[llvm] [AMDGPU] Fix canonicalization of truncated values. (PR #83054)

[Harald van Dijk via llvm-commits]

================
@@ -2944,6 +2944,34 @@ def : GCNPat<
    (V_BFREV_B32_e64 (i32 (EXTRACT_SUBREG VReg_64:$a, sub1))), sub0,
    (V_BFREV_B32_e64 (i32 (EXTRACT_SUBREG VReg_64:$a, sub0))), sub1)>;
 
+// If fcanonicalize's operand is implicitly canonicalized, we only need a copy.
+let AddedComplexity = 1000 in {
+def : GCNPat<
+  (is_canonicalized_1<fcanonicalize> f16:$src),
----------------
hvdijk wrote:

I had a look and couldn't find any custom predicate logic other than for instructions, which means we cannot do it on the operand level as the operand need not be an instruction. It's possible I missed something simple, but I think I would prefer to just leave it like this over making big changes to make TableGen support operand-level predicates.

https://github.com/llvm/llvm-project/pull/83054