TTFX regression on 1.11

[thofma]

I noticed a substantial TTFX regression on 1.11. I bisected one of those to #54428:

On 480dfab:

  | | |_| | | | (_| |  |  Version 1.11.0-beta1.60 (2024-05-23)
 _/ |\__'_|_|_|\__'_|  |  HEAD/480dfabb67 (fork: 176 commits, 287 days)
|__/                   |

julia> @time @eval using Hecke; @time @eval maximal_order(quadratic_field(-3)[1])
[...]
  2.524573 seconds (4.49 M allocations: 262.166 MiB, 10.80% gc time, 19.21% compilation time: 91% of which was recompilation)
 10.920348 seconds (42.05 M allocations: 2.076 GiB, 7.97% gc time, 99.99% compilation time: 40% of which was recompilation)

where as on the commit before (d186d8a) it is:

  2.126242 seconds (2.52 M allocations: 165.732 MiB, 11.79% gc time, 2.81% compilation time: 50% of which was recompilation)
  0.012258 seconds (22.56 k allocations: 1.142 MiB, 92.89% compilation time)

Another datapoint: To reproduce it julia -e '@time @eval ... does not work, but one has to do julia -e 'using REPL; ...'.

On 1.11 it is even worse due to some other change, but this here is the first one.

I looked at some other regression reports, but could not identify a duplicate.

[giordano]

Can you confirm this is only about v1.11 and the latency is vastly improved on master already? Then it should be a matter of identifying, and possibly backport, what reduced the latency.

[thofma]

Yes, on master it is back to

  2.930722 seconds (5.74 M allocations: 300.170 MiB, 12.11% gc time, 8.95% compilation time: 64% of which was recompilation)
  0.042835 seconds (65.44 k allocations: 3.345 MiB, 97.71% compilation time: 83% of which was recompilation)

which is still a regression, but acceptable. I'll try to do a bisect.

[ViralBShah]

Can we close this one?