YJIT: incorrect code size accounting?

[maximecb]

I was looking at the stats for chunky-png yesterday and I noticed that the code_region_overhead metric that I recently added was negative on this benchmark:

From ruby/master, as of ff329ce

inline_code_size:            619,828
outlined_code_size:          447,776
code_region_size:            999,424
code_region_overhead:        -68,180 (-6.8%)

This failure mode doesn't happen on every run, but for some runs, we get a negative number. The overhead is calculated as follows in yjit.rb:

code_region_overhead = stats[:code_region_size] - (stats[:inline_code_size] + stats[:outlined_code_size])

It seems like we must be doing something wrong somewhere, somehow overcounting inline/outlined code size, or undercounting the code region size, because AFAIK this number should never be negative. It would be good to identify the cause of the error, because it means the metrics we've used so far may have been somewhat off, and the code region overhead value currently being reported is obviously wrong and useless in its current state.

@XrXr could you take a look at this and try to determine where the issue is? 🤔

[k0kubun]

It seems like we could run self.past_page_bytes += self.current_page_bytes(); multiple times for the same page when we're moving from an old page to the next page. I'll work on a patch.