1 //===-- SystemZTargetMachine.cpp - Define TargetMachine for SystemZ -------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "MCTargetDesc/SystemZMCTargetDesc.h" 11 #include "SystemZ.h" 12 #include "SystemZMachineScheduler.h" 13 #include "SystemZTargetMachine.h" 14 #include "SystemZTargetTransformInfo.h" 15 #include "llvm/ADT/Optional.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/Analysis/TargetTransformInfo.h" 20 #include "llvm/CodeGen/Passes.h" 21 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" 22 #include "llvm/CodeGen/TargetPassConfig.h" 23 #include "llvm/IR/DataLayout.h" 24 #include "llvm/Support/CodeGen.h" 25 #include "llvm/Support/TargetRegistry.h" 26 #include "llvm/Target/TargetLoweringObjectFile.h" 27 #include "llvm/Transforms/Scalar.h" 28 #include <string> 29 30 using namespace llvm; 31 32 extern "C" void LLVMInitializeSystemZTarget() { 33 // Register the target. 34 RegisterTargetMachine<SystemZTargetMachine> X(getTheSystemZTarget()); 35 } 36 37 // Determine whether we use the vector ABI. 38 static bool UsesVectorABI(StringRef CPU, StringRef FS) { 39 // We use the vector ABI whenever the vector facility is avaiable. 40 // This is the case by default if CPU is z13 or later, and can be 41 // overridden via "[+-]vector" feature string elements. 42 bool VectorABI = true; 43 if (CPU.empty() || CPU == "generic" || 44 CPU == "z10" || CPU == "z196" || CPU == "zEC12") 45 VectorABI = false; 46 47 SmallVector<StringRef, 3> Features; 48 FS.split(Features, ',', -1, false /* KeepEmpty */); 49 for (auto &Feature : Features) { 50 if (Feature == "vector" || Feature == "+vector") 51 VectorABI = true; 52 if (Feature == "-vector") 53 VectorABI = false; 54 } 55 56 return VectorABI; 57 } 58 59 static std::string computeDataLayout(const Triple &TT, StringRef CPU, 60 StringRef FS) { 61 bool VectorABI = UsesVectorABI(CPU, FS); 62 std::string Ret; 63 64 // Big endian. 65 Ret += "E"; 66 67 // Data mangling. 68 Ret += DataLayout::getManglingComponent(TT); 69 70 // Make sure that global data has at least 16 bits of alignment by 71 // default, so that we can refer to it using LARL. We don't have any 72 // special requirements for stack variables though. 73 Ret += "-i1:8:16-i8:8:16"; 74 75 // 64-bit integers are naturally aligned. 76 Ret += "-i64:64"; 77 78 // 128-bit floats are aligned only to 64 bits. 79 Ret += "-f128:64"; 80 81 // When using the vector ABI, 128-bit vectors are also aligned to 64 bits. 82 if (VectorABI) 83 Ret += "-v128:64"; 84 85 // We prefer 16 bits of aligned for all globals; see above. 86 Ret += "-a:8:16"; 87 88 // Integer registers are 32 or 64 bits. 89 Ret += "-n32:64"; 90 91 return Ret; 92 } 93 94 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) { 95 // Static code is suitable for use in a dynamic executable; there is no 96 // separate DynamicNoPIC model. 97 if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC) 98 return Reloc::Static; 99 return *RM; 100 } 101 102 SystemZTargetMachine::SystemZTargetMachine(const Target &T, const Triple &TT, 103 StringRef CPU, StringRef FS, 104 const TargetOptions &Options, 105 Optional<Reloc::Model> RM, 106 CodeModel::Model CM, 107 CodeGenOpt::Level OL) 108 : LLVMTargetMachine(T, computeDataLayout(TT, CPU, FS), TT, CPU, FS, Options, 109 getEffectiveRelocModel(RM), CM, OL), 110 TLOF(llvm::make_unique<TargetLoweringObjectFileELF>()), 111 Subtarget(TT, CPU, FS, *this) { 112 initAsmInfo(); 113 } 114 115 SystemZTargetMachine::~SystemZTargetMachine() = default; 116 117 namespace { 118 119 /// SystemZ Code Generator Pass Configuration Options. 120 class SystemZPassConfig : public TargetPassConfig { 121 public: 122 SystemZPassConfig(SystemZTargetMachine *TM, PassManagerBase &PM) 123 : TargetPassConfig(TM, PM) {} 124 125 SystemZTargetMachine &getSystemZTargetMachine() const { 126 return getTM<SystemZTargetMachine>(); 127 } 128 129 ScheduleDAGInstrs * 130 createPostMachineScheduler(MachineSchedContext *C) const override { 131 return new ScheduleDAGMI(C, 132 llvm::make_unique<SystemZPostRASchedStrategy>(C), 133 /*RemoveKillFlags=*/true); 134 } 135 136 void addIRPasses() override; 137 bool addInstSelector() override; 138 bool addILPOpts() override; 139 void addPreSched2() override; 140 void addPreEmitPass() override; 141 }; 142 143 } // end anonymous namespace 144 145 void SystemZPassConfig::addIRPasses() { 146 if (getOptLevel() != CodeGenOpt::None) 147 addPass(createSystemZTDCPass()); 148 149 TargetPassConfig::addIRPasses(); 150 } 151 152 bool SystemZPassConfig::addInstSelector() { 153 addPass(createSystemZISelDag(getSystemZTargetMachine(), getOptLevel())); 154 155 if (getOptLevel() != CodeGenOpt::None) 156 addPass(createSystemZLDCleanupPass(getSystemZTargetMachine())); 157 158 return false; 159 } 160 161 bool SystemZPassConfig::addILPOpts() { 162 addPass(&EarlyIfConverterID); 163 return true; 164 } 165 166 void SystemZPassConfig::addPreSched2() { 167 addPass(createSystemZExpandPseudoPass(getSystemZTargetMachine())); 168 169 if (getOptLevel() != CodeGenOpt::None) 170 addPass(&IfConverterID); 171 } 172 173 void SystemZPassConfig::addPreEmitPass() { 174 // Do instruction shortening before compare elimination because some 175 // vector instructions will be shortened into opcodes that compare 176 // elimination recognizes. 177 if (getOptLevel() != CodeGenOpt::None) 178 addPass(createSystemZShortenInstPass(getSystemZTargetMachine()), false); 179 180 // We eliminate comparisons here rather than earlier because some 181 // transformations can change the set of available CC values and we 182 // generally want those transformations to have priority. This is 183 // especially true in the commonest case where the result of the comparison 184 // is used by a single in-range branch instruction, since we will then 185 // be able to fuse the compare and the branch instead. 186 // 187 // For example, two-address NILF can sometimes be converted into 188 // three-address RISBLG. NILF produces a CC value that indicates whether 189 // the low word is zero, but RISBLG does not modify CC at all. On the 190 // other hand, 64-bit ANDs like NILL can sometimes be converted to RISBG. 191 // The CC value produced by NILL isn't useful for our purposes, but the 192 // value produced by RISBG can be used for any comparison with zero 193 // (not just equality). So there are some transformations that lose 194 // CC values (while still being worthwhile) and others that happen to make 195 // the CC result more useful than it was originally. 196 // 197 // Another reason is that we only want to use BRANCH ON COUNT in cases 198 // where we know that the count register is not going to be spilled. 199 // 200 // Doing it so late makes it more likely that a register will be reused 201 // between the comparison and the branch, but it isn't clear whether 202 // preventing that would be a win or not. 203 if (getOptLevel() != CodeGenOpt::None) 204 addPass(createSystemZElimComparePass(getSystemZTargetMachine()), false); 205 addPass(createSystemZLongBranchPass(getSystemZTargetMachine())); 206 207 // Do final scheduling after all other optimizations, to get an 208 // optimal input for the decoder (branch relaxation must happen 209 // after block placement). 210 if (getOptLevel() != CodeGenOpt::None) 211 addPass(&PostMachineSchedulerID); 212 } 213 214 TargetPassConfig *SystemZTargetMachine::createPassConfig(PassManagerBase &PM) { 215 return new SystemZPassConfig(this, PM); 216 } 217 218 TargetIRAnalysis SystemZTargetMachine::getTargetIRAnalysis() { 219 return TargetIRAnalysis([this](const Function &F) { 220 return TargetTransformInfo(SystemZTTIImpl(this, F)); 221 }); 222 } 223