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/CompilerConfig.h" 18 19 #include "Log.h" 20 #include "Properties.h" 21 22 #include "bcc/Config.h" 23 24 #include <llvm/CodeGen/SchedulerRegistry.h> 25 #include <llvm/MC/SubtargetFeature.h> 26 #include <llvm/Support/Host.h> 27 #include <llvm/Support/TargetRegistry.h> 28 29 using namespace bcc; 30 31 #if defined (PROVIDE_X86_CODEGEN) && !defined(__HOST__) 32 33 namespace { 34 35 // Utility function to add feature flags supported by the running CPU. 36 // This function is only needed for on-device bcc for x86. 37 void AddX86NativeCPUFeatures(std::vector<std::string>* attributes) { 38 llvm::StringMap<bool> features; 39 if (llvm::sys::getHostCPUFeatures(features)) { 40 for (const auto& f : features) 41 attributes->push_back((f.second ? '+' : '-') + f.first().str()); 42 } 43 44 // LLVM generates AVX code that treats a long3 as 256 bits, while 45 // RenderScript considers a long3 192 bits (http://b/28879581) 46 attributes->push_back("-avx"); 47 } 48 49 } 50 #endif // (PROVIDE_X86_CODEGEN) && !defined(__HOST__) 51 52 CompilerConfig::CompilerConfig(const std::string &pTriple) 53 : mTriple(pTriple), mFullPrecision(true), mTarget(nullptr) { 54 //===--------------------------------------------------------------------===// 55 // Default setting of target options 56 //===--------------------------------------------------------------------===// 57 58 // Use soft-float ABI. This only selects the ABI (and is applicable only to 59 // ARM targets). Codegen still uses hardware FPU by default. To use software 60 // floating point, add 'soft-float' feature to mFeatureString below. 61 mTargetOpts.FloatABIType = llvm::FloatABI::Soft; 62 63 //===--------------------------------------------------------------------===// 64 // Default setting for code model 65 //===--------------------------------------------------------------------===// 66 mCodeModel = llvm::CodeModel::Small; 67 68 //===--------------------------------------------------------------------===// 69 // Default setting for optimization level (-O2) 70 //===--------------------------------------------------------------------===// 71 mOptLevel = llvm::CodeGenOpt::Default; 72 73 //===--------------------------------------------------------------------===// 74 // Default setting for architecture type 75 //===--------------------------------------------------------------------===// 76 mArchType = llvm::Triple::UnknownArch; 77 78 initializeTarget(); 79 initializeArch(); 80 81 return; 82 } 83 84 bool CompilerConfig::initializeTarget() { 85 std::string error; 86 mTarget = llvm::TargetRegistry::lookupTarget(mTriple, error); 87 if (mTarget != nullptr) { 88 return true; 89 } else { 90 ALOGE("Cannot initialize llvm::Target for given triple '%s'! (%s)", 91 mTriple.c_str(), error.c_str()); 92 return false; 93 } 94 } 95 96 bool CompilerConfig::initializeArch() { 97 if (mTarget != nullptr) { 98 mArchType = llvm::Triple::getArchTypeForLLVMName(mTarget->getName()); 99 } else { 100 mArchType = llvm::Triple::UnknownArch; 101 return false; 102 } 103 104 // Configure each architecture for any necessary additional flags. 105 std::vector<std::string> attributes; 106 switch (mArchType) { 107 #if defined(PROVIDE_ARM_CODEGEN) 108 case llvm::Triple::arm: { 109 llvm::StringMap<bool> features; 110 llvm::sys::getHostCPUFeatures(features); 111 112 #if defined(__HOST__) || defined(ARCH_ARM_HAVE_VFP) 113 attributes.push_back("+vfp3"); 114 #if !defined(__HOST__) && !defined(ARCH_ARM_HAVE_VFP_D32) 115 attributes.push_back("+d16"); 116 #endif // !__HOST__ && !ARCH_ARM_HAVE_VFP_D32 117 #endif // __HOST__ || ARCH_ARM_HAVE_VFP 118 119 #if defined(__HOST__) || defined(ARCH_ARM_HAVE_NEON) 120 // Only enable NEON on ARM if we have relaxed precision floats. 121 if (!mFullPrecision) { 122 attributes.push_back("+neon"); 123 } else { 124 #endif // __HOST__ || ARCH_ARM_HAVE_NEON 125 attributes.push_back("-neon"); 126 attributes.push_back("-neonfp"); 127 #if defined(__HOST__) || defined(ARCH_ARM_HAVE_NEON) 128 } 129 #endif // __HOST__ || ARCH_ARM_HAVE_NEON 130 131 if (!getProperty("debug.rs.arm-no-hwdiv")) { 132 if (features.count("hwdiv-arm") && features["hwdiv-arm"]) 133 attributes.push_back("+hwdiv-arm"); 134 135 if (features.count("hwdiv") && features["hwdiv"]) 136 attributes.push_back("+hwdiv"); 137 } 138 139 // Enable fp16 attribute if available in the feature list. This feature 140 // will not be added in the host version of bcc or bcc_compat since 141 // 'features' would correspond to features in an x86 host. 142 if (features.count("fp16") && features["fp16"]) 143 attributes.push_back("+fp16"); 144 145 #if defined(PROVIDE_ARM64_CODEGEN) 146 // On AArch64, asimd in /proc/cpuinfo signals the presence of hardware 147 // half-precision conversion instructions. getHostCPUFeatures translates 148 // this to "neon". If PROVIDE_ARM64_CODEGEN is set, enable "+fp16" for ARM 149 // codegen if "neon" is present in features. 150 if (features.count("neon") && features["neon"]) 151 attributes.push_back("+fp16"); 152 #endif // PROVIDE_ARM64_CODEGEN 153 154 #if defined(TARGET_BUILD) 155 if (!getProperty("debug.rs.arm-no-tune-for-cpu")) { 156 #ifdef DEFAULT_ARM_CODEGEN 157 setCPU(llvm::sys::getHostCPUName()); 158 #endif 159 } 160 #endif // TARGET_BUILD 161 162 break; 163 } 164 #endif // PROVIDE_ARM_CODEGEN 165 166 #if defined(PROVIDE_ARM64_CODEGEN) 167 case llvm::Triple::aarch64: 168 #if defined(TARGET_BUILD) 169 if (!getProperty("debug.rs.arm-no-tune-for-cpu")) { 170 #ifdef DEFAULT_ARM64_CODEGEN 171 setCPU(llvm::sys::getHostCPUName()); 172 #endif 173 } 174 #endif // TARGET_BUILD 175 break; 176 #endif // PROVIDE_ARM64_CODEGEN 177 178 #if defined (PROVIDE_MIPS_CODEGEN) 179 case llvm::Triple::mips: 180 case llvm::Triple::mipsel: 181 if (!mRelocModel.hasValue()) { 182 mRelocModel = llvm::Reloc::Static; 183 } 184 break; 185 #endif // PROVIDE_MIPS_CODEGEN 186 187 #if defined (PROVIDE_MIPS64_CODEGEN) 188 case llvm::Triple::mips64: 189 case llvm::Triple::mips64el: 190 // Default revision for MIPS64 Android is R6. 191 setCPU("mips64r6"); 192 break; 193 #endif // PROVIDE_MIPS64_CODEGEN 194 195 #if defined (PROVIDE_X86_CODEGEN) 196 case llvm::Triple::x86: 197 getTargetOptions().UseInitArray = true; 198 #if defined (DEFAULT_X86_CODEGEN) && !defined (__HOST__) 199 setCPU(llvm::sys::getHostCPUName()); 200 AddX86NativeCPUFeatures(&attributes); 201 #else 202 // generic fallback for 32bit x86 targets 203 setCPU("atom"); 204 #endif // DEFAULT_X86_CODEGEN && !__HOST__ 205 break; 206 #endif // PROVIDE_X86_CODEGEN 207 208 #if defined (PROVIDE_X86_CODEGEN) 209 // PROVIDE_X86_CODEGEN is defined for both x86 and x86_64 210 case llvm::Triple::x86_64: 211 #if defined(DEFAULT_X86_64_CODEGEN) && !defined(__HOST__) 212 setCPU(llvm::sys::getHostCPUName()); 213 AddX86NativeCPUFeatures(&attributes); 214 #else 215 // generic fallback for 64bit x86 targets 216 setCPU("core2"); 217 #endif 218 // x86_64 needs small CodeModel if use PIC_ reloc, or else dlopen failed with TEXTREL. 219 if (mRelocModel.hasValue() && mRelocModel.getValue() == llvm::Reloc::PIC_) { 220 setCodeModel(llvm::CodeModel::Small); 221 } else { 222 setCodeModel(llvm::CodeModel::Medium); 223 } 224 getTargetOptions().UseInitArray = true; 225 break; 226 #endif // PROVIDE_X86_CODEGEN 227 228 default: 229 ALOGE("Unsupported architecture type: %s", mTarget->getName()); 230 return false; 231 } 232 233 setFeatureString(attributes); 234 return true; 235 } 236 237 void CompilerConfig::setFeatureString(const std::vector<std::string> &pAttrs) { 238 llvm::SubtargetFeatures f; 239 240 for (std::vector<std::string>::const_iterator attr_iter = pAttrs.begin(), 241 attr_end = pAttrs.end(); 242 attr_iter != attr_end; attr_iter++) { 243 f.AddFeature(*attr_iter); 244 } 245 246 mFeatureString = f.getString(); 247 return; 248 } 249