Add a framework for eliminating instructions that produces undemanded bits.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25945 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 3124da5..8287f2d 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -858,7 +858,7 @@
SDOperand DAGCombiner::visitAND(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
- SDOperand LL, LR, RL, RR, CC0, CC1;
+ SDOperand LL, LR, RL, RR, CC0, CC1, Old, New;
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
MVT::ValueType VT = N1.getValueType();
@@ -884,13 +884,6 @@
SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
if (RAND.Val != 0)
return RAND;
- // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
- if (N1C && N0.getOpcode() == ISD::SIGN_EXTEND_INREG) {
- unsigned ExtendBits =
- MVT::getSizeInBits(cast<VTSDNode>(N0.getOperand(1))->getVT());
- if (ExtendBits == 64 || ((N1C->getValue() & (~0ULL << ExtendBits)) == 0))
- return DAG.getNode(ISD::AND, VT, N0.getOperand(0), N1);
- }
// fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
if (N1C && N0.getOpcode() == ISD::OR)
if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
@@ -966,6 +959,26 @@
WorkList.push_back(ANDNode.Val);
return DAG.getNode(N0.getOpcode(), VT, ANDNode, N0.getOperand(1));
}
+ // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
+ // fold (and (sra)) -> (and (srl)) when possible.
+ if (TLI.DemandedBitsAreZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits), Old,
+ New, DAG)) {
+ WorkList.push_back(N);
+ CombineTo(Old.Val, New);
+ return SDOperand();
+ }
+ // FIXME: DemandedBitsAreZero cannot currently handle AND with non-constant
+ // RHS and propagate known cleared bits to LHS. For this reason, we must keep
+ // this fold, for now, for the following testcase:
+ //
+ //int %test2(uint %mode.0.i.0) {
+ // %tmp.79 = cast uint %mode.0.i.0 to int
+ // %tmp.80 = shr int %tmp.79, ubyte 15
+ // %tmp.81 = shr uint %mode.0.i.0, ubyte 16
+ // %tmp.82 = cast uint %tmp.81 to int
+ // %tmp.83 = and int %tmp.80, %tmp.82
+ // ret int %tmp.83
+ //}
// fold (and (sra)) -> (and (srl)) when possible.
if (N0.getOpcode() == ISD::SRA && N0.Val->hasOneUse()) {
if (ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
@@ -1240,6 +1253,8 @@
SDOperand DAGCombiner::visitSHL(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
+ SDOperand Old = SDOperand();
+ SDOperand New = SDOperand();
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
MVT::ValueType VT = N0.getValueType();
@@ -1260,6 +1275,12 @@
// if (shl x, c) is known to be zero, return 0
if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
return DAG.getConstant(0, VT);
+ if (N1C && TLI.DemandedBitsAreZero(SDOperand(N,0), ~0ULL >> (64-OpSizeInBits),
+ Old, New, DAG)) {
+ WorkList.push_back(N);
+ CombineTo(Old.Val, New);
+ return SDOperand();
+ }
// fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
if (N1C && N0.getOpcode() == ISD::SHL &&
N0.getOperand(1).getOpcode() == ISD::Constant) {
@@ -1650,8 +1671,7 @@
return N0;
// fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
- return DAG.getNode(ISD::AND, N0.getValueType(), N0,
- DAG.getConstant(~0ULL >> (64-EVTBits), VT));
+ return DAG.getZeroExtendInReg(N0, EVT);
// fold (sext_in_reg (srl x)) -> sra x
if (N0.getOpcode() == ISD::SRL &&
N0.getOperand(1).getOpcode() == ISD::Constant &&
diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index c79045b..8951490 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -131,7 +131,63 @@
return NULL;
}
-
+/// DemandedBitsAreZero - Return true if 'Op & Mask' demands no bits from a bit
+/// set operation such as a sign extend or or/xor with constant whose only
+/// use is Op. If it returns true, the old node that sets bits which are
+/// not demanded is returned in Old, and its replacement node is returned in
+/// New, such that callers of SetBitsAreZero may call CombineTo on them if
+/// desired.
+bool TargetLowering::DemandedBitsAreZero(const SDOperand &Op, uint64_t Mask,
+ SDOperand &Old, SDOperand &New,
+ SelectionDAG &DAG) {
+ // If the operation has more than one use, we're not interested in it.
+ // Tracking down and checking all uses would be problematic and slow.
+ if (!Op.hasOneUse())
+ return false;
+
+ switch (Op.getOpcode()) {
+ case ISD::AND:
+ // (X & C1) & C2 == 0 iff C1 & C2 == 0.
+ if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ uint64_t NewVal = Mask & AndRHS->getValue();
+ return DemandedBitsAreZero(Op.getOperand(0), NewVal, Old, New, DAG);
+ }
+ break;
+ case ISD::SHL:
+ // (ushl X, C1) & C2 == 0 iff X & (C2 >> C1) == 0
+ if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ uint64_t NewVal = Mask >> ShAmt->getValue();
+ return DemandedBitsAreZero(Op.getOperand(0), NewVal, Old, New, DAG);
+ }
+ break;
+ case ISD::SIGN_EXTEND_INREG: {
+ MVT::ValueType EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
+ unsigned ExtendBits = MVT::getSizeInBits(EVT);
+ // If we're extending from something smaller than MVT::i64 and all of the
+ // sign extension bits are masked, return true and set New to be a zero
+ // extend inreg from the same type.
+ if (ExtendBits < 64 && ((Mask & (~0ULL << ExtendBits)) == 0)) {
+ Old = Op;
+ New = DAG.getZeroExtendInReg(Op.getOperand(0), EVT);
+ return true;
+ }
+ break;
+ }
+ case ISD::SRA:
+ if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ unsigned OpBits = MVT::getSizeInBits(Op.getValueType());
+ unsigned SH = ShAmt->getValue();
+ if (SH && ((Mask & (~0ULL << (OpBits-SH))) == 0)) {
+ Old = Op;
+ New = DAG.getNode(ISD::SRL, Op.getValueType(), Op.getOperand(0),
+ Op.getOperand(1));
+ return true;
+ }
+ }
+ break;
+ }
+ return false;
+}
/// MaskedValueIsZero - Return true if 'Op & Mask' is known to be zero. We use
/// this predicate to simplify operations downstream. Op and Mask are known to