NVPTXAsmPrinter.h revision 37ed9c199ca639565f6ce88105f9e39e898d82d0 
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 LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H
16#define LLVM_LIB_TARGET_NVPTX_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 numSymbols;   // number of symbol addresses
90
91  private:
92    const unsigned size;   // size of the buffer in bytes
93    std::vector<unsigned char> buffer; // the buffer
94    SmallVector<unsigned, 4> symbolPosInBuffer;
95    SmallVector<const Value *, 4> Symbols;
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        : size(_size), buffer(_size), O(_O), AP(_AP) {
104      curpos = 0;
105      numSymbols = 0;
106      EmitGeneric = AP.EmitGeneric;
107    }
108    unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) {
109      assert((curpos + Num) <= size);
110      assert((curpos + Bytes) <= size);
111      for (int i = 0; i < Num; ++i) {
112        buffer[curpos] = Ptr[i];
113        curpos++;
114      }
115      for (int i = Num; i < Bytes; ++i) {
116        buffer[curpos] = 0;
117        curpos++;
118      }
119      return curpos;
120    }
121    unsigned addZeros(int Num) {
122      assert((curpos + Num) <= size);
123      for (int i = 0; i < Num; ++i) {
124        buffer[curpos] = 0;
125        curpos++;
126      }
127      return curpos;
128    }
129    void addSymbol(const Value *GVar) {
130      symbolPosInBuffer.push_back(curpos);
131      Symbols.push_back(GVar);
132      numSymbols++;
133    }
134    void print() {
135      if (numSymbols == 0) {
136        // print out in bytes
137        for (unsigned i = 0; i < size; i++) {
138          if (i)
139            O << ", ";
140          O << (unsigned int) buffer[i];
141        }
142      } else {
143        // print out in 4-bytes or 8-bytes
144        unsigned int pos = 0;
145        unsigned int nSym = 0;
146        unsigned int nextSymbolPos = symbolPosInBuffer[nSym];
147        unsigned int nBytes = 4;
148        if (AP.nvptxSubtarget.is64Bit())
149          nBytes = 8;
150        for (pos = 0; pos < size; pos += nBytes) {
151          if (pos)
152            O << ", ";
153          if (pos == nextSymbolPos) {
154            const Value *v = Symbols[nSym];
155            if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
156              MCSymbol *Name = AP.getSymbol(GVar);
157              PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
158              bool IsNonGenericPointer = false;
159              if (PTy && PTy->getAddressSpace() != 0) {
160                IsNonGenericPointer = true;
161              }
162              if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
163                O << "generic(";
164                O << *Name;
165                O << ")";
166              } else {
167                O << *Name;
168              }
169            } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) {
170              O << *nvptx::LowerConstant(Cexpr, AP);
171            } else
172              llvm_unreachable("symbol type unknown");
173            nSym++;
174            if (nSym >= numSymbols)
175              nextSymbolPos = size + 1;
176            else
177              nextSymbolPos = symbolPosInBuffer[nSym];
178          } else if (nBytes == 4)
179            O << *(unsigned int *)(&buffer[pos]);
180          else
181            O << *(unsigned long long *)(&buffer[pos]);
182        }
183      }
184    }
185  };
186
187  friend class AggBuffer;
188
189  void emitSrcInText(StringRef filename, unsigned line);
190
191private:
192  const char *getPassName() const override { return "NVPTX Assembly Printer"; }
193
194  const Function *F;
195  std::string CurrentFnName;
196
197  void EmitFunctionEntryLabel() override;
198  void EmitFunctionBodyStart() override;
199  void EmitFunctionBodyEnd() override;
200  void emitImplicitDef(const MachineInstr *MI) const override;
201
202  void EmitInstruction(const MachineInstr *) override;
203  void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI);
204  bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp);
205  MCOperand GetSymbolRef(const MCSymbol *Symbol);
206  unsigned encodeVirtualRegister(unsigned Reg);
207
208  void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {}
209
210  void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier,
211                                 raw_ostream &O);
212  void printMemOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
213                       const char *Modifier = nullptr);
214  void printImplicitDef(const MachineInstr *MI, raw_ostream &O) const;
215  void printModuleLevelGV(const GlobalVariable *GVar, raw_ostream &O,
216                          bool = false);
217  void printParamName(int paramIndex, raw_ostream &O);
218  void printParamName(Function::const_arg_iterator I, int paramIndex,
219                      raw_ostream &O);
220  void emitGlobals(const Module &M);
221  void emitHeader(Module &M, raw_ostream &O);
222  void emitKernelFunctionDirectives(const Function &F, raw_ostream &O) const;
223  void emitVirtualRegister(unsigned int vr, raw_ostream &);
224  void emitFunctionExternParamList(const MachineFunction &MF);
225  void emitFunctionParamList(const Function *, raw_ostream &O);
226  void emitFunctionParamList(const MachineFunction &MF, raw_ostream &O);
227  void setAndEmitFunctionVirtualRegisters(const MachineFunction &MF);
228  void emitFunctionTempData(const MachineFunction &MF, unsigned &FrameSize);
229  bool isImageType(const Type *Ty);
230  void printReturnValStr(const Function *, raw_ostream &O);
231  void printReturnValStr(const MachineFunction &MF, raw_ostream &O);
232  bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
233                       unsigned AsmVariant, const char *ExtraCode,
234                       raw_ostream &) override;
235  void printOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
236                    const char *Modifier = nullptr);
237  bool PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo,
238                             unsigned AsmVariant, const char *ExtraCode,
239                             raw_ostream &) override;
240protected:
241  bool doInitialization(Module &M) override;
242  bool doFinalization(Module &M) override;
243
244private:
245  std::string CurrentBankselLabelInBasicBlock;
246
247  bool GlobalsEmitted;
248
249  // This is specific per MachineFunction.
250  const MachineRegisterInfo *MRI;
251  // The contents are specific for each
252  // MachineFunction. But the size of the
253  // array is not.
254  typedef DenseMap<unsigned, unsigned> VRegMap;
255  typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap;
256  VRegRCMap VRegMapping;
257  // cache the subtarget here.
258  const NVPTXSubtarget &nvptxSubtarget;
259  // Build the map between type name and ID based on module's type
260  // symbol table.
261  std::map<const Type *, std::string> TypeNameMap;
262
263  // List of variables demoted to a function scope.
264  std::map<const Function *, std::vector<const GlobalVariable *> > localDecls;
265
266  // To record filename to ID mapping
267  std::map<std::string, unsigned> filenameMap;
268  void recordAndEmitFilenames(Module &);
269
270  void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O);
271  void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const;
272  std::string getPTXFundamentalTypeStr(const Type *Ty, bool = true) const;
273  void printScalarConstant(const Constant *CPV, raw_ostream &O);
274  void printFPConstant(const ConstantFP *Fp, raw_ostream &O);
275  void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer);
276  void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer);
277
278  void printOperandProper(const MachineOperand &MO);
279
280  void emitLinkageDirective(const GlobalValue *V, raw_ostream &O);
281  void emitDeclarations(const Module &, raw_ostream &O);
282  void emitDeclaration(const Function *, raw_ostream &O);
283
284  static const char *getRegisterName(unsigned RegNo);
285  void emitDemotedVars(const Function *, raw_ostream &);
286
287  bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
288                               MCOperand &MCOp);
289  void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
290
291  LineReader *reader;
292  LineReader *getReader(std::string);
293
294  // Used to control the need to emit .generic() in the initializer of
295  // module scope variables.
296  // Although ptx supports the hybrid mode like the following,
297  //    .global .u32 a;
298  //    .global .u32 b;
299  //    .global .u32 addr[] = {a, generic(b)}
300  // we have difficulty representing the difference in the NVVM IR.
301  //
302  // Since the address value should always be generic in CUDA C and always
303  // be specific in OpenCL, we use this simple control here.
304  //
305  bool EmitGeneric;
306
307public:
308  NVPTXAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
309      : AsmPrinter(TM, Streamer),
310        nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
311    CurrentBankselLabelInBasicBlock = "";
312    reader = nullptr;
313    EmitGeneric = (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA);
314  }
315
316  ~NVPTXAsmPrinter() {
317    if (!reader)
318      delete reader;
319  }
320
321  bool ignoreLoc(const MachineInstr &);
322
323  std::string getVirtualRegisterName(unsigned) const;
324
325  DebugLoc prevDebugLoc;
326  void emitLineNumberAsDotLoc(const MachineInstr &);
327};
328} // end of namespace
329
330#endif
331