[llvm] [AMDGPU][GIsel] Fix constrainRegAttrs to retain whichever of the two types is more specific instead of always adopting ConstrainingRegTy (PR #227226)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-globalisel

Author: adeshcom14

<details>
<summary>Changes</summary>

On extended-LLT targets, a scalar type can be either an untyped s32 or a specifically typed i32/f32. When a combine merges two equivalent registers, code that unifies their types treated these as interchangeable and overwrote the specifically typed registers type with the untyped one, even though its defining instruction never changed. Since GlobalISel's compiled matcher tables only match the specific type, which makes the register as unselectable and crashes.
This patch fixes it to keep the more specific type instead of always taking whichever side happens to be the replacement register.

Fixes: LCOMPILER-2715

Assisted by: Claude Sonnet 

---
Full diff: https://github.com/llvm/llvm-project/pull/227226.diff


2 Files Affected:

- (modified) llvm/lib/CodeGen/MachineRegisterInfo.cpp (+10-1) 
- (added) llvm/test/CodeGen/AMDGPU/GlobalISel/regbank-combiner-constrain-reg-attrs.ll (+174) 


``````````diff
The server is unavailable at this time. Please wait a few minutes before you try again.
``````````

</details>


https://github.com/llvm/llvm-project/pull/227226