This commit contains a few changes that had to go in together.
1. Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
(and also scalar_to_vector).
2. Xor/and/or are indifferent to the swizzle operation (shuffle of one src).
Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A, B))
3. Optimize swizzles of shuffles: shuff(shuff(x, y), undef) -> shuff(x, y).
4. Fix an X86ISelLowering optimization which was very bitcast-sensitive.
Code which was previously compiled to this:
movd (%rsi), %xmm0
movdqa .LCPI0_0(%rip), %xmm2
pshufb %xmm2, %xmm0
movd (%rdi), %xmm1
pshufb %xmm2, %xmm1
pxor %xmm0, %xmm1
pshufb .LCPI0_1(%rip), %xmm1
movd %xmm1, (%rdi)
ret
Now compiles to this:
movl (%rsi), %eax
xorl %eax, (%rdi)
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153848 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index bac644a..5e88fcb 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -2336,6 +2336,68 @@
ORNode, N0.getOperand(1));
}
+ // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
+ // Only perform this optimization after type legalization and before
+ // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
+ // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
+ // we don't want to undo this promotion.
+ // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
+ // on scalars.
+ if ((N0.getOpcode() == ISD::BITCAST || N0.getOpcode() == ISD::SCALAR_TO_VECTOR)
+ && Level == AfterLegalizeVectorOps) {
+ SDValue In0 = N0.getOperand(0);
+ SDValue In1 = N1.getOperand(0);
+ EVT In0Ty = In0.getValueType();
+ EVT In1Ty = In1.getValueType();
+ // If both incoming values are integers, and the original types are the same.
+ if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
+ SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), In0Ty, In0, In1);
+ SDValue BC = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, Op);
+ AddToWorkList(Op.getNode());
+ return BC;
+ }
+ }
+
+ // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
+ // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
+ // If both shuffles use the same mask, and both shuffle within a single
+ // vector, then it is worthwhile to move the swizzle after the operation.
+ // The type-legalizer generates this pattern when loading illegal
+ // vector types from memory. In many cases this allows additional shuffle
+ // optimizations.
+ if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
+ ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
+ ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
+ SDValue In0 = SVN0->getOperand(0);
+ SDValue In1 = SVN1->getOperand(0);
+ EVT In0Ty = In0.getValueType();
+ EVT In1Ty = In1.getValueType();
+
+ unsigned NumElts = VT.getVectorNumElements();
+ // Check that both shuffles are swizzles.
+ bool SingleVecShuff = (N0.getOperand(1).getOpcode() == ISD::UNDEF &&
+ N1.getOperand(1).getOpcode() == ISD::UNDEF);
+
+ // Check that both shuffles use the same mask. The masks are known to be of
+ // the same length because the result vector type is the same.
+ bool SameMask = true;
+ for (unsigned i = 0; i != NumElts; ++i) {
+ int Idx0 = SVN0->getMaskElt(i);
+ int Idx1 = SVN1->getMaskElt(i);
+ if (Idx0 != Idx1) {
+ SameMask = false;
+ break;
+ }
+ }
+
+ if (SameMask && SingleVecShuff && In0Ty == In1Ty) {
+ SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT, In0, In1);
+ SDValue Shuff = DAG.getVectorShuffle(VT, N->getDebugLoc(), Op,
+ DAG.getUNDEF(VT), &SVN0->getMask()[0]);
+ AddToWorkList(Op.getNode());
+ return Shuff;
+ }
+ }
return SDValue();
}
@@ -7721,6 +7783,36 @@
return N0;
}
}
+
+ // If this shuffle node is simply a swizzle of another shuffle node,
+ // optimize shuffle(shuffle(x, y), undef) -> shuffle(x, y).
+ if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
+ N1.getOpcode() == ISD::UNDEF) {
+
+ SmallVector<int, 8> NewMask;
+ ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
+
+ EVT InVT = N0.getValueType();
+ int InNumElts = InVT.getVectorNumElements();
+
+ for (unsigned i = 0; i != NumElts; ++i) {
+ int Idx = SVN->getMaskElt(i);
+ // If we access the second (undef) operand then this index can be
+ // canonicalized to undef as well.
+ if (Idx >= InNumElts)
+ Idx = -1;
+ // Next, this index comes from the first value, which is the incoming
+ // shuffle. Adopt the incoming index.
+ if (Idx >= 0)
+ Idx = OtherSV->getMaskElt(Idx);
+
+ NewMask.push_back(Idx);
+ }
+
+ return DAG.getVectorShuffle(VT, N->getDebugLoc(), OtherSV->getOperand(0),
+ OtherSV->getOperand(1), &NewMask[0]);
+ }
+
return SDValue();
}