[DAGCombiner, PowerPC] allow X - (fpext(-Y) --> X + fpext(Y) with multiple uses

[spatel]

Summary

This is a transform that I limited in instcombine in rL329821 because it was creating more instructions in IR when the cast has multiple uses.

However, in the post-commit mailing list thread for that change, @escha wrote that an out-of-tree target had regressed from that change because fpext is free on that target. Since I don't have access to that target, I'm using PowerPC as a stand-in to show the benefit of the transform here in the backend. PowerPC is the only in-tree target that sets isFPExtFree() to 'true'.

If this looks right, then we may need to add another hook for isFPTruncFree() to fully recover from the IR change because it affected fptrunc casts too.

[nemanjai]

From the PPC perspective, this is fine since the sequences are equivalent. It also seems perfectly reasonable to look through a free operation here.
However, I imagine that @escha might want to look at this as well and give the final approval.

[spatel]

In D45598#1067909, @nemanjai wrote:

From the PPC perspective, this is fine since the sequences are equivalent. It also seems perfectly reasonable to look through a free operation here.

The sequence with fadd is better in theory because we've reduced the dependency chain (the fneg can execute in parallel now).

However, I imagine that @escha might want to look at this as well and give the final approval.

There are follow-up comments in the post-commit thread for rL329821. The source of the problem for the original motivating case is unknown. But I think this change can't hurt that target, so I'll proceed with this one.

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

 /// form more cheaply than the expression itself.
 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
                                const TargetLowering &TLI,
                                const TargetOptions *Options,
                                unsigned Depth = 0) {
   // fneg is removable even if it has multiple uses.
   if (Op.getOpcode() == ISD::FNEG) return 2;
 
-  // Don't allow anything with multiple uses.
-  if (!Op.hasOneUse()) return 0;
+  // Don't allow anything with multiple uses unless we know it is free.
+  EVT VT = Op.getValueType();
+  if (!Op.hasOneUse())
+    if (!(Op.getOpcode() == ISD::FP_EXTEND &&
+          TLI.isFPExtFree(VT, Op.getOperand(0).getValueType())))
+      return 0;
 
   // Don't recurse exponentially.
   if (Depth > 6) return 0;
 
-  EVT VT = Op.getValueType();
   switch (Op.getOpcode()) {
   default: return false;
   case ISD::ConstantFP: {
     if (!LegalOperations)
       return 1;
 
     // Don't invert constant FP values after legalization unless the target says
     // the negated constant is legal.

 
 /// If isNegatibleForFree returns true, return the newly negated expression.
 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
                                     bool LegalOperations, unsigned Depth = 0) {
   const TargetOptions &Options = DAG.getTarget().Options;
   // fneg is removable even if it has multiple uses.
   if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
 
-  // Don't allow anything with multiple uses.
-  assert(Op.hasOneUse() && "Unknown reuse!");
-
   assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
 
   const SDNodeFlags Flags = Op.getNode()->getFlags();
 
   switch (Op.getOpcode()) {
   default: llvm_unreachable("Unknown code");
   case ISD::ConstantFP: {
     APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();

llvm/trunk/test/CodeGen/PowerPC/fsub-fneg.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -verify-machineinstrs -mtriple=powerpc64le < %s | FileCheck %s
 
-; FIXME: When fpext is free, we should look through it for optimizations
+; When fpext is free, we should look through it for optimizations
 ; even if it has multiple uses and produce an 'fadd' here.
 ; Y - (fpext(-X)) --> Y + fpext(X)
 
 define double @neg_ext_op1_extra_use(float %x, double %y) nounwind {
 ; CHECK-LABEL: neg_ext_op1_extra_use:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    fneg 0, 1
-; CHECK-NEXT:    xssubdp 1, 2, 0
-; CHECK-NEXT:    xsdivdp 1, 0, 1
+; CHECK-NEXT:    xsadddp 0, 2, 1
+; CHECK-NEXT:    fneg 13, 1
+; CHECK-NEXT:    xsdivdp 1, 13, 0
 ; CHECK-NEXT:    blr
   %t1 = fsub float -0.0, %x
   %t2 = fpext float %t1 to double
   %t3 = fsub double %y, %t2
   %t4 = fdiv double %t2, %t3
   ret double %t4
 }