Teach codegen to lower llvm.powi to an efficient (but not optimal)
multiply sequence when the power is a constant integer. Before, our
codegen for std::pow(.., int) always turned into a libcall, which was
really inefficient.
This should also make many gfortran programs happier I'd imagine.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92388 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 7747825..db127f2 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -462,7 +462,8 @@
// The number of parts is a power of 2. Repeatedly bisect the value using
// EXTRACT_ELEMENT.
Parts[0] = DAG.getNode(ISD::BIT_CONVERT, dl,
- EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()),
+ EVT::getIntegerVT(*DAG.getContext(),
+ ValueVT.getSizeInBits()),
Val);
if (DisableScheduling)
@@ -4261,6 +4262,59 @@
setValue(&I, result);
}
+
+/// ExpandPowI - Expand a llvm.powi intrinsic.
+static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
+ SelectionDAG &DAG) {
+ // If RHS is a constant, we can expand this out to a multiplication tree,
+ // otherwise we end up lowering to a call to __powidf2 (for example). When
+ // optimizing for size, we only want to do this if the expansion would produce
+ // a small number of multiplies, otherwise we do the full expansion.
+ if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
+ // Get the exponent as a positive value.
+ unsigned Val = RHSC->getSExtValue();
+ if ((int)Val < 0) Val = -Val;
+
+ // powi(x, 0) -> 1.0
+ if (Val == 0)
+ return DAG.getConstantFP(1.0, LHS.getValueType());
+
+ Function *F = DAG.getMachineFunction().getFunction();
+ if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
+ // If optimizing for size, don't insert too many multiplies. This
+ // inserts up to 5 multiplies.
+ CountPopulation_32(Val)+Log2_32(Val) < 7) {
+ // We use the simple binary decomposition method to generate the multiply
+ // sequence. There are more optimal ways to do this (for example,
+ // powi(x,15) generates one more multiply than it should), but this has
+ // the benefit of being both really simple and much better than a libcall.
+ SDValue Res; // Logically starts equal to 1.0
+ SDValue CurSquare = LHS;
+ while (Val) {
+ if (Val & 1)
+ if (Res.getNode())
+ Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
+ else
+ Res = CurSquare; // 1.0*CurSquare.
+
+ CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
+ CurSquare, CurSquare);
+ Val >>= 1;
+ }
+
+ // If the original was negative, invert the result, producing 1/(x*x*x).
+ if (RHSC->getSExtValue() < 0)
+ Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
+ DAG.getConstantFP(1.0, LHS.getValueType()), Res);
+ return Res;
+ }
+ }
+
+ // Otherwise, expand to a libcall.
+ return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
+}
+
+
/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
/// we want to emit this as a call to a named external function, return the name
/// otherwise lower it and return null.
@@ -4536,10 +4590,8 @@
DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
return 0;
case Intrinsic::powi:
- Res = DAG.getNode(ISD::FPOWI, dl,
- getValue(I.getOperand(1)).getValueType(),
- getValue(I.getOperand(1)),
- getValue(I.getOperand(2)));
+ Res = ExpandPowI(dl, getValue(I.getOperand(1)), getValue(I.getOperand(2)),
+ DAG);
setValue(&I, Res);
if (DisableScheduling)
DAG.AssignOrdering(Res.getNode(), SDNodeOrder);