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/Config/Config.h"
36 #include "bcc/Renderscript/RSScript.h"
37 #include "bcc/Renderscript/RSTransforms.h"
38 #include "bcc/Renderscript/RSUtils.h"
39 #include "bcc/Script.h"
40 #include "bcc/Source.h"
41 #include "bcc/Support/CompilerConfig.h"
42 #include "bcc/Support/Log.h"
43 #include "bcc/Support/OutputFile.h"
44 #include "bcinfo/MetadataExtractor.h"
45 #include "rsDefines.h"
46
47 #include <string>
48
49 using namespace bcc;
50
GetErrorString(enum ErrorCode pErrCode)51 const char *Compiler::GetErrorString(enum ErrorCode pErrCode) {
52 switch (pErrCode) {
53 case kSuccess:
54 return "Successfully compiled.";
55 case kInvalidConfigNoTarget:
56 return "Invalid compiler config supplied (getTarget() returns nullptr.) "
57 "(missing call to CompilerConfig::initialize()?)";
58 case kErrCreateTargetMachine:
59 return "Failed to create llvm::TargetMachine.";
60 case kErrSwitchTargetMachine:
61 return "Failed to switch llvm::TargetMachine.";
62 case kErrNoTargetMachine:
63 return "Failed to compile the script since there's no available "
64 "TargetMachine. (missing call to Compiler::config()?)";
65 case kErrMaterialization:
66 return "Failed to materialize the module.";
67 case kErrInvalidOutputFileState:
68 return "Supplied output file was invalid (in the error state.)";
69 case kErrPrepareOutput:
70 return "Failed to prepare file for output.";
71 case kPrepareCodeGenPass:
72 return "Failed to construct pass list for code-generation.";
73 case kErrCustomPasses:
74 return "Error occurred while adding custom passes.";
75 case kErrInvalidSource:
76 return "Error loading input bitcode";
77 case kIllegalGlobalFunction:
78 return "Use of undefined external function";
79 case kErrInvalidTargetMachine:
80 return "Invalid/unexpected llvm::TargetMachine.";
81 }
82
83 // This assert should never be reached as the compiler verifies that the
84 // above switch coveres all enum values.
85 bccAssert(false && "Unknown error code encountered");
86 return "";
87 }
88
89 //===----------------------------------------------------------------------===//
90 // Instance Methods
91 //===----------------------------------------------------------------------===//
Compiler()92 Compiler::Compiler() : mTarget(nullptr), mEnableOpt(true) {
93 return;
94 }
95
Compiler(const CompilerConfig & pConfig)96 Compiler::Compiler(const CompilerConfig &pConfig) : mTarget(nullptr),
97 mEnableOpt(true) {
98 const std::string &triple = pConfig.getTriple();
99
100 enum ErrorCode err = config(pConfig);
101 if (err != kSuccess) {
102 ALOGE("%s (%s, features: %s)", GetErrorString(err),
103 triple.c_str(), pConfig.getFeatureString().c_str());
104 return;
105 }
106
107 return;
108 }
109
config(const CompilerConfig & pConfig)110 enum Compiler::ErrorCode Compiler::config(const CompilerConfig &pConfig) {
111 if (pConfig.getTarget() == nullptr) {
112 return kInvalidConfigNoTarget;
113 }
114
115 llvm::TargetMachine *new_target =
116 (pConfig.getTarget())->createTargetMachine(pConfig.getTriple(),
117 pConfig.getCPU(),
118 pConfig.getFeatureString(),
119 pConfig.getTargetOptions(),
120 pConfig.getRelocationModel(),
121 pConfig.getCodeModel(),
122 pConfig.getOptimizationLevel());
123
124 if (new_target == nullptr) {
125 return ((mTarget != nullptr) ? kErrSwitchTargetMachine :
126 kErrCreateTargetMachine);
127 }
128
129 // Replace the old TargetMachine.
130 delete mTarget;
131 mTarget = new_target;
132
133 // Adjust register allocation policy according to the optimization level.
134 // createFastRegisterAllocator: fast but bad quality
135 // createLinearScanRegisterAllocator: not so fast but good quality
136 if ((pConfig.getOptimizationLevel() == llvm::CodeGenOpt::None)) {
137 llvm::RegisterRegAlloc::setDefault(llvm::createFastRegisterAllocator);
138 } else {
139 llvm::RegisterRegAlloc::setDefault(llvm::createGreedyRegisterAllocator);
140 }
141
142 return kSuccess;
143 }
144
~Compiler()145 Compiler::~Compiler() {
146 delete mTarget;
147 }
148
149
150 // This function has complete responsibility for creating and executing the
151 // exact list of compiler passes.
runPasses(Script & pScript,llvm::raw_pwrite_stream & pResult)152 enum Compiler::ErrorCode Compiler::runPasses(Script &pScript,
153 llvm::raw_pwrite_stream &pResult) {
154 // Pass manager for link-time optimization
155 llvm::legacy::PassManager transformPasses;
156
157 // Empty MCContext.
158 llvm::MCContext *mc_context = nullptr;
159
160 transformPasses.add(
161 createTargetTransformInfoWrapperPass(mTarget->getTargetIRAnalysis()));
162
163 // Add some initial custom passes.
164 addInvokeHelperPass(transformPasses);
165 addExpandKernelPass(transformPasses);
166 addDebugInfoPass(pScript, transformPasses);
167 addInvariantPass(transformPasses);
168 if (mTarget->getOptLevel() != llvm::CodeGenOpt::None) {
169 if (!addInternalizeSymbolsPass(pScript, transformPasses))
170 return kErrCustomPasses;
171 }
172 addGlobalInfoPass(pScript, transformPasses);
173
174 if (mTarget->getOptLevel() == llvm::CodeGenOpt::None) {
175 transformPasses.add(llvm::createGlobalOptimizerPass());
176 transformPasses.add(llvm::createConstantMergePass());
177
178 } else {
179 // FIXME: Figure out which passes should be executed.
180 llvm::PassManagerBuilder Builder;
181 Builder.Inliner = llvm::createFunctionInliningPass();
182 Builder.populateLTOPassManager(transformPasses);
183
184 /* FIXME: Reenable autovectorization after rebase.
185 bug 19324423
186 // Add vectorization passes after LTO passes are in
187 // additional flag: -unroll-runtime
188 transformPasses.add(llvm::createLoopUnrollPass(-1, 16, 0, 1));
189 // Need to pass appropriate flags here: -scalarize-load-store
190 transformPasses.add(llvm::createScalarizerPass());
191 transformPasses.add(llvm::createCFGSimplificationPass());
192 transformPasses.add(llvm::createScopedNoAliasAAPass());
193 transformPasses.add(llvm::createScalarEvolutionAliasAnalysisPass());
194 // additional flags: -slp-vectorize-hor -slp-vectorize-hor-store (unnecessary?)
195 transformPasses.add(llvm::createSLPVectorizerPass());
196 transformPasses.add(llvm::createDeadCodeEliminationPass());
197 transformPasses.add(llvm::createInstructionCombiningPass());
198 */
199 }
200
201 // These passes have to come after LTO, since we don't want to examine
202 // functions that are never actually called.
203 if (llvm::Triple(getTargetMachine().getTargetTriple()).getArch() == llvm::Triple::x86_64)
204 transformPasses.add(createRSX86_64CallConvPass()); // Add pass to correct calling convention for X86-64.
205 transformPasses.add(createRSIsThreadablePass()); // Add pass to mark script as threadable.
206
207 // RSEmbedInfoPass needs to come after we have scanned for non-threadable
208 // functions.
209 // Script passed to RSCompiler must be a RSScript.
210 RSScript &script = static_cast<RSScript &>(pScript);
211 if (script.getEmbedInfo())
212 transformPasses.add(createRSEmbedInfoPass());
213
214 // Execute the passes.
215 transformPasses.run(pScript.getSource().getModule());
216
217 // Run backend separately to avoid interference between debug metadata
218 // generation and backend initialization.
219 llvm::legacy::PassManager codeGenPasses;
220
221 // Add passes to the pass manager to emit machine code through MC layer.
222 if (mTarget->addPassesToEmitMC(codeGenPasses, mc_context, pResult,
223 /* DisableVerify */false)) {
224 return kPrepareCodeGenPass;
225 }
226
227 // Execute the passes.
228 codeGenPasses.run(pScript.getSource().getModule());
229
230 return kSuccess;
231 }
232
compile(Script & pScript,llvm::raw_pwrite_stream & pResult,llvm::raw_ostream * IRStream)233 enum Compiler::ErrorCode Compiler::compile(Script &pScript,
234 llvm::raw_pwrite_stream &pResult,
235 llvm::raw_ostream *IRStream) {
236 llvm::Module &module = pScript.getSource().getModule();
237 enum ErrorCode err;
238
239 if (mTarget == nullptr) {
240 return kErrNoTargetMachine;
241 }
242
243 const std::string &triple = module.getTargetTriple();
244 const llvm::DataLayout dl = getTargetMachine().createDataLayout();
245 unsigned int pointerSize = dl.getPointerSizeInBits();
246 if (triple == "armv7-none-linux-gnueabi") {
247 if (pointerSize != 32) {
248 return kErrInvalidSource;
249 }
250 } else if (triple == "aarch64-none-linux-gnueabi") {
251 if (pointerSize != 64) {
252 return kErrInvalidSource;
253 }
254 } else {
255 return kErrInvalidSource;
256 }
257
258 if (getTargetMachine().getTargetTriple().getArch() == llvm::Triple::x86) {
259 // Detect and fail if TargetMachine datalayout is different than what we
260 // expect. This is to detect changes in default target layout for x86 and
261 // update X86_CUSTOM_DL_STRING in include/bcc/Config/Config.h appropriately.
262 if (dl.getStringRepresentation().compare(X86_DEFAULT_DL_STRING) != 0) {
263 return kErrInvalidTargetMachine;
264 }
265 }
266
267 // Sanitize module's target information.
268 module.setTargetTriple(getTargetMachine().getTargetTriple().str());
269 module.setDataLayout(getTargetMachine().createDataLayout());
270
271 // Materialize the bitcode module.
272 if (module.getMaterializer() != nullptr) {
273 // A module with non-null materializer means that it is a lazy-load module.
274 // Materialize it now. This function returns false when the materialization
275 // is successful.
276 std::error_code ec = module.materializeAll();
277 if (ec) {
278 ALOGE("Failed to materialize the module `%s'! (%s)",
279 module.getModuleIdentifier().c_str(), ec.message().c_str());
280 return kErrMaterialization;
281 }
282 }
283
284 if ((err = runPasses(pScript, pResult)) != kSuccess) {
285 return err;
286 }
287
288 if (IRStream) {
289 *IRStream << module;
290 }
291
292 return kSuccess;
293 }
294
compile(Script & pScript,OutputFile & pResult,llvm::raw_ostream * IRStream)295 enum Compiler::ErrorCode Compiler::compile(Script &pScript,
296 OutputFile &pResult,
297 llvm::raw_ostream *IRStream) {
298 // Check the state of the specified output file.
299 if (pResult.hasError()) {
300 return kErrInvalidOutputFileState;
301 }
302
303 // Open the output file decorated in llvm::raw_ostream.
304 llvm::raw_pwrite_stream *out = pResult.dup();
305 if (out == nullptr) {
306 return kErrPrepareOutput;
307 }
308
309 // Delegate the request.
310 enum Compiler::ErrorCode err = compile(pScript, *out, IRStream);
311
312 // Close the output before return.
313 delete out;
314
315 return err;
316 }
317
addInternalizeSymbolsPass(Script & pScript,llvm::legacy::PassManager & pPM)318 bool Compiler::addInternalizeSymbolsPass(Script &pScript, llvm::legacy::PassManager &pPM) {
319 // Add a pass to internalize the symbols that don't need to have global
320 // visibility.
321 RSScript &script = static_cast<RSScript &>(pScript);
322 llvm::Module &module = script.getSource().getModule();
323 bcinfo::MetadataExtractor me(&module);
324 if (!me.extract()) {
325 bccAssert(false && "Could not extract metadata for module!");
326 return false;
327 }
328
329 // The vector contains the symbols that should not be internalized.
330 std::vector<const char *> export_symbols;
331
332 const char *sf[] = {
333 kRoot, // Graphics drawing function or compute kernel.
334 kInit, // Initialization routine called implicitly on startup.
335 kRsDtor, // Static global destructor for a script instance.
336 kRsInfo, // Variable containing string of RS metadata info.
337 kRsGlobalEntries, // Optional number of global variables.
338 kRsGlobalNames, // Optional global variable name info.
339 kRsGlobalAddresses, // Optional global variable address info.
340 kRsGlobalSizes, // Optional global variable size info.
341 kRsGlobalProperties, // Optional global variable properties.
342 nullptr // Must be nullptr-terminated.
343 };
344 const char **special_functions = sf;
345 // Special RS functions should always be global symbols.
346 while (*special_functions != nullptr) {
347 export_symbols.push_back(*special_functions);
348 special_functions++;
349 }
350
351 // Visibility of symbols appeared in rs_export_var and rs_export_func should
352 // also be preserved.
353 size_t exportVarCount = me.getExportVarCount();
354 size_t exportFuncCount = me.getExportFuncCount();
355 size_t exportForEachCount = me.getExportForEachSignatureCount();
356 size_t exportReduceCount = me.getExportReduceCount();
357 const char **exportVarNameList = me.getExportVarNameList();
358 const char **exportFuncNameList = me.getExportFuncNameList();
359 const char **exportForEachNameList = me.getExportForEachNameList();
360 const bcinfo::MetadataExtractor::Reduce *exportReduceList = me.getExportReduceList();
361 size_t i;
362
363 for (i = 0; i < exportVarCount; ++i) {
364 export_symbols.push_back(exportVarNameList[i]);
365 }
366
367 for (i = 0; i < exportFuncCount; ++i) {
368 export_symbols.push_back(exportFuncNameList[i]);
369 }
370
371 // Expanded foreach functions should not be internalized; nor should
372 // general reduction initializer, combiner, and outconverter
373 // functions. keep_funcs keeps the names of these functions around
374 // until createInternalizePass() is finished making its own copy of
375 // the visible symbols.
376 std::vector<std::string> keep_funcs;
377 keep_funcs.reserve(exportForEachCount + exportReduceCount*4);
378
379 for (i = 0; i < exportForEachCount; ++i) {
380 keep_funcs.push_back(std::string(exportForEachNameList[i]) + ".expand");
381 }
382 auto keepFuncsPushBackIfPresent = [&keep_funcs](const char *Name) {
383 if (Name) keep_funcs.push_back(Name);
384 };
385 for (i = 0; i < exportReduceCount; ++i) {
386 keep_funcs.push_back(std::string(exportReduceList[i].mAccumulatorName) + ".expand");
387 keepFuncsPushBackIfPresent(exportReduceList[i].mInitializerName);
388 if (exportReduceList[i].mCombinerName != nullptr) {
389 keep_funcs.push_back(exportReduceList[i].mCombinerName);
390 } else {
391 keep_funcs.push_back(nameReduceCombinerFromAccumulator(exportReduceList[i].mAccumulatorName));
392 }
393 keepFuncsPushBackIfPresent(exportReduceList[i].mOutConverterName);
394 }
395
396 for (auto &symbol_name : keep_funcs) {
397 export_symbols.push_back(symbol_name.c_str());
398 }
399
400 // http://b/26165616 - WAR for this bug defines the __truncxfhf2 function in
401 // frameworks/rs/driver/runtime. Don't internalize this function for x86, so
402 // that a script can find and link against it.
403 llvm::Triple triple(getTargetMachine().getTargetTriple());
404 if (triple.getArch() == llvm::Triple::x86) {
405 export_symbols.push_back("__truncxfhf2");
406 }
407
408 pPM.add(llvm::createInternalizePass(export_symbols));
409
410 return true;
411 }
412
addInvokeHelperPass(llvm::legacy::PassManager & pPM)413 void Compiler::addInvokeHelperPass(llvm::legacy::PassManager &pPM) {
414 llvm::Triple arch(getTargetMachine().getTargetTriple());
415 if (arch.isArch64Bit()) {
416 pPM.add(createRSInvokeHelperPass());
417 }
418 }
419
addDebugInfoPass(Script & pScript,llvm::legacy::PassManager & pPM)420 void Compiler::addDebugInfoPass(Script &pScript, llvm::legacy::PassManager &pPM) {
421 if (pScript.getSource().getDebugInfoEnabled())
422 pPM.add(createRSAddDebugInfoPass());
423 }
424
addExpandKernelPass(llvm::legacy::PassManager & pPM)425 void Compiler::addExpandKernelPass(llvm::legacy::PassManager &pPM) {
426 // Expand ForEach and reduce on CPU path to reduce launch overhead.
427 bool pEnableStepOpt = true;
428 pPM.add(createRSKernelExpandPass(pEnableStepOpt));
429 }
430
addGlobalInfoPass(Script & pScript,llvm::legacy::PassManager & pPM)431 void Compiler::addGlobalInfoPass(Script &pScript, llvm::legacy::PassManager &pPM) {
432 // Add additional information about RS global variables inside the Module.
433 RSScript &script = static_cast<RSScript &>(pScript);
434 if (script.getEmbedGlobalInfo()) {
435 pPM.add(createRSGlobalInfoPass(script.getEmbedGlobalInfoSkipConstant()));
436 }
437 }
438
addInvariantPass(llvm::legacy::PassManager & pPM)439 void Compiler::addInvariantPass(llvm::legacy::PassManager &pPM) {
440 // Mark Loads from RsExpandKernelDriverInfo as "load.invariant".
441 // Should run after ExpandForEach and before inlining.
442 pPM.add(createRSInvariantPass());
443 }
444
screenGlobalFunctions(Script & pScript)445 enum Compiler::ErrorCode Compiler::screenGlobalFunctions(Script &pScript) {
446 llvm::Module &module = pScript.getSource().getModule();
447
448 // Materialize the bitcode module in case this is a lazy-load module. Do not
449 // clear the materializer by calling materializeAllPermanently since the
450 // runtime library has not been merged into the module yet.
451 if (module.getMaterializer() != nullptr) {
452 std::error_code ec = module.materializeAll();
453 if (ec) {
454 ALOGE("Failed to materialize module `%s' when screening globals! (%s)",
455 module.getModuleIdentifier().c_str(), ec.message().c_str());
456 return kErrMaterialization;
457 }
458 }
459
460 // Add pass to check for illegal function calls.
461 llvm::legacy::PassManager pPM;
462 pPM.add(createRSScreenFunctionsPass());
463 pPM.run(module);
464
465 return kSuccess;
466
467 }
468
translateGEPs(Script & pScript)469 void Compiler::translateGEPs(Script &pScript) {
470 llvm::legacy::PassManager pPM;
471 pPM.add(createRSX86TranslateGEPPass());
472
473 // Materialization done in screenGlobalFunctions above.
474 pPM.run(pScript.getSource().getModule());
475 }
476