[llvm] [LoopInterchange] Bail out when memory instruction ratio is high (PR #192954)

[Madhur Amilkanthwar via llvm-commits]

madhur13490 wrote:

This is a tricky territory of heuristics. I am not impressed by compile-time improvements but I am definitely worried about compromises it may end up making in doing loop interchange. 

So, with this patch, how many new loops are interchanged? How many loops those were interchanging will *not* be interchanged now? Do you have any data on these? 

I believe if this patch end up regressing number of loop interchanged by huge amount then it is not really worth the small compile-time improvement we are seeing.

https://github.com/llvm/llvm-project/pull/192954