1 //===-- NVPTXAsmPrinter.h - NVPTX LLVM assembly writer --------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains a printer that converts from our internal representation
11 // of machine-dependent LLVM code to NVPTX assembly language.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H
16 #define LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_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"
33 // The ptx syntax and format is very different from that usually seem in a .s
35 // therefore we are not able to use the MCAsmStreamer interface here.
37 // We are handcrafting the output method here.
39 // A better approach is to clone the MCAsmStreamer to a MCPTXAsmStreamer
40 // (subclass of MCStreamer).
42 // This is defined in AsmPrinter.cpp.
43 // Used to process the constant expressions in initializers.
46 LowerConstant(const llvm::Constant *CV, llvm::AsmPrinter &AP);
56 std::string theFileName;
57 SmallVector<unsigned, 32> lineOffset;
59 LineReader(std::string filename) {
61 fstr.open(filename.c_str());
62 theFileName = filename;
64 std::string fileName() { return theFileName; }
65 ~LineReader() { fstr.close(); }
66 std::string readLine(unsigned line);
69 class LLVM_LIBRARY_VISIBILITY NVPTXAsmPrinter : public AsmPrinter {
72 // Used to buffer the emitted string for initializing global
75 // Normally an aggregate (array, vector or structure) is emitted
76 // as a u8[]. However, if one element/field of the aggregate
77 // is a non-NULL address, then the aggregate is emitted as u32[]
80 // We first layout the aggregate in 'buffer' in bytes, except for
81 // those symbol addresses. For the i-th symbol address in the
82 //aggregate, its corresponding 4-byte or 8-byte elements in 'buffer'
83 // are filled with 0s. symbolPosInBuffer[i-1] records its position
84 // in 'buffer', and Symbols[i-1] records the Value*.
86 // Once we have this AggBuffer setup, we can choose how to print
89 unsigned size; // size of the buffer in bytes
90 unsigned char *buffer; // the buffer
91 unsigned numSymbols; // number of symbol addresses
92 SmallVector<unsigned, 4> symbolPosInBuffer;
93 SmallVector<const Value *, 4> Symbols;
102 AggBuffer(unsigned _size, raw_ostream &_O, NVPTXAsmPrinter &_AP)
104 buffer = new unsigned char[_size];
108 EmitGeneric = AP.EmitGeneric;
110 ~AggBuffer() { delete[] buffer; }
111 unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) {
112 assert((curpos + Num) <= size);
113 assert((curpos + Bytes) <= size);
114 for (int i = 0; i < Num; ++i) {
115 buffer[curpos] = Ptr[i];
118 for (int i = Num; i < Bytes; ++i) {
124 unsigned addZeros(int Num) {
125 assert((curpos + Num) <= size);
126 for (int i = 0; i < Num; ++i) {
132 void addSymbol(const Value *GVar) {
133 symbolPosInBuffer.push_back(curpos);
134 Symbols.push_back(GVar);
138 if (numSymbols == 0) {
139 // print out in bytes
140 for (unsigned i = 0; i < size; i++) {
143 O << (unsigned int) buffer[i];
146 // print out in 4-bytes or 8-bytes
147 unsigned int pos = 0;
148 unsigned int nSym = 0;
149 unsigned int nextSymbolPos = symbolPosInBuffer[nSym];
150 unsigned int nBytes = 4;
151 if (AP.nvptxSubtarget.is64Bit())
153 for (pos = 0; pos < size; pos += nBytes) {
156 if (pos == nextSymbolPos) {
157 const Value *v = Symbols[nSym];
158 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
159 MCSymbol *Name = AP.getSymbol(GVar);
160 PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
161 bool IsNonGenericPointer = false;
162 if (PTy && PTy->getAddressSpace() != 0) {
163 IsNonGenericPointer = true;
165 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
172 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) {
173 O << *nvptx::LowerConstant(Cexpr, AP);
175 llvm_unreachable("symbol type unknown");
177 if (nSym >= numSymbols)
178 nextSymbolPos = size + 1;
180 nextSymbolPos = symbolPosInBuffer[nSym];
181 } else if (nBytes == 4)
182 O << *(unsigned int *)(buffer + pos);
184 O << *(unsigned long long *)(buffer + pos);
190 friend class AggBuffer;
192 void emitSrcInText(StringRef filename, unsigned line);
195 const char *getPassName() const override { return "NVPTX Assembly Printer"; }
198 std::string CurrentFnName;
200 void EmitFunctionEntryLabel() override;
201 void EmitFunctionBodyStart() override;
202 void EmitFunctionBodyEnd() override;
203 void emitImplicitDef(const MachineInstr *MI) const override;
205 void EmitInstruction(const MachineInstr *) override;
206 void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI);
207 bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp);
208 MCOperand GetSymbolRef(const MCSymbol *Symbol);
209 unsigned encodeVirtualRegister(unsigned Reg);
211 void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {}
213 void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier,
215 void printMemOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
216 const char *Modifier = nullptr);
217 void printImplicitDef(const MachineInstr *MI, raw_ostream &O) const;
218 void printModuleLevelGV(const GlobalVariable *GVar, raw_ostream &O,
220 void printParamName(int paramIndex, raw_ostream &O);
221 void printParamName(Function::const_arg_iterator I, int paramIndex,
223 void emitGlobals(const Module &M);
224 void emitHeader(Module &M, raw_ostream &O);
225 void emitKernelFunctionDirectives(const Function &F, raw_ostream &O) const;
226 void emitVirtualRegister(unsigned int vr, raw_ostream &);
227 void emitFunctionExternParamList(const MachineFunction &MF);
228 void emitFunctionParamList(const Function *, raw_ostream &O);
229 void emitFunctionParamList(const MachineFunction &MF, raw_ostream &O);
230 void setAndEmitFunctionVirtualRegisters(const MachineFunction &MF);
231 void emitFunctionTempData(const MachineFunction &MF, unsigned &FrameSize);
232 bool isImageType(const Type *Ty);
233 void printReturnValStr(const Function *, raw_ostream &O);
234 void printReturnValStr(const MachineFunction &MF, raw_ostream &O);
235 bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
236 unsigned AsmVariant, const char *ExtraCode,
237 raw_ostream &) override;
238 void printOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
239 const char *Modifier = nullptr);
240 bool PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo,
241 unsigned AsmVariant, const char *ExtraCode,
242 raw_ostream &) override;
244 bool doInitialization(Module &M) override;
245 bool doFinalization(Module &M) override;
248 std::string CurrentBankselLabelInBasicBlock;
252 // This is specific per MachineFunction.
253 const MachineRegisterInfo *MRI;
254 // The contents are specific for each
255 // MachineFunction. But the size of the
257 typedef DenseMap<unsigned, unsigned> VRegMap;
258 typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap;
259 VRegRCMap VRegMapping;
260 // cache the subtarget here.
261 const NVPTXSubtarget &nvptxSubtarget;
262 // Build the map between type name and ID based on module's type
264 std::map<const Type *, std::string> TypeNameMap;
266 // List of variables demoted to a function scope.
267 std::map<const Function *, std::vector<const GlobalVariable *> > localDecls;
269 // To record filename to ID mapping
270 std::map<std::string, unsigned> filenameMap;
271 void recordAndEmitFilenames(Module &);
273 void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O);
274 void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const;
275 std::string getPTXFundamentalTypeStr(const Type *Ty, bool = true) const;
276 void printScalarConstant(const Constant *CPV, raw_ostream &O);
277 void printFPConstant(const ConstantFP *Fp, raw_ostream &O);
278 void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer);
279 void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer);
281 void printOperandProper(const MachineOperand &MO);
283 void emitLinkageDirective(const GlobalValue *V, raw_ostream &O);
284 void emitDeclarations(const Module &, raw_ostream &O);
285 void emitDeclaration(const Function *, raw_ostream &O);
287 static const char *getRegisterName(unsigned RegNo);
288 void emitDemotedVars(const Function *, raw_ostream &);
290 bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
292 void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
295 LineReader *getReader(std::string);
297 // Used to control the need to emit .generic() in the initializer of
298 // module scope variables.
299 // Although ptx supports the hybrid mode like the following,
302 // .global .u32 addr[] = {a, generic(b)}
303 // we have difficulty representing the difference in the NVVM IR.
305 // Since the address value should always be generic in CUDA C and always
306 // be specific in OpenCL, we use this simple control here.
311 NVPTXAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
312 : AsmPrinter(TM, Streamer),
313 nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
314 CurrentBankselLabelInBasicBlock = "";
316 EmitGeneric = (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA);
324 bool ignoreLoc(const MachineInstr &);
326 std::string getVirtualRegisterName(unsigned) const;
328 DebugLoc prevDebugLoc;
329 void emitLineNumberAsDotLoc(const MachineInstr &);
331 } // end of namespace