Ignore inlining costs for precompilation with --compile=all
#53798
Conversation
This helps to minimize the amount of missed precompiles that occur by eliding the inlining cost check when julia is run under `--compile=all`. This will lead to slightly larger sysimage sizes. However, in practice it leads to more extensive and successful precompilation, which can minimize the number of JIT events in call sites with dynamic dispatch. This is an alternative to JuliaLang#53547, where removing the inlining cost check was made universal. However, that lead to ~15% larger sysimage sizes by default. This implements Jeff's suggestion that this mode be enabled under `--compile=all`.
|
It does seem like there's something going awry with the logic in the default case, but this looks to be a very reasonable solution. This is impacting a real-world use-case where we need to ensure 100% precompilation. |
|
I think this is a clear W, since it is compile=all mode and already making a massive sysimage, so we shouldn't then exclude small things |
| if (inferred && | ||
| inferred != jl_nothing && | ||
| (jl_ir_inlining_cost(inferred) == UINT16_MAX)) { | ||
| (jl_options.compile_enabled != JL_OPTIONS_COMPILE_ALL && jl_ir_inlining_cost(inferred) == UINT16_MAX)) { |
There was a problem hiding this comment.
did you mean to write jl_options.compile_enabled == JL_OPTIONS_COMPILE_ALL || here instead?
There was a problem hiding this comment.
Oooooh, good catch. Yes, I do believe that's what we intended. This was meant to be the "softer" #53547, where we only considered the inlineability heuristics when we're not compile_all'ing... intending to otherwise compile even small functions.
There was a problem hiding this comment.
(For posterity: Fixed in 59d74d4.)
…ang#53798) This helps to minimize the amount of missed precompiles that occur by eliding the inlining cost check when julia is run under `--compile=all`. This will lead to slightly larger sysimage sizes. However, in practice it leads to more extensive and successful precompilation, which can minimize the number of JIT events in call sites with dynamic dispatch. This is an alternative to JuliaLang#53547, where removing the inlining cost check was made universal. However, that lead to ~15% larger sysimage sizes by default. This implements Jeff's suggestion that this mode be enabled under `--compile=all`.
This helps to minimize the amount of missed precompiles
that occur by eliding the inlining cost check when julia is
run under
--compile=all. This will lead toslightly larger sysimage sizes. However, in practice it leads
to more extensive and successful precompilation, which can
minimize the number of JIT events in
call sites with dynamic dispatch.
This is an alternative to #53547, where removing the inlining
cost check was made universal. However, that lead to ~15% larger
sysimage sizes by default. This implements Jeff's suggestion that
this mode be enabled under
--compile=all.