NVPTXAsmPrinter.h revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f 
1//===-- NVPTXAsmPrinter.h - NVPTX LLVM assembly writer --------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file contains a printer that converts from our internal representation
11// of machine-dependent LLVM code to NVPTX assembly language.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef NVPTXASMPRINTER_H
16#define NVPTXASMPRINTER_H
17
18#include "NVPTX.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"
31#include <fstream>
32
33// The ptx syntax and format is very different from that usually seem in a .s
34// file,
35// therefore we are not able to use the MCAsmStreamer interface here.
36//
37// We are handcrafting the output method here.
38//
39// A better approach is to clone the MCAsmStreamer to a MCPTXAsmStreamer
40// (subclass of MCStreamer).
41
42// This is defined in AsmPrinter.cpp.
43// Used to process the constant expressions in initializers.
44namespace nvptx {
45const llvm::MCExpr *
46LowerConstant(const llvm::Constant *CV, llvm::AsmPrinter &AP);
47}
48
49namespace llvm {
50
51class LineReader {
52private:
53  unsigned theCurLine;
54  std::ifstream fstr;
55  char buff[512];
56  std::string theFileName;
57  SmallVector<unsigned, 32> lineOffset;
58public:
59  LineReader(std::string filename) {
60    theCurLine = 0;
61    fstr.open(filename.c_str());
62    theFileName = filename;
63  }
64  std::string fileName() { return theFileName; }
65  ~LineReader() { fstr.close(); }
66  std::string readLine(unsigned line);
67};
68
69class LLVM_LIBRARY_VISIBILITY NVPTXAsmPrinter : public AsmPrinter {
70
71  class AggBuffer {
72    // Used to buffer the emitted string for initializing global
73    // aggregates.
74    //
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[]
78    // or u64[].
79    //
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*.
85    //
86    // Once we have this AggBuffer setup, we can choose how to print
87    // it out.
88  public:
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;
94
95  private:
96    unsigned curpos;
97    raw_ostream &O;
98    NVPTXAsmPrinter &AP;
99    bool EmitGeneric;
100
101  public:
102    AggBuffer(unsigned _size, raw_ostream &_O, NVPTXAsmPrinter &_AP)
103        : O(_O), AP(_AP) {
104      buffer = new unsigned char[_size];
105      size = _size;
106      curpos = 0;
107      numSymbols = 0;
108      EmitGeneric = AP.EmitGeneric;
109    }
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];
116        curpos++;
117      }
118      for (int i = Num; i < Bytes; ++i) {
119        buffer[curpos] = 0;
120        curpos++;
121      }
122      return curpos;
123    }
124    unsigned addZeros(int Num) {
125      assert((curpos + Num) <= size);
126      for (int i = 0; i < Num; ++i) {
127        buffer[curpos] = 0;
128        curpos++;
129      }
130      return curpos;
131    }
132    void addSymbol(const Value *GVar) {
133      symbolPosInBuffer.push_back(curpos);
134      Symbols.push_back(GVar);
135      numSymbols++;
136    }
137    void print() {
138      if (numSymbols == 0) {
139        // print out in bytes
140        for (unsigned i = 0; i < size; i++) {
141          if (i)
142            O << ", ";
143          O << (unsigned int) buffer[i];
144        }
145      } else {
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())
152          nBytes = 8;
153        for (pos = 0; pos < size; pos += nBytes) {
154          if (pos)
155            O << ", ";
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;
164              }
165              if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
166                O << "generic(";
167                O << *Name;
168                O << ")";
169              } else {
170                O << *Name;
171              }
172            } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) {
173              O << *nvptx::LowerConstant(Cexpr, AP);
174            } else
175              llvm_unreachable("symbol type unknown");
176            nSym++;
177            if (nSym >= numSymbols)
178              nextSymbolPos = size + 1;
179            else
180              nextSymbolPos = symbolPosInBuffer[nSym];
181          } else if (nBytes == 4)
182            O << *(unsigned int *)(buffer + pos);
183          else
184            O << *(unsigned long long *)(buffer + pos);
185        }
186      }
187    }
188  };
189
190  friend class AggBuffer;
191
192  void emitSrcInText(StringRef filename, unsigned line);
193
194private:
195  const char *getPassName() const override { return "NVPTX Assembly Printer"; }
196
197  const Function *F;
198  std::string CurrentFnName;
199
200  void EmitFunctionEntryLabel() override;
201  void EmitFunctionBodyStart() override;
202  void EmitFunctionBodyEnd() override;
203  void emitImplicitDef(const MachineInstr *MI) const override;
204
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);
210
211  void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {}
212
213  void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier,
214                                 raw_ostream &O);
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,
219                          bool = false);
220  void printParamName(int paramIndex, raw_ostream &O);
221  void printParamName(Function::const_arg_iterator I, int paramIndex,
222                      raw_ostream &O);
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;
243protected:
244  bool doInitialization(Module &M) override;
245  bool doFinalization(Module &M) override;
246
247private:
248  std::string CurrentBankselLabelInBasicBlock;
249
250  bool GlobalsEmitted;
251
252  // This is specific per MachineFunction.
253  const MachineRegisterInfo *MRI;
254  // The contents are specific for each
255  // MachineFunction. But the size of the
256  // array is not.
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
263  // symbol table.
264  std::map<const Type *, std::string> TypeNameMap;
265
266  // List of variables demoted to a function scope.
267  std::map<const Function *, std::vector<const GlobalVariable *> > localDecls;
268
269  // To record filename to ID mapping
270  std::map<std::string, unsigned> filenameMap;
271  void recordAndEmitFilenames(Module &);
272
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);
280
281  void printOperandProper(const MachineOperand &MO);
282
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);
286
287  static const char *getRegisterName(unsigned RegNo);
288  void emitDemotedVars(const Function *, raw_ostream &);
289
290  bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
291                               MCOperand &MCOp);
292  void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
293
294  LineReader *reader;
295  LineReader *getReader(std::string);
296
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,
300  //    .global .u32 a;
301  //    .global .u32 b;
302  //    .global .u32 addr[] = {a, generic(b)}
303  // we have difficulty representing the difference in the NVVM IR.
304  //
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.
307  //
308  bool EmitGeneric;
309
310public:
311  NVPTXAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
312      : AsmPrinter(TM, Streamer),
313        nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
314    CurrentBankselLabelInBasicBlock = "";
315    reader = nullptr;
316    EmitGeneric = (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA);
317  }
318
319  ~NVPTXAsmPrinter() {
320    if (!reader)
321      delete reader;
322  }
323
324  bool ignoreLoc(const MachineInstr &);
325
326  std::string getVirtualRegisterName(unsigned) const;
327
328  DebugLoc prevDebugLoc;
329  void emitLineNumberAsDotLoc(const MachineInstr &);
330};
331} // end of namespace
332
333#endif
334