[llvm] [LoopIdiom] Add a range attribute to formed memset/memcpy/memmove (PR #226801)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Kuba (Brecka) Mracek (kubamracek)

<details>
<summary>Changes</summary>

Part 1 of optimizing a recognized memset (see example below): Adding a range attribute to the memset call so that the upper bound is carried in the IR. Part 2 will prevent lowering a (non-constant) llvm.memset with a small upper bound into a memset libcall (expensive) and instead emit a branch or per-size branches instead (cheap).

This PR:

> Attach a range attribute to the length (or, for memset.pattern, the count) operand. The upper bound comes from the loop's constant maximum backedge-taken count and conditions guarding the loop, which also hold at the new call in the preheader. For a loop like this...
> 
>   for (unsigned i = n; i < 4; ++i)
>     dst[i] = -1;
> 
> ...the memset gets `i64 range(i64 0, 17)`.
> 
> LoopIdiomRecognize applies this to memset, memset.pattern, memcpy and memmove calls.

Assisted-by: Claude


---
Full diff: https://github.com/llvm/llvm-project/pull/226801.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp (+50-1) 
- (added) llvm/test/Transforms/LoopIdiom/mem-intrinsic-length-range.ll (+74) 


``````````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/226801