[llvm] [SystemZ] SLP reductions: cost functions of reductions and scalarization (PR #112491)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> A slight regression (mcf) has gotten in the way of this patch being committed. It seems that the problem is that when SLP thinks element loads are free, it creates VLREP; VLEG; VST  (two element loads and a vector store), out of two separate loads and stores. However, the SystemZ backend prefers MVC (memcopy) even of 8 bytes, so the result is that this sequence replaces two MVCs, and not four instructions.
> 
> 
> 
> Not sure how to best avoid this:
> 
> - Somehow make SLP avoid these cases where memcpy would be preferred (would that be all load-> store sequences?). New hook like "preferMemCpy()"? There is the TLI->getMaxStoresPerMemcpy() which is kind of similar.
> 
> - SystemZ backend could optimize this VST pattern into MVC.
> 
> - Make SLP pass the actual Load Instruction pointers to the  getScalarizationOverhead() so that the SystemZ implementation can recognize loads used by stores.
> 
> 

Prefer option 3

https://github.com/llvm/llvm-project/pull/112491