187    transformPasses.add(llvm::createCFGSimplificationPass());
188    transformPasses.add(llvm::createScopedNoAliasAAPass());
189    transformPasses.add(llvm::createScalarEvolutionAliasAnalysisPass());
190    // additional flags: -slp-vectorize-hor -slp-vectorize-hor-store (unnecessary?)
191    transformPasses.add(llvm::createSLPVectorizerPass());
192    transformPasses.add(llvm::createDeadCodeEliminationPass());
193    transformPasses.add(llvm::createInstructionCombiningPass());
194    */
195  }
196
197  // These passes have to come after LTO, since we don't want to examine
198  // functions that are never actually called.
199  if (llvm::Triple(getTargetMachine().getTargetTriple()).getArch() == llvm::Triple::x86_64)
200    transformPasses.add(createRSX86_64CallConvPass());  // Add pass to correct calling convention for X86-64.
201  transformPasses.add(createRSIsThreadablePass());      // Add pass to mark script as threadable.
202
203  // RSEmbedInfoPass needs to come after we have scanned for non-threadable
204  // functions.
205  // Script passed to RSCompiler must be a RSScript.
206  RSScript &script = static_cast<RSScript &>(pScript);
207  if (script.getEmbedInfo())
208    transformPasses.add(createRSEmbedInfoPass());
209
210  // Execute the passes.
211  transformPasses.run(pScript.getSource().getModule());
212
213  // Run backend separately to avoid interference between debug metadata
214  // generation and backend initialization.
215  llvm::legacy::PassManager codeGenPasses;
216
217  // Add passes to the pass manager to emit machine code through MC layer.
218  if (mTarget->addPassesToEmitMC(codeGenPasses, mc_context, pResult,
219                                 /* DisableVerify */false)) {
220    return kPrepareCodeGenPass;
221  }
222
223  // Execute the passes.
224  codeGenPasses.run(pScript.getSource().getModule());
225
226  return kSuccess;
227}
228
229enum Compiler::ErrorCode Compiler::compile(Script &pScript,
230                                           llvm::raw_pwrite_stream &pResult,
231                                           llvm::raw_ostream *IRStream) {
232  llvm::Module &module = pScript.getSource().getModule();
233  enum ErrorCode err;
234
235  if (mTarget == nullptr) {
236    return kErrNoTargetMachine;
237  }
238
239  const std::string &triple = module.getTargetTriple();
240  const llvm::DataLayout dl = getTargetMachine().createDataLayout();
241  unsigned int pointerSize = dl.getPointerSizeInBits();
242  if (triple == "armv7-none-linux-gnueabi") {
243    if (pointerSize != 32) {
244      return kErrInvalidSource;
245    }
246  } else if (triple == "aarch64-none-linux-gnueabi") {
247    if (pointerSize != 64) {
248      return kErrInvalidSource;
249    }
250  } else {
251    return kErrInvalidSource;
252  }
253
254  // Sanitize module's target information.
255  module.setTargetTriple(getTargetMachine().getTargetTriple().str());
256  module.setDataLayout(getTargetMachine().createDataLayout());
257
258  // Materialize the bitcode module.
259  if (module.getMaterializer() != nullptr) {
260    // A module with non-null materializer means that it is a lazy-load module.
261    // Materialize it now.  This function returns false when the materialization
262    // is successful.
263    std::error_code ec = module.materializeAll();
264    if (ec) {
265      ALOGE("Failed to materialize the module `%s'! (%s)",
266            module.getModuleIdentifier().c_str(), ec.message().c_str());
267      return kErrMaterialization;
268    }
269  }
270
271  if ((err = runPasses(pScript, pResult)) != kSuccess) {
272    return err;
273  }
274
275  if (IRStream) {
276    *IRStream << module;
277  }
278
279  return kSuccess;
280}
281
282enum Compiler::ErrorCode Compiler::compile(Script &pScript,
283                                           OutputFile &pResult,
284                                           llvm::raw_ostream *IRStream) {
285  // Check the state of the specified output file.
286  if (pResult.hasError()) {
287    return kErrInvalidOutputFileState;
288  }
289
290  // Open the output file decorated in llvm::raw_ostream.
291  llvm::raw_pwrite_stream *out = pResult.dup();
292  if (out == nullptr) {
293    return kErrPrepareOutput;
294  }
295
296  // Delegate the request.
297  enum Compiler::ErrorCode err = compile(pScript, *out, IRStream);
298
299  // Close the output before return.
300  delete out;
301
302  return err;
303}
304
305bool Compiler::addInternalizeSymbolsPass(Script &pScript, llvm::legacy::PassManager &pPM) {
306  // Add a pass to internalize the symbols that don't need to have global
307  // visibility.
308  RSScript &script = static_cast<RSScript &>(pScript);
309  llvm::Module &module = script.getSource().getModule();
310  bcinfo::MetadataExtractor me(&module);
311  if (!me.extract()) {
312    bccAssert(false && "Could not extract metadata for module!");
313    return false;
314  }
315
316  // The vector contains the symbols that should not be internalized.
317  std::vector<const char *> export_symbols;
318
319  const char *sf[] = {
320    kRoot,               // Graphics drawing function or compute kernel.
321    kInit,               // Initialization routine called implicitly on startup.
322    kRsDtor,             // Static global destructor for a script instance.
323    kRsInfo,             // Variable containing string of RS metadata info.
324    kRsGlobalEntries,    // Optional number of global variables.
325    kRsGlobalNames,      // Optional global variable name info.
326    kRsGlobalAddresses,  // Optional global variable address info.
327    kRsGlobalSizes,      // Optional global variable size info.
328    kRsGlobalProperties, // Optional global variable properties.
329    nullptr              // Must be nullptr-terminated.
330  };
331  const char **special_functions = sf;
332  // Special RS functions should always be global symbols.
333  while (*special_functions != nullptr) {
334    export_symbols.push_back(*special_functions);
335    special_functions++;
336  }
337
338  // Visibility of symbols appeared in rs_export_var and rs_export_func should
339  // also be preserved.
340  size_t exportVarCount = me.getExportVarCount();
341  size_t exportFuncCount = me.getExportFuncCount();
342  size_t exportForEachCount = me.getExportForEachSignatureCount();
343  size_t exportReduceCount = me.getExportReduceCount();
344  size_t exportReduceNewCount = me.getExportReduceNewCount();
345  const char **exportVarNameList = me.getExportVarNameList();
346  const char **exportFuncNameList = me.getExportFuncNameList();
347  const char **exportForEachNameList = me.getExportForEachNameList();
348  const char **exportReduceNameList = me.getExportReduceNameList();
349  const bcinfo::MetadataExtractor::ReduceNew *exportReduceNewList = me.getExportReduceNewList();
350  size_t i;
351
352  for (i = 0; i < exportVarCount; ++i) {
353    export_symbols.push_back(exportVarNameList[i]);
354  }
355
356  for (i = 0; i < exportFuncCount; ++i) {
357    export_symbols.push_back(exportFuncNameList[i]);
358  }
359
360  // Expanded foreach and reduce functions should not be internalized;
361  // nor should general reduction initializer and outconverter
362  // functions. keep_funcs keeps the names of these functions around
363  // until createInternalizePass() is finished making its own copy of
364  // the visible symbols.
365  std::vector<std::string> keep_funcs;
366  keep_funcs.reserve(exportForEachCount + exportReduceCount + exportReduceNewCount*3);
367
368  for (i = 0; i < exportForEachCount; ++i) {
369    keep_funcs.push_back(std::string(exportForEachNameList[i]) + ".expand");
370  }
371  for (i = 0; i < exportReduceCount; ++i) {
372    keep_funcs.push_back(std::string(exportReduceNameList[i]) + ".expand");
373  }
374  auto keepFuncsPushBackIfPresent = [&keep_funcs](const char *Name) {
375    if (Name) keep_funcs.push_back(Name);
376  };
377  for (i = 0; i < exportReduceNewCount; ++i) {
378    keep_funcs.push_back(std::string(exportReduceNewList[i].mAccumulatorName) + ".expand");
379    // Note: driver does not currently use the combiner function
380    keepFuncsPushBackIfPresent(exportReduceNewList[i].mInitializerName);
381    keepFuncsPushBackIfPresent(exportReduceNewList[i].mOutConverterName);
382  }
383
384  for (auto &symbol_name : keep_funcs) {
385    export_symbols.push_back(symbol_name.c_str());
386  }
387
388  // http://b/26165616 - WAR for this bug defines the __truncxfhf2 function in
389  // frameworks/rs/driver/runtime.  Don't internalize this function for x86, so
390  // that a script can find and link against it.
391  llvm::Triple triple(getTargetMachine().getTargetTriple());
392  if (triple.getArch() == llvm::Triple::x86) {
393    export_symbols.push_back("__truncxfhf2");
394  }
395
396  pPM.add(llvm::createInternalizePass(export_symbols));
397
398  return true;
399}
400
401void Compiler::addInvokeHelperPass(llvm::legacy::PassManager &pPM) {
402  llvm::Triple arch(getTargetMachine().getTargetTriple());
403  if (arch.isArch64Bit()) {
404    pPM.add(createRSInvokeHelperPass());
405  }
406}
407
408void Compiler::addDebugInfoPass(Script &pScript, llvm::legacy::PassManager &pPM) {
409  if (pScript.getSource().getDebugInfoEnabled())
410    pPM.add(createRSAddDebugInfoPass());
411}
412
413void Compiler::addExpandKernelPass(llvm::legacy::PassManager &pPM) {
414  // Expand ForEach and reduce on CPU path to reduce launch overhead.
415  bool pEnableStepOpt = true;
416  pPM.add(createRSKernelExpandPass(pEnableStepOpt));
417}
418
419void Compiler::addGlobalInfoPass(Script &pScript, llvm::legacy::PassManager &pPM) {
420  // Add additional information about RS global variables inside the Module.
421  RSScript &script = static_cast<RSScript &>(pScript);
422  if (script.getEmbedGlobalInfo()) {
423    pPM.add(createRSGlobalInfoPass(script.getEmbedGlobalInfoSkipConstant()));
424  }
425}
426
427void Compiler::addInvariantPass(llvm::legacy::PassManager &pPM) {
428  // Mark Loads from RsExpandKernelDriverInfo as "load.invariant".
429  // Should run after ExpandForEach and before inlining.
430  pPM.add(createRSInvariantPass());
431}
432
433enum Compiler::ErrorCode Compiler::screenGlobalFunctions(Script &pScript) {
434  llvm::Module &module = pScript.getSource().getModule();
435
436  // Materialize the bitcode module in case this is a lazy-load module.  Do not
437  // clear the materializer by calling materializeAllPermanently since the
438  // runtime library has not been merged into the module yet.
439  if (module.getMaterializer() != nullptr) {
440    std::error_code ec = module.materializeAll();
441    if (ec) {
442      ALOGE("Failed to materialize module `%s' when screening globals! (%s)",
443            module.getModuleIdentifier().c_str(), ec.message().c_str());
444      return kErrMaterialization;
445    }
446  }
447
448  // Add pass to check for illegal function calls.
449  llvm::legacy::PassManager pPM;
450  pPM.add(createRSScreenFunctionsPass());
451  pPM.run(module);
452
453  return kSuccess;
454
455}
456