You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
perf regression for doc generation from rollup 90416 #90512
I looked through the rollup diff. None of the compiler changes (and certainly not the rustdoc) changes jump out at me as potentially causing this regression. #90377 doesn't seem likely either, since I don't think it'd affect rustdoc differently from the compiler.
However, I wonder if the regression could be due to #89876. It makes some core internal macros more complex and adds attributes to a bunch of core traits, which could be slowing rustdoc down (especially since impls and traits are one of the slowest parts of rustdoc). It still seems like a bit of a stretch though to cause such huge regressions.
Also, the most significant regressions are for doc builds, but there are non-trivial regressions on check (and maybe other) builds too (EDIT: the check regressions show up even with "Filter out very small changes" checked):
I wonder if it makes sense not to investigate too heavily for helloworld regressions. Rustdoc has lots of global analysis which I've been meaning to look at to see if it's truly necessary, but I don't think there's much we can do to avoid regressions like this in the short-term.
Activity
pnkfelix commentedon Nov 2, 2021
(spoke to @GuillaumeGomez , who hypothesizes based on the contents of Pull Request #90416 that this is a spurious regression.)
camelid commentedon Nov 2, 2021
I wonder if it's possible that something in rustc changed that is only used by rustdoc, so the regression only showed up in doc builds?
camelid commentedon Nov 2, 2021
I looked through the rollup diff. None of the compiler changes (and certainly not the rustdoc) changes jump out at me as potentially causing this regression. #90377 doesn't seem likely either, since I don't think it'd affect rustdoc differently from the compiler.
However, I wonder if the regression could be due to #89876. It makes some
coreinternal macros more complex and adds attributes to a bunch ofcoretraits, which could be slowing rustdoc down (especially since impls and traits are one of the slowest parts of rustdoc). It still seems like a bit of a stretch though to cause such huge regressions.Also, the most significant regressions are for doc builds, but there are non-trivial regressions on check (and maybe other) builds too (EDIT: the check regressions show up even with "Filter out very small changes" checked):
jyn514 commentedon Nov 2, 2021
I wonder if it makes sense not to investigate too heavily for helloworld regressions. Rustdoc has lots of global analysis which I've been meaning to look at to see if it's truly necessary, but I don't think there's much we can do to avoid regressions like this in the short-term.
camelid commentedon Nov 2, 2021
I agree about helloworld, but real-world crates regressed a lot too. E.g., ripgrep and futures regressed by about 1.75%.
camelid commentedon Nov 9, 2021
Looking at the query diff view,
build_extern_trait_implandmetadata_decode_entry_item_attrsboth seem to have regressed by several percentage points.