1 //===- CodeEmitterGen.cpp - Code Emitter Generator ------------------------===// 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 // CodeEmitterGen uses the descriptions of instructions and their fields to 11 // construct an automated code emitter: a function that, given a MachineInstr, 12 // returns the (currently, 32-bit unsigned) value of the instruction. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "CodeEmitterGen.h" 17 #include "CodeGenTarget.h" 18 #include "Record.h" 19 #include "llvm/ADT/StringExtras.h" 20 #include "llvm/Support/CommandLine.h" 21 #include "llvm/Support/Debug.h" 22 #include <map> 23 using namespace llvm; 24 25 // FIXME: Somewhat hackish to use a command line option for this. There should 26 // be a CodeEmitter class in the Target.td that controls this sort of thing 27 // instead. 28 static cl::opt<bool> 29 MCEmitter("mc-emitter", 30 cl::desc("Generate CodeEmitter for use with the MC library."), 31 cl::init(false)); 32 33 void CodeEmitterGen::reverseBits(std::vector<Record*> &Insts) { 34 for (std::vector<Record*>::iterator I = Insts.begin(), E = Insts.end(); 35 I != E; ++I) { 36 Record *R = *I; 37 if (R->getValueAsString("Namespace") == "TargetOpcode") 38 continue; 39 40 BitsInit *BI = R->getValueAsBitsInit("Inst"); 41 42 unsigned numBits = BI->getNumBits(); 43 BitsInit *NewBI = new BitsInit(numBits); 44 for (unsigned bit = 0, end = numBits / 2; bit != end; ++bit) { 45 unsigned bitSwapIdx = numBits - bit - 1; 46 Init *OrigBit = BI->getBit(bit); 47 Init *BitSwap = BI->getBit(bitSwapIdx); 48 NewBI->setBit(bit, BitSwap); 49 NewBI->setBit(bitSwapIdx, OrigBit); 50 } 51 if (numBits % 2) { 52 unsigned middle = (numBits + 1) / 2; 53 NewBI->setBit(middle, BI->getBit(middle)); 54 } 55 56 // Update the bits in reversed order so that emitInstrOpBits will get the 57 // correct endianness. 58 R->getValue("Inst")->setValue(NewBI); 59 } 60 } 61 62 // If the VarBitInit at position 'bit' matches the specified variable then 63 // return the variable bit position. Otherwise return -1. 64 int CodeEmitterGen::getVariableBit(const std::string &VarName, 65 BitsInit *BI, int bit) { 66 if (VarBitInit *VBI = dynamic_cast<VarBitInit*>(BI->getBit(bit))) 67 if (VarInit *VI = dynamic_cast<VarInit*>(VBI->getVariable())) 68 if (VI->getName() == VarName) 69 return VBI->getBitNum(); 70 71 return -1; 72 } 73 74 void CodeEmitterGen:: 75 AddCodeToMergeInOperand(Record *R, BitsInit *BI, const std::string &VarName, 76 unsigned &NumberedOp, 77 std::string &Case, CodeGenTarget &Target) { 78 CodeGenInstruction &CGI = Target.getInstruction(R); 79 80 // Determine if VarName actually contributes to the Inst encoding. 81 int bit = BI->getNumBits()-1; 82 83 // Scan for a bit that this contributed to. 84 for (; bit >= 0; ) { 85 if (getVariableBit(VarName, BI, bit) != -1) 86 break; 87 88 --bit; 89 } 90 91 // If we found no bits, ignore this value, otherwise emit the call to get the 92 // operand encoding. 93 if (bit < 0) return; 94 95 // If the operand matches by name, reference according to that 96 // operand number. Non-matching operands are assumed to be in 97 // order. 98 unsigned OpIdx; 99 if (CGI.Operands.hasOperandNamed(VarName, OpIdx)) { 100 // Get the machine operand number for the indicated operand. 101 OpIdx = CGI.Operands[OpIdx].MIOperandNo; 102 assert(!CGI.Operands.isFlatOperandNotEmitted(OpIdx) && 103 "Explicitly used operand also marked as not emitted!"); 104 } else { 105 /// If this operand is not supposed to be emitted by the 106 /// generated emitter, skip it. 107 while (CGI.Operands.isFlatOperandNotEmitted(NumberedOp)) 108 ++NumberedOp; 109 OpIdx = NumberedOp++; 110 } 111 112 std::pair<unsigned, unsigned> SO = CGI.Operands.getSubOperandNumber(OpIdx); 113 std::string &EncoderMethodName = CGI.Operands[SO.first].EncoderMethodName; 114 115 // If the source operand has a custom encoder, use it. This will 116 // get the encoding for all of the suboperands. 117 if (!EncoderMethodName.empty()) { 118 // A custom encoder has all of the information for the 119 // sub-operands, if there are more than one, so only 120 // query the encoder once per source operand. 121 if (SO.second == 0) { 122 Case += " // op: " + VarName + "\n" + 123 " op = " + EncoderMethodName + "(MI, " + utostr(OpIdx); 124 if (MCEmitter) 125 Case += ", Fixups"; 126 Case += ");\n"; 127 } 128 } else { 129 Case += " // op: " + VarName + "\n" + 130 " op = getMachineOpValue(MI, MI.getOperand(" + utostr(OpIdx) + ")"; 131 if (MCEmitter) 132 Case += ", Fixups"; 133 Case += ");\n"; 134 } 135 136 for (; bit >= 0; ) { 137 int varBit = getVariableBit(VarName, BI, bit); 138 139 // If this bit isn't from a variable, skip it. 140 if (varBit == -1) { 141 --bit; 142 continue; 143 } 144 145 // Figure out the consequtive range of bits covered by this operand, in 146 // order to generate better encoding code. 147 int beginInstBit = bit; 148 int beginVarBit = varBit; 149 int N = 1; 150 for (--bit; bit >= 0;) { 151 varBit = getVariableBit(VarName, BI, bit); 152 if (varBit == -1 || varBit != (beginVarBit - N)) break; 153 ++N; 154 --bit; 155 } 156 157 unsigned opMask = ~0U >> (32-N); 158 int opShift = beginVarBit - N + 1; 159 opMask <<= opShift; 160 opShift = beginInstBit - beginVarBit; 161 162 if (opShift > 0) { 163 Case += " Value |= (op & " + utostr(opMask) + "U) << " + 164 itostr(opShift) + ";\n"; 165 } else if (opShift < 0) { 166 Case += " Value |= (op & " + utostr(opMask) + "U) >> " + 167 itostr(-opShift) + ";\n"; 168 } else { 169 Case += " Value |= op & " + utostr(opMask) + "U;\n"; 170 } 171 } 172 } 173 174 175 std::string CodeEmitterGen::getInstructionCase(Record *R, 176 CodeGenTarget &Target) { 177 std::string Case; 178 179 BitsInit *BI = R->getValueAsBitsInit("Inst"); 180 const std::vector<RecordVal> &Vals = R->getValues(); 181 unsigned NumberedOp = 0; 182 183 // Loop over all of the fields in the instruction, determining which are the 184 // operands to the instruction. 185 for (unsigned i = 0, e = Vals.size(); i != e; ++i) { 186 // Ignore fixed fields in the record, we're looking for values like: 187 // bits<5> RST = { ?, ?, ?, ?, ? }; 188 if (Vals[i].getPrefix() || Vals[i].getValue()->isComplete()) 189 continue; 190 191 AddCodeToMergeInOperand(R, BI, Vals[i].getName(), NumberedOp, Case, Target); 192 } 193 194 std::string PostEmitter = R->getValueAsString("PostEncoderMethod"); 195 if (!PostEmitter.empty()) 196 Case += " Value = " + PostEmitter + "(MI, Value);\n"; 197 198 return Case; 199 } 200 201 void CodeEmitterGen::run(raw_ostream &o) { 202 CodeGenTarget Target(Records); 203 std::vector<Record*> Insts = Records.getAllDerivedDefinitions("Instruction"); 204 205 // For little-endian instruction bit encodings, reverse the bit order 206 if (Target.isLittleEndianEncoding()) reverseBits(Insts); 207 208 EmitSourceFileHeader("Machine Code Emitter", o); 209 std::string Namespace = Insts[0]->getValueAsString("Namespace") + "::"; 210 211 const std::vector<const CodeGenInstruction*> &NumberedInstructions = 212 Target.getInstructionsByEnumValue(); 213 214 // Emit function declaration 215 o << "unsigned " << Target.getName(); 216 if (MCEmitter) 217 o << "MCCodeEmitter::getBinaryCodeForInstr(const MCInst &MI,\n" 218 << " SmallVectorImpl<MCFixup> &Fixups) const {\n"; 219 else 220 o << "CodeEmitter::getBinaryCodeForInstr(const MachineInstr &MI) const {\n"; 221 222 // Emit instruction base values 223 o << " static const unsigned InstBits[] = {\n"; 224 for (std::vector<const CodeGenInstruction*>::const_iterator 225 IN = NumberedInstructions.begin(), 226 EN = NumberedInstructions.end(); 227 IN != EN; ++IN) { 228 const CodeGenInstruction *CGI = *IN; 229 Record *R = CGI->TheDef; 230 231 if (R->getValueAsString("Namespace") == "TargetOpcode") { 232 o << " 0U,\n"; 233 continue; 234 } 235 236 BitsInit *BI = R->getValueAsBitsInit("Inst"); 237 238 // Start by filling in fixed values. 239 unsigned Value = 0; 240 for (unsigned i = 0, e = BI->getNumBits(); i != e; ++i) { 241 if (BitInit *B = dynamic_cast<BitInit*>(BI->getBit(e-i-1))) 242 Value |= B->getValue() << (e-i-1); 243 } 244 o << " " << Value << "U," << '\t' << "// " << R->getName() << "\n"; 245 } 246 o << " 0U\n };\n"; 247 248 // Map to accumulate all the cases. 249 std::map<std::string, std::vector<std::string> > CaseMap; 250 251 // Construct all cases statement for each opcode 252 for (std::vector<Record*>::iterator IC = Insts.begin(), EC = Insts.end(); 253 IC != EC; ++IC) { 254 Record *R = *IC; 255 if (R->getValueAsString("Namespace") == "TargetOpcode") 256 continue; 257 const std::string &InstName = R->getName(); 258 std::string Case = getInstructionCase(R, Target); 259 260 CaseMap[Case].push_back(InstName); 261 } 262 263 // Emit initial function code 264 o << " const unsigned opcode = MI.getOpcode();\n" 265 << " unsigned Value = InstBits[opcode];\n" 266 << " unsigned op = 0;\n" 267 << " (void)op; // suppress warning\n" 268 << " switch (opcode) {\n"; 269 270 // Emit each case statement 271 std::map<std::string, std::vector<std::string> >::iterator IE, EE; 272 for (IE = CaseMap.begin(), EE = CaseMap.end(); IE != EE; ++IE) { 273 const std::string &Case = IE->first; 274 std::vector<std::string> &InstList = IE->second; 275 276 for (int i = 0, N = InstList.size(); i < N; i++) { 277 if (i) o << "\n"; 278 o << " case " << Namespace << InstList[i] << ":"; 279 } 280 o << " {\n"; 281 o << Case; 282 o << " break;\n" 283 << " }\n"; 284 } 285 286 // Default case: unhandled opcode 287 o << " default:\n" 288 << " std::string msg;\n" 289 << " raw_string_ostream Msg(msg);\n" 290 << " Msg << \"Not supported instr: \" << MI;\n" 291 << " report_fatal_error(Msg.str());\n" 292 << " }\n" 293 << " return Value;\n" 294 << "}\n\n"; 295 } 296