[llvm] MachineBlockPlacement: Add tolerance to comparisons (PR #67197)

[via llvm-commits]

spupyrev wrote:

> Why not using the "emulated/estimated" counts? You get the same rounding effect, as suggested in the diff, but in addition the raw counts have an intuitive interpretation ("how many times a block/jump was executed?"). 

Any other thoughts on whether we should/could use profile counts in MachineBlockPlacement (and perhaps other optimizations), instead of frequencies? Intuitively it makes sense to me and it would resolve some of the stability issues discussed in the diff. It seems there is already such an option for tail duplication (https://reviews.llvm.org/D83265). I'm curious why it is not applied for other placement heuristics? cc @david-xl 

https://github.com/llvm/llvm-project/pull/67197