//
//===----------------------------------------------------------------------===//
-#ifndef NVPTXASMPRINTER_H
-#define NVPTXASMPRINTER_H
+#ifndef LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H
+#define LLVM_LIB_TARGET_NVPTX_NVPTXASMPRINTER_H
#include "NVPTX.h"
#include "NVPTXSubtarget.h"
#include "llvm/MC/MCSymbol.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/FormattedStream.h"
-#include "llvm/Target/Mangler.h"
#include "llvm/Target/TargetMachine.h"
#include <fstream>
// A better approach is to clone the MCAsmStreamer to a MCPTXAsmStreamer
// (subclass of MCStreamer).
-// This is defined in AsmPrinter.cpp.
-// Used to process the constant expressions in initializers.
-namespace nvptx {
-const llvm::MCExpr *
-LowerConstant(const llvm::Constant *CV, llvm::AsmPrinter &AP);
-}
-
namespace llvm {
class LineReader {
// Once we have this AggBuffer setup, we can choose how to print
// it out.
public:
- unsigned size; // size of the buffer in bytes
- unsigned char *buffer; // the buffer
unsigned numSymbols; // number of symbol addresses
- SmallVector<unsigned, 4> symbolPosInBuffer;
- SmallVector<Value *, 4> Symbols;
private:
+ const unsigned size; // size of the buffer in bytes
+ std::vector<unsigned char> buffer; // the buffer
+ SmallVector<unsigned, 4> symbolPosInBuffer;
+ SmallVector<const Value *, 4> Symbols;
unsigned curpos;
raw_ostream &O;
NVPTXAsmPrinter &AP;
+ bool EmitGeneric;
public:
AggBuffer(unsigned _size, raw_ostream &_O, NVPTXAsmPrinter &_AP)
- : O(_O), AP(_AP) {
- buffer = new unsigned char[_size];
- size = _size;
+ : size(_size), buffer(_size), O(_O), AP(_AP) {
curpos = 0;
numSymbols = 0;
+ EmitGeneric = AP.EmitGeneric;
}
- ~AggBuffer() { delete[] buffer; }
unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) {
assert((curpos + Num) <= size);
assert((curpos + Bytes) <= size);
}
return curpos;
}
- void addSymbol(Value *GVar) {
+ void addSymbol(const Value *GVar) {
symbolPosInBuffer.push_back(curpos);
Symbols.push_back(GVar);
numSymbols++;
unsigned int nSym = 0;
unsigned int nextSymbolPos = symbolPosInBuffer[nSym];
unsigned int nBytes = 4;
- if (AP.nvptxSubtarget.is64Bit())
+ if (static_cast<const NVPTXTargetMachine &>(AP.TM).is64Bit())
nBytes = 8;
for (pos = 0; pos < size; pos += nBytes) {
if (pos)
O << ", ";
if (pos == nextSymbolPos) {
- Value *v = Symbols[nSym];
- if (GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
- MCSymbol *Name = AP.Mang->getSymbol(GVar);
- O << *Name;
- } else if (ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) {
- O << *nvptx::LowerConstant(Cexpr, AP);
+ const Value *v = Symbols[nSym];
+ if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
+ MCSymbol *Name = AP.getSymbol(GVar);
+ PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
+ bool IsNonGenericPointer = false;
+ if (PTy && PTy->getAddressSpace() != 0) {
+ IsNonGenericPointer = true;
+ }
+ if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
+ O << "generic(";
+ O << *Name;
+ O << ")";
+ } else {
+ O << *Name;
+ }
+ } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) {
+ O << *AP.lowerConstant(Cexpr);
} else
llvm_unreachable("symbol type unknown");
nSym++;
else
nextSymbolPos = symbolPosInBuffer[nSym];
} else if (nBytes == 4)
- O << *(unsigned int *)(buffer + pos);
+ O << *(unsigned int *)(&buffer[pos]);
else
- O << *(unsigned long long *)(buffer + pos);
+ O << *(unsigned long long *)(&buffer[pos]);
}
}
}
friend class AggBuffer;
- virtual void emitSrcInText(StringRef filename, unsigned line);
+ void emitSrcInText(StringRef filename, unsigned line);
private:
- virtual const char *getPassName() const { return "NVPTX Assembly Printer"; }
+ const char *getPassName() const override { return "NVPTX Assembly Printer"; }
const Function *F;
std::string CurrentFnName;
- void EmitFunctionEntryLabel();
- void EmitFunctionBodyStart();
- void EmitFunctionBodyEnd();
+ void EmitBasicBlockStart(const MachineBasicBlock &MBB) const override;
+ void EmitFunctionEntryLabel() override;
+ void EmitFunctionBodyStart() override;
+ void EmitFunctionBodyEnd() override;
+ void emitImplicitDef(const MachineInstr *MI) const override;
- void EmitInstruction(const MachineInstr *);
+ void EmitInstruction(const MachineInstr *) override;
+ void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI);
+ bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp);
+ MCOperand GetSymbolRef(const MCSymbol *Symbol);
+ unsigned encodeVirtualRegister(unsigned Reg);
- void EmitAlignment(unsigned NumBits, const GlobalValue *GV = 0) const {}
+ void EmitAlignment(unsigned NumBits, const GlobalValue *GV = nullptr) const {}
- void printGlobalVariable(const GlobalVariable *GVar);
- void printOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
- const char *Modifier = 0);
- void printLdStCode(const MachineInstr *MI, int opNum, raw_ostream &O,
- const char *Modifier = 0);
void printVecModifiedImmediate(const MachineOperand &MO, const char *Modifier,
raw_ostream &O);
void printMemOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
- const char *Modifier = 0);
+ const char *Modifier = nullptr);
void printImplicitDef(const MachineInstr *MI, raw_ostream &O) const;
- // definition autogenerated.
- void printInstruction(const MachineInstr *MI, raw_ostream &O);
- void printModuleLevelGV(GlobalVariable *GVar, raw_ostream &O, bool = false);
+ void printModuleLevelGV(const GlobalVariable *GVar, raw_ostream &O,
+ bool = false);
void printParamName(int paramIndex, raw_ostream &O);
void printParamName(Function::const_arg_iterator I, int paramIndex,
raw_ostream &O);
- void emitHeader(Module &M, raw_ostream &O);
+ void emitGlobals(const Module &M);
+ void emitHeader(Module &M, raw_ostream &O, const NVPTXSubtarget &STI);
void emitKernelFunctionDirectives(const Function &F, raw_ostream &O) const;
- void emitVirtualRegister(unsigned int vr, bool isVec, raw_ostream &O);
+ void emitVirtualRegister(unsigned int vr, raw_ostream &);
void emitFunctionExternParamList(const MachineFunction &MF);
void emitFunctionParamList(const Function *, raw_ostream &O);
void emitFunctionParamList(const MachineFunction &MF, raw_ostream &O);
void setAndEmitFunctionVirtualRegisters(const MachineFunction &MF);
void emitFunctionTempData(const MachineFunction &MF, unsigned &FrameSize);
bool isImageType(const Type *Ty);
+ void printReturnValStr(const Function *, raw_ostream &O);
+ void printReturnValStr(const MachineFunction &MF, raw_ostream &O);
bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
unsigned AsmVariant, const char *ExtraCode,
- raw_ostream &);
+ raw_ostream &) override;
+ void printOperand(const MachineInstr *MI, int opNum, raw_ostream &O,
+ const char *Modifier = nullptr);
bool PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo,
unsigned AsmVariant, const char *ExtraCode,
- raw_ostream &);
- void printReturnValStr(const Function *, raw_ostream &O);
- void printReturnValStr(const MachineFunction &MF, raw_ostream &O);
-
+ raw_ostream &) override;
protected:
- bool doInitialization(Module &M);
- bool doFinalization(Module &M);
+ bool doInitialization(Module &M) override;
+ bool doFinalization(Module &M) override;
private:
std::string CurrentBankselLabelInBasicBlock;
+ bool GlobalsEmitted;
+
// This is specific per MachineFunction.
const MachineRegisterInfo *MRI;
// The contents are specific for each
// MachineFunction. But the size of the
// array is not.
- std::map<unsigned, unsigned> *VRidGlobal2LocalMap;
- // cache the subtarget here.
- const NVPTXSubtarget &nvptxSubtarget;
+ typedef DenseMap<unsigned, unsigned> VRegMap;
+ typedef DenseMap<const TargetRegisterClass *, VRegMap> VRegRCMap;
+ VRegRCMap VRegMapping;
+
+ // Cache the subtarget here.
+ const NVPTXSubtarget *nvptxSubtarget;
+
// Build the map between type name and ID based on module's type
// symbol table.
std::map<const Type *, std::string> TypeNameMap;
// List of variables demoted to a function scope.
- std::map<const Function *, std::vector<GlobalVariable *> > localDecls;
+ std::map<const Function *, std::vector<const GlobalVariable *> > localDecls;
// To record filename to ID mapping
std::map<std::string, unsigned> filenameMap;
void emitPTXGlobalVariable(const GlobalVariable *GVar, raw_ostream &O);
void emitPTXAddressSpace(unsigned int AddressSpace, raw_ostream &O) const;
std::string getPTXFundamentalTypeStr(const Type *Ty, bool = true) const;
- void printScalarConstant(Constant *CPV, raw_ostream &O);
+ void printScalarConstant(const Constant *CPV, raw_ostream &O);
void printFPConstant(const ConstantFP *Fp, raw_ostream &O);
- void bufferLEByte(Constant *CPV, int Bytes, AggBuffer *aggBuffer);
- void bufferAggregateConstant(Constant *CV, AggBuffer *aggBuffer);
+ void bufferLEByte(const Constant *CPV, int Bytes, AggBuffer *aggBuffer);
+ void bufferAggregateConstant(const Constant *CV, AggBuffer *aggBuffer);
void printOperandProper(const MachineOperand &MO);
void emitLinkageDirective(const GlobalValue *V, raw_ostream &O);
- void emitDeclarations(Module &, raw_ostream &O);
+ void emitDeclarations(const Module &, raw_ostream &O);
void emitDeclaration(const Function *, raw_ostream &O);
static const char *getRegisterName(unsigned RegNo);
void emitDemotedVars(const Function *, raw_ostream &);
+ bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
+ MCOperand &MCOp);
+ void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
+
+ bool isLoopHeaderOfNoUnroll(const MachineBasicBlock &MBB) const;
+
LineReader *reader;
LineReader *getReader(std::string);
+
+ // Used to control the need to emit .generic() in the initializer of
+ // module scope variables.
+ // Although ptx supports the hybrid mode like the following,
+ // .global .u32 a;
+ // .global .u32 b;
+ // .global .u32 addr[] = {a, generic(b)}
+ // we have difficulty representing the difference in the NVVM IR.
+ //
+ // Since the address value should always be generic in CUDA C and always
+ // be specific in OpenCL, we use this simple control here.
+ //
+ bool EmitGeneric;
+
public:
- NVPTXAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
- : AsmPrinter(TM, Streamer),
- nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
+ NVPTXAsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer)
+ : AsmPrinter(TM, std::move(Streamer)),
+ EmitGeneric(static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() ==
+ NVPTX::CUDA) {
CurrentBankselLabelInBasicBlock = "";
- VRidGlobal2LocalMap = NULL;
- reader = NULL;
+ reader = nullptr;
}
~NVPTXAsmPrinter() {
delete reader;
}
+ bool runOnMachineFunction(MachineFunction &F) override {
+ nvptxSubtarget = &F.getSubtarget<NVPTXSubtarget>();
+ return AsmPrinter::runOnMachineFunction(F);
+ }
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.addRequired<MachineLoopInfo>();
+ AsmPrinter::getAnalysisUsage(AU);
+ }
+
bool ignoreLoc(const MachineInstr &);
- virtual void getVirtualRegisterName(unsigned, bool, raw_ostream &);
+ std::string getVirtualRegisterName(unsigned) const;
DebugLoc prevDebugLoc;
void emitLineNumberAsDotLoc(const MachineInstr &);