[Costs] Increase costs for rem by const operations

[MaxGraey]

Reminder by constant should have slightly more cost due to x % C usually lower to
x - C * (x * M' >> C")
while div by const to:
x * M' >> C"
on VM engines.

[kripken]

I think we should have a methodology for deciding on these constants before we change any. Ah, I see you recently opened WebAssembly/design#1434 ... that would really be useful to get.

I don't have time to look into this myself. But I'd try to find existing documentation, either in compilers or for CPUs. Failing that, actually running microbenchmarks would be useful.

[MaxGraey]

I think we should have a methodology for deciding on these constants before we change any

Yes, I agree. That's why I opened design issue. But it seems that not everyone is interested in taking the time to do it. In the meantime, I just look at what the godbolt tools for specific operation codegen and specify the costs. This is the best we can do right now, because I'm afraid a standardized costs table won't appear anytime soon, if ever.

[kripken]

Another option aside from godbolt is to find where LLVM has such a table - I assume it must have one somewhere?

[MaxGraey]

Another option aside from godbolt is to find where LLVM has such a table - I assume it must have one somewhere?

I was checked the cost table from LLVM. Unfortunately it is not very accurate and less oriented to modern platforms. In addition, it's formed empirically as well

[MaxGraey]

For example:

div / rem costs:
https://github.com/llvm/llvm-project/blob/e356027016c6365b3d8924f54c33e2c63d931492/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h#L453

and range of costs quite limited:
https://github.com/llvm/llvm-project/blob/main/llvm/include/llvm/Analysis/TargetTransformInfo.h#L261