6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-01-28 06:32:09 +00:00
Code Issues Projects Releases Wiki Activity

R600/SI: improve post ISel folding

Browse Source 
Not only fold immediates, but avoid unnecessary copies as well.

Signed-off-by: Christian König <christian.koenig@amd.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178024 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Christian Konig 2013-03-26 14:04:17 +00:00
parent 3851e9869f
commit 0c4e61ff0a
1 changed files with 44 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

76
lib/Target/R600/SIISelLowering.cpp
View File

@ -484,22 +484,23 @@ bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, SDValue &Op,
MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
SDNode *Node = Op.getNode();
int OpClass;
const TargetRegisterClass *OpClass;
if (MachineSDNode *MN = dyn_cast<MachineSDNode>(Node)) {
const MCInstrDesc &Desc = TII->get(MN->getMachineOpcode());
OpClass = Desc.OpInfo[Op.getResNo()].RegClass;
int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
if (OpClassID == -1)
OpClass = getRegClassFor(Op.getSimpleValueType());
else
OpClass = TRI->getRegClass(OpClassID);
} else if (Node->getOpcode() == ISD::CopyFromReg) {
RegisterSDNode *Reg = cast<RegisterSDNode>(Node->getOperand(1).getNode());
OpClass = MRI.getRegClass(Reg->getReg())->getID();
OpClass = MRI.getRegClass(Reg->getReg());
} else
return false;
if (OpClass == -1)
return false;
return TRI->getRegClass(RegClass)->hasSubClassEq(TRI->getRegClass(OpClass));
return TRI->getRegClass(RegClass)->hasSubClassEq(OpClass);
}
/// \brief Make sure that we don't exeed the number of allowed scalars
@ -595,41 +596,52 @@ SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
// Is this a VSrc or SSrc operand ?
unsigned RegClass = Desc->OpInfo[Op].RegClass;
if (!isVSrc(RegClass) && !isSSrc(RegClass)) {
if (isVSrc(RegClass) || isSSrc(RegClass)) {
// Try to fold the immediates
if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
// Folding didn't worked, make sure we don't hit the SReg limit
ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
}
continue;
}
if (i == 1 && Desc->isCommutable() &&
fitsRegClass(DAG, Ops[0], RegClass) &&
foldImm(Ops[1], Immediate, ScalarSlotUsed)) {
if (i == 1 && Desc->isCommutable() &&
fitsRegClass(DAG, Ops[0], RegClass)) {
assert(isVSrc(Desc->OpInfo[NumDefs].RegClass) ||
isSSrc(Desc->OpInfo[NumDefs].RegClass));
unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
// Test if it makes sense to swap operands
if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
(!fitsRegClass(DAG, Ops[1], RegClass) &&
fitsRegClass(DAG, Ops[1], OtherRegClass))) {
// Swap commutable operands
SDValue Tmp = Ops[1];
Ops[1] = Ops[0];
Ops[0] = Tmp;
} else if (DescE64 && !Immediate) {
// Test if it makes sense to switch to e64 encoding
RegClass = DescE64->OpInfo[Op].RegClass;
int32_t TmpImm = -1;
if ((isVSrc(RegClass) || isSSrc(RegClass)) &&
foldImm(Ops[i], TmpImm, ScalarSlotUsed)) {
Immediate = -1;
Promote2e64 = true;
Desc = DescE64;
DescE64 = 0;
}
continue;
}
continue;
}
// Try to fold the immediates
if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
// Folding didn't worked, make sure we don't hit the SReg limit
ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
if (DescE64 && !Immediate) {
// Test if it makes sense to switch to e64 encoding
unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
continue;
int32_t TmpImm = -1;
if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
(!fitsRegClass(DAG, Ops[i], RegClass) &&
fitsRegClass(DAG, Ops[1], OtherRegClass))) {
// Switch to e64 encoding
Immediate = -1;
Promote2e64 = true;
Desc = DescE64;
DescE64 = 0;
}
}
}