aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorSimon Pilgrim <llvm-dev@redking.me.uk>2022-07-23 13:11:07 +0100
committerSimon Pilgrim <llvm-dev@redking.me.uk>2022-07-23 13:17:24 +0100
commit5f89d2bae964e21d8b7e9226c594d15d1c2e8325 (patch)
tree02c6c247383a97ad5a3f9f09b206cced46405cfd
parent559f07b872116fb85ea7d493768a6eb450329fa0 (diff)
downloadllvm-5f89d2bae964e21d8b7e9226c594d15d1c2e8325.zip
llvm-5f89d2bae964e21d8b7e9226c594d15d1c2e8325.tar.gz
llvm-5f89d2bae964e21d8b7e9226c594d15d1c2e8325.tar.bz2
[DAG] Move OR(AND(X,C1),AND(OR(X,Y),C2)) -> OR(AND(X,OR(C1,C2)),AND(Y,C2)) fold to SimplifyDemandedBits
This will fix the SystemZ v3i31 memcpy regression in D77804 (with the help of D129765 as well....). It should also allow us to /bend/ the oneuse limitation for cases where we can use demanded bits to safely peek though multiple uses of the AND ops.
-rw-r--r--llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp21
-rw-r--r--llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp27
2 files changed, 27 insertions, 21 deletions
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index ebc145b..edb0756 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -6888,27 +6888,6 @@ static SDValue visitORCommutative(SelectionDAG &DAG, SDValue N0, SDValue N1,
if (getBitwiseNotOperand(N00, N01,
/* AllowUndefs */ false) == N1)
return DAG.getNode(ISD::OR, SDLoc(N), VT, N01, N1);
-
- // (or (and X, C1), (and (or X, Y), C2)) -> (or (and X, C1|C2), (and Y, C2))
- if (N1.getOpcode() == ISD::AND) {
- SDValue N10 = N1.getOperand(0);
- if (N10.getOpcode() == ISD::OR) {
- SDValue N11 = N1.getOperand(1);
- SDValue N100 = N10.getOperand(0);
- SDValue N101 = N10.getOperand(1);
- if (((N00 == N100) || (N00 == N101)) && N0->hasOneUse() &&
- N1->hasOneUse()) {
- SDLoc DL(N);
- if (SDValue C12 =
- DAG.FoldConstantArithmetic(ISD::OR, DL, VT, {N01, N11})) {
- SDValue Y = (N00 == N100 ? N101 : N100);
- return DAG.getNode(ISD::OR, DL, VT,
- DAG.getNode(ISD::AND, DL, VT, N00, C12),
- DAG.getNode(ISD::AND, DL, VT, Y, N11));
- }
- }
- }
- }
}
if (SDValue R = foldLogicOfShifts(N, N0, N1, DAG))
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 31ea506..8f71b9f 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -1467,6 +1467,33 @@ bool TargetLowering::SimplifyDemandedBits(
}
}
+ // (or (and X, C1), (and (or X, Y), C2)) -> (or (and X, C1|C2), (and Y, C2))
+ // TODO: Use SimplifyMultipleUseDemandedBits to peek through masks.
+ if (Op0.getOpcode() == ISD::AND && Op1.getOpcode() == ISD::AND &&
+ Op0->hasOneUse() && Op1->hasOneUse()) {
+ // Attempt to match all commutations - m_c_Or would've been useful!
+ for (int I = 0; I != 2; ++I) {
+ SDValue X = Op.getOperand(I).getOperand(0);
+ SDValue C1 = Op.getOperand(I).getOperand(1);
+ SDValue Alt = Op.getOperand(1 - I).getOperand(0);
+ SDValue C2 = Op.getOperand(1 - I).getOperand(1);
+ if (Alt.getOpcode() == ISD::OR) {
+ for (int J = 0; J != 2; ++J) {
+ if (X == Alt.getOperand(J)) {
+ SDValue Y = Alt.getOperand(1 - J);
+ if (SDValue C12 = TLO.DAG.FoldConstantArithmetic(ISD::OR, dl, VT,
+ {C1, C2})) {
+ SDValue MaskX = TLO.DAG.getNode(ISD::AND, dl, VT, X, C12);
+ SDValue MaskY = TLO.DAG.getNode(ISD::AND, dl, VT, Y, C2);
+ return TLO.CombineTo(
+ Op, TLO.DAG.getNode(ISD::OR, dl, VT, MaskX, MaskY));
+ }
+ }
+ }
+ }
+ }
+ }
+
Known |= Known2;
break;
}