X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=utils%2FTableGen%2FAsmWriterEmitter.cpp;h=4b543d3f9fc65b49c09ad38bf650b24df7cc6460;hb=47f0e3f434e2e43f951c3a826c40906cb15b7285;hp=448ebad91f09194e255bcd264b78a97a2c55e97a;hpb=67db883487fca3472fdde51e931657e22d4d0495;p=oota-llvm.git diff --git a/utils/TableGen/AsmWriterEmitter.cpp b/utils/TableGen/AsmWriterEmitter.cpp index 448ebad91f0..4b543d3f9fc 100644 --- a/utils/TableGen/AsmWriterEmitter.cpp +++ b/utils/TableGen/AsmWriterEmitter.cpp @@ -12,16 +12,57 @@ // //===----------------------------------------------------------------------===// -#include "AsmWriterEmitter.h" #include "AsmWriterInst.h" #include "CodeGenTarget.h" -#include "Record.h" -#include "StringToOffsetTable.h" +#include "SequenceToOffsetTable.h" +#include "llvm/ADT/SmallString.h" +#include "llvm/ADT/StringExtras.h" +#include "llvm/ADT/Twine.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/Format.h" #include "llvm/Support/MathExtras.h" +#include "llvm/TableGen/Error.h" +#include "llvm/TableGen/Record.h" +#include "llvm/TableGen/TableGenBackend.h" #include +#include +#include +#include using namespace llvm; +#define DEBUG_TYPE "asm-writer-emitter" + +namespace { +class AsmWriterEmitter { + RecordKeeper &Records; + CodeGenTarget Target; + std::map CGIAWIMap; + const std::vector *NumberedInstructions; + std::vector Instructions; + std::vector PrintMethods; +public: + AsmWriterEmitter(RecordKeeper &R); + + void run(raw_ostream &o); + +private: + void EmitPrintInstruction(raw_ostream &o); + void EmitGetRegisterName(raw_ostream &o); + void EmitPrintAliasInstruction(raw_ostream &O); + + AsmWriterInst *getAsmWriterInstByID(unsigned ID) const { + assert(ID < NumberedInstructions->size()); + std::map::const_iterator I = + CGIAWIMap.find(NumberedInstructions->at(ID)); + assert(I != CGIAWIMap.end() && "Didn't find inst!"); + return I->second; + } + void FindUniqueOperandCommands(std::vector &UOC, + std::vector &InstIdxs, + std::vector &InstOpsUsed) const; +}; +} // end anonymous namespace + static void PrintCases(std::vector > &OpsToPrint, raw_ostream &O) { O << " case " << OpsToPrint.back().first << ": "; @@ -104,7 +145,7 @@ void AsmWriterEmitter:: FindUniqueOperandCommands(std::vector &UniqueOperandCommands, std::vector &InstIdxs, std::vector &InstOpsUsed) const { - InstIdxs.assign(NumberedInstructions.size(), ~0U); + InstIdxs.assign(NumberedInstructions->size(), ~0U); // This vector parallels UniqueOperandCommands, keeping track of which // instructions each case are used for. It is a comma separated string of @@ -113,9 +154,10 @@ FindUniqueOperandCommands(std::vector &UniqueOperandCommands, InstrsForCase.resize(UniqueOperandCommands.size()); InstOpsUsed.assign(UniqueOperandCommands.size(), 0); - for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) { + for (unsigned i = 0, e = NumberedInstructions->size(); i != e; ++i) { const AsmWriterInst *Inst = getAsmWriterInstByID(i); - if (Inst == 0) continue; // PHI, INLINEASM, PROLOG_LABEL, etc. + if (!Inst) + continue; // PHI, INLINEASM, CFI_INSTRUCTION, etc. std::string Command; if (Inst->Operands.empty()) @@ -166,9 +208,6 @@ FindUniqueOperandCommands(std::vector &UniqueOperandCommands, // Otherwise, scan to see if all of the other instructions in this command // set share the operand. bool AllSame = true; - // Keep track of the maximum, number of operands or any - // instruction we see in the group. - size_t MaxSize = FirstInst->Operands.size(); for (NIT = std::find(NIT+1, InstIdxs.end(), CommandIdx); NIT != InstIdxs.end(); @@ -178,10 +217,6 @@ FindUniqueOperandCommands(std::vector &UniqueOperandCommands, const AsmWriterInst *OtherInst = getAsmWriterInstByID(NIT-InstIdxs.begin()); - if (OtherInst && - OtherInst->Operands.size() > FirstInst->Operands.size()) - MaxSize = std::max(MaxSize, OtherInst->Operands.size()); - if (!OtherInst || OtherInst->Operands.size() == Op || OtherInst->Operands[Op] != FirstInst->Operands[Op]) { AllSame = false; @@ -238,65 +273,62 @@ static void UnescapeString(std::string &Str) { } /// EmitPrintInstruction - Generate the code for the "printInstruction" method -/// implementation. +/// implementation. Destroys all instances of AsmWriterInst information, by +/// clearing the Instructions vector. void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { - CodeGenTarget Target(Records); Record *AsmWriter = Target.getAsmWriter(); std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName"); - bool isMC = AsmWriter->getValueAsBit("isMCAsmWriter"); - const char *MachineInstrClassName = isMC ? "MCInst" : "MachineInstr"; + unsigned PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget"); O << "/// printInstruction - This method is automatically generated by tablegen\n" "/// from the instruction set description.\n" "void " << Target.getName() << ClassName - << "::printInstruction(const " << MachineInstrClassName - << " *MI, raw_ostream &O) {\n"; - - std::vector Instructions; - - for (CodeGenTarget::inst_iterator I = Target.inst_begin(), - E = Target.inst_end(); I != E; ++I) - if (!(*I)->AsmString.empty() && - (*I)->TheDef->getName() != "PHI") - Instructions.push_back( - AsmWriterInst(**I, - AsmWriter->getValueAsInt("Variant"), - AsmWriter->getValueAsInt("FirstOperandColumn"), - AsmWriter->getValueAsInt("OperandSpacing"))); - - // Get the instruction numbering. - NumberedInstructions = Target.getInstructionsByEnumValue(); - - // Compute the CodeGenInstruction -> AsmWriterInst mapping. Note that not - // all machine instructions are necessarily being printed, so there may be - // target instructions not in this map. - for (unsigned i = 0, e = Instructions.size(); i != e; ++i) - CGIAWIMap.insert(std::make_pair(Instructions[i].CGI, &Instructions[i])); + << "::printInstruction(const MCInst *MI, " + << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "") + << "raw_ostream &O) {\n"; // Build an aggregate string, and build a table of offsets into it. - StringToOffsetTable StringTable; + SequenceToOffsetTable StringTable; /// OpcodeInfo - This encodes the index of the string to use for the first /// chunk of the output as well as indices used for operand printing. - std::vector OpcodeInfo; + /// To reduce the number of unhandled cases, we expand the size from 32-bit + /// to 32+16 = 48-bit. + std::vector OpcodeInfo; + + // Add all strings to the string table upfront so it can generate an optimized + // representation. + for (unsigned i = 0, e = NumberedInstructions->size(); i != e; ++i) { + AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions->at(i)]; + if (AWI && + AWI->Operands[0].OperandType == + AsmWriterOperand::isLiteralTextOperand && + !AWI->Operands[0].Str.empty()) { + std::string Str = AWI->Operands[0].Str; + UnescapeString(Str); + StringTable.add(Str); + } + } + + StringTable.layout(); unsigned MaxStringIdx = 0; - for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) { - AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions[i]]; + for (unsigned i = 0, e = NumberedInstructions->size(); i != e; ++i) { + AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions->at(i)]; unsigned Idx; - if (AWI == 0) { + if (!AWI) { // Something not handled by the asmwriter printer. Idx = ~0U; } else if (AWI->Operands[0].OperandType != AsmWriterOperand::isLiteralTextOperand || AWI->Operands[0].Str.empty()) { // Something handled by the asmwriter printer, but with no leading string. - Idx = StringTable.GetOrAddStringOffset(""); + Idx = StringTable.get(""); } else { std::string Str = AWI->Operands[0].Str; UnescapeString(Str); - Idx = StringTable.GetOrAddStringOffset(Str); + Idx = StringTable.get(Str); MaxStringIdx = std::max(MaxStringIdx, Idx); // Nuke the string from the operand list. It is now handled! @@ -312,9 +344,9 @@ void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { // To reduce code size, we compactify common instructions into a few bits // in the opcode-indexed table. - unsigned BitsLeft = 32-AsmStrBits; + unsigned BitsLeft = 64-AsmStrBits; - std::vector > TableDrivenOperandPrinters; + std::vector> TableDrivenOperandPrinters; while (1) { std::vector UniqueOperandCommands; @@ -338,13 +370,14 @@ void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { } // Otherwise, we can include this in the initial lookup table. Add it in. - BitsLeft -= NumBits; for (unsigned i = 0, e = InstIdxs.size(); i != e; ++i) - if (InstIdxs[i] != ~0U) - OpcodeInfo[i] |= InstIdxs[i] << (BitsLeft+AsmStrBits); + if (InstIdxs[i] != ~0U) { + OpcodeInfo[i] |= (uint64_t)InstIdxs[i] << (64-BitsLeft); + } + BitsLeft -= NumBits; // Remove the info about this operand. - for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) { + for (unsigned i = 0, e = NumberedInstructions->size(); i != e; ++i) { if (AsmWriterInst *Inst = getAsmWriterInstByID(i)) if (!Inst->Operands.empty()) { unsigned NumOps = NumInstOpsHandled[InstIdxs[i]]; @@ -356,34 +389,59 @@ void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { } // Remember the handlers for this set of operands. - TableDrivenOperandPrinters.push_back(UniqueOperandCommands); + TableDrivenOperandPrinters.push_back(std::move(UniqueOperandCommands)); } - - O<<" static const unsigned OpInfo[] = {\n"; - for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) { - O << " " << OpcodeInfo[i] << "U,\t// " - << NumberedInstructions[i]->TheDef->getName() << "\n"; + // We always emit at least one 32-bit table. A second table is emitted if + // more bits are needed. + O<<" static const uint32_t OpInfo[] = {\n"; + for (unsigned i = 0, e = NumberedInstructions->size(); i != e; ++i) { + O << " " << (OpcodeInfo[i] & 0xffffffff) << "U,\t// " + << NumberedInstructions->at(i)->TheDef->getName() << "\n"; } // Add a dummy entry so the array init doesn't end with a comma. O << " 0U\n"; O << " };\n\n"; + if (BitsLeft < 32) { + // Add a second OpInfo table only when it is necessary. + // Adjust the type of the second table based on the number of bits needed. + O << " static const uint" + << ((BitsLeft < 16) ? "32" : (BitsLeft < 24) ? "16" : "8") + << "_t OpInfo2[] = {\n"; + for (unsigned i = 0, e = NumberedInstructions->size(); i != e; ++i) { + O << " " << (OpcodeInfo[i] >> 32) << "U,\t// " + << NumberedInstructions->at(i)->TheDef->getName() << "\n"; + } + // Add a dummy entry so the array init doesn't end with a comma. + O << " 0U\n"; + O << " };\n\n"; + } + // Emit the string itself. - O << " const char *AsmStrs = \n"; - StringTable.EmitString(O); - O << ";\n\n"; + O << " static const char AsmStrs[] = {\n"; + StringTable.emit(O, printChar); + O << " };\n\n"; O << " O << \"\\t\";\n\n"; - O << " // Emit the opcode for the instruction.\n" - << " unsigned Bits = OpInfo[MI->getOpcode()];\n" - << " assert(Bits != 0 && \"Cannot print this instruction.\");\n" + O << " // Emit the opcode for the instruction.\n"; + if (BitsLeft < 32) { + // If we have two tables then we need to perform two lookups and combine + // the results into a single 64-bit value. + O << " uint64_t Bits1 = OpInfo[MI->getOpcode()];\n" + << " uint64_t Bits2 = OpInfo2[MI->getOpcode()];\n" + << " uint64_t Bits = (Bits2 << 32) | Bits1;\n"; + } else { + // If only one table is used we just need to perform a single lookup. + O << " uint32_t Bits = OpInfo[MI->getOpcode()];\n"; + } + O << " assert(Bits != 0 && \"Cannot print this instruction.\");\n" << " O << AsmStrs+(Bits & " << (1 << AsmStrBits)-1 << ")-1;\n\n"; // Output the table driven operand information. - BitsLeft = 32-AsmStrBits; + BitsLeft = 64-AsmStrBits; for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) { std::vector &Commands = TableDrivenOperandPrinters[i]; @@ -393,14 +451,13 @@ void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { assert(NumBits <= BitsLeft && "consistency error"); // Emit code to extract this field from Bits. - BitsLeft -= NumBits; - O << "\n // Fragment " << i << " encoded into " << NumBits << " bits for " << Commands.size() << " unique commands.\n"; if (Commands.size() == 2) { // Emit two possibilitys with if/else. - O << " if ((Bits >> " << (BitsLeft+AsmStrBits) << ") & " + O << " if ((Bits >> " + << (64-BitsLeft) << ") & " << ((1 << NumBits)-1) << ") {\n" << Commands[1] << " } else {\n" @@ -410,9 +467,10 @@ void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { // Emit a single possibility. O << Commands[0] << "\n\n"; } else { - O << " switch ((Bits >> " << (BitsLeft+AsmStrBits) << ") & " + O << " switch ((Bits >> " + << (64-BitsLeft) << ") & " << ((1 << NumBits)-1) << ") {\n" - << " default: // unreachable.\n"; + << " default: llvm_unreachable(\"Invalid command number.\");\n"; // Print out all the cases. for (unsigned i = 0, e = Commands.size(); i != e; ++i) { @@ -422,6 +480,7 @@ void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { } O << " }\n\n"; } + BitsLeft -= NumBits; } // Okay, delete instructions with no operand info left. @@ -457,98 +516,611 @@ void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) { O << "}\n"; } +static const char *getMinimalTypeForRange(uint64_t Range) { + assert(Range < 0xFFFFFFFFULL && "Enum too large"); + if (Range > 0xFFFF) + return "uint32_t"; + if (Range > 0xFF) + return "uint16_t"; + return "uint8_t"; +} + +static void +emitRegisterNameString(raw_ostream &O, StringRef AltName, + const std::deque &Registers) { + SequenceToOffsetTable StringTable; + SmallVector AsmNames(Registers.size()); + unsigned i = 0; + for (const auto &Reg : Registers) { + std::string &AsmName = AsmNames[i++]; + + // "NoRegAltName" is special. We don't need to do a lookup for that, + // as it's just a reference to the default register name. + if (AltName == "" || AltName == "NoRegAltName") { + AsmName = Reg.TheDef->getValueAsString("AsmName"); + if (AsmName.empty()) + AsmName = Reg.getName(); + } else { + // Make sure the register has an alternate name for this index. + std::vector AltNameList = + Reg.TheDef->getValueAsListOfDefs("RegAltNameIndices"); + unsigned Idx = 0, e; + for (e = AltNameList.size(); + Idx < e && (AltNameList[Idx]->getName() != AltName); + ++Idx) + ; + // If the register has an alternate name for this index, use it. + // Otherwise, leave it empty as an error flag. + if (Idx < e) { + std::vector AltNames = + Reg.TheDef->getValueAsListOfStrings("AltNames"); + if (AltNames.size() <= Idx) + PrintFatalError(Reg.TheDef->getLoc(), + "Register definition missing alt name for '" + + AltName + "'."); + AsmName = AltNames[Idx]; + } + } + StringTable.add(AsmName); + } + + StringTable.layout(); + O << " static const char AsmStrs" << AltName << "[] = {\n"; + StringTable.emit(O, printChar); + O << " };\n\n"; + + O << " static const " << getMinimalTypeForRange(StringTable.size()-1) + << " RegAsmOffset" << AltName << "[] = {"; + for (unsigned i = 0, e = Registers.size(); i != e; ++i) { + if ((i % 14) == 0) + O << "\n "; + O << StringTable.get(AsmNames[i]) << ", "; + } + O << "\n };\n" + << "\n"; +} void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) { - CodeGenTarget Target(Records); Record *AsmWriter = Target.getAsmWriter(); std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName"); - const std::vector &Registers = Target.getRegisters(); + const auto &Registers = Target.getRegBank().getRegisters(); + std::vector AltNameIndices = Target.getRegAltNameIndices(); + bool hasAltNames = AltNameIndices.size() > 1; - StringToOffsetTable StringTable; O << "\n\n/// getRegisterName - This method is automatically generated by tblgen\n" "/// from the register set description. This returns the assembler name\n" "/// for the specified register.\n" - "const char *" << Target.getName() << ClassName - << "::getRegisterName(unsigned RegNo) {\n" - << " assert(RegNo && RegNo < " << (Registers.size()+1) - << " && \"Invalid register number!\");\n" - << "\n" - << " static const unsigned RegAsmOffset[] = {"; - for (unsigned i = 0, e = Registers.size(); i != e; ++i) { - const CodeGenRegister &Reg = Registers[i]; + "const char *" << Target.getName() << ClassName << "::"; + if (hasAltNames) + O << "\ngetRegisterName(unsigned RegNo, unsigned AltIdx) {\n"; + else + O << "getRegisterName(unsigned RegNo) {\n"; + O << " assert(RegNo && RegNo < " << (Registers.size()+1) + << " && \"Invalid register number!\");\n" + << "\n"; - std::string AsmName = Reg.TheDef->getValueAsString("AsmName"); - if (AsmName.empty()) - AsmName = Reg.getName(); + if (hasAltNames) { + for (unsigned i = 0, e = AltNameIndices.size(); i < e; ++i) + emitRegisterNameString(O, AltNameIndices[i]->getName(), Registers); + } else + emitRegisterNameString(O, "", Registers); + + if (hasAltNames) { + O << " switch(AltIdx) {\n" + << " default: llvm_unreachable(\"Invalid register alt name index!\");\n"; + for (unsigned i = 0, e = AltNameIndices.size(); i < e; ++i) { + std::string Namespace = AltNameIndices[1]->getValueAsString("Namespace"); + std::string AltName(AltNameIndices[i]->getName()); + O << " case " << Namespace << "::" << AltName << ":\n" + << " assert(*(AsmStrs" << AltName << "+RegAsmOffset" + << AltName << "[RegNo-1]) &&\n" + << " \"Invalid alt name index for register!\");\n" + << " return AsmStrs" << AltName << "+RegAsmOffset" + << AltName << "[RegNo-1];\n"; + } + O << " }\n"; + } else { + O << " assert (*(AsmStrs+RegAsmOffset[RegNo-1]) &&\n" + << " \"Invalid alt name index for register!\");\n" + << " return AsmStrs+RegAsmOffset[RegNo-1];\n"; + } + O << "}\n"; +} +namespace { +// IAPrinter - Holds information about an InstAlias. Two InstAliases match if +// they both have the same conditionals. In which case, we cannot print out the +// alias for that pattern. +class IAPrinter { + std::vector Conds; + std::map> OpMap; + SmallVector ReqFeatures; + + std::string Result; + std::string AsmString; +public: + IAPrinter(std::string R, std::string AS) : Result(R), AsmString(AS) {} + + void addCond(const std::string &C) { Conds.push_back(C); } + + void addOperand(StringRef Op, int OpIdx, int PrintMethodIdx = -1) { + assert(OpIdx >= 0 && OpIdx < 0xFE && "Idx out of range"); + assert(PrintMethodIdx >= -1 && PrintMethodIdx < 0xFF && + "Idx out of range"); + OpMap[Op] = std::make_pair(OpIdx, PrintMethodIdx); + } - if ((i % 14) == 0) - O << "\n "; + bool isOpMapped(StringRef Op) { return OpMap.find(Op) != OpMap.end(); } + int getOpIndex(StringRef Op) { return OpMap[Op].first; } + std::pair &getOpData(StringRef Op) { return OpMap[Op]; } + + std::pair parseName(StringRef::iterator Start, + StringRef::iterator End) { + StringRef::iterator I = Start; + StringRef::iterator Next; + if (*I == '{') { + // ${some_name} + Start = ++I; + while (I != End && *I != '}') + ++I; + Next = I; + // eat the final '}' + if (Next != End) + ++Next; + } else { + // $name, just eat the usual suspects. + while (I != End && + ((*I >= 'a' && *I <= 'z') || (*I >= 'A' && *I <= 'Z') || + (*I >= '0' && *I <= '9') || *I == '_')) + ++I; + Next = I; + } - O << StringTable.GetOrAddStringOffset(AsmName) << ", "; + return std::make_pair(StringRef(Start, I - Start), Next); } - O << "0\n" - << " };\n" - << "\n"; - O << " const char *AsmStrs =\n"; - StringTable.EmitString(O); - O << ";\n"; + void print(raw_ostream &O) { + if (Conds.empty() && ReqFeatures.empty()) { + O.indent(6) << "return true;\n"; + return; + } + + O << "if ("; + + for (std::vector::iterator + I = Conds.begin(), E = Conds.end(); I != E; ++I) { + if (I != Conds.begin()) { + O << " &&\n"; + O.indent(8); + } - O << " return AsmStrs+RegAsmOffset[RegNo-1];\n" - << "}\n"; + O << *I; + } + + O << ") {\n"; + O.indent(6) << "// " << Result << "\n"; + + // Directly mangle mapped operands into the string. Each operand is + // identified by a '$' sign followed by a byte identifying the number of the + // operand. We add one to the index to avoid zero bytes. + StringRef ASM(AsmString); + SmallString<128> OutString; + raw_svector_ostream OS(OutString); + for (StringRef::iterator I = ASM.begin(), E = ASM.end(); I != E;) { + OS << *I; + if (*I == '$') { + StringRef Name; + std::tie(Name, I) = parseName(++I, E); + assert(isOpMapped(Name) && "Unmapped operand!"); + + int OpIndex, PrintIndex; + std::tie(OpIndex, PrintIndex) = getOpData(Name); + if (PrintIndex == -1) { + // Can use the default printOperand route. + OS << format("\\x%02X", (unsigned char)OpIndex + 1); + } else + // 3 bytes if a PrintMethod is needed: 0xFF, the MCInst operand + // number, and which of our pre-detected Methods to call. + OS << format("\\xFF\\x%02X\\x%02X", OpIndex + 1, PrintIndex + 1); + } else { + ++I; + } + } + + // Emit the string. + O.indent(6) << "AsmString = \"" << OutString << "\";\n"; + + O.indent(6) << "break;\n"; + O.indent(4) << '}'; + } + + bool operator==(const IAPrinter &RHS) const { + if (Conds.size() != RHS.Conds.size()) + return false; + + unsigned Idx = 0; + for (const auto &str : Conds) + if (str != RHS.Conds[Idx++]) + return false; + + return true; + } +}; + +} // end anonymous namespace + +static unsigned CountNumOperands(StringRef AsmString, unsigned Variant) { + std::string FlatAsmString = + CodeGenInstruction::FlattenAsmStringVariants(AsmString, Variant); + AsmString = FlatAsmString; + + return AsmString.count(' ') + AsmString.count('\t'); +} + +namespace { +struct AliasPriorityComparator { + typedef std::pair ValueType; + bool operator()(const ValueType &LHS, const ValueType &RHS) { + if (LHS.second == RHS.second) { + // We don't actually care about the order, but for consistency it + // shouldn't depend on pointer comparisons. + return LHS.first.TheDef->getName() < RHS.first.TheDef->getName(); + } + + // Aliases with larger priorities should be considered first. + return LHS.second > RHS.second; + } +}; } -void AsmWriterEmitter::EmitGetInstructionName(raw_ostream &O) { - CodeGenTarget Target(Records); + +void AsmWriterEmitter::EmitPrintAliasInstruction(raw_ostream &O) { Record *AsmWriter = Target.getAsmWriter(); + + O << "\n#ifdef PRINT_ALIAS_INSTR\n"; + O << "#undef PRINT_ALIAS_INSTR\n\n"; + + ////////////////////////////// + // Gather information about aliases we need to print + ////////////////////////////// + + // Emit the method that prints the alias instruction. std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName"); + unsigned Variant = AsmWriter->getValueAsInt("Variant"); + unsigned PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget"); + + std::vector AllInstAliases = + Records.getAllDerivedDefinitions("InstAlias"); + + // Create a map from the qualified name to a list of potential matches. + typedef std::set, AliasPriorityComparator> + AliasWithPriority; + std::map AliasMap; + for (std::vector::iterator + I = AllInstAliases.begin(), E = AllInstAliases.end(); I != E; ++I) { + const Record *R = *I; + int Priority = R->getValueAsInt("EmitPriority"); + if (Priority < 1) + continue; // Aliases with priority 0 are never emitted. + + const DagInit *DI = R->getValueAsDag("ResultInst"); + const DefInit *Op = cast(DI->getOperator()); + AliasMap[getQualifiedName(Op->getDef())].insert( + std::make_pair(CodeGenInstAlias(*I, Variant, Target), Priority)); + } - const std::vector &NumberedInstructions = - Target.getInstructionsByEnumValue(); + // A map of which conditions need to be met for each instruction operand + // before it can be matched to the mnemonic. + std::map> IAPrinterMap; + + // A list of MCOperandPredicates for all operands in use, and the reverse map + std::vector MCOpPredicates; + DenseMap MCOpPredicateMap; + + for (auto &Aliases : AliasMap) { + for (auto &Alias : Aliases.second) { + const CodeGenInstAlias &CGA = Alias.first; + unsigned LastOpNo = CGA.ResultInstOperandIndex.size(); + unsigned NumResultOps = + CountNumOperands(CGA.ResultInst->AsmString, Variant); + + // Don't emit the alias if it has more operands than what it's aliasing. + if (NumResultOps < CountNumOperands(CGA.AsmString, Variant)) + continue; + + IAPrinter IAP(CGA.Result->getAsString(), CGA.AsmString); + + unsigned NumMIOps = 0; + for (auto &Operand : CGA.ResultOperands) + NumMIOps += Operand.getMINumOperands(); + + std::string Cond; + Cond = std::string("MI->getNumOperands() == ") + llvm::utostr(NumMIOps); + IAP.addCond(Cond); + + bool CantHandle = false; + + unsigned MIOpNum = 0; + for (unsigned i = 0, e = LastOpNo; i != e; ++i) { + std::string Op = "MI->getOperand(" + llvm::utostr(MIOpNum) + ")"; + + const CodeGenInstAlias::ResultOperand &RO = CGA.ResultOperands[i]; + + switch (RO.Kind) { + case CodeGenInstAlias::ResultOperand::K_Record: { + const Record *Rec = RO.getRecord(); + StringRef ROName = RO.getName(); + int PrintMethodIdx = -1; + + // These two may have a PrintMethod, which we want to record (if it's + // the first time we've seen it) and provide an index for the aliasing + // code to use. + if (Rec->isSubClassOf("RegisterOperand") || + Rec->isSubClassOf("Operand")) { + std::string PrintMethod = Rec->getValueAsString("PrintMethod"); + if (PrintMethod != "" && PrintMethod != "printOperand") { + PrintMethodIdx = std::find(PrintMethods.begin(), + PrintMethods.end(), PrintMethod) - + PrintMethods.begin(); + if (static_cast(PrintMethodIdx) == PrintMethods.size()) + PrintMethods.push_back(PrintMethod); + } + } + + if (Rec->isSubClassOf("RegisterOperand")) + Rec = Rec->getValueAsDef("RegClass"); + if (Rec->isSubClassOf("RegisterClass")) { + IAP.addCond(Op + ".isReg()"); + + if (!IAP.isOpMapped(ROName)) { + IAP.addOperand(ROName, MIOpNum, PrintMethodIdx); + Record *R = CGA.ResultOperands[i].getRecord(); + if (R->isSubClassOf("RegisterOperand")) + R = R->getValueAsDef("RegClass"); + Cond = std::string("MRI.getRegClass(") + Target.getName() + "::" + + R->getName() + "RegClassID)" + ".contains(" + Op + ".getReg())"; + } else { + Cond = Op + ".getReg() == MI->getOperand(" + + llvm::utostr(IAP.getOpIndex(ROName)) + ").getReg()"; + } + } else { + // Assume all printable operands are desired for now. This can be + // overridden in the InstAlias instantiation if necessary. + IAP.addOperand(ROName, MIOpNum, PrintMethodIdx); + + // There might be an additional predicate on the MCOperand + unsigned Entry = MCOpPredicateMap[Rec]; + if (!Entry) { + if (!Rec->isValueUnset("MCOperandPredicate")) { + MCOpPredicates.push_back(Rec); + Entry = MCOpPredicates.size(); + MCOpPredicateMap[Rec] = Entry; + } else + break; // No conditions on this operand at all + } + Cond = Target.getName() + ClassName + "ValidateMCOperand(" + + Op + ", " + llvm::utostr(Entry) + ")"; + } + // for all subcases of ResultOperand::K_Record: + IAP.addCond(Cond); + break; + } + case CodeGenInstAlias::ResultOperand::K_Imm: { + // Just because the alias has an immediate result, doesn't mean the + // MCInst will. An MCExpr could be present, for example. + IAP.addCond(Op + ".isImm()"); + + Cond = Op + ".getImm() == " + + llvm::utostr(CGA.ResultOperands[i].getImm()); + IAP.addCond(Cond); + break; + } + case CodeGenInstAlias::ResultOperand::K_Reg: + // If this is zero_reg, something's playing tricks we're not + // equipped to handle. + if (!CGA.ResultOperands[i].getRegister()) { + CantHandle = true; + break; + } + + Cond = Op + ".getReg() == " + Target.getName() + "::" + + CGA.ResultOperands[i].getRegister()->getName(); + IAP.addCond(Cond); + break; + } - StringToOffsetTable StringTable; - O << -"\n\n#ifdef GET_INSTRUCTION_NAME\n" -"#undef GET_INSTRUCTION_NAME\n\n" -"/// getInstructionName: This method is automatically generated by tblgen\n" -"/// from the instruction set description. This returns the enum name of the\n" -"/// specified instruction.\n" - "const char *" << Target.getName() << ClassName - << "::getInstructionName(unsigned Opcode) {\n" - << " assert(Opcode < " << NumberedInstructions.size() - << " && \"Invalid instruction number!\");\n" - << "\n" - << " static const unsigned InstAsmOffset[] = {"; - for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) { - const CodeGenInstruction &Inst = *NumberedInstructions[i]; - - std::string AsmName = Inst.TheDef->getName(); - if ((i % 14) == 0) - O << "\n "; + MIOpNum += RO.getMINumOperands(); + } + + if (CantHandle) continue; + IAPrinterMap[Aliases.first].push_back(std::move(IAP)); + } + } + + ////////////////////////////// + // Write out the printAliasInstr function + ////////////////////////////// - O << StringTable.GetOrAddStringOffset(AsmName) << ", "; + std::string Header; + raw_string_ostream HeaderO(Header); + + HeaderO << "bool " << Target.getName() << ClassName + << "::printAliasInstr(const MCInst" + << " *MI, " << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "") + << "raw_ostream &OS) {\n"; + + std::string Cases; + raw_string_ostream CasesO(Cases); + + for (auto &Entry : IAPrinterMap) { + std::vector &IAPs = Entry.second; + std::vector UniqueIAPs; + + for (auto &LHS : IAPs) { + bool IsDup = false; + for (const auto &RHS : IAPs) { + if (&LHS != &RHS && LHS == RHS) { + IsDup = true; + break; + } + } + + if (!IsDup) + UniqueIAPs.push_back(&LHS); + } + + if (UniqueIAPs.empty()) continue; + + CasesO.indent(2) << "case " << Entry.first << ":\n"; + + for (std::vector::iterator + II = UniqueIAPs.begin(), IE = UniqueIAPs.end(); II != IE; ++II) { + IAPrinter *IAP = *II; + CasesO.indent(4); + IAP->print(CasesO); + CasesO << '\n'; + } + + CasesO.indent(4) << "return false;\n"; + } + + if (CasesO.str().empty()) { + O << HeaderO.str(); + O << " return false;\n"; + O << "}\n\n"; + O << "#endif // PRINT_ALIAS_INSTR\n"; + return; } - O << "0\n" - << " };\n" - << "\n"; - O << " const char *Strs =\n"; - StringTable.EmitString(O); - O << ";\n"; + if (!MCOpPredicates.empty()) + O << "static bool " << Target.getName() << ClassName + << "ValidateMCOperand(\n" + << " const MCOperand &MCOp, unsigned PredicateIndex);\n"; + + O << HeaderO.str(); + O.indent(2) << "const char *AsmString;\n"; + O.indent(2) << "switch (MI->getOpcode()) {\n"; + O.indent(2) << "default: return false;\n"; + O << CasesO.str(); + O.indent(2) << "}\n\n"; + + // Code that prints the alias, replacing the operands with the ones from the + // MCInst. + O << " unsigned I = 0;\n"; + O << " while (AsmString[I] != ' ' && AsmString[I] != '\t' &&\n"; + O << " AsmString[I] != '\\0')\n"; + O << " ++I;\n"; + O << " OS << '\\t' << StringRef(AsmString, I);\n"; + + O << " if (AsmString[I] != '\\0') {\n"; + O << " OS << '\\t';\n"; + O << " do {\n"; + O << " if (AsmString[I] == '$') {\n"; + O << " ++I;\n"; + O << " if (AsmString[I] == (char)0xff) {\n"; + O << " ++I;\n"; + O << " int OpIdx = AsmString[I++] - 1;\n"; + O << " int PrintMethodIdx = AsmString[I++] - 1;\n"; + O << " printCustomAliasOperand(MI, OpIdx, PrintMethodIdx, "; + O << (PassSubtarget ? "STI, " : ""); + O << "OS);\n"; + O << " } else\n"; + O << " printOperand(MI, unsigned(AsmString[I++]) - 1, "; + O << (PassSubtarget ? "STI, " : ""); + O << "OS);\n"; + O << " } else {\n"; + O << " OS << AsmString[I++];\n"; + O << " }\n"; + O << " } while (AsmString[I] != '\\0');\n"; + O << " }\n\n"; + + O << " return true;\n"; + O << "}\n\n"; + + ////////////////////////////// + // Write out the printCustomAliasOperand function + ////////////////////////////// + + O << "void " << Target.getName() << ClassName << "::" + << "printCustomAliasOperand(\n" + << " const MCInst *MI, unsigned OpIdx,\n" + << " unsigned PrintMethodIdx,\n" + << (PassSubtarget ? " const MCSubtargetInfo &STI,\n" : "") + << " raw_ostream &OS) {\n"; + if (PrintMethods.empty()) + O << " llvm_unreachable(\"Unknown PrintMethod kind\");\n"; + else { + O << " switch (PrintMethodIdx) {\n" + << " default:\n" + << " llvm_unreachable(\"Unknown PrintMethod kind\");\n" + << " break;\n"; + + for (unsigned i = 0; i < PrintMethods.size(); ++i) { + O << " case " << i << ":\n" + << " " << PrintMethods[i] << "(MI, OpIdx, " + << (PassSubtarget ? "STI, " : "") << "OS);\n" + << " break;\n"; + } + O << " }\n"; + } + O << "}\n\n"; + + if (!MCOpPredicates.empty()) { + O << "static bool " << Target.getName() << ClassName + << "ValidateMCOperand(\n" + << " const MCOperand &MCOp, unsigned PredicateIndex) {\n" + << " switch (PredicateIndex) {\n" + << " default:\n" + << " llvm_unreachable(\"Unknown MCOperandPredicate kind\");\n" + << " break;\n"; + + for (unsigned i = 0; i < MCOpPredicates.size(); ++i) { + Init *MCOpPred = MCOpPredicates[i]->getValueInit("MCOperandPredicate"); + if (StringInit *SI = dyn_cast(MCOpPred)) { + O << " case " << i + 1 << ": {\n" + << SI->getValue() << "\n" + << " }\n"; + } else + llvm_unreachable("Unexpected MCOperandPredicate field!"); + } + O << " }\n" + << "}\n\n"; + } - O << " return Strs+InstAsmOffset[Opcode];\n" - << "}\n\n#endif\n"; + O << "#endif // PRINT_ALIAS_INSTR\n"; } +AsmWriterEmitter::AsmWriterEmitter(RecordKeeper &R) : Records(R), Target(R) { + Record *AsmWriter = Target.getAsmWriter(); + for (const CodeGenInstruction *I : Target.instructions()) + if (!I->AsmString.empty() && I->TheDef->getName() != "PHI") + Instructions.emplace_back(*I, AsmWriter->getValueAsInt("Variant"), + AsmWriter->getValueAsInt("PassSubtarget")); + // Get the instruction numbering. + NumberedInstructions = &Target.getInstructionsByEnumValue(); -void AsmWriterEmitter::run(raw_ostream &O) { - EmitSourceFileHeader("Assembly Writer Source Fragment", O); + // Compute the CodeGenInstruction -> AsmWriterInst mapping. Note that not + // all machine instructions are necessarily being printed, so there may be + // target instructions not in this map. + for (unsigned i = 0, e = Instructions.size(); i != e; ++i) + CGIAWIMap.insert(std::make_pair(Instructions[i].CGI, &Instructions[i])); +} +void AsmWriterEmitter::run(raw_ostream &O) { EmitPrintInstruction(O); EmitGetRegisterName(O); - EmitGetInstructionName(O); + EmitPrintAliasInstruction(O); +} + + +namespace llvm { + +void EmitAsmWriter(RecordKeeper &RK, raw_ostream &OS) { + emitSourceFileHeader("Assembly Writer Source Fragment", OS); + AsmWriterEmitter(RK).run(OS); } +} // End llvm namespace