1 //===-- NVPTXAsmPrinter.h - NVPTX LLVM assembly writer --------------------===// 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 // This file contains a printer that converts from our internal representation 11 // of machine-dependent LLVM code to NVPTX assembly language. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H 16 #define LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H 17 18 #include "NVPTX.h" 19 #include "NVPTXSubtarget.h" 20 #include "NVPTXTargetMachine.h" 21 #include "llvm/ADT/SmallString.h" 22 #include "llvm/ADT/StringExtras.h" 23 #include "llvm/CodeGen/AsmPrinter.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/MC/MCAsmInfo.h" 26 #include "llvm/MC/MCExpr.h" 27 #include "llvm/MC/MCSymbol.h" 28 #include "llvm/Support/CommandLine.h" 29 #include "llvm/Support/FormattedStream.h" 30 #include "llvm/Target/TargetMachine.h" 31 #include <fstream> 32 33 // The ptx syntax and format is very different from that usually seem in a .s 34 // file, 35 // therefore we are not able to use the MCAsmStreamer interface here. 36 // 37 // We are handcrafting the output method here. 38 // 39 // A better approach is to clone the MCAsmStreamer to a MCPTXAsmStreamer 40 // (subclass of MCStreamer). 41 42 namespace llvm { 43 44 class LineReader { 45 private: 46 unsigned theCurLine; 47 std::ifstream fstr; 48 char buff[512]; 49 std::string theFileName; 50 SmallVector<unsigned, 32> lineOffset; 51 public: LineReader(std::string filename)52 LineReader(std::string filename) { 53 theCurLine = 0; 54 fstr.open(filename.c_str()); 55 theFileName = filename; 56 } fileName()57 std::string fileName() { return theFileName; } ~LineReader()58 ~LineReader() { fstr.close(); } 59 std::string readLine(unsigned line); 60 }; 61 62 class LLVM_LIBRARY_VISIBILITY NVPTXAsmPrinter : public AsmPrinter { 63 64 class AggBuffer { 65 // Used to buffer the emitted string for initializing global 66 // aggregates. 67 // 68 // Normally an aggregate (array, vector or structure) is emitted 69 // as a u8[]. However, if one element/field of the aggregate 70 // is a non-NULL address, then the aggregate is emitted as u32[] 71 // or u64[]. 72 // 73 // We first layout the aggregate in 'buffer' in bytes, except for 74 // those symbol addresses. For the i-th symbol address in the 75 //aggregate, its corresponding 4-byte or 8-byte elements in 'buffer' 76 // are filled with 0s. symbolPosInBuffer[i-1] records its position 77 // in 'buffer', and Symbols[i-1] records the Value*. 78 // 79 // Once we have this AggBuffer setup, we can choose how to print 80 // it out. 81 public: 82 unsigned numSymbols; // number of symbol addresses 83 84 private: 85 const unsigned size; // size of the buffer in bytes 86 std::vector<unsigned char> buffer; // the buffer 87 SmallVector<unsigned, 4> symbolPosInBuffer; 88 SmallVector<const Value *, 4> Symbols; 89 unsigned curpos; 90 raw_ostream &O; 91 NVPTXAsmPrinter &AP; 92 bool EmitGeneric; 93 94 public: AggBuffer(unsigned size,raw_ostream & O,NVPTXAsmPrinter & AP)95 AggBuffer(unsigned size, raw_ostream &O, NVPTXAsmPrinter &AP) 96 : size(size), buffer(size), O(O), AP(AP) { 97 curpos = 0; 98 numSymbols = 0; 99 EmitGeneric = AP.EmitGeneric; 100 } addBytes(unsigned char * Ptr,int Num,int Bytes)101 unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) { 102 assert((curpos + Num) <= size); 103 assert((curpos + Bytes) <= size); 104 for (int i = 0; i < Num; ++i) { 105 buffer[curpos] = Ptr[i]; 106 curpos++; 107 } 108 for (int i = Num; i < Bytes; ++i) { 109 buffer[curpos] = 0; 110 curpos++; 111 } 112 return curpos; 113 } addZeros(int Num)114 unsigned addZeros(int Num) { 115 assert((curpos + Num) <= size); 116 for (int i = 0; i < Num; ++i) { 117 buffer[curpos] = 0; 118 curpos++; 119 } 120 return curpos; 121 } addSymbol(const Value * GVar)122 void addSymbol(const Value *GVar) { 123 symbolPosInBuffer.push_back(curpos); 124 Symbols.push_back(GVar); 125 numSymbols++; 126 } print()127 void print() { 128 if (numSymbols == 0) { 129 // print out in bytes 130 for (unsigned i = 0; i < size; i++) { 131 if (i) 132 O << ", "; 133 O << (unsigned int) buffer[i]; 134 } 135 } else { 136 // print out in 4-bytes or 8-bytes 137 unsigned int pos = 0; 138 unsigned int nSym = 0; 139 unsigned int nextSymbolPos = symbolPosInBuffer[nSym]; 140 unsigned int nBytes = 4; 141 if (static_cast<const NVPTXTargetMachine &>(AP.TM).is64Bit()) 142 nBytes = 8; 143 for (pos = 0; pos < size; pos += nBytes) { 144 if (pos) 145 O << ", "; 146 if (pos == nextSymbolPos) { 147 const Value *v = Symbols[nSym]; 148 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) { 149 MCSymbol *Name = AP.getSymbol(GVar); 150 PointerType *PTy = dyn_cast<PointerType>(GVar->getType()); 151 bool IsNonGenericPointer = false; 152 if (PTy && PTy->getAddressSpace() != 0) { 153 IsNonGenericPointer = true; 154 } 155 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) { 156 O << "generic("; 157 O << *Name; 158 O << ")"; 159 } else { 160 O << *Name; 161 } 162 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) { 163 O << *AP.lowerConstant(Cexpr); 164 } else 165 llvm_unreachable("symbol type unknown"); 166 nSym++; 167 if (nSym >= numSymbols) 168 nextSymbolPos = size + 1; 169 else 170 nextSymbolPos = symbolPosInBuffer[nSym]; 171 } else if (nBytes == 4) 172 O << *(unsigned int *)(&buffer[pos]); 173 else 174 O << *(unsigned long long *)(&buffer[pos]); 175 } 176 } 177 } 178 }; 179 180 friend class AggBuffer; 181 182 void emitSrcInText(StringRef filename, unsigned line); 183 184 private: getPassName()185 const char *getPassName() const override { return "NVPTX Assembly Printer"; } 186 187 const Function *F; 188 std::string CurrentFnName; 189 190 void EmitBasicBlockStart(const MachineBasicBlock &MBB) const override; 191 void EmitFunctionEntryLabel() override; 192 void EmitFunctionBodyStart() override; 193 void EmitFunctionBodyEnd() override; 194 void emitImplicitDef(const MachineInstr *MI) const override; 195 196 void EmitInstruction(const MachineInstr *) override; 197 void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI); 198 bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp); 199 MCOperand GetSymbolRef(const MCSymbol *Symbol); 200 unsigned encodeVirtualRegister(unsigned Reg); 201 202 void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {} 203 204 void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier, 205 raw_ostream &O); 206 void printMemOperand(const MachineInstr *MI, int opNum, raw_ostream &O, 207 const char *Modifier = nullptr); 208 void printImplicitDef(const MachineInstr *MI, raw_ostream &O) const; 209 void printModuleLevelGV(const GlobalVariable *GVar, raw_ostream &O, 210 bool = false); 211 void printParamName(int paramIndex, raw_ostream &O); 212 void printParamName(Function::const_arg_iterator I, int paramIndex, 213 raw_ostream &O); 214 void emitGlobals(const Module &M); 215 void emitHeader(Module &M, raw_ostream &O, const NVPTXSubtarget &STI); 216 void emitKernelFunctionDirectives(const Function &F, raw_ostream &O) const; 217 void emitVirtualRegister(unsigned int vr, raw_ostream &); 218 void emitFunctionExternParamList(const MachineFunction &MF); 219 void emitFunctionParamList(const Function *, raw_ostream &O); 220 void emitFunctionParamList(const MachineFunction &MF, raw_ostream &O); 221 void setAndEmitFunctionVirtualRegisters(const MachineFunction &MF); 222 void emitFunctionTempData(const MachineFunction &MF, unsigned &FrameSize); 223 bool isImageType(const Type *Ty); 224 void printReturnValStr(const Function *, raw_ostream &O); 225 void printReturnValStr(const MachineFunction &MF, raw_ostream &O); 226 bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo, 227 unsigned AsmVariant, const char *ExtraCode, 228 raw_ostream &) override; 229 void printOperand(const MachineInstr *MI, int opNum, raw_ostream &O, 230 const char *Modifier = nullptr); 231 bool PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo, 232 unsigned AsmVariant, const char *ExtraCode, 233 raw_ostream &) override; 234 protected: 235 bool doInitialization(Module &M) override; 236 bool doFinalization(Module &M) override; 237 238 private: 239 std::string CurrentBankselLabelInBasicBlock; 240 241 bool GlobalsEmitted; 242 243 // This is specific per MachineFunction. 244 const MachineRegisterInfo *MRI; 245 // The contents are specific for each 246 // MachineFunction. But the size of the 247 // array is not. 248 typedef DenseMap<unsigned, unsigned> VRegMap; 249 typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap; 250 VRegRCMap VRegMapping; 251 252 // Cache the subtarget here. 253 const NVPTXSubtarget *nvptxSubtarget; 254 255 // Build the map between type name and ID based on module's type 256 // symbol table. 257 std::map<const Type *, std::string> TypeNameMap; 258 259 // List of variables demoted to a function scope. 260 std::map<const Function *, std::vector<const GlobalVariable *> > localDecls; 261 262 // To record filename to ID mapping 263 std::map<std::string, unsigned> filenameMap; 264 void recordAndEmitFilenames(Module &); 265 266 void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O); 267 void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const; 268 std::string getPTXFundamentalTypeStr(const Type *Ty, bool = true) const; 269 void printScalarConstant(const Constant *CPV, raw_ostream &O); 270 void printFPConstant(const ConstantFP *Fp, raw_ostream &O); 271 void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer); 272 void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer); 273 274 void printOperandProper(const MachineOperand &MO); 275 276 void emitLinkageDirective(const GlobalValue *V, raw_ostream &O); 277 void emitDeclarations(const Module &, raw_ostream &O); 278 void emitDeclaration(const Function *, raw_ostream &O); 279 280 static const char *getRegisterName(unsigned RegNo); 281 void emitDemotedVars(const Function *, raw_ostream &); 282 283 bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo, 284 MCOperand &MCOp); 285 void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp); 286 287 bool isLoopHeaderOfNoUnroll(const MachineBasicBlock &MBB) const; 288 289 LineReader *reader; 290 LineReader *getReader(std::string); 291 292 // Used to control the need to emit .generic() in the initializer of 293 // module scope variables. 294 // Although ptx supports the hybrid mode like the following, 295 // .global .u32 a; 296 // .global .u32 b; 297 // .global .u32 addr[] = {a, generic(b)} 298 // we have difficulty representing the difference in the NVVM IR. 299 // 300 // Since the address value should always be generic in CUDA C and always 301 // be specific in OpenCL, we use this simple control here. 302 // 303 bool EmitGeneric; 304 305 public: NVPTXAsmPrinter(TargetMachine & TM,std::unique_ptr<MCStreamer> Streamer)306 NVPTXAsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer) 307 : AsmPrinter(TM, std::move(Streamer)), 308 EmitGeneric(static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == 309 NVPTX::CUDA) { 310 CurrentBankselLabelInBasicBlock = ""; 311 reader = nullptr; 312 } 313 ~NVPTXAsmPrinter()314 ~NVPTXAsmPrinter() { 315 if (!reader) 316 delete reader; 317 } 318 runOnMachineFunction(MachineFunction & F)319 bool runOnMachineFunction(MachineFunction &F) override { 320 nvptxSubtarget = &F.getSubtarget<NVPTXSubtarget>(); 321 return AsmPrinter::runOnMachineFunction(F); 322 } getAnalysisUsage(AnalysisUsage & AU)323 void getAnalysisUsage(AnalysisUsage &AU) const override { 324 AU.addRequired<MachineLoopInfo>(); 325 AsmPrinter::getAnalysisUsage(AU); 326 } 327 328 bool ignoreLoc(const MachineInstr &); 329 330 std::string getVirtualRegisterName(unsigned) const; 331 332 DebugLoc prevDebugLoc; 333 void emitLineNumberAsDotLoc(const MachineInstr &); 334 }; 335 } // end of namespace 336 337 #endif 338