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;
89 // SymbolsBeforeStripping[i] is the original form of Symbols[i] before
90 // stripping pointer casts, i.e.,
91 // Symbols[i] == SymbolsBeforeStripping[i]->stripPointerCasts().
93 // We need to keep these values because AggBuffer::print decides whether to
94 // emit a "generic()" cast for Symbols[i] depending on the address space of
95 // SymbolsBeforeStripping[i].
96 SmallVector<const Value *, 4> SymbolsBeforeStripping;
103 AggBuffer(unsigned size, raw_ostream &O, NVPTXAsmPrinter &AP)
104 : size(size), buffer(size), O(O), AP(AP) {
107 EmitGeneric = AP.EmitGeneric;
109 unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) {
110 assert((curpos + Num) <= size);
111 assert((curpos + Bytes) <= size);
112 for (int i = 0; i < Num; ++i) {
113 buffer[curpos] = Ptr[i];
116 for (int i = Num; i < Bytes; ++i) {
122 unsigned addZeros(int Num) {
123 assert((curpos + Num) <= size);
124 for (int i = 0; i < Num; ++i) {
130 void addSymbol(const Value *GVar, const Value *GVarBeforeStripping) {
131 symbolPosInBuffer.push_back(curpos);
132 Symbols.push_back(GVar);
133 SymbolsBeforeStripping.push_back(GVarBeforeStripping);
137 if (numSymbols == 0) {
138 // print out in bytes
139 for (unsigned i = 0; i < size; i++) {
142 O << (unsigned int) buffer[i];
145 // print out in 4-bytes or 8-bytes
146 unsigned int pos = 0;
147 unsigned int nSym = 0;
148 unsigned int nextSymbolPos = symbolPosInBuffer[nSym];
149 unsigned int nBytes = 4;
150 if (static_cast<const NVPTXTargetMachine &>(AP.TM).is64Bit())
152 for (pos = 0; pos < size; pos += nBytes) {
155 if (pos == nextSymbolPos) {
156 const Value *v = Symbols[nSym];
157 const Value *v0 = SymbolsBeforeStripping[nSym];
158 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
159 MCSymbol *Name = AP.getSymbol(GVar);
160 PointerType *PTy = dyn_cast<PointerType>(v0->getType());
161 bool IsNonGenericPointer = false; // Is v0 a non-generic pointer?
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 << *AP.lowerConstant(Cexpr);
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 EmitBasicBlockStart(const MachineBasicBlock &MBB) const override;
201 void EmitFunctionEntryLabel() override;
202 void EmitFunctionBodyStart() override;
203 void EmitFunctionBodyEnd() override;
204 void emitImplicitDef(const MachineInstr *MI) const override;
206 void EmitInstruction(const MachineInstr *) override;
207 void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI);
208 bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp);
209 MCOperand GetSymbolRef(const MCSymbol *Symbol);
210 unsigned encodeVirtualRegister(unsigned Reg);
212 void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {}
214 void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier,
216 void printMemOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
217 const char *Modifier = nullptr);
218 void printImplicitDef(const MachineInstr *MI, raw_ostream &O) const;
219 void printModuleLevelGV(const GlobalVariable *GVar, raw_ostream &O,
221 void printParamName(int paramIndex, raw_ostream &O);
222 void printParamName(Function::const_arg_iterator I, int paramIndex,
224 void emitGlobals(const Module &M);
225 void emitHeader(Module &M, raw_ostream &O, const NVPTXSubtarget &STI);
226 void emitKernelFunctionDirectives(const Function &F, raw_ostream &O) const;
227 void emitVirtualRegister(unsigned int vr, raw_ostream &);
228 void emitFunctionExternParamList(const MachineFunction &MF);
229 void emitFunctionParamList(const Function *, raw_ostream &O);
230 void emitFunctionParamList(const MachineFunction &MF, raw_ostream &O);
231 void setAndEmitFunctionVirtualRegisters(const MachineFunction &MF);
232 void emitFunctionTempData(const MachineFunction &MF, unsigned &FrameSize);
233 bool isImageType(const Type *Ty);
234 void printReturnValStr(const Function *, raw_ostream &O);
235 void printReturnValStr(const MachineFunction &MF, raw_ostream &O);
236 bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
237 unsigned AsmVariant, const char *ExtraCode,
238 raw_ostream &) override;
239 void printOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
240 const char *Modifier = nullptr);
241 bool PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo,
242 unsigned AsmVariant, const char *ExtraCode,
243 raw_ostream &) override;
245 bool doInitialization(Module &M) override;
246 bool doFinalization(Module &M) override;
249 std::string CurrentBankselLabelInBasicBlock;
253 // This is specific per MachineFunction.
254 const MachineRegisterInfo *MRI;
255 // The contents are specific for each
256 // MachineFunction. But the size of the
258 typedef DenseMap<unsigned, unsigned> VRegMap;
259 typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap;
260 VRegRCMap VRegMapping;
262 // Cache the subtarget here.
263 const NVPTXSubtarget *nvptxSubtarget;
265 // Build the map between type name and ID based on module's type
267 std::map<const Type *, std::string> TypeNameMap;
269 // List of variables demoted to a function scope.
270 std::map<const Function *, std::vector<const GlobalVariable *> > localDecls;
272 // To record filename to ID mapping
273 std::map<std::string, unsigned> filenameMap;
274 void recordAndEmitFilenames(Module &);
276 void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O);
277 void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const;
278 std::string getPTXFundamentalTypeStr(const Type *Ty, bool = true) const;
279 void printScalarConstant(const Constant *CPV, raw_ostream &O);
280 void printFPConstant(const ConstantFP *Fp, raw_ostream &O);
281 void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer);
282 void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer);
284 void printOperandProper(const MachineOperand &MO);
286 void emitLinkageDirective(const GlobalValue *V, raw_ostream &O);
287 void emitDeclarations(const Module &, raw_ostream &O);
288 void emitDeclaration(const Function *, raw_ostream &O);
290 static const char *getRegisterName(unsigned RegNo);
291 void emitDemotedVars(const Function *, raw_ostream &);
293 bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
295 void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
297 bool isLoopHeaderOfNoUnroll(const MachineBasicBlock &MBB) const;
300 LineReader *getReader(std::string);
302 // Used to control the need to emit .generic() in the initializer of
303 // module scope variables.
304 // Although ptx supports the hybrid mode like the following,
307 // .global .u32 addr[] = {a, generic(b)}
308 // we have difficulty representing the difference in the NVVM IR.
310 // Since the address value should always be generic in CUDA C and always
311 // be specific in OpenCL, we use this simple control here.
316 NVPTXAsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer)
317 : AsmPrinter(TM, std::move(Streamer)),
318 EmitGeneric(static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() ==
320 CurrentBankselLabelInBasicBlock = "";
329 bool runOnMachineFunction(MachineFunction &F) override {
330 nvptxSubtarget = &F.getSubtarget<NVPTXSubtarget>();
331 return AsmPrinter::runOnMachineFunction(F);
333 void getAnalysisUsage(AnalysisUsage &AU) const override {
334 AU.addRequired<MachineLoopInfo>();
335 AsmPrinter::getAnalysisUsage(AU);
338 bool ignoreLoc(const MachineInstr &);
340 std::string getVirtualRegisterName(unsigned) const;
342 DebugLoc prevDebugLoc;
343 void emitLineNumberAsDotLoc(const MachineInstr &);
345 } // end of namespace