Page MenuHomePhabricator
Differential D138736 Diff 478716 llvm/lib/Target/PowerPC/PPCISelLowering.cpp

Changeset View

Standalone View

View Options

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 10,180 Lines • ▼ Show 20 Lines if (Subtarget.isISA3_0() && (V1->hasOneUse() || V2->hasOneUse())) {
// if V2 is dead, then we swap V1 and V2 so we can // if V2 is dead, then we swap V1 and V2 so we can
// use V2 as the destination instead. // use V2 as the destination instead.
if (!V1->hasOneUse() && V2->hasOneUse()) { if (!V1->hasOneUse() && V2->hasOneUse()) {
std::swap(V1, V2); std::swap(V1, V2);
NeedSwap = !NeedSwap; NeedSwap = !NeedSwap;
} }
} }
bool V1HasXXSWAPD = V1->getOperand(0)->getOpcode() == PPCISD::XXSWAPD;
bool V2HasXXSWAPD = V2->getOperand(0)->getOpcode() == PPCISD::XXSWAPD;
// The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
// that it is in input element units, not in bytes. Convert now. // that it is in input element units, not in bytes. Convert now.
// For little endian, the order of the input vectors is reversed, and // For little endian, the order of the input vectors is reversed, and
// the permutation mask is complemented with respect to 31. This is // the permutation mask is complemented with respect to 31. This is
// necessary to produce proper semantics with the big-endian-based vperm // necessary to produce proper semantics with the big-endian-based vperm
// instruction. // instruction.
EVT EltVT = V1.getValueType().getVectorElementType(); EVT EltVT = V1.getValueType().getVectorElementType();
unsigned BytesPerElement = EltVT.getSizeInBits() / 8; unsigned BytesPerElement = EltVT.getSizeInBits() / 8;
bool V1HasXXSWAPD = V1->getOperand(0)->getOpcode() == PPCISD::XXSWAPD;
bool V2HasXXSWAPD = V2->getOperand(0)->getOpcode() == PPCISD::XXSWAPD;
/* /*
Vectors will be appended like so: [ V1 | v2 ] Vectors will be appended like so: [ V1 | v2 ]
XXSWAPD on V1: XXSWAPD on V1:
stefanpUnsubmitted
Not Done
ReplyInline Actions

I don't think you need to bother with this extra ternary operator.
Everywhere you use V1HasXXSWAPD or V2HasXXSWAPD seems to already be guarded by Opcode == PPCISD::XXPERM so it doesn't matter what value those two bool values have in cases when you are not using the XXPERM oeprand.

stefanp: I don't think you need to bother with this extra ternary operator. Everywhere you use…
[ A | B | C | D ] -> [ C | D | A | B ] [ A | B | C | D ] -> [ C | D | A | B ]
0-3 4-7 8-11 12-15 0-3 4-7 8-11 12-15 0-3 4-7 8-11 12-15 0-3 4-7 8-11 12-15
i.e. index of A, B += 8, and index of C, D -= 8. i.e. index of A, B += 8, and index of C, D -= 8.
XXSWAPD on V2: XXSWAPD on V2:
[ E | F | G | H ] -> [ G | H | E | F ] [ E | F | G | H ] -> [ G | H | E | F ]
16-19 20-23 24-27 28-31 16-19 20-23 24-27 28-31 16-19 20-23 24-27 28-31 16-19 20-23 24-27 28-31
i.e. index of E, F += 8, index of G, H -= 8 i.e. index of E, F += 8, index of G, H -= 8
Swap V1 and V2: Swap V1 and V2:
[ V1 | V2 ] -> [ V2 | V1 ] [ V1 | V2 ] -> [ V2 | V1 ]
0-15 16-31 0-15 16-31 0-15 16-31 0-15 16-31
i.e. index of V1 += 16, index of V2 -= 16 i.e. index of V1 += 16, index of V2 -= 16
*/ */
SmallVector<SDValue, 16> ResultMask; SmallVector<SDValue, 16> ResultMask;
for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) { for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i]; unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
if (Opcode == PPCISD::XXPERM) {
if (V1HasXXSWAPD) { if (V1HasXXSWAPD) {
if (SrcElt < 8) if (SrcElt < 8)
SrcElt += 8; SrcElt += 8;
else if (SrcElt < 16) else if (SrcElt < 16)
SrcElt -= 8; SrcElt -= 8;
} }
if (V2HasXXSWAPD) { if (V2HasXXSWAPD) {
if (SrcElt > 23) if (SrcElt > 23)
SrcElt -= 8; SrcElt -= 8;
else if (SrcElt > 15) else if (SrcElt > 15)
SrcElt += 8; SrcElt += 8;
} }
if (NeedSwap) { if (NeedSwap) {
if (SrcElt < 16) if (SrcElt < 16)
SrcElt += 16; SrcElt += 16;
else else
SrcElt -= 16; SrcElt -= 16;
} }
}
for (unsigned j = 0; j != BytesPerElement; ++j) for (unsigned j = 0; j != BytesPerElement; ++j)
if (isLittleEndian) if (isLittleEndian)
ResultMask.push_back( ResultMask.push_back(
DAG.getConstant(31 - (SrcElt * BytesPerElement + j), dl, MVT::i32)); DAG.getConstant(31 - (SrcElt * BytesPerElement + j), dl, MVT::i32));
else else
ResultMask.push_back( ResultMask.push_back(
DAG.getConstant(SrcElt * BytesPerElement + j, dl, MVT::i32)); DAG.getConstant(SrcElt * BytesPerElement + j, dl, MVT::i32));
} }
if (Opcode == PPCISD::XXPERM && (V1HasXXSWAPD || V2HasXXSWAPD)) {
if (V1HasXXSWAPD) { if (V1HasXXSWAPD) {
dl = SDLoc(V1->getOperand(0)); dl = SDLoc(V1->getOperand(0));
V1 = V1->getOperand(0)->getOperand(1); V1 = V1->getOperand(0)->getOperand(1);
} }
if (V2HasXXSWAPD) { if (V2HasXXSWAPD) {
dl = SDLoc(V2->getOperand(0)); dl = SDLoc(V2->getOperand(0));
V2 = V2->getOperand(0)->getOperand(1); V2 = V2->getOperand(0)->getOperand(1);
} }
if (V1HasXXSWAPD || V2HasXXSWAPD || Opcode == PPCISD::XXPERM) {
if (isPPC64 && ValType != MVT::v2f64) if (isPPC64 && ValType != MVT::v2f64)
V1 = DAG.getBitcast(MVT::v2f64, V1); V1 = DAG.getBitcast(MVT::v2f64, V1);
if (isPPC64 && V2.getValueType() != MVT::v2f64) if (isPPC64 && V2.getValueType() != MVT::v2f64)
V2 = DAG.getBitcast(MVT::v2f64, V2); V2 = DAG.getBitcast(MVT::v2f64, V2);
} }
stefanpUnsubmitted
Not Done
ReplyInline Actions

For this section how about:

if (Opcode == PPCISD::XXPERM && (V1HasXXSWAPD || V2HasXXSWAPD)) {
  /// Same code as above but you don't need to check that (V1HasXXSWAPD || V2HasXXSWAPD)
}

You don't need to go into the if statement at all if there isn't at least one XXSWAPD.

stefanp: For this section how about: ``` if (Opcode == PPCISD::XXPERM && (V1HasXXSWAPD || V2HasXXSWAPD))…
ShufflesHandledWithVPERM++; ShufflesHandledWithVPERM++;
SDValue VPermMask = DAG.getBuildVector(MVT::v16i8, dl, ResultMask); SDValue VPermMask = DAG.getBuildVector(MVT::v16i8, dl, ResultMask);
LLVM_DEBUG({ LLVM_DEBUG({
stefanpUnsubmitted
Not Done
ReplyInline Actions

Is it possible to have a swap on both sides?

stefanp: Is it possible to have a swap on both sides?
maryammoAuthorUnsubmitted
Done
ReplyInline Actions

I have not see such a case.
btw, this patch is to address the vperm issue, so we can look at the swap on both sides for xxperm if needed separately. Thanks.

maryammo: I have not see such a case. btw, this patch is to address the vperm issue, so we can look at…
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
if (Opcode == PPCISD::XXPERM) { if (Opcode == PPCISD::XXPERM) {
dbgs() << "Emitting a XXPERM for the following shuffle:\n"; dbgs() << "Emitting a XXPERM for the following shuffle:\n";
} else { } else {
dbgs() << "Emitting a VPERM for the following shuffle:\n"; dbgs() << "Emitting a VPERM for the following shuffle:\n";
} }
SVOp->dump(); SVOp->dump();
dbgs() << "With the following permute control vector:\n"; dbgs() << "With the following permute control vector:\n";
▲ Show 20 Lines • Show All 8,136 Lines • Show Last 20 Lines