llvm-project/llvm/utils/TableGen/CodeEmitterGen.cpp

327 lines
11 KiB
C++
Raw Normal View History

//===- CodeEmitterGen.cpp - Code Emitter Generator ------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// CodeEmitterGen uses the descriptions of instructions and their fields to
// construct an automated code emitter: a function that, given a MachineInstr,
// returns the (currently, 32-bit unsigned) value of the instruction.
//
//===----------------------------------------------------------------------===//
#include "CodeGenTarget.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/TableGen/Record.h"
#include "llvm/TableGen/TableGenBackend.h"
#include <map>
#include <string>
#include <vector>
using namespace llvm;
namespace {
class CodeEmitterGen {
RecordKeeper &Records;
public:
CodeEmitterGen(RecordKeeper &R) : Records(R) {}
void run(raw_ostream &o);
private:
int getVariableBit(const std::string &VarName, BitsInit *BI, int bit);
std::string getInstructionCase(Record *R, CodeGenTarget &Target);
void AddCodeToMergeInOperand(Record *R, BitsInit *BI,
const std::string &VarName,
unsigned &NumberedOp,
[TableGen] Optionally forbid overlap between named and positional operands There are currently two schemes for mapping instruction operands to instruction-format variables for generating the instruction encoders and decoders for the assembler and disassembler respectively: a) to map by name and b) to map by position. In the long run, we'd like to remove the position-based scheme and use only name-based mapping. Unfortunately, the name-based scheme currently cannot deal with complex operands (those with suboperands), and so we currently must use the position-based scheme for those. On the other hand, the position-based scheme cannot deal with (register) variables that are split into multiple ranges. An upcoming commit to the PowerPC backend (adding VSX support) will require this capability. While we could teach the position-based scheme to handle that, since we'd like to move away from the position-based mapping generally, it seems silly to teach it new tricks now. What makes more sense is to allow for partial transitioning: use the name-based mapping when possible, and only use the position-based scheme when necessary. Now the problem is that mixing the two sensibly was not possible: the position-based mapping would map based on position, but would not skip those variables that were mapped by name. Instead, the two sets of assignments would overlap. However, I cannot currently change the current behavior, because there are some backends that rely on it [I think mistakenly, but I'll send a message to llvmdev about that]. So I've added a new TableGen bit variable: noNamedPositionallyEncodedOperands, that can be used to cause the position-based mapping to skip variables mapped by name. llvm-svn: 203767
2014-03-13 15:57:54 +08:00
std::set<unsigned> &NamedOpIndices,
std::string &Case, CodeGenTarget &Target);
};
2006-07-14 06:17:08 +08:00
// If the VarBitInit at position 'bit' matches the specified variable then
// return the variable bit position. Otherwise return -1.
int CodeEmitterGen::getVariableBit(const std::string &VarName,
BitsInit *BI, int bit) {
if (VarBitInit *VBI = dyn_cast<VarBitInit>(BI->getBit(bit))) {
if (VarInit *VI = dyn_cast<VarInit>(VBI->getBitVar()))
if (VI->getName() == VarName)
return VBI->getBitNum();
} else if (VarInit *VI = dyn_cast<VarInit>(BI->getBit(bit))) {
if (VI->getName() == VarName)
return 0;
}
2010-10-08 00:56:28 +08:00
return -1;
2010-10-08 00:56:28 +08:00
}
void CodeEmitterGen::
AddCodeToMergeInOperand(Record *R, BitsInit *BI, const std::string &VarName,
unsigned &NumberedOp,
[TableGen] Optionally forbid overlap between named and positional operands There are currently two schemes for mapping instruction operands to instruction-format variables for generating the instruction encoders and decoders for the assembler and disassembler respectively: a) to map by name and b) to map by position. In the long run, we'd like to remove the position-based scheme and use only name-based mapping. Unfortunately, the name-based scheme currently cannot deal with complex operands (those with suboperands), and so we currently must use the position-based scheme for those. On the other hand, the position-based scheme cannot deal with (register) variables that are split into multiple ranges. An upcoming commit to the PowerPC backend (adding VSX support) will require this capability. While we could teach the position-based scheme to handle that, since we'd like to move away from the position-based mapping generally, it seems silly to teach it new tricks now. What makes more sense is to allow for partial transitioning: use the name-based mapping when possible, and only use the position-based scheme when necessary. Now the problem is that mixing the two sensibly was not possible: the position-based mapping would map based on position, but would not skip those variables that were mapped by name. Instead, the two sets of assignments would overlap. However, I cannot currently change the current behavior, because there are some backends that rely on it [I think mistakenly, but I'll send a message to llvmdev about that]. So I've added a new TableGen bit variable: noNamedPositionallyEncodedOperands, that can be used to cause the position-based mapping to skip variables mapped by name. llvm-svn: 203767
2014-03-13 15:57:54 +08:00
std::set<unsigned> &NamedOpIndices,
std::string &Case, CodeGenTarget &Target) {
CodeGenInstruction &CGI = Target.getInstruction(R);
// Determine if VarName actually contributes to the Inst encoding.
int bit = BI->getNumBits()-1;
// Scan for a bit that this contributed to.
for (; bit >= 0; ) {
if (getVariableBit(VarName, BI, bit) != -1)
break;
--bit;
}
// If we found no bits, ignore this value, otherwise emit the call to get the
// operand encoding.
if (bit < 0) return;
// If the operand matches by name, reference according to that
// operand number. Non-matching operands are assumed to be in
// order.
unsigned OpIdx;
if (CGI.Operands.hasOperandNamed(VarName, OpIdx)) {
// Get the machine operand number for the indicated operand.
OpIdx = CGI.Operands[OpIdx].MIOperandNo;
assert(!CGI.Operands.isFlatOperandNotEmitted(OpIdx) &&
"Explicitly used operand also marked as not emitted!");
} else {
unsigned NumberOps = CGI.Operands.size();
/// If this operand is not supposed to be emitted by the
/// generated emitter, skip it.
while (NumberedOp < NumberOps &&
[TableGen] Optionally forbid overlap between named and positional operands There are currently two schemes for mapping instruction operands to instruction-format variables for generating the instruction encoders and decoders for the assembler and disassembler respectively: a) to map by name and b) to map by position. In the long run, we'd like to remove the position-based scheme and use only name-based mapping. Unfortunately, the name-based scheme currently cannot deal with complex operands (those with suboperands), and so we currently must use the position-based scheme for those. On the other hand, the position-based scheme cannot deal with (register) variables that are split into multiple ranges. An upcoming commit to the PowerPC backend (adding VSX support) will require this capability. While we could teach the position-based scheme to handle that, since we'd like to move away from the position-based mapping generally, it seems silly to teach it new tricks now. What makes more sense is to allow for partial transitioning: use the name-based mapping when possible, and only use the position-based scheme when necessary. Now the problem is that mixing the two sensibly was not possible: the position-based mapping would map based on position, but would not skip those variables that were mapped by name. Instead, the two sets of assignments would overlap. However, I cannot currently change the current behavior, because there are some backends that rely on it [I think mistakenly, but I'll send a message to llvmdev about that]. So I've added a new TableGen bit variable: noNamedPositionallyEncodedOperands, that can be used to cause the position-based mapping to skip variables mapped by name. llvm-svn: 203767
2014-03-13 15:57:54 +08:00
(CGI.Operands.isFlatOperandNotEmitted(NumberedOp) ||
(NamedOpIndices.size() && NamedOpIndices.count(
CGI.Operands.getSubOperandNumber(NumberedOp).first)))) {
++NumberedOp;
if (NumberedOp >= CGI.Operands.back().MIOperandNo +
CGI.Operands.back().MINumOperands) {
errs() << "Too few operands in record " << R->getName() <<
" (no match for variable " << VarName << "):\n";
errs() << *R;
errs() << '\n';
return;
}
}
OpIdx = NumberedOp++;
}
std::pair<unsigned, unsigned> SO = CGI.Operands.getSubOperandNumber(OpIdx);
std::string &EncoderMethodName = CGI.Operands[SO.first].EncoderMethodName;
// If the source operand has a custom encoder, use it. This will
// get the encoding for all of the suboperands.
if (!EncoderMethodName.empty()) {
// A custom encoder has all of the information for the
// sub-operands, if there are more than one, so only
// query the encoder once per source operand.
if (SO.second == 0) {
Case += " // op: " + VarName + "\n" +
" op = " + EncoderMethodName + "(MI, " + utostr(OpIdx);
Case += ", Fixups, STI";
Case += ");\n";
}
} else {
Case += " // op: " + VarName + "\n" +
" op = getMachineOpValue(MI, MI.getOperand(" + utostr(OpIdx) + ")";
Case += ", Fixups, STI";
Case += ");\n";
}
for (; bit >= 0; ) {
int varBit = getVariableBit(VarName, BI, bit);
// If this bit isn't from a variable, skip it.
if (varBit == -1) {
--bit;
continue;
}
2011-01-28 07:08:52 +08:00
// Figure out the consecutive range of bits covered by this operand, in
// order to generate better encoding code.
int beginInstBit = bit;
int beginVarBit = varBit;
int N = 1;
for (--bit; bit >= 0;) {
varBit = getVariableBit(VarName, BI, bit);
if (varBit == -1 || varBit != (beginVarBit - N)) break;
++N;
--bit;
}
uint64_t opMask = ~(uint64_t)0 >> (64-N);
int opShift = beginVarBit - N + 1;
opMask <<= opShift;
opShift = beginInstBit - beginVarBit;
if (opShift > 0) {
Case += " Value |= (op & UINT64_C(" + utostr(opMask) + ")) << " +
itostr(opShift) + ";\n";
} else if (opShift < 0) {
Case += " Value |= (op & UINT64_C(" + utostr(opMask) + ")) >> " +
itostr(-opShift) + ";\n";
} else {
Case += " Value |= op & UINT64_C(" + utostr(opMask) + ");\n";
}
}
}
std::string CodeEmitterGen::getInstructionCase(Record *R,
CodeGenTarget &Target) {
std::string Case;
BitsInit *BI = R->getValueAsBitsInit("Inst");
const std::vector<RecordVal> &Vals = R->getValues();
unsigned NumberedOp = 0;
[TableGen] Optionally forbid overlap between named and positional operands There are currently two schemes for mapping instruction operands to instruction-format variables for generating the instruction encoders and decoders for the assembler and disassembler respectively: a) to map by name and b) to map by position. In the long run, we'd like to remove the position-based scheme and use only name-based mapping. Unfortunately, the name-based scheme currently cannot deal with complex operands (those with suboperands), and so we currently must use the position-based scheme for those. On the other hand, the position-based scheme cannot deal with (register) variables that are split into multiple ranges. An upcoming commit to the PowerPC backend (adding VSX support) will require this capability. While we could teach the position-based scheme to handle that, since we'd like to move away from the position-based mapping generally, it seems silly to teach it new tricks now. What makes more sense is to allow for partial transitioning: use the name-based mapping when possible, and only use the position-based scheme when necessary. Now the problem is that mixing the two sensibly was not possible: the position-based mapping would map based on position, but would not skip those variables that were mapped by name. Instead, the two sets of assignments would overlap. However, I cannot currently change the current behavior, because there are some backends that rely on it [I think mistakenly, but I'll send a message to llvmdev about that]. So I've added a new TableGen bit variable: noNamedPositionallyEncodedOperands, that can be used to cause the position-based mapping to skip variables mapped by name. llvm-svn: 203767
2014-03-13 15:57:54 +08:00
std::set<unsigned> NamedOpIndices;
// Collect the set of operand indices that might correspond to named
// operand, and skip these when assigning operands based on position.
if (Target.getInstructionSet()->
getValueAsBit("noNamedPositionallyEncodedOperands")) {
CodeGenInstruction &CGI = Target.getInstruction(R);
for (unsigned i = 0, e = Vals.size(); i != e; ++i) {
unsigned OpIdx;
if (!CGI.Operands.hasOperandNamed(Vals[i].getName(), OpIdx))
continue;
NamedOpIndices.insert(OpIdx);
}
}
// Loop over all of the fields in the instruction, determining which are the
// operands to the instruction.
for (unsigned i = 0, e = Vals.size(); i != e; ++i) {
// Ignore fixed fields in the record, we're looking for values like:
// bits<5> RST = { ?, ?, ?, ?, ? };
if (Vals[i].getPrefix() || Vals[i].getValue()->isComplete())
continue;
[TableGen] Optionally forbid overlap between named and positional operands There are currently two schemes for mapping instruction operands to instruction-format variables for generating the instruction encoders and decoders for the assembler and disassembler respectively: a) to map by name and b) to map by position. In the long run, we'd like to remove the position-based scheme and use only name-based mapping. Unfortunately, the name-based scheme currently cannot deal with complex operands (those with suboperands), and so we currently must use the position-based scheme for those. On the other hand, the position-based scheme cannot deal with (register) variables that are split into multiple ranges. An upcoming commit to the PowerPC backend (adding VSX support) will require this capability. While we could teach the position-based scheme to handle that, since we'd like to move away from the position-based mapping generally, it seems silly to teach it new tricks now. What makes more sense is to allow for partial transitioning: use the name-based mapping when possible, and only use the position-based scheme when necessary. Now the problem is that mixing the two sensibly was not possible: the position-based mapping would map based on position, but would not skip those variables that were mapped by name. Instead, the two sets of assignments would overlap. However, I cannot currently change the current behavior, because there are some backends that rely on it [I think mistakenly, but I'll send a message to llvmdev about that]. So I've added a new TableGen bit variable: noNamedPositionallyEncodedOperands, that can be used to cause the position-based mapping to skip variables mapped by name. llvm-svn: 203767
2014-03-13 15:57:54 +08:00
AddCodeToMergeInOperand(R, BI, Vals[i].getName(), NumberedOp,
NamedOpIndices, Case, Target);
}
std::string PostEmitter = R->getValueAsString("PostEncoderMethod");
if (!PostEmitter.empty()) {
Case += " Value = " + PostEmitter + "(MI, Value";
Case += ", STI";
Case += ");\n";
}
return Case;
}
void CodeEmitterGen::run(raw_ostream &o) {
CodeGenTarget Target(Records);
std::vector<Record*> Insts = Records.getAllDerivedDefinitions("Instruction");
2010-10-08 00:56:28 +08:00
// For little-endian instruction bit encodings, reverse the bit order
Target.reverseBitsForLittleEndianEncoding();
2010-10-08 00:56:28 +08:00
const std::vector<const CodeGenInstruction*> &NumberedInstructions =
Target.getInstructionsByEnumValue();
// Emit function declaration
o << "uint64_t " << Target.getName();
o << "MCCodeEmitter::getBinaryCodeForInstr(const MCInst &MI,\n"
<< " SmallVectorImpl<MCFixup> &Fixups,\n"
<< " const MCSubtargetInfo &STI) const {\n";
// Emit instruction base values
o << " static const uint64_t InstBits[] = {\n";
for (std::vector<const CodeGenInstruction*>::const_iterator
IN = NumberedInstructions.begin(),
EN = NumberedInstructions.end();
IN != EN; ++IN) {
const CodeGenInstruction *CGI = *IN;
Record *R = CGI->TheDef;
2010-10-08 00:56:28 +08:00
if (R->getValueAsString("Namespace") == "TargetOpcode" ||
R->getValueAsBit("isPseudo")) {
o << " UINT64_C(0),\n";
continue;
}
2010-10-08 00:56:28 +08:00
BitsInit *BI = R->getValueAsBitsInit("Inst");
// Start by filling in fixed values.
uint64_t Value = 0;
for (unsigned i = 0, e = BI->getNumBits(); i != e; ++i) {
if (BitInit *B = dyn_cast<BitInit>(BI->getBit(e-i-1)))
Value |= (uint64_t)B->getValue() << (e-i-1);
}
o << " UINT64_C(" << Value << ")," << '\t' << "// " << R->getName() << "\n";
}
o << " UINT64_C(0)\n };\n";
2010-10-08 00:56:28 +08:00
// Map to accumulate all the cases.
std::map<std::string, std::vector<std::string> > CaseMap;
2010-10-08 00:56:28 +08:00
// Construct all cases statement for each opcode
for (std::vector<Record*>::iterator IC = Insts.begin(), EC = Insts.end();
IC != EC; ++IC) {
Record *R = *IC;
if (R->getValueAsString("Namespace") == "TargetOpcode" ||
R->getValueAsBit("isPseudo"))
continue;
2011-02-04 07:26:36 +08:00
const std::string &InstName = R->getValueAsString("Namespace") + "::"
+ R->getName();
std::string Case = getInstructionCase(R, Target);
CaseMap[Case].push_back(InstName);
}
// Emit initial function code
o << " const unsigned opcode = MI.getOpcode();\n"
<< " uint64_t Value = InstBits[opcode];\n"
<< " uint64_t op = 0;\n"
<< " (void)op; // suppress warning\n"
<< " switch (opcode) {\n";
// Emit each case statement
std::map<std::string, std::vector<std::string> >::iterator IE, EE;
for (IE = CaseMap.begin(), EE = CaseMap.end(); IE != EE; ++IE) {
const std::string &Case = IE->first;
std::vector<std::string> &InstList = IE->second;
for (int i = 0, N = InstList.size(); i < N; i++) {
if (i) o << "\n";
2011-02-04 07:26:36 +08:00
o << " case " << InstList[i] << ":";
}
o << " {\n";
o << Case;
o << " break;\n"
<< " }\n";
}
// Default case: unhandled opcode
o << " default:\n"
<< " std::string msg;\n"
<< " raw_string_ostream Msg(msg);\n"
<< " Msg << \"Not supported instr: \" << MI;\n"
<< " report_fatal_error(Msg.str());\n"
<< " }\n"
<< " return Value;\n"
<< "}\n\n";
}
} // End anonymous namespace
namespace llvm {
void EmitCodeEmitter(RecordKeeper &RK, raw_ostream &OS) {
emitSourceFileHeader("Machine Code Emitter", OS);
CodeEmitterGen(RK).run(OS);
}
} // End llvm namespace