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).
50 std::string theFileName;
51 SmallVector<unsigned, 32> lineOffset;
53 LineReader(std::string filename) {
55 fstr.open(filename.c_str());
56 theFileName = filename;
58 std::string fileName() { return theFileName; }
59 ~LineReader() { fstr.close(); }
60 std::string readLine(unsigned line);
63 class LLVM_LIBRARY_VISIBILITY NVPTXAsmPrinter : public AsmPrinter {
66 // Used to buffer the emitted string for initializing global
69 // Normally an aggregate (array, vector or structure) is emitted
70 // as a u8[]. However, if one element/field of the aggregate
71 // is a non-NULL address, then the aggregate is emitted as u32[]
74 // We first layout the aggregate in 'buffer' in bytes, except for
75 // those symbol addresses. For the i-th symbol address in the
76 //aggregate, its corresponding 4-byte or 8-byte elements in 'buffer'
77 // are filled with 0s. symbolPosInBuffer[i-1] records its position
78 // in 'buffer', and Symbols[i-1] records the Value*.
80 // Once we have this AggBuffer setup, we can choose how to print
83 unsigned numSymbols; // number of symbol addresses
86 const unsigned size; // size of the buffer in bytes
87 std::vector<unsigned char> buffer; // the buffer
88 SmallVector<unsigned, 4> symbolPosInBuffer;
89 SmallVector<const Value *, 4> Symbols;
90 // SymbolsBeforeStripping[i] is the original form of Symbols[i] before
91 // stripping pointer casts, i.e.,
92 // Symbols[i] == SymbolsBeforeStripping[i]->stripPointerCasts().
94 // We need to keep these values because AggBuffer::print decides whether to
95 // emit a "generic()" cast for Symbols[i] depending on the address space of
96 // SymbolsBeforeStripping[i].
97 SmallVector<const Value *, 4> SymbolsBeforeStripping;
104 AggBuffer(unsigned size, raw_ostream &O, NVPTXAsmPrinter &AP)
105 : size(size), buffer(size), O(O), AP(AP) {
108 EmitGeneric = AP.EmitGeneric;
110 unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) {
111 assert((curpos + Num) <= size);
112 assert((curpos + Bytes) <= size);
113 for (int i = 0; i < Num; ++i) {
114 buffer[curpos] = Ptr[i];
117 for (int i = Num; i < Bytes; ++i) {
123 unsigned addZeros(int Num) {
124 assert((curpos + Num) <= size);
125 for (int i = 0; i < Num; ++i) {
131 void addSymbol(const Value *GVar, const Value *GVarBeforeStripping) {
132 symbolPosInBuffer.push_back(curpos);
133 Symbols.push_back(GVar);
134 SymbolsBeforeStripping.push_back(GVarBeforeStripping);
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 (static_cast<const NVPTXTargetMachine &>(AP.TM).is64Bit())
153 for (pos = 0; pos < size; pos += nBytes) {
156 if (pos == nextSymbolPos) {
157 const Value *v = Symbols[nSym];
158 const Value *v0 = SymbolsBeforeStripping[nSym];
159 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
160 MCSymbol *Name = AP.getSymbol(GVar);
161 PointerType *PTy = dyn_cast<PointerType>(v0->getType());
162 bool IsNonGenericPointer = false; // Is v0 a non-generic pointer?
163 if (PTy && PTy->getAddressSpace() != 0) {
164 IsNonGenericPointer = true;
166 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
168 Name->print(O, AP.MAI);
171 Name->print(O, AP.MAI);
173 } else if (const ConstantExpr *CExpr = dyn_cast<ConstantExpr>(v0)) {
175 AP.lowerConstantForGV(cast<Constant>(CExpr), false);
176 AP.printMCExpr(*Expr, O);
178 llvm_unreachable("symbol type unknown");
180 if (nSym >= numSymbols)
181 nextSymbolPos = size + 1;
183 nextSymbolPos = symbolPosInBuffer[nSym];
184 } else if (nBytes == 4)
185 O << *(unsigned int *)(&buffer[pos]);
187 O << *(unsigned long long *)(&buffer[pos]);
193 friend class AggBuffer;
195 void emitSrcInText(StringRef filename, unsigned line);
198 const char *getPassName() const override { return "NVPTX Assembly Printer"; }
201 std::string CurrentFnName;
203 void EmitBasicBlockStart(const MachineBasicBlock &MBB) const override;
204 void EmitFunctionEntryLabel() override;
205 void EmitFunctionBodyStart() override;
206 void EmitFunctionBodyEnd() override;
207 void emitImplicitDef(const MachineInstr *MI) const override;
209 void EmitInstruction(const MachineInstr *) override;
210 void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI);
211 bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp);
212 MCOperand GetSymbolRef(const MCSymbol *Symbol);
213 unsigned encodeVirtualRegister(unsigned Reg);
215 void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {}
217 void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier,
219 void printMemOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
220 const char *Modifier = nullptr);
221 void printImplicitDef(const MachineInstr *MI, raw_ostream &O) const;
222 void printModuleLevelGV(const GlobalVariable *GVar, raw_ostream &O,
224 void printParamName(int paramIndex, raw_ostream &O);
225 void printParamName(Function::const_arg_iterator I, int paramIndex,
227 void emitGlobals(const Module &M);
228 void emitHeader(Module &M, raw_ostream &O, const NVPTXSubtarget &STI);
229 void emitKernelFunctionDirectives(const Function &F, raw_ostream &O) const;
230 void emitVirtualRegister(unsigned int vr, raw_ostream &);
231 void emitFunctionExternParamList(const MachineFunction &MF);
232 void emitFunctionParamList(const Function *, raw_ostream &O);
233 void emitFunctionParamList(const MachineFunction &MF, raw_ostream &O);
234 void setAndEmitFunctionVirtualRegisters(const MachineFunction &MF);
235 void emitFunctionTempData(const MachineFunction &MF, unsigned &FrameSize);
236 bool isImageType(const Type *Ty);
237 void printReturnValStr(const Function *, raw_ostream &O);
238 void printReturnValStr(const MachineFunction &MF, raw_ostream &O);
239 bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
240 unsigned AsmVariant, const char *ExtraCode,
241 raw_ostream &) override;
242 void printOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
243 const char *Modifier = nullptr);
244 bool PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo,
245 unsigned AsmVariant, const char *ExtraCode,
246 raw_ostream &) override;
248 const MCExpr *lowerConstantForGV(const Constant *CV, bool ProcessingGeneric);
249 void printMCExpr(const MCExpr &Expr, raw_ostream &OS);
252 bool doInitialization(Module &M) override;
253 bool doFinalization(Module &M) override;
256 std::string CurrentBankselLabelInBasicBlock;
260 // This is specific per MachineFunction.
261 const MachineRegisterInfo *MRI;
262 // The contents are specific for each
263 // MachineFunction. But the size of the
265 typedef DenseMap<unsigned, unsigned> VRegMap;
266 typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap;
267 VRegRCMap VRegMapping;
269 // Cache the subtarget here.
270 const NVPTXSubtarget *nvptxSubtarget;
272 // Build the map between type name and ID based on module's type
274 std::map<const Type *, std::string> TypeNameMap;
276 // List of variables demoted to a function scope.
277 std::map<const Function *, std::vector<const GlobalVariable *> > localDecls;
279 // To record filename to ID mapping
280 std::map<std::string, unsigned> filenameMap;
281 void recordAndEmitFilenames(Module &);
283 void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O);
284 void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const;
285 std::string getPTXFundamentalTypeStr(const Type *Ty, bool = true) const;
286 void printScalarConstant(const Constant *CPV, raw_ostream &O);
287 void printFPConstant(const ConstantFP *Fp, raw_ostream &O);
288 void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer);
289 void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer);
291 void printOperandProper(const MachineOperand &MO);
293 void emitLinkageDirective(const GlobalValue *V, raw_ostream &O);
294 void emitDeclarations(const Module &, raw_ostream &O);
295 void emitDeclaration(const Function *, raw_ostream &O);
297 static const char *getRegisterName(unsigned RegNo);
298 void emitDemotedVars(const Function *, raw_ostream &);
300 bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
302 void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
304 bool isLoopHeaderOfNoUnroll(const MachineBasicBlock &MBB) const;
307 LineReader *getReader(std::string);
309 // Used to control the need to emit .generic() in the initializer of
310 // module scope variables.
311 // Although ptx supports the hybrid mode like the following,
314 // .global .u32 addr[] = {a, generic(b)}
315 // we have difficulty representing the difference in the NVVM IR.
317 // Since the address value should always be generic in CUDA C and always
318 // be specific in OpenCL, we use this simple control here.
323 NVPTXAsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer)
324 : AsmPrinter(TM, std::move(Streamer)),
325 EmitGeneric(static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() ==
327 CurrentBankselLabelInBasicBlock = "";
336 bool runOnMachineFunction(MachineFunction &F) override {
337 nvptxSubtarget = &F.getSubtarget<NVPTXSubtarget>();
338 return AsmPrinter::runOnMachineFunction(F);
340 void getAnalysisUsage(AnalysisUsage &AU) const override {
341 AU.addRequired<MachineLoopInfo>();
342 AsmPrinter::getAnalysisUsage(AU);
345 bool ignoreLoc(const MachineInstr &);
347 std::string getVirtualRegisterName(unsigned) const;
349 DebugLoc prevDebugLoc;
350 void emitLineNumberAsDotLoc(const MachineInstr &);
352 } // end of namespace