fix a long standing fixme, which required fixing a bunch of other

issues to get here.  We now trim the result type list of the
CompleteMatch or MorphNodeTo operation to be the same size as the
thing we're matching.  this means that if you match (add GPR, GPR)
with an instruction that produces a normal result and a flag that
we now trim the result in tblgen instead of having to do it 
dynamically.  This exposed a bunch of inconsistencies in result
counting that happened to be getting lucky since the days of the
old isel.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@99728 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Chris Lattner 2010-03-27 20:45:15 +00:00
parent 0d7952ed5a
commit b4a52b04bd

View File

@ -729,7 +729,7 @@ EmitResultInstructionAsOperand(const TreePatternNode *N,
// (set AL, (mul ...))
// This also handles implicit results like:
// (implicit EFLAGS)
if (isRoot && Pattern.getDstRegs().size() != 0) {
if (isRoot && !Pattern.getDstRegs().empty()) {
// If the root came from an implicit def in the instruction handling stuff,
// don't re-add it.
Record *HandledReg = 0;
@ -837,33 +837,35 @@ void MatcherGen::EmitResultCode() {
// At this point, we have however many values the result pattern produces.
// However, the input pattern might not need all of these. If there are
// excess values at the end (such as condition codes etc) just lop them off.
// This doesn't need to worry about flags or chains, just explicit results.
// excess values at the end (such as implicit defs of condition codes etc)
// just lop them off. This doesn't need to worry about flags or chains, just
// explicit results.
//
// FIXME2: This doesn't work because there is currently no way to get an
// accurate count of the # results the source pattern sets. This is because
// of the "parallel" construct in X86 land, which looks like this:
//
//def : Pat<(parallel (X86and_flag GR8:$src1, GR8:$src2),
// (implicit EFLAGS)),
// (AND8rr GR8:$src1, GR8:$src2)>;
//
// This idiom means to match the two-result node X86and_flag (which is
// declared as returning a single result, because we can't match multi-result
// nodes yet). In this case, we would have to know that the input has two
// results. However, mul8r is modelled exactly the same way, but without
// implicit defs included. The fix is to support multiple results directly
// and eliminate 'parallel'.
//
// FIXME2: When this is fixed, we should revert the terrible hack in the
// OPC_EmitNode code in the interpreter.
#if 0
const TreePatternNode *Src = Pattern.getSrcPattern();
unsigned NumSrcResults = Src->getTypeNum(0) != MVT::isVoid ? 1 : 0;
NumSrcResults += Pattern.getDstRegs().size();
unsigned NumSrcResults = Pattern.getSrcPattern()->getNumTypes();
// If the pattern also has (implicit) results, count them as well.
if (!Pattern.getDstRegs().empty()) {
// If the root came from an implicit def in the instruction handling stuff,
// don't re-add it.
Record *HandledReg = 0;
const TreePatternNode *DstPat = Pattern.getDstPattern();
if (!DstPat->isLeaf() &&DstPat->getOperator()->isSubClassOf("Instruction")){
const CodeGenTarget &CGT = CGP.getTargetInfo();
CodeGenInstruction &II = CGT.getInstruction(DstPat->getOperator());
if (II.HasOneImplicitDefWithKnownVT(CGT) != MVT::Other)
HandledReg = II.ImplicitDefs[0];
}
for (unsigned i = 0; i != Pattern.getDstRegs().size(); ++i) {
Record *Reg = Pattern.getDstRegs()[i];
if (!Reg->isSubClassOf("Register") || Reg == HandledReg) continue;
++NumSrcResults;
}
}
assert(Ops.size() >= NumSrcResults && "Didn't provide enough results");
Ops.resize(NumSrcResults);
#endif
// If the matched pattern covers nodes which define a flag result, emit a node
// that tells the matcher about them so that it can update their results.