1 //===-- NVPTXAsmPrinter.h - NVPTX LLVM assembly writer ----------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains a printer that converts from our internal representation 10 // of machine-dependent LLVM code to NVPTX assembly language. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H 15 #define LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H 16 17 #include "NVPTX.h" 18 #include "NVPTXSubtarget.h" 19 #include "NVPTXTargetMachine.h" 20 #include "llvm/ADT/DenseMap.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/StringRef.h" 23 #include "llvm/CodeGen/AsmPrinter.h" 24 #include "llvm/CodeGen/MachineFunction.h" 25 #include "llvm/CodeGen/MachineLoopInfo.h" 26 #include "llvm/IR/Constants.h" 27 #include "llvm/IR/DebugLoc.h" 28 #include "llvm/IR/DerivedTypes.h" 29 #include "llvm/IR/Function.h" 30 #include "llvm/IR/GlobalAlias.h" 31 #include "llvm/IR/GlobalValue.h" 32 #include "llvm/IR/Value.h" 33 #include "llvm/MC/MCExpr.h" 34 #include "llvm/MC/MCStreamer.h" 35 #include "llvm/MC/MCSymbol.h" 36 #include "llvm/Pass.h" 37 #include "llvm/Support/Casting.h" 38 #include "llvm/Support/Compiler.h" 39 #include "llvm/Support/ErrorHandling.h" 40 #include "llvm/Support/raw_ostream.h" 41 #include "llvm/Target/TargetMachine.h" 42 #include <algorithm> 43 #include <cassert> 44 #include <map> 45 #include <memory> 46 #include <string> 47 #include <vector> 48 49 // The ptx syntax and format is very different from that usually seem in a .s 50 // file, 51 // therefore we are not able to use the MCAsmStreamer interface here. 52 // 53 // We are handcrafting the output method here. 54 // 55 // A better approach is to clone the MCAsmStreamer to a MCPTXAsmStreamer 56 // (subclass of MCStreamer). 57 58 namespace llvm { 59 60 class MCOperand; 61 62 class LLVM_LIBRARY_VISIBILITY NVPTXAsmPrinter : public AsmPrinter { 63 64 class AggBuffer { 65 // Used to buffer the emitted string for initializing global aggregates. 66 // 67 // Normally an aggregate (array, vector, or structure) is emitted as a u8[]. 68 // However, if either element/field of the aggregate is a non-NULL address, 69 // and all such addresses are properly aligned, then the aggregate is 70 // emitted as u32[] or u64[]. In the case of unaligned addresses, the 71 // aggregate is emitted as u8[], and the mask() operator is used for all 72 // pointers. 73 // 74 // We first layout the aggregate in 'buffer' in bytes, except for those 75 // symbol addresses. For the i-th symbol address in the aggregate, its 76 // corresponding 4-byte or 8-byte elements in 'buffer' are filled with 0s. 77 // symbolPosInBuffer[i-1] records its position in 'buffer', and Symbols[i-1] 78 // records the Value*. 79 // 80 // Once we have this AggBuffer setup, we can choose how to print it out. 81 public: 82 // number of symbol addresses 83 unsigned numSymbols() const { return Symbols.size(); } 84 85 bool allSymbolsAligned(unsigned ptrSize) const { 86 return llvm::all_of(symbolPosInBuffer, 87 [=](unsigned pos) { return pos % ptrSize == 0; }); 88 } 89 90 private: 91 const unsigned size; // size of the buffer in bytes 92 std::vector<unsigned char> buffer; // the buffer 93 SmallVector<unsigned, 4> symbolPosInBuffer; 94 SmallVector<const Value *, 4> Symbols; 95 // SymbolsBeforeStripping[i] is the original form of Symbols[i] before 96 // stripping pointer casts, i.e., 97 // Symbols[i] == SymbolsBeforeStripping[i]->stripPointerCasts(). 98 // 99 // We need to keep these values because AggBuffer::print decides whether to 100 // emit a "generic()" cast for Symbols[i] depending on the address space of 101 // SymbolsBeforeStripping[i]. 102 SmallVector<const Value *, 4> SymbolsBeforeStripping; 103 unsigned curpos; 104 NVPTXAsmPrinter &AP; 105 bool EmitGeneric; 106 107 public: 108 AggBuffer(unsigned size, NVPTXAsmPrinter &AP) 109 : size(size), buffer(size), AP(AP) { 110 curpos = 0; 111 EmitGeneric = AP.EmitGeneric; 112 } 113 114 // Copy Num bytes from Ptr. 115 // if Bytes > Num, zero fill up to Bytes. 116 unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) { 117 assert((curpos + Num) <= size); 118 assert((curpos + Bytes) <= size); 119 for (int i = 0; i < Num; ++i) { 120 buffer[curpos] = Ptr[i]; 121 curpos++; 122 } 123 for (int i = Num; i < Bytes; ++i) { 124 buffer[curpos] = 0; 125 curpos++; 126 } 127 return curpos; 128 } 129 130 unsigned addZeros(int Num) { 131 assert((curpos + Num) <= size); 132 for (int i = 0; i < Num; ++i) { 133 buffer[curpos] = 0; 134 curpos++; 135 } 136 return curpos; 137 } 138 139 void addSymbol(const Value *GVar, const Value *GVarBeforeStripping) { 140 symbolPosInBuffer.push_back(curpos); 141 Symbols.push_back(GVar); 142 SymbolsBeforeStripping.push_back(GVarBeforeStripping); 143 } 144 145 void printBytes(raw_ostream &os); 146 void printWords(raw_ostream &os); 147 148 private: 149 void printSymbol(unsigned nSym, raw_ostream &os); 150 }; 151 152 friend class AggBuffer; 153 154 private: 155 StringRef getPassName() const override { return "NVPTX Assembly Printer"; } 156 157 const Function *F; 158 std::string CurrentFnName; 159 160 void emitStartOfAsmFile(Module &M) override; 161 void emitBasicBlockStart(const MachineBasicBlock &MBB) override; 162 void emitFunctionEntryLabel() override; 163 void emitFunctionBodyStart() override; 164 void emitFunctionBodyEnd() override; 165 void emitImplicitDef(const MachineInstr *MI) const override; 166 167 void emitInstruction(const MachineInstr *) override; 168 void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI); 169 bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp); 170 MCOperand GetSymbolRef(const MCSymbol *Symbol); 171 unsigned encodeVirtualRegister(unsigned Reg); 172 173 void printMemOperand(const MachineInstr *MI, unsigned OpNum, raw_ostream &O, 174 const char *Modifier = nullptr); 175 void printModuleLevelGV(const GlobalVariable *GVar, raw_ostream &O, 176 bool processDemoted, const NVPTXSubtarget &STI); 177 void emitGlobals(const Module &M); 178 void emitGlobalAlias(const Module &M, const GlobalAlias &GA) override; 179 void emitHeader(Module &M, raw_ostream &O, const NVPTXSubtarget &STI); 180 void emitKernelFunctionDirectives(const Function &F, raw_ostream &O) const; 181 void emitVirtualRegister(unsigned int vr, raw_ostream &); 182 void emitFunctionParamList(const Function *, raw_ostream &O); 183 void setAndEmitFunctionVirtualRegisters(const MachineFunction &MF); 184 void printReturnValStr(const Function *, raw_ostream &O); 185 void printReturnValStr(const MachineFunction &MF, raw_ostream &O); 186 bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo, 187 const char *ExtraCode, raw_ostream &) override; 188 void printOperand(const MachineInstr *MI, unsigned OpNum, raw_ostream &O); 189 bool PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo, 190 const char *ExtraCode, raw_ostream &) override; 191 192 const MCExpr *lowerConstantForGV(const Constant *CV, bool ProcessingGeneric); 193 void printMCExpr(const MCExpr &Expr, raw_ostream &OS); 194 195 protected: 196 bool doInitialization(Module &M) override; 197 bool doFinalization(Module &M) override; 198 199 private: 200 bool GlobalsEmitted; 201 202 // This is specific per MachineFunction. 203 const MachineRegisterInfo *MRI; 204 // The contents are specific for each 205 // MachineFunction. But the size of the 206 // array is not. 207 typedef DenseMap<unsigned, unsigned> VRegMap; 208 typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap; 209 VRegRCMap VRegMapping; 210 211 // List of variables demoted to a function scope. 212 std::map<const Function *, std::vector<const GlobalVariable *>> localDecls; 213 214 void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O, 215 const NVPTXSubtarget &STI); 216 void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const; 217 std::string getPTXFundamentalTypeStr(Type *Ty, bool = true) const; 218 void printScalarConstant(const Constant *CPV, raw_ostream &O); 219 void printFPConstant(const ConstantFP *Fp, raw_ostream &O); 220 void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer); 221 void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer); 222 223 void emitLinkageDirective(const GlobalValue *V, raw_ostream &O); 224 void emitDeclarations(const Module &, raw_ostream &O); 225 void emitDeclaration(const Function *, raw_ostream &O); 226 void emitAliasDeclaration(const GlobalAlias *, raw_ostream &O); 227 void emitDeclarationWithName(const Function *, MCSymbol *, raw_ostream &O); 228 void emitDemotedVars(const Function *, raw_ostream &); 229 230 bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo, 231 MCOperand &MCOp); 232 void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp); 233 234 bool isLoopHeaderOfNoUnroll(const MachineBasicBlock &MBB) const; 235 236 // Used to control the need to emit .generic() in the initializer of 237 // module scope variables. 238 // Although ptx supports the hybrid mode like the following, 239 // .global .u32 a; 240 // .global .u32 b; 241 // .global .u32 addr[] = {a, generic(b)} 242 // we have difficulty representing the difference in the NVVM IR. 243 // 244 // Since the address value should always be generic in CUDA C and always 245 // be specific in OpenCL, we use this simple control here. 246 // 247 bool EmitGeneric; 248 249 public: 250 NVPTXAsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer) 251 : AsmPrinter(TM, std::move(Streamer)), 252 EmitGeneric(static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == 253 NVPTX::CUDA) {} 254 255 bool runOnMachineFunction(MachineFunction &F) override; 256 257 void getAnalysisUsage(AnalysisUsage &AU) const override { 258 AU.addRequired<MachineLoopInfoWrapperPass>(); 259 AsmPrinter::getAnalysisUsage(AU); 260 } 261 262 std::string getVirtualRegisterName(unsigned) const; 263 264 const MCSymbol *getFunctionFrameSymbol() const override; 265 266 // Make emitGlobalVariable() no-op for NVPTX. 267 // Global variables have been already emitted by the time the base AsmPrinter 268 // attempts to do so in doFinalization() (see NVPTXAsmPrinter::emitGlobals()). 269 void emitGlobalVariable(const GlobalVariable *GV) override {} 270 }; 271 272 } // end namespace llvm 273 274 #endif // LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H 275