[compiler-rt] [llvm] [MC/DC][Coverage] Introduce "Bitmap Bias" for continuous mode (PR #96126)

[NAKAMURA Takumi via llvm-commits]

chapuni wrote:

@petrhosek Thanks for the comment.

I don't think this would introduce serious performance regression, and I don't think coverage-instrumented codes would be sensitive of performance so much.

I know it'd be better to use the single bias offset between counters and bitmaps. We might do in the future. This is suboptimal but practical, I suppose.

It'd be the best if we could map and override the profraw into the existing profc and profbm, as Darwin is doing.
For it, I guess we would modify the layout of profc and profbm as aligned to phys page.
(Darwin's impl doesn't use contiguous mapping but individual mappings to profc and profbm)

I'll propose the bitmap for coverage in the near future. It will not require counters.
At the moment, I think 2nd bias offset would be practical for supporting mcdc with continuous mode.

https://github.com/llvm/llvm-project/pull/96126