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).
49 std::string theFileName;
50 SmallVector<unsigned, 32> lineOffset;
52 LineReader(std::string filename) {
54 fstr.open(filename.c_str());
55 theFileName = filename;
57 std::string fileName() { return theFileName; }
58 ~LineReader() { fstr.close(); }
59 std::string readLine(unsigned line);
62 class LLVM_LIBRARY_VISIBILITY NVPTXAsmPrinter : public AsmPrinter {
65 // Used to buffer the emitted string for initializing global
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[]
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*.
79 // Once we have this AggBuffer setup, we can choose how to print
82 unsigned numSymbols; // number of symbol addresses
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;
95 AggBuffer(unsigned _size, raw_ostream &_O, NVPTXAsmPrinter &_AP)
96 : size(_size), buffer(_size), O(_O), AP(_AP) {
99 EmitGeneric = AP.EmitGeneric;
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];
108 for (int i = Num; i < Bytes; ++i) {
114 unsigned addZeros(int Num) {
115 assert((curpos + Num) <= size);
116 for (int i = 0; i < Num; ++i) {
122 void addSymbol(const Value *GVar) {
123 symbolPosInBuffer.push_back(curpos);
124 Symbols.push_back(GVar);
128 if (numSymbols == 0) {
129 // print out in bytes
130 for (unsigned i = 0; i < size; i++) {
133 O << (unsigned int) buffer[i];
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 (AP.nvptxSubtarget.is64Bit())
143 for (pos = 0; pos < size; pos += nBytes) {
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;
155 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
162 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) {
163 O << *AP.lowerConstant(Cexpr);
165 llvm_unreachable("symbol type unknown");
167 if (nSym >= numSymbols)
168 nextSymbolPos = size + 1;
170 nextSymbolPos = symbolPosInBuffer[nSym];
171 } else if (nBytes == 4)
172 O << *(unsigned int *)(&buffer[pos]);
174 O << *(unsigned long long *)(&buffer[pos]);
180 friend class AggBuffer;
182 void emitSrcInText(StringRef filename, unsigned line);
185 const char *getPassName() const override { return "NVPTX Assembly Printer"; }
188 std::string CurrentFnName;
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;
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);
202 void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {}
204 void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier,
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,
211 void printParamName(int paramIndex, raw_ostream &O);
212 void printParamName(Function::const_arg_iterator I, int paramIndex,
214 void emitGlobals(const Module &M);
215 void emitHeader(Module &M, raw_ostream &O);
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;
235 bool doInitialization(Module &M) override;
236 bool doFinalization(Module &M) override;
239 std::string CurrentBankselLabelInBasicBlock;
243 // This is specific per MachineFunction.
244 const MachineRegisterInfo *MRI;
245 // The contents are specific for each
246 // MachineFunction. But the size of the
248 typedef DenseMap<unsigned, unsigned> VRegMap;
249 typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap;
250 VRegRCMap VRegMapping;
251 // cache the subtarget here.
252 const NVPTXSubtarget &nvptxSubtarget;
253 // Build the map between type name and ID based on module's type
255 std::map<const Type *, std::string> TypeNameMap;
257 // List of variables demoted to a function scope.
258 std::map<const Function *, std::vector<const GlobalVariable *> > localDecls;
260 // To record filename to ID mapping
261 std::map<std::string, unsigned> filenameMap;
262 void recordAndEmitFilenames(Module &);
264 void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O);
265 void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const;
266 std::string getPTXFundamentalTypeStr(const Type *Ty, bool = true) const;
267 void printScalarConstant(const Constant *CPV, raw_ostream &O);
268 void printFPConstant(const ConstantFP *Fp, raw_ostream &O);
269 void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer);
270 void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer);
272 void printOperandProper(const MachineOperand &MO);
274 void emitLinkageDirective(const GlobalValue *V, raw_ostream &O);
275 void emitDeclarations(const Module &, raw_ostream &O);
276 void emitDeclaration(const Function *, raw_ostream &O);
278 static const char *getRegisterName(unsigned RegNo);
279 void emitDemotedVars(const Function *, raw_ostream &);
281 bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
283 void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
285 bool isLoopHeaderOfNoUnroll(const MachineBasicBlock &MBB) const;
288 LineReader *getReader(std::string);
290 // Used to control the need to emit .generic() in the initializer of
291 // module scope variables.
292 // Although ptx supports the hybrid mode like the following,
295 // .global .u32 addr[] = {a, generic(b)}
296 // we have difficulty representing the difference in the NVVM IR.
298 // Since the address value should always be generic in CUDA C and always
299 // be specific in OpenCL, we use this simple control here.
304 NVPTXAsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer)
305 : AsmPrinter(TM, std::move(Streamer)),
306 nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
307 CurrentBankselLabelInBasicBlock = "";
309 EmitGeneric = (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA);
317 void getAnalysisUsage(AnalysisUsage &AU) const override {
318 AU.addRequired<MachineLoopInfo>();
319 AsmPrinter::getAnalysisUsage(AU);
322 bool ignoreLoc(const MachineInstr &);
324 std::string getVirtualRegisterName(unsigned) const;
326 DebugLoc prevDebugLoc;
327 void emitLineNumberAsDotLoc(const MachineInstr &);
329 } // end of namespace