5% regression in some Inference benchmarks due to #51149 (Set temporary env whenever Julia is started with --project=@temp)

[Zentrik]

I'm confused as to how 583981f has caused this, but it caused a 6% regression in min wall time, allocations and memory for the [["inference", "allinference", "REPL.REPLCompletions.completions"]] benchmark and a 3% regression for the [["inference", "allinference", "Base.init_stdio(::Ptr{Cvoid})"]] benchmark.

I would presume this regression should be fixable given how little the pr changed but I haven't been able to figure out what caused the regression in the first place.

Nanosoldier report: https://github.com/JuliaCI/NanosoldierReports/blob/master/benchmark/by_date/2024-06/10/report.md

[oscardssmith]

Are you sure this isn't noise?

[Zentrik]

Yes, I can reproduce it reliably and it reproduced on the only other nanosoldier run, see https://tealquaternion.camdvr.org/compare.html?start=2024-06-04&end=2024-06-13&stat=min-wall-time&name=allinferen.

[IanButterworth]

Sorry if I'm being slow, but what from the reports indicates that commit? There's a huge number of commits between that report and the previous?

[Zentrik]

Sorry, I bisected the regression to that commit.

[KristofferC]

and a 3% regression for the [["inference", "allinference", "Base.init_stdio(::Ptr{Cvoid})"]] benchmark.

This seems like dup of the reason in #53695, #53459 (comment)

[Zentrik]

Yes, for both benchmarks it's because it now needs to infer mktempdir in load_path_expand.