[llvm] [AMDGPU] Improve combining on v_ldexp (PR #190236)

[Petar Avramovic via llvm-commits]

================
@@ -232,9 +232,18 @@ def AMDGPUPostLegalizerCombiner: GICombiner<
   let CombineAllMethodName = "tryCombineAllImpl";
 }
 
+// Like combine_fmul_with_select_to_fldexp but with VGPR detection for RegBankCombiner.
+def combine_fmul_select_to_fldexp_regbank : GICombineRule<
+  (defs root:$root, build_fn_matchinfo:$matchinfo),
+  (match  (G_FMUL $dst, $x, $select):$root,
+          (G_SELECT $select, $y, $A, $B):$sel,
+          [{ return matchFmulSelectToFldexpVgpr(*${root}, *${sel}, ${matchinfo}); }]),
+  (apply  [{ Helper.applyBuildFn(*${root}, ${matchinfo}); }])>;
+
----------------
petar-avramovic wrote:

Move this up where other combines are described

https://github.com/llvm/llvm-project/pull/190236