[llvm] [CodeGen] [AMDGPU] Attempt DAGCombine for fmul with select to ldexp (PR #111109)

[Matt Arsenault via llvm-commits]

================
@@ -14476,6 +14477,60 @@ SDValue SITargetLowering::performFDivCombine(SDNode *N,
   return SDValue();
 }
 
+SDValue SITargetLowering::performFMulCombine(SDNode *N,
+                                             DAGCombinerInfo &DCI) const {
+  SelectionDAG &DAG = DCI.DAG;
+  EVT VT = N->getValueType(0);
+  EVT i32VT = VT.changeElementType(MVT::i32);
+
+  SDLoc SL(N);
+  SDValue LHS = N->getOperand(0);
+  SDValue RHS = N->getOperand(1);
+
+  // ldexp(x, zext(i1 y)) -> fmul x, (select y, 2.0, 1.0)
+  // ldexp(x, sext(i1 y)) -> fmul x, (select y, 0.5, 1.0)
+  //
+  // The above mentioned ldexp folding works fine for
+  // bf16/f32, but as for f64 it creates f64 select which
+  // is costly to materialize as compared to f64 ldexp
+  // so here we undo the transform for f64 datatype.
+  // Also in case of f16, its cheaper to materialize inline
+  // 32 bit-constant (via ldexp use) rather than using fmul.
----------------
arsenm wrote:

Comment is verbose. Just state that it's cheaper to use i32 inline constants than to materialize f16 or f64 values 

https://github.com/llvm/llvm-project/pull/111109