[llvm] [AMDGPU][True16][MC] true16 for v_alignbit_b32 (PR #119409)

[Ivan Kosarev via llvm-commits]

================
@@ -3003,6 +3060,40 @@ def : GCNPat <
   sub1)
 >;
 
+let True16Predicate = UseFakeTrue16Insts in
+def : GCNPat <
+  (i64 (bswap i64:$a)),
+  (REG_SEQUENCE VReg_64,
+  (V_BFI_B32_e64 (S_MOV_B32 (i32 0x00ff00ff)),
----------------
kosarev wrote:

It seems it's `(inst)` DAGs here, not `(ops)`, so I guess the idea is to have something like:
```
  defvar NoMods = !if(hasTrue16, (inst 0), (inst));
  dag ALIGNBIT32_INST1 = !con(NoMods, (inst operand1), NoMods, (inst operand1),
                              NoMods, (inst (i32 24)), NoMods, NoMods);

```

And then there's still only one operand varies, so probably no  need to repeat all the operands every time.

Also, can you turn the temporary values into `defvar`s?

https://github.com/llvm/llvm-project/pull/119409