Magic division by constant for OptimizeForJS. Does it make sense?

[MaxGraey]

I'm working on a new rule for OptimizeForJS that optimizes the division / modulo by constant for 64-bit types. Since for this particular case it is too expensive to call and execute a helper (polyfill) function for full division. According to my simple tests this gives up to 7x performance gain if the high part of dividend is not zero. The PR will come out pretty huge, so I'm wondering if it's worth it.

Rule will rewrite this i64(x) / C to something like this in simplified form mulh64(x, mag.M) >> mag.S

[tlively]

This sounds fairly straightforward, but I don't understand everything involved. What will make the PR large?

[MaxGraey]

See #4055. It's just half of PR

[MaxGraey]

So I expect it will be 1000-1200 LOCs including lit tests.