[llvm] [AArch64] Improve scalar fixed-point int-to-fp codegen (PR #220299)

[Vimal Patel via llvm-commits]

================
@@ -9666,6 +9638,56 @@ def gi_fixedpoint_scalar_xform
     : GICustomOperandRenderer<"renderFixedPointScalarXForm">,
       GISDNodeXFormEquiv<fixedpoint_scalar_xform>;
 
+multiclass IntegerToFPScalarPats<SDPatternOperator OpNode, string INST> {
+  // TODO: Support these patterns for GlobalISel as well.
+  let GISelShouldIgnore = 1 in {
+  // Expect the integer operand to be in GPR register.
+  def : Pat<(f16 (fmul (f16 (OpNode (i32 GPR32:$Rn))), fixedpoint_recip_f16_i32:$scale)),
----------------
pvimal816a wrote:

Seems like I missed those. You're right that the `fmul` patterns are not needed.

https://github.com/llvm/llvm-project/pull/220299