1 /*
2 * Copyright 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/Renderscript/RSCompilerDriver.h"
18
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/IR/AssemblyAnnotationWriter.h"
21 #include <llvm/IR/Module.h>
22 #include "llvm/Linker/Linker.h"
23 #include <llvm/Support/CommandLine.h>
24 #include <llvm/Support/Path.h>
25 #include <llvm/Support/raw_ostream.h>
26 #include <llvm/Target/TargetMachine.h>
27
28 #include "bcinfo/BitcodeWrapper.h"
29 #include "bcc/Assert.h"
30 #include "bcinfo/MetadataExtractor.h"
31 #include "bcc/BCCContext.h"
32 #include "bcc/Compiler.h"
33 #include "bcc/Config/Config.h"
34 #include "bcc/Renderscript/RSScript.h"
35 #include "bcc/Renderscript/RSScriptGroupFusion.h"
36 #include "bcc/Support/CompilerConfig.h"
37 #include "bcc/Source.h"
38 #include "bcc/Support/FileMutex.h"
39 #include "bcc/Support/Log.h"
40 #include "bcc/Support/InputFile.h"
41 #include "bcc/Support/Initialization.h"
42 #include "bcc/Support/OutputFile.h"
43
44 #include <sstream>
45 #include <string>
46
47 #ifdef __ANDROID__
48 #include <cutils/properties.h>
49 #endif
50 #include <utils/StopWatch.h>
51
52 using namespace bcc;
53
RSCompilerDriver(bool pUseCompilerRT)54 RSCompilerDriver::RSCompilerDriver(bool pUseCompilerRT) :
55 mConfig(nullptr), mCompiler(), mDebugContext(false),
56 mLinkRuntimeCallback(nullptr), mEnableGlobalMerge(true),
57 mEmbedGlobalInfo(false), mEmbedGlobalInfoSkipConstant(false) {
58 init::Initialize();
59 }
60
~RSCompilerDriver()61 RSCompilerDriver::~RSCompilerDriver() {
62 delete mConfig;
63 }
64
65
66 #if defined(PROVIDE_ARM_CODEGEN)
67 extern llvm::cl::opt<bool> EnableGlobalMerge;
68 #endif
69
setupConfig(const RSScript & pScript)70 bool RSCompilerDriver::setupConfig(const RSScript &pScript) {
71 bool changed = false;
72
73 const llvm::CodeGenOpt::Level script_opt_level =
74 static_cast<llvm::CodeGenOpt::Level>(pScript.getOptimizationLevel());
75
76 #if defined(PROVIDE_ARM_CODEGEN)
77 EnableGlobalMerge = mEnableGlobalMerge;
78 #endif
79
80 if (mConfig != nullptr) {
81 // Renderscript bitcode may have their optimization flag configuration
82 // different than the previous run of RS compilation.
83 if (mConfig->getOptimizationLevel() != script_opt_level) {
84 mConfig->setOptimizationLevel(script_opt_level);
85 changed = true;
86 }
87 } else {
88 // Haven't run the compiler ever.
89 mConfig = new (std::nothrow) CompilerConfig(DEFAULT_TARGET_TRIPLE_STRING);
90 if (mConfig == nullptr) {
91 // Return false since mConfig remains NULL and out-of-memory.
92 return false;
93 }
94 mConfig->setOptimizationLevel(script_opt_level);
95 changed = true;
96 }
97
98 #if defined(PROVIDE_ARM_CODEGEN)
99 bcinfo::MetadataExtractor me(&pScript.getSource().getModule());
100 if (!me.extract()) {
101 bccAssert("Could not extract RS pragma metadata for module!");
102 }
103
104 bool script_full_prec = (me.getRSFloatPrecision() == bcinfo::RS_FP_Full);
105 if (mConfig->getFullPrecision() != script_full_prec) {
106 mConfig->setFullPrecision(script_full_prec);
107 changed = true;
108 }
109 #endif
110
111 return changed;
112 }
113
compileScript(RSScript & pScript,const char * pScriptName,const char * pOutputPath,const char * pRuntimePath,const char * pBuildChecksum,bool pDumpIR)114 Compiler::ErrorCode RSCompilerDriver::compileScript(RSScript& pScript, const char* pScriptName,
115 const char* pOutputPath,
116 const char* pRuntimePath,
117 const char* pBuildChecksum,
118 bool pDumpIR) {
119 // embed build checksum metadata into the source
120 if (pBuildChecksum != nullptr && strlen(pBuildChecksum) > 0) {
121 pScript.getSource().addBuildChecksumMetadata(pBuildChecksum);
122 }
123
124 // Verify that the only external functions in pScript are Renderscript
125 // functions. Fail if verification returns an error.
126 if (mCompiler.screenGlobalFunctions(pScript) != Compiler::kSuccess) {
127 return Compiler::kErrInvalidSource;
128 }
129
130 // For (32-bit) x86, translate GEPs on structs or arrays of structs to GEPs on
131 // int8* with byte offsets. This is to ensure that layout of structs with
132 // 64-bit scalar fields matches frontend-generated code that adheres to ARM
133 // data layout.
134 //
135 // The translation is done before RenderScript runtime library is linked
136 // (during LinkRuntime below) to ensure that RenderScript-driver-provided
137 // structs (like Allocation_t) don't get forced into using the ARM layout
138 // rules.
139 if (mCompiler.getTargetMachine().getTargetTriple().getArch() == llvm::Triple::x86) {
140 mCompiler.translateGEPs(pScript);
141 }
142
143 //===--------------------------------------------------------------------===//
144 // Link RS script with Renderscript runtime.
145 //===--------------------------------------------------------------------===//
146 if (!RSScript::LinkRuntime(pScript, pRuntimePath)) {
147 ALOGE("Failed to link script '%s' with Renderscript runtime %s!",
148 pScriptName, pRuntimePath);
149 return Compiler::kErrInvalidSource;
150 }
151
152 {
153 // FIXME(srhines): Windows compilation can't use locking like this, but
154 // we also don't need to worry about concurrent writers of the same file.
155 #ifndef USE_MINGW
156 //===------------------------------------------------------------------===//
157 // Acquire the write lock for writing output object file.
158 //===------------------------------------------------------------------===//
159 FileMutex<FileBase::kWriteLock> write_output_mutex(pOutputPath);
160
161 if (write_output_mutex.hasError() || !write_output_mutex.lock()) {
162 ALOGE("Unable to acquire the lock for writing %s! (%s)",
163 pOutputPath, write_output_mutex.getErrorMessage().c_str());
164 return Compiler::kErrInvalidSource;
165 }
166 #endif
167
168 // Open the output file for write.
169 OutputFile output_file(pOutputPath,
170 FileBase::kTruncate | FileBase::kBinary);
171
172 if (output_file.hasError()) {
173 ALOGE("Unable to open %s for write! (%s)", pOutputPath,
174 output_file.getErrorMessage().c_str());
175 return Compiler::kErrInvalidSource;
176 }
177
178 // Setup the config to the compiler.
179 bool compiler_need_reconfigure = setupConfig(pScript);
180
181 if (mConfig == nullptr) {
182 ALOGE("Failed to setup config for RS compiler to compile %s!",
183 pOutputPath);
184 return Compiler::kErrInvalidSource;
185 }
186
187 if (compiler_need_reconfigure) {
188 Compiler::ErrorCode err = mCompiler.config(*mConfig);
189 if (err != Compiler::kSuccess) {
190 ALOGE("Failed to config the RS compiler for %s! (%s)",pOutputPath,
191 Compiler::GetErrorString(err));
192 return Compiler::kErrInvalidSource;
193 }
194 }
195
196 OutputFile *ir_file = nullptr;
197 llvm::raw_fd_ostream *IRStream = nullptr;
198 if (pDumpIR) {
199 std::string path(pOutputPath);
200 path.append(".ll");
201 ir_file = new OutputFile(path.c_str(), FileBase::kTruncate);
202 IRStream = ir_file->dup();
203 }
204
205 // Run the compiler.
206 Compiler::ErrorCode compile_result =
207 mCompiler.compile(pScript, output_file, IRStream);
208
209 if (ir_file) {
210 ir_file->close();
211 delete ir_file;
212 }
213
214 if (compile_result != Compiler::kSuccess) {
215 ALOGE("Unable to compile the source to file %s! (%s)", pOutputPath,
216 Compiler::GetErrorString(compile_result));
217 return Compiler::kErrInvalidSource;
218 }
219 }
220
221 return Compiler::kSuccess;
222 }
223
build(BCCContext & pContext,const char * pCacheDir,const char * pResName,const char * pBitcode,size_t pBitcodeSize,const char * pBuildChecksum,const char * pRuntimePath,RSLinkRuntimeCallback pLinkRuntimeCallback,bool pDumpIR)224 bool RSCompilerDriver::build(BCCContext &pContext,
225 const char *pCacheDir,
226 const char *pResName,
227 const char *pBitcode,
228 size_t pBitcodeSize,
229 const char *pBuildChecksum,
230 const char *pRuntimePath,
231 RSLinkRuntimeCallback pLinkRuntimeCallback,
232 bool pDumpIR) {
233 // android::StopWatch build_time("bcc: RSCompilerDriver::build time");
234 //===--------------------------------------------------------------------===//
235 // Check parameters.
236 //===--------------------------------------------------------------------===//
237 if ((pCacheDir == nullptr) || (pResName == nullptr)) {
238 ALOGE("Invalid parameter passed to RSCompilerDriver::build()! (cache dir: "
239 "%s, resource name: %s)", ((pCacheDir) ? pCacheDir : "(null)"),
240 ((pResName) ? pResName : "(null)"));
241 return false;
242 }
243
244 if ((pBitcode == nullptr) || (pBitcodeSize <= 0)) {
245 ALOGE("No bitcode supplied! (bitcode: %p, size of bitcode: %u)",
246 pBitcode, static_cast<unsigned>(pBitcodeSize));
247 return false;
248 }
249
250 //===--------------------------------------------------------------------===//
251 // Construct output path.
252 // {pCacheDir}/{pResName}.o
253 //===--------------------------------------------------------------------===//
254 llvm::SmallString<80> output_path(pCacheDir);
255 llvm::sys::path::append(output_path, pResName);
256 llvm::sys::path::replace_extension(output_path, ".o");
257
258 //===--------------------------------------------------------------------===//
259 // Load the bitcode and create script.
260 //===--------------------------------------------------------------------===//
261 Source *source = Source::CreateFromBuffer(pContext, pResName,
262 pBitcode, pBitcodeSize);
263 if (source == nullptr) {
264 return false;
265 }
266
267 RSScript script(*source, getConfig());
268 if (pLinkRuntimeCallback) {
269 setLinkRuntimeCallback(pLinkRuntimeCallback);
270 }
271
272 script.setLinkRuntimeCallback(getLinkRuntimeCallback());
273
274 script.setEmbedGlobalInfo(mEmbedGlobalInfo);
275 script.setEmbedGlobalInfoSkipConstant(mEmbedGlobalInfoSkipConstant);
276
277 // Read information from bitcode wrapper.
278 bcinfo::BitcodeWrapper wrapper(pBitcode, pBitcodeSize);
279 script.setCompilerVersion(wrapper.getCompilerVersion());
280 script.setOptimizationLevel(static_cast<RSScript::OptimizationLevel>(
281 wrapper.getOptimizationLevel()));
282
283 // Assertion-enabled builds can't compile legacy bitcode (due to the use of
284 // getName() with anonymous structure definitions).
285 #ifdef FORCE_BUILD_LLVM_DISABLE_NDEBUG
286 static const uint32_t kSlangMinimumFixedStructureNames = 2310;
287 uint32_t version = wrapper.getCompilerVersion();
288 if (version < kSlangMinimumFixedStructureNames) {
289 ALOGE("Found invalid legacy bitcode compiled with a version %u llvm-rs-cc "
290 "used with an assertion build", version);
291 ALOGE("Please recompile this apk with a more recent llvm-rs-cc "
292 "(at least %u)", kSlangMinimumFixedStructureNames);
293 return false;
294 }
295 #endif
296
297 //===--------------------------------------------------------------------===//
298 // Compile the script
299 //===--------------------------------------------------------------------===//
300 Compiler::ErrorCode status = compileScript(script, pResName,
301 output_path.c_str(),
302 pRuntimePath,
303 pBuildChecksum,
304 pDumpIR);
305
306 return status == Compiler::kSuccess;
307 }
308
buildScriptGroup(BCCContext & Context,const char * pOutputFilepath,const char * pRuntimePath,const char * pRuntimeRelaxedPath,bool dumpIR,const char * buildChecksum,const std::vector<Source * > & sources,const std::list<std::list<std::pair<int,int>>> & toFuse,const std::list<std::string> & fused,const std::list<std::list<std::pair<int,int>>> & invokes,const std::list<std::string> & invokeBatchNames)309 bool RSCompilerDriver::buildScriptGroup(
310 BCCContext& Context, const char* pOutputFilepath, const char* pRuntimePath,
311 const char* pRuntimeRelaxedPath, bool dumpIR, const char* buildChecksum,
312 const std::vector<Source*>& sources,
313 const std::list<std::list<std::pair<int, int>>>& toFuse,
314 const std::list<std::string>& fused,
315 const std::list<std::list<std::pair<int, int>>>& invokes,
316 const std::list<std::string>& invokeBatchNames) {
317
318 // Read and store metadata before linking the modules together
319 std::vector<bcinfo::MetadataExtractor*> metadata;
320 for (Source* source : sources) {
321 if (!source->extractMetadata()) {
322 ALOGE("Cannot extract metadata from module");
323 return false;
324 }
325 }
326
327 // ---------------------------------------------------------------------------
328 // Link all input modules into a single module
329 // ---------------------------------------------------------------------------
330
331 llvm::LLVMContext& context = Context.getLLVMContext();
332 llvm::Module module("Merged Script Group", context);
333
334 llvm::Linker linker(module);
335 for (Source* source : sources) {
336 std::unique_ptr<llvm::Module> sourceModule(&source->getModule());
337 if (linker.linkInModule(std::move(sourceModule))) {
338 ALOGE("Linking for module in source failed.");
339 return false;
340 }
341 // source->getModule() is destroyed after linking.
342 source->markModuleDestroyed();
343 }
344
345 // ---------------------------------------------------------------------------
346 // Create fused kernels
347 // ---------------------------------------------------------------------------
348
349 auto inputIter = toFuse.begin();
350 for (const std::string& nameOfFused : fused) {
351 auto inputKernels = *inputIter++;
352 std::vector<Source*> sourcesToFuse;
353 std::vector<int> slots;
354
355 for (auto p : inputKernels) {
356 sourcesToFuse.push_back(sources[p.first]);
357 slots.push_back(p.second);
358 }
359
360 if (!fuseKernels(Context, sourcesToFuse, slots, nameOfFused, &module)) {
361 return false;
362 }
363 }
364
365 // ---------------------------------------------------------------------------
366 // Rename invokes
367 // ---------------------------------------------------------------------------
368
369 auto invokeIter = invokes.begin();
370 for (const std::string& newName : invokeBatchNames) {
371 auto inputInvoke = *invokeIter++;
372 auto p = inputInvoke.front();
373 Source* source = sources[p.first];
374 int slot = p.second;
375
376 if (!renameInvoke(Context, source, slot, newName, &module)) {
377 return false;
378 }
379 }
380
381 // ---------------------------------------------------------------------------
382 // Compile the new module with fused kernels
383 // ---------------------------------------------------------------------------
384
385 const std::unique_ptr<Source> source(
386 Source::CreateFromModule(Context, pOutputFilepath, module, true));
387 RSScript script(*source);
388
389 // Embed the info string directly in the ELF
390 script.setEmbedInfo(true);
391 script.setOptimizationLevel(RSScript::kOptLvl3);
392 script.setEmbedGlobalInfo(mEmbedGlobalInfo);
393 script.setEmbedGlobalInfoSkipConstant(mEmbedGlobalInfoSkipConstant);
394
395 llvm::SmallString<80> output_path(pOutputFilepath);
396 llvm::sys::path::replace_extension(output_path, ".o");
397
398 // Pick the right runtime lib
399 const char* coreLibPath = pRuntimePath;
400 if (strcmp(pRuntimeRelaxedPath, "")) {
401 bcinfo::MetadataExtractor me(&module);
402 me.extract();
403 if (me.getRSFloatPrecision() == bcinfo::RS_FP_Relaxed) {
404 coreLibPath = pRuntimeRelaxedPath;
405 }
406 }
407
408 compileScript(script, pOutputFilepath, output_path.c_str(), coreLibPath,
409 buildChecksum, dumpIR);
410
411 return true;
412 }
413
buildForCompatLib(RSScript & pScript,const char * pOut,const char * pBuildChecksum,const char * pRuntimePath,bool pDumpIR)414 bool RSCompilerDriver::buildForCompatLib(RSScript &pScript, const char *pOut,
415 const char *pBuildChecksum,
416 const char *pRuntimePath,
417 bool pDumpIR) {
418 // Embed the info string directly in the ELF, since this path is for an
419 // offline (host) compilation.
420 pScript.setEmbedInfo(true);
421
422 pScript.setEmbedGlobalInfo(mEmbedGlobalInfo);
423 pScript.setEmbedGlobalInfoSkipConstant(mEmbedGlobalInfoSkipConstant);
424 pScript.setLinkRuntimeCallback(getLinkRuntimeCallback());
425
426 Compiler::ErrorCode status = compileScript(pScript, pOut, pOut, pRuntimePath,
427 pBuildChecksum, pDumpIR);
428 if (status != Compiler::kSuccess) {
429 return false;
430 }
431
432 return true;
433 }
434