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 } else if (VarInit *VI = dynamic_cast<VarInit*>(BI->getBit(bit))) { 71 if (VI->getName() == VarName) 72 return 0; 73 } 74 75 return -1; 76 } 77 78 void CodeEmitterGen:: 79 AddCodeToMergeInOperand(Record *R, BitsInit *BI, const std::string &VarName, 80 unsigned &NumberedOp, 81 std::string &Case, CodeGenTarget &Target) { 82 CodeGenInstruction &CGI = Target.getInstruction(R); 83 84 // Determine if VarName actually contributes to the Inst encoding. 85 int bit = BI->getNumBits()-1; 86 87 // Scan for a bit that this contributed to. 88 for (; bit >= 0; ) { 89 if (getVariableBit(VarName, BI, bit) != -1) 90 break; 91 92 --bit; 93 } 94 95 // If we found no bits, ignore this value, otherwise emit the call to get the 96 // operand encoding. 97 if (bit < 0) return; 98 99 // If the operand matches by name, reference according to that 100 // operand number. Non-matching operands are assumed to be in 101 // order. 102 unsigned OpIdx; 103 if (CGI.Operands.hasOperandNamed(VarName, OpIdx)) { 104 // Get the machine operand number for the indicated operand. 105 OpIdx = CGI.Operands[OpIdx].MIOperandNo; 106 assert(!CGI.Operands.isFlatOperandNotEmitted(OpIdx) && 107 "Explicitly used operand also marked as not emitted!"); 108 } else { 109 /// If this operand is not supposed to be emitted by the 110 /// generated emitter, skip it. 111 while (CGI.Operands.isFlatOperandNotEmitted(NumberedOp)) 112 ++NumberedOp; 113 OpIdx = NumberedOp++; 114 } 115 116 std::pair<unsigned, unsigned> SO = CGI.Operands.getSubOperandNumber(OpIdx); 117 std::string &EncoderMethodName = CGI.Operands[SO.first].EncoderMethodName; 118 119 // If the source operand has a custom encoder, use it. This will 120 // get the encoding for all of the suboperands. 121 if (!EncoderMethodName.empty()) { 122 // A custom encoder has all of the information for the 123 // sub-operands, if there are more than one, so only 124 // query the encoder once per source operand. 125 if (SO.second == 0) { 126 Case += " // op: " + VarName + "\n" + 127 " op = " + EncoderMethodName + "(MI, " + utostr(OpIdx); 128 if (MCEmitter) 129 Case += ", Fixups"; 130 Case += ");\n"; 131 } 132 } else { 133 Case += " // op: " + VarName + "\n" + 134 " op = getMachineOpValue(MI, MI.getOperand(" + utostr(OpIdx) + ")"; 135 if (MCEmitter) 136 Case += ", Fixups"; 137 Case += ");\n"; 138 } 139 140 for (; bit >= 0; ) { 141 int varBit = getVariableBit(VarName, BI, bit); 142 143 // If this bit isn't from a variable, skip it. 144 if (varBit == -1) { 145 --bit; 146 continue; 147 } 148 149 // Figure out the consecutive range of bits covered by this operand, in 150 // order to generate better encoding code. 151 int beginInstBit = bit; 152 int beginVarBit = varBit; 153 int N = 1; 154 for (--bit; bit >= 0;) { 155 varBit = getVariableBit(VarName, BI, bit); 156 if (varBit == -1 || varBit != (beginVarBit - N)) break; 157 ++N; 158 --bit; 159 } 160 161 unsigned opMask = ~0U >> (32-N); 162 int opShift = beginVarBit - N + 1; 163 opMask <<= opShift; 164 opShift = beginInstBit - beginVarBit; 165 166 if (opShift > 0) { 167 Case += " Value |= (op & " + utostr(opMask) + "U) << " + 168 itostr(opShift) + ";\n"; 169 } else if (opShift < 0) { 170 Case += " Value |= (op & " + utostr(opMask) + "U) >> " + 171 itostr(-opShift) + ";\n"; 172 } else { 173 Case += " Value |= op & " + utostr(opMask) + "U;\n"; 174 } 175 } 176 } 177 178 179 std::string CodeEmitterGen::getInstructionCase(Record *R, 180 CodeGenTarget &Target) { 181 std::string Case; 182 183 BitsInit *BI = R->getValueAsBitsInit("Inst"); 184 const std::vector<RecordVal> &Vals = R->getValues(); 185 unsigned NumberedOp = 0; 186 187 // Loop over all of the fields in the instruction, determining which are the 188 // operands to the instruction. 189 for (unsigned i = 0, e = Vals.size(); i != e; ++i) { 190 // Ignore fixed fields in the record, we're looking for values like: 191 // bits<5> RST = { ?, ?, ?, ?, ? }; 192 if (Vals[i].getPrefix() || Vals[i].getValue()->isComplete()) 193 continue; 194 195 AddCodeToMergeInOperand(R, BI, Vals[i].getName(), NumberedOp, Case, Target); 196 } 197 198 std::string PostEmitter = R->getValueAsString("PostEncoderMethod"); 199 if (!PostEmitter.empty()) 200 Case += " Value = " + PostEmitter + "(MI, Value);\n"; 201 202 return Case; 203 } 204 205 void CodeEmitterGen::run(raw_ostream &o) { 206 CodeGenTarget Target(Records); 207 std::vector<Record*> Insts = Records.getAllDerivedDefinitions("Instruction"); 208 209 // For little-endian instruction bit encodings, reverse the bit order 210 if (Target.isLittleEndianEncoding()) reverseBits(Insts); 211 212 EmitSourceFileHeader("Machine Code Emitter", o); 213 214 const std::vector<const CodeGenInstruction*> &NumberedInstructions = 215 Target.getInstructionsByEnumValue(); 216 217 // Emit function declaration 218 o << "unsigned " << Target.getName(); 219 if (MCEmitter) 220 o << "MCCodeEmitter::getBinaryCodeForInstr(const MCInst &MI,\n" 221 << " SmallVectorImpl<MCFixup> &Fixups) const {\n"; 222 else 223 o << "CodeEmitter::getBinaryCodeForInstr(const MachineInstr &MI) const {\n"; 224 225 // Emit instruction base values 226 o << " static const unsigned InstBits[] = {\n"; 227 for (std::vector<const CodeGenInstruction*>::const_iterator 228 IN = NumberedInstructions.begin(), 229 EN = NumberedInstructions.end(); 230 IN != EN; ++IN) { 231 const CodeGenInstruction *CGI = *IN; 232 Record *R = CGI->TheDef; 233 234 if (R->getValueAsString("Namespace") == "TargetOpcode") { 235 o << " 0U,\n"; 236 continue; 237 } 238 239 BitsInit *BI = R->getValueAsBitsInit("Inst"); 240 241 // Start by filling in fixed values. 242 unsigned Value = 0; 243 for (unsigned i = 0, e = BI->getNumBits(); i != e; ++i) { 244 if (BitInit *B = dynamic_cast<BitInit*>(BI->getBit(e-i-1))) 245 Value |= B->getValue() << (e-i-1); 246 } 247 o << " " << Value << "U," << '\t' << "// " << R->getName() << "\n"; 248 } 249 o << " 0U\n };\n"; 250 251 // Map to accumulate all the cases. 252 std::map<std::string, std::vector<std::string> > CaseMap; 253 254 // Construct all cases statement for each opcode 255 for (std::vector<Record*>::iterator IC = Insts.begin(), EC = Insts.end(); 256 IC != EC; ++IC) { 257 Record *R = *IC; 258 if (R->getValueAsString("Namespace") == "TargetOpcode") 259 continue; 260 const std::string &InstName = R->getValueAsString("Namespace") + "::" 261 + R->getName(); 262 std::string Case = getInstructionCase(R, Target); 263 264 CaseMap[Case].push_back(InstName); 265 } 266 267 // Emit initial function code 268 o << " const unsigned opcode = MI.getOpcode();\n" 269 << " unsigned Value = InstBits[opcode];\n" 270 << " unsigned op = 0;\n" 271 << " (void)op; // suppress warning\n" 272 << " switch (opcode) {\n"; 273 274 // Emit each case statement 275 std::map<std::string, std::vector<std::string> >::iterator IE, EE; 276 for (IE = CaseMap.begin(), EE = CaseMap.end(); IE != EE; ++IE) { 277 const std::string &Case = IE->first; 278 std::vector<std::string> &InstList = IE->second; 279 280 for (int i = 0, N = InstList.size(); i < N; i++) { 281 if (i) o << "\n"; 282 o << " case " << InstList[i] << ":"; 283 } 284 o << " {\n"; 285 o << Case; 286 o << " break;\n" 287 << " }\n"; 288 } 289 290 // Default case: unhandled opcode 291 o << " default:\n" 292 << " std::string msg;\n" 293 << " raw_string_ostream Msg(msg);\n" 294 << " Msg << \"Not supported instr: \" << MI;\n" 295 << " report_fatal_error(Msg.str());\n" 296 << " }\n" 297 << " return Value;\n" 298 << "}\n\n"; 299 } 300