[llvm] [LegalizeVectorTypes] Preserve original MemoryOperand and MemVT when widening fixed vector load to vp_load. (PR #109473)

Craig Topper via llvm-commits llvm-commits at lists.llvm.org
Sun Sep 22 23:54:13 PDT 2024


topperc wrote:

> LGTM, I just quickly checked to see what we do currently and we're increasing the MMO size at least for fixed vectors
> 
> ```
> t5: v3i64,ch = load<(load (s192) from %ir.p, align 32)> t0, t2, undef:i64
> ```
> 
> ```
> t19: v4i64,ch = vp_load<(load (s256) from %ir.p)> t0, t2, undef:i64, t17, Constant:i64<3>
> ```

I guess should have mentioned we already do the right thing when widening fixed vector stores.

https://github.com/llvm/llvm-project/pull/109473


More information about the llvm-commits mailing list