/**
* Returns an aligned/unaligned move instruction depending on the operand and, if the operand is
* an address, on the displacement. The method assumes that any addresses (both base addresses and
* indices) are aligned at vector boundaries.
*
* @param arch The architecture description
* @param specDatatype A datatype or <code>null</code> if a generic data type is to be used (if
* all SIMD vectors have the same length in bits, <code>null</code> is fine)
* @param rgOperands The operands to examine and from which to determine what move instruction to
* use
* @return The name of the move instruction to use to move the operands <code>rgOperands</code>
*/
public static TypeBaseIntrinsicEnum getMovFpr(
CodeGeneratorSharedObjects data,
boolean bIsLoad,
Specifier specDatatype,
IOperand... rgOperands) {
if (data.getOptions().isAlwaysUseNonalignedMoves())
return bIsLoad
? TypeBaseIntrinsicEnum.LOAD_FPR_UNALIGNED
: TypeBaseIntrinsicEnum.STORE_FPR_UNALIGNED;
int nVectorLength = -1;
for (IOperand op : rgOperands) {
if (op instanceof IOperand.Address) {
// assume that base addresses are aligned at vector boundaries
// so if the displacement is not a multiple of the vector length, we need to do an
// unaligned load
if (nVectorLength == -1) {
Specifier specType = specDatatype == null ? Globals.BASE_DATATYPES[0] : specDatatype;
nVectorLength =
data.getArchitectureDescription().getSIMDVectorLength(specType)
* ArchitectureDescriptionManager.getTypeSize(specType);
}
if ((((IOperand.Address) op).getDisplacement() % nVectorLength) != 0)
return bIsLoad
? TypeBaseIntrinsicEnum.LOAD_FPR_UNALIGNED
: TypeBaseIntrinsicEnum.STORE_FPR_UNALIGNED;
}
}
return bIsLoad
? TypeBaseIntrinsicEnum.LOAD_FPR_ALIGNED
: TypeBaseIntrinsicEnum.STORE_FPR_ALIGNED;
}
/**
* Runs the instruction list translator.
*
* @return The translated instruction list
*/
private InstructionList run() {
IArchitectureDescription arch = m_data.getArchitectureDescription();
// don't use register reusing for now (screws up the instruction scheduler)
final boolean bUseRegisterRemover = /*true*/ false && arch.hasNonDestructiveOperations();
InstructionList ilIn =
bUseRegisterRemover
? new UnneededPseudoRegistersRemover(
arch, InstructionList.EInstructionListType.GENERIC, m_setReusedRegisters)
.optimize(m_ilIn)
: m_ilIn;
for (IInstruction instruction : ilIn) translateInstruction(instruction);
return m_ilOut;
}
/**
* Creates the architecture-specific instructions to implement the generic instruction <code>
* instruction</code> for the intrinsic <code>intrinsic</code>.
*
* @param instruction The generic instruction
* @param intrinsic The intrinsic corresponding to <code>instruction</code>
* @param rgSourceOps The array of operands of the generic instruction
* @param rgDestArgs The array of intrinsic arguments
* @param rgPermSourceToDest The argument permutation source → destination (where source is
* the generic instruction, destination is the target architecture specific instruction)
* @param rgPermDestToSource The argument permutation destination → source (where source is
* the generic instruction, destination is the target architecture specific instruction)
* @param nOutputArgDestIndex The index of the output argument in the array of intrinsic
* arguments, <code>rgDestArgs</code>
* @param bIntrinsicHasSharedResult <code>true</code> iff the intrinsic requires that an argument
* is a shared in/out
*/
private void createInstructions(
Instruction instruction,
Intrinsic intrinsic,
IOperand[] rgSourceOps,
Argument[] rgDestArgs,
int[] rgPermSourceToDest,
int[] rgPermDestToSource,
int nOutputArgDestIndex,
boolean bIntrinsicHasSharedResult) {
// maps operands to substitute operands within the actual generated computation instruction
Map<IOperand, IOperand> mapSubstitutions = new HashMap<>();
IOperand[] rgDestOps = new IOperand[rgDestArgs.length];
IOperand opSourceOutput = rgSourceOps[rgSourceOps.length - 1];
boolean bHasNonCompatibleResultOperand = false;
IOperand opTmpResultOperand = null;
if (bIntrinsicHasSharedResult) {
// find the operand which, in the intrinsic, is both input and output
IOperand opShared = rgSourceOps[rgPermDestToSource[nOutputArgDestIndex]];
// if the respective input and the output arguments are different, move the value of the input
// to the result
// the result will then be overwritten by the intrinsic
if (!opSourceOutput.equals(opShared)) {
IOperand opOut = opSourceOutput;
boolean bIsOneOfNonSharedInputArgsResult =
InstructionListTranslator.getIndexOfNonSharedInputArgsResult(rgSourceOps, opShared)
!= -1;
if (!(opSourceOutput instanceof IOperand.IRegisterOperand)
|| bIsOneOfNonSharedInputArgsResult) {
bHasNonCompatibleResultOperand = true;
opTmpResultOperand = new IOperand.PseudoRegister(TypeRegisterType.SIMD);
opOut = opTmpResultOperand;
}
// opOut can replace both opShared (the input operand, which in the architecture-specific
// intrinsic
// is also an output argument) and opOut (the operand, to which the result is written)
mapSubstitutions.put(opShared, opOut);
if (!bIsOneOfNonSharedInputArgsResult) mapSubstitutions.put(opSourceOutput, opOut);
Instruction instrNewMov =
new Instruction(
getMovFpr(opShared instanceof IOperand.Address, opShared), opShared, opOut);
instrNewMov.setParameterAssignment(instruction.getParameterAssignment());
translateInstruction(instrNewMov);
}
}
// gather operands and issue move instructions for non-compatible operands
for (int i = 0; i < rgSourceOps.length; i++) {
if (rgPermSourceToDest[i] != UNDEFINED) {
boolean bIsResultOperand = i == rgSourceOps.length - 1;
IOperand opSubstitute = mapSubstitutions.get(rgSourceOps[i]);
if (opSubstitute != null) {
// if already a non-compatible result operand has been found,
// substitute the corresponding operand with the temporary one
rgDestOps[rgPermSourceToDest[i]] = opSubstitute;
} else {
boolean bIsCompatible = isCompatible(rgSourceOps[i], rgDestArgs[rgPermSourceToDest[i]]);
rgDestOps[rgPermSourceToDest[i]] =
bIsCompatible ? rgSourceOps[i] : new IOperand.PseudoRegister(TypeRegisterType.SIMD);
if (!bIsCompatible) {
if (bIsResultOperand) {
// this is the result operand
// move instruction will be generated after issuing the main instruction
bHasNonCompatibleResultOperand = true;
opTmpResultOperand = rgDestOps[rgPermSourceToDest[i]];
} else {
// mov arg_i, tmp
mapSubstitutions.put(rgSourceOps[i], rgDestOps[rgPermSourceToDest[i]]);
Instruction instrNewMov =
new Instruction(
getMovFpr(rgSourceOps[i] instanceof IOperand.Address, rgSourceOps[i]),
rgSourceOps[i],
rgDestOps[rgPermSourceToDest[i]]);
instrNewMov.setParameterAssignment(instruction.getParameterAssignment());
translateInstruction(instrNewMov);
}
}
}
}
}
// add the main instruction
////
// if (instruction.getIntrinsicBaseName ().equals (TypeBaseIntrinsicEnum.DIVIDE.value ()))
// {
// IOperand.PseudoRegister opTmp = new IOperand.PseudoRegister (TypeRegisterType.SIMD);
// m_ilOut.addInstruction (new Instruction ("vrcpps", rgDestOps[0], opTmp));
// m_ilOut.addInstruction (new Instruction ("vmulps", rgDestOps[1], opTmp, rgDestOps[2]));
// }
// else
////
String strInstruction = intrinsic.getName();
boolean bIsLoad =
intrinsic.getBaseName().equals(TypeBaseIntrinsicEnum.LOAD_FPR_ALIGNED.value());
boolean bIsStore =
intrinsic.getBaseName().equals(TypeBaseIntrinsicEnum.STORE_FPR_ALIGNED.value());
if (bIsLoad || bIsStore) {
Intrinsic i =
m_data
.getArchitectureDescription()
.getIntrinsic(getMovFpr(bIsLoad, rgDestOps).value(), m_specDatatype);
if (i != null) strInstruction = i.getName();
}
Instruction instrNew = new Instruction(strInstruction, instruction.getIntrinsic(), rgDestOps);
instrNew.setParameterAssignment(instruction.getParameterAssignment());
m_ilOut.addInstruction(instrNew);
// add a move-result instruction if needed
if (bHasNonCompatibleResultOperand) {
// mov tmp, result
Instruction instrNewMov =
new Instruction(
getMovFpr(opTmpResultOperand instanceof IOperand.Address, opTmpResultOperand),
opTmpResultOperand,
rgSourceOps[rgSourceOps.length - 1]);
instrNewMov.setParameterAssignment(instruction.getParameterAssignment());
translateInstruction(instrNewMov);
}
}
/**
* Finds the intrinsic for the instruction <code>instruction</code>.
*
* @param instruction The instruction for which to find the corresponding intrinsic.
* @return The instrinsic corresponding to the instruction <code>instruction</code>
*/
private Intrinsic getIntrinsicForInstruction(Instruction instruction) {
return m_data
.getArchitectureDescription()
.getIntrinsic(instruction.getInstructionName(), m_specDatatype);
}
