[llvm] [GIsel] Add combiner for constant-folding G_BITCAST (PR #193060)

[Alex MacLean via llvm-commits]

================
@@ -1390,6 +1394,23 @@ def constant_fold_cast_op : GICombineRule<
    [{ return Helper.matchConstantFoldCastOp(*${d}, ${matchinfo}); }]),
   (apply [{ Helper.replaceInstWithConstant(*${d}, ${matchinfo}); }])>;
 
+
+def constant_fold_bitcast_of_const : BuildFnCombineRule<
+  (match (G_CONSTANT $src, $imm),
+         (G_BITCAST $d, $src):$root,
+   [{ return Helper.matchConstantFoldBitcast(
+        ${d}, ${imm}.getCImm()->getValue(), ${matchinfo}); }])>;
----------------
AlexMaclean wrote:

For a case where a constant integer is cast to a vector, It seems like folding that to a BUILD_VECTOR of constants adds more instructions and has the potential to introduce regressions if all-constant build vectors are not handled well elsewhere. Do you think it is a good idea to always canonicalize to this? 

https://github.com/llvm/llvm-project/pull/193060