return false;
}
+// Insert a node into the DAG at least before the Pos node's position. This
+// will reposition the node as needed, and will assign it a node ID that is <=
+// the Pos node's ID. Note that this does *not* preserve the uniqueness of node
+// IDs! The selection DAG must no longer depend on their uniqueness when this
+// is used.
+static void InsertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
+ if (N.getNode()->getNodeId() == -1 ||
+ N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
+ DAG.RepositionNode(Pos.getNode(), N.getNode());
+ N.getNode()->setNodeId(Pos.getNode()->getNodeId());
+ }
+}
+
+// Transform "(X >> (8-C1)) & C2" to "(X >> 8) & 0xff)" if safe. This
+// allows us to convert the shift and and into an h-register extract and
+// a scaled index. Returns false if the simplification is performed.
+static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
+ uint64_t Mask,
+ SDValue Shift, SDValue X,
+ X86ISelAddressMode &AM) {
+ if (Shift.getOpcode() != ISD::SRL ||
+ !isa<ConstantSDNode>(Shift.getOperand(1)) ||
+ !Shift.hasOneUse())
+ return true;
+
+ int ScaleLog = 8 - Shift.getConstantOperandVal(1);
+ if (ScaleLog <= 0 || ScaleLog >= 4 ||
+ Mask != (0xffu << ScaleLog))
+ return true;
+
+ EVT VT = N.getValueType();
+ DebugLoc DL = N.getDebugLoc();
+ SDValue Eight = DAG.getConstant(8, MVT::i8);
+ SDValue NewMask = DAG.getConstant(0xff, VT);
+ SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
+ SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
+ SDValue ShlCount = DAG.getConstant(ScaleLog, MVT::i8);
+ SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
+
+ // Insert the new nodes into the topological ordering. We must do this in
+ // a valid topological ordering as nothing is going to go back and re-sort
+ // these nodes. We continually insert before 'N' in sequence as this is
+ // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
+ // hierarchy left to express.
+ InsertDAGNode(DAG, N, Eight);
+ InsertDAGNode(DAG, N, Srl);
+ InsertDAGNode(DAG, N, NewMask);
+ InsertDAGNode(DAG, N, And);
+ InsertDAGNode(DAG, N, ShlCount);
+ InsertDAGNode(DAG, N, Shl);
+ DAG.ReplaceAllUsesWith(N, Shl);
+ AM.IndexReg = And;
+ AM.Scale = (1 << ScaleLog);
+ return false;
+}
+
+// Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
+// allows us to fold the shift into this addressing mode. Returns false if the
+// transform succeeded.
+static bool FoldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
+ uint64_t Mask,
+ SDValue Shift, SDValue X,
+ X86ISelAddressMode &AM) {
+ if (Shift.getOpcode() != ISD::SHL ||
+ !isa<ConstantSDNode>(Shift.getOperand(1)))
+ return true;
+
+ // Not likely to be profitable if either the AND or SHIFT node has more
+ // than one use (unless all uses are for address computation). Besides,
+ // isel mechanism requires their node ids to be reused.
+ if (!N.hasOneUse() || !Shift.hasOneUse())
+ return true;
+
+ // Verify that the shift amount is something we can fold.
+ unsigned ShiftAmt = Shift.getConstantOperandVal(1);
+ if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
+ return true;
+
+ EVT VT = N.getValueType();
+ DebugLoc DL = N.getDebugLoc();
+ SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, VT);
+ SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
+ SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
+
+ // Insert the new nodes into the topological ordering. We must do this in
+ // a valid topological ordering as nothing is going to go back and re-sort
+ // these nodes. We continually insert before 'N' in sequence as this is
+ // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
+ // hierarchy left to express.
+ InsertDAGNode(DAG, N, NewMask);
+ InsertDAGNode(DAG, N, NewAnd);
+ InsertDAGNode(DAG, N, NewShift);
+ DAG.ReplaceAllUsesWith(N, NewShift);
+
+ AM.Scale = 1 << ShiftAmt;
+ AM.IndexReg = NewAnd;
+ return false;
+}
+
+// Implement some heroics to detect shifts of masked values where the mask can
+// be replaced by extending the shift and undoing that in the addressing mode
+// scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
+// (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
+// the addressing mode. This results in code such as:
+//
+// int f(short *y, int *lookup_table) {
+// ...
+// return *y + lookup_table[*y >> 11];
+// }
+//
+// Turning into:
+// movzwl (%rdi), %eax
+// movl %eax, %ecx
+// shrl $11, %ecx
+// addl (%rsi,%rcx,4), %eax
+//
+// Instead of:
+// movzwl (%rdi), %eax
+// movl %eax, %ecx
+// shrl $9, %ecx
+// andl $124, %rcx
+// addl (%rsi,%rcx), %eax
+//
+// Note that this function assumes the mask is provided as a mask *after* the
+// value is shifted. The input chain may or may not match that, but computing
+// such a mask is trivial.
+static bool FoldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
+ uint64_t Mask,
+ SDValue Shift, SDValue X,
+ X86ISelAddressMode &AM) {
+ if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
+ !isa<ConstantSDNode>(Shift.getOperand(1)))
+ return true;
+
+ unsigned ShiftAmt = Shift.getConstantOperandVal(1);
+ unsigned MaskLZ = CountLeadingZeros_64(Mask);
+ unsigned MaskTZ = CountTrailingZeros_64(Mask);
+
+ // The amount of shift we're trying to fit into the addressing mode is taken
+ // from the trailing zeros of the mask.
+ unsigned AMShiftAmt = MaskTZ;
+
+ // There is nothing we can do here unless the mask is removing some bits.
+ // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
+ if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
+
+ // We also need to ensure that mask is a continuous run of bits.
+ if (CountTrailingOnes_64(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
+
+ // Scale the leading zero count down based on the actual size of the value.
+ // Also scale it down based on the size of the shift.
+ MaskLZ -= (64 - X.getValueSizeInBits()) + ShiftAmt;
+
+ // The final check is to ensure that any masked out high bits of X are
+ // already known to be zero. Otherwise, the mask has a semantic impact
+ // other than masking out a couple of low bits. Unfortunately, because of
+ // the mask, zero extensions will be removed from operands in some cases.
+ // This code works extra hard to look through extensions because we can
+ // replace them with zero extensions cheaply if necessary.
+ bool ReplacingAnyExtend = false;
+ if (X.getOpcode() == ISD::ANY_EXTEND) {
+ unsigned ExtendBits =
+ X.getValueSizeInBits() - X.getOperand(0).getValueSizeInBits();
+ // Assume that we'll replace the any-extend with a zero-extend, and
+ // narrow the search to the extended value.
+ X = X.getOperand(0);
+ MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
+ ReplacingAnyExtend = true;
+ }
+ APInt MaskedHighBits = APInt::getHighBitsSet(X.getValueSizeInBits(),
+ MaskLZ);
+ APInt KnownZero, KnownOne;
+ DAG.ComputeMaskedBits(X, MaskedHighBits, KnownZero, KnownOne);
+ if (MaskedHighBits != KnownZero) return true;
+
+ // We've identified a pattern that can be transformed into a single shift
+ // and an addressing mode. Make it so.
+ EVT VT = N.getValueType();
+ if (ReplacingAnyExtend) {
+ assert(X.getValueType() != VT);
+ // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
+ SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, X.getDebugLoc(), VT, X);
+ InsertDAGNode(DAG, N, NewX);
+ X = NewX;
+ }
+ DebugLoc DL = N.getDebugLoc();
+ SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, MVT::i8);
+ SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
+ SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, MVT::i8);
+ SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
+
+ // Insert the new nodes into the topological ordering. We must do this in
+ // a valid topological ordering as nothing is going to go back and re-sort
+ // these nodes. We continually insert before 'N' in sequence as this is
+ // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
+ // hierarchy left to express.
+ InsertDAGNode(DAG, N, NewSRLAmt);
+ InsertDAGNode(DAG, N, NewSRL);
+ InsertDAGNode(DAG, N, NewSHLAmt);
+ InsertDAGNode(DAG, N, NewSHL);
+ DAG.ReplaceAllUsesWith(N, NewSHL);
+
+ AM.Scale = 1 << AMShiftAmt;
+ AM.IndexReg = NewSRL;
+ return false;
+}
+
bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
unsigned Depth) {
DebugLoc dl = N.getDebugLoc();
break;
}
+ case ISD::SRL: {
+ // Scale must not be used already.
+ if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
+
+ SDValue And = N.getOperand(0);
+ if (And.getOpcode() != ISD::AND) break;
+ SDValue X = And.getOperand(0);
+
+ // We only handle up to 64-bit values here as those are what matter for
+ // addressing mode optimizations.
+ if (X.getValueSizeInBits() > 64) break;
+
+ // The mask used for the transform is expected to be post-shift, but we
+ // found the shift first so just apply the shift to the mask before passing
+ // it down.
+ if (!isa<ConstantSDNode>(N.getOperand(1)) ||
+ !isa<ConstantSDNode>(And.getOperand(1)))
+ break;
+ uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
+
+ // Try to fold the mask and shift into the scale, and return false if we
+ // succeed.
+ if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
+ return false;
+ break;
+ }
+
case ISD::SMUL_LOHI:
case ISD::UMUL_LOHI:
// A mul_lohi where we need the low part can be folded as a plain multiply.
AM.Scale = 1;
// Insert the new nodes into the topological ordering.
- if (Zero.getNode()->getNodeId() == -1 ||
- Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), Zero.getNode());
- Zero.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- if (Neg.getNode()->getNodeId() == -1 ||
- Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), Neg.getNode());
- Neg.getNode()->setNodeId(N.getNode()->getNodeId());
- }
+ InsertDAGNode(*CurDAG, N, Zero);
+ InsertDAGNode(*CurDAG, N, Neg);
return false;
}
// Perform some heroic transforms on an and of a constant-count shift
// with a constant to enable use of the scaled offset field.
- SDValue Shift = N.getOperand(0);
- if (Shift.getNumOperands() != 2) break;
-
// Scale must not be used already.
if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
+ SDValue Shift = N.getOperand(0);
+ if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
SDValue X = Shift.getOperand(0);
- ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
- ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
- if (!C1 || !C2) break;
-
- // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
- // allows us to convert the shift and and into an h-register extract and
- // a scaled index.
- if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
- unsigned ScaleLog = 8 - C1->getZExtValue();
- if (ScaleLog > 0 && ScaleLog < 4 &&
- C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
- SDValue Eight = CurDAG->getConstant(8, MVT::i8);
- SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
- SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
- X, Eight);
- SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
- Srl, Mask);
- SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
- SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
- And, ShlCount);
-
- // Insert the new nodes into the topological ordering.
- if (Eight.getNode()->getNodeId() == -1 ||
- Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), Eight.getNode());
- Eight.getNode()->setNodeId(X.getNode()->getNodeId());
- }
- if (Mask.getNode()->getNodeId() == -1 ||
- Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), Mask.getNode());
- Mask.getNode()->setNodeId(X.getNode()->getNodeId());
- }
- if (Srl.getNode()->getNodeId() == -1 ||
- Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
- CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
- Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
- }
- if (And.getNode()->getNodeId() == -1 ||
- And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), And.getNode());
- And.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- if (ShlCount.getNode()->getNodeId() == -1 ||
- ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
- ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- if (Shl.getNode()->getNodeId() == -1 ||
- Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), Shl.getNode());
- Shl.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- CurDAG->ReplaceAllUsesWith(N, Shl);
- AM.IndexReg = And;
- AM.Scale = (1 << ScaleLog);
- return false;
- }
- }
- // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
- // allows us to fold the shift into this addressing mode.
- if (Shift.getOpcode() != ISD::SHL) break;
+ // We only handle up to 64-bit values here as those are what matter for
+ // addressing mode optimizations.
+ if (X.getValueSizeInBits() > 64) break;
- // Not likely to be profitable if either the AND or SHIFT node has more
- // than one use (unless all uses are for address computation). Besides,
- // isel mechanism requires their node ids to be reused.
- if (!N.hasOneUse() || !Shift.hasOneUse())
+ if (!isa<ConstantSDNode>(N.getOperand(1)))
break;
-
- // Verify that the shift amount is something we can fold.
- unsigned ShiftCst = C1->getZExtValue();
- if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
- break;
-
- // Get the new AND mask, this folds to a constant.
- SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
- SDValue(C2, 0), SDValue(C1, 0));
- SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
- NewANDMask);
- SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
- NewAND, SDValue(C1, 0));
+ uint64_t Mask = N.getConstantOperandVal(1);
- // Insert the new nodes into the topological ordering.
- if (C1->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), C1);
- C1->setNodeId(X.getNode()->getNodeId());
- }
- if (NewANDMask.getNode()->getNodeId() == -1 ||
- NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
- NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
- }
- if (NewAND.getNode()->getNodeId() == -1 ||
- NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
- CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
- NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
- }
- if (NewSHIFT.getNode()->getNodeId() == -1 ||
- NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
- NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
- }
+ // Try to fold the mask and shift into an extract and scale.
+ if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
+ return false;
- CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
-
- AM.Scale = 1 << ShiftCst;
- AM.IndexReg = NewAND;
- return false;
+ // Try to fold the mask and shift directly into the scale.
+ if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
+ return false;
+
+ // Try to swap the mask and shift to place shifts which can be done as
+ // a scale on the outside of the mask.
+ if (!FoldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
+ return false;
+ break;
}
}
SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
getI8Imm(ShlVal));
- break;
}
case X86ISD::UMUL: {
SDValue N0 = Node->getOperand(0);
break;
}
case ISD::STORE: {
+ // The DEC64m tablegen pattern is currently not able to match the case where
+ // the EFLAGS on the original DEC are used.
+ // we'll need to improve tablegen to allow flags to be transferred from a
+ // node in the pattern to the result node. probably with a new keyword
+ // for example, we have this
+ // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
+ // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
+ // (implicit EFLAGS)]>;
+ // but maybe need something like this
+ // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
+ // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
+ // (transferrable EFLAGS)]>;
StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
SDValue Chain = StoreNode->getOperand(0);
SDValue StoredVal = StoreNode->getOperand(1);