================
@@ -6995,6 +7002,122 @@ static SDValue performSINT_TO_FPCombine(SDNode *N,
SelectionDAG &DAG,
return SDValue();
}
+// Try to widen AND, OR and XOR nodes to VT in order to remove casts around
+// logical operations, like in the example below.
+// or (and (truncate x, truncate y)),
+// (xor (truncate z, build_vector (constants)))
+// Given a target type \p VT, we generate
+// or (and x, y), (xor z, zext(build_vector (constants)))
+// given x, y and z are of type \p VT. We can do so, if operands are either
+// truncates from VT types, the second operand is a vector of constants, can
+// be recursively promoted or is an existing extension we can extend further.
+static SDValue PromoteMaskArithmetic(SDValue N, const SDLoc &DL, EVT VT,
+ SelectionDAG &DAG,
+ const LoongArchSubtarget &Subtarget,
+ unsigned Depth) {
+ // Limit recursion to avoid excessive compile times.
+ if (Depth >= SelectionDAG::MaxRecursionDepth)
+ return SDValue();
+
+ if (!ISD::isBitwiseLogicOp(N.getOpcode()))
+ return SDValue();
+
+ SDValue N0 = N.getOperand(0);
+ SDValue N1 = N.getOperand(1);
+
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (!TLI.isOperationLegalOrPromote(N.getOpcode(), VT))
+ return SDValue();
+
+ if (SDValue NN0 =
+ PromoteMaskArithmetic(N0, DL, VT, DAG, Subtarget, Depth + 1))
+ N0 = NN0;
+ else {
+ // The left side has to be a 'trunc'.
+ bool LHSTrunc = N0.getOpcode() == ISD::TRUNCATE &&
+ N0.getOperand(0).getValueType() == VT;
+ if (LHSTrunc)
+ N0 = N0.getOperand(0);
+ else
+ return SDValue();
+ }
+
+ if (SDValue NN1 =
+ PromoteMaskArithmetic(N1, DL, VT, DAG, Subtarget, Depth + 1))
+ N1 = NN1;
+ else {
+ // The right side has to be a 'trunc', a (foldable) constant or an
+ // existing extension we can extend further.
+ bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE &&
+ N1.getOperand(0).getValueType() == VT;
+ if (RHSTrunc)
+ N1 = N1.getOperand(0);
+ else if (ISD::isExtVecInRegOpcode(N1.getOpcode()) && VT.is256BitVector() &&
+ Subtarget.hasExtLASX() && N1.hasOneUse())
+ N1 = DAG.getNode(N1.getOpcode(), DL, VT, N1.getOperand(0));
+ // On 32-bit platform, i64 is an illegal integer scalar type, and
+ // FoldConstantArithmetic will fail for v4i64. This may be optimized in the
+ // future.
+ else if (SDValue Cst =
+ DAG.FoldConstantArithmetic(ISD::ZERO_EXTEND, DL, VT, {N1}))
+ N1 = Cst;
+ else
+ return SDValue();
+ }
+
+ return DAG.getNode(N.getOpcode(), DL, VT, N0, N1);
+}
+
+// On LASX the type v4i1/v8i1/v16i1 may be legalized to v4i32/v8i16/v16i8,
which
+// is LSX-sized register. In most cases we actually compare or select
LASX-sized
+// registers and mixing the two types creates horrible code. This method
+// optimizes some of the transition sequences.
+static SDValue PromoteMaskArithmetic(SDValue N, const SDLoc &DL,
+ SelectionDAG &DAG,
+ const LoongArchSubtarget &Subtarget) {
+ EVT VT = N.getValueType();
+ assert(VT.isVector() && "Expected vector type");
+ assert((N.getOpcode() == ISD::ANY_EXTEND ||
+ N.getOpcode() == ISD::ZERO_EXTEND ||
+ N.getOpcode() == ISD::SIGN_EXTEND) &&
+ "Invalid Node");
+
+ if (!Subtarget.hasExtLASX() || !VT.is256BitVector())
+ return SDValue();
+
+ SDValue Narrow = N.getOperand(0);
+ EVT NarrowVT = Narrow.getValueType();
+
+ // Generate the wide operation.
+ SDValue Op = PromoteMaskArithmetic(Narrow, DL, VT, DAG, Subtarget, 0);
+ if (!Op)
+ return SDValue();
+ switch (N.getOpcode()) {
+ default:
+ llvm_unreachable("Unexpected opcode");
+ case ISD::ANY_EXTEND:
+ return Op;
+ case ISD::ZERO_EXTEND:
+ return DAG.getZeroExtendInReg(Op, DL, NarrowVT);
+ case ISD::SIGN_EXTEND:
+ return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Op,
+ DAG.getValueType(NarrowVT));
+ }
+}
+
+static SDValue performANY_EXTENDCombine(SDNode *N, SelectionDAG &DAG,
----------------
wangleiat wrote:
Small naming suggestion: `performANY_EXTENDCombine` handles `ANY`, `ZERO`, and
`SIGN` extends, not just `ANY`. Renaming it to `performExtendCombine` would
better reflect what it does.
https://github.com/llvm/llvm-project/pull/163523
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
