[DAGCombiner] try to convert pow(x, 0.25) to sqrt(sqrt(x))

This was proposed as an IR transform in D49306, but it was not clearly justifiable as a canonicalization. 
Here, we only do the transform when the target tells us that sqrt can be lowered with inline code.

This is the basic case. Some potential enhancements are in the TODO comments:

1. Generalize the transform for other exponents (allow more than 2 sqrt calcs if that's really cheaper).
2. If we have less fast-math-flags, generate code to avoid -0.0 and/or INF.
3. Allow the transform when optimizing/minimizing size (might require a target hook to get that right).

Note that by default, x86 converts single-precision sqrt calcs into sqrt reciprocal estimate with 
refinement. That codegen is controlled by CPU attributes and can be manually overridden. We have plenty 
of test coverage for that already, so I didn't bother to include extra testing for that here. AArch uses 
its full-precision ops in all cases (not sure if that's the intended behavior or not, but that should 
also be covered by existing tests).

Differential Revision: https://reviews.llvm.org/D51630 

llvm-svn: 341481

1 files changed, 41 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index ce4297def73..26b9bc2b025 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -350,6 +350,7 @@ namespace {
     SDValue visitFREM(SDNode *N);
     SDValue visitFSQRT(SDNode *N);
     SDValue visitFCOPYSIGN(SDNode *N);
+    SDValue visitFPOW(SDNode *N);
     SDValue visitSINT_TO_FP(SDNode *N);
     SDValue visitUINT_TO_FP(SDNode *N);
     SDValue visitFP_TO_SINT(SDNode *N);
@@ -1568,6 +1569,7 @@ SDValue DAGCombiner::visit(SDNode *N) {
   case ISD::FREM:               return visitFREM(N);
   case ISD::FSQRT:              return visitFSQRT(N);
   case ISD::FCOPYSIGN:          return visitFCOPYSIGN(N);
+  case ISD::FPOW:               return visitFPOW(N);
   case ISD::SINT_TO_FP:         return visitSINT_TO_FP(N);
   case ISD::UINT_TO_FP:         return visitUINT_TO_FP(N);
   case ISD::FP_TO_SINT:         return visitFP_TO_SINT(N);
@@ -11566,6 +11568,45 @@ SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
   return SDValue();
 }
 
+SDValue DAGCombiner::visitFPOW(SDNode *N) {
+  ConstantFPSDNode *ExponentC = isConstOrConstSplatFP(N->getOperand(1));
+  if (!ExponentC)
+    return SDValue();
+
+  // Try to convert x ** (1/4) into square roots.
+  // x ** (1/2) is canonicalized to sqrt, so we do not bother with that case.
+  // TODO: This could be extended (using a target hook) to handle smaller
+  // power-of-2 fractional exponents.
+  if (ExponentC->getValueAPF().isExactlyValue(0.25)) {
+    // pow(-0.0, 0.25) = +0.0; sqrt(sqrt(-0.0)) = -0.0.
+    // pow(-inf, 0.25) = +inf; sqrt(sqrt(-inf)) =  NaN.
+    // For regular numbers, rounding may cause the results to differ.
+    // Therefore, we require { nsz ninf afn } for this transform.
+    // TODO: We could select out the special cases if we don't have nsz/ninf.
+    SDNodeFlags Flags = N->getFlags();
+    if (!Flags.hasNoSignedZeros() || !Flags.hasNoInfs() ||
+        !Flags.hasApproximateFuncs())
+      return SDValue();
+
+    // Don't double the number of libcalls. We are trying to inline fast code.
+    EVT VT = N->getValueType(0);
+    if (!DAG.getTargetLoweringInfo().isOperationLegalOrCustom(ISD::FSQRT, VT))
+      return SDValue();
+
+    // Assume that libcalls are the smallest code.
+    // TODO: This restriction should probably be lifted for vectors.
+    if (DAG.getMachineFunction().getFunction().optForSize())
+      return SDValue();
+
+    // pow(X, 0.25) --> sqrt(sqrt(X))
+    SDLoc DL(N);
+    SDValue Sqrt = DAG.getNode(ISD::FSQRT, DL, VT, N->getOperand(0), Flags);
+    return DAG.getNode(ISD::FSQRT, DL, VT, Sqrt, Flags);
+  }
+
+  return SDValue();
+}
+
 static SDValue foldFPToIntToFP(SDNode *N, SelectionDAG &DAG,
                                const TargetLowering &TLI) {
   // This optimization is guarded by a function attribute because it may produce