obviously, allowing -opt:inline:** breaks -Wconf:cat=foo:s

I thought I wasn't supposed to update the build, but I must have tossed a bad coin.

[error] -opt-inline-from is deprecated: use -opt:inline:**
[error] one error found

Edit: setupPublishCore needs old options

++ sbt -sbt-version 1.5.6 -Dsbt.override.build.repos=true -Dsbt.repository.config=/home/jenkins/workspace/scala-2.13.x-validate-main/scripts/sbt-repositories-config -warn 'setupPublishCore https://scala-ci.typesafe.com/artifactory/scala-pr-validation-snapshots/' publish
[error] 'all' is not a valid choice for '-opt'
[error] 'inline:scala/**' is not a valid choice for '-opt'
[error] 'all' is not a valid choice for '-opt'
[error] 'inline:scala/**' is not a valid choice for '-opt'
[error] 'all' is not a valid choice for '-opt'
[error] 'inline:scala/**' is not a valid choice for '-opt'

Edit: undeprecate the old options because no point in 2.13.

Also, some usages of -opt:l:inline in tests require only -opt:inline, others really do need -opt:all or a subset.

Thank you for taking this on!

I see two issues with the current state:

• -opt:inline:pattern only enables the inliner, no local optimizations. There is no situtaion I can think of for anyone to ever want this, local optimizations should always be on (except maybe us for testing the inliner). It's too easy to make that mistake - even this PR changes many -opt:l:inline -opt-inline-from:** to -opt:inline:**.
• -opt:inline and -opt:all default to -opt:inline:**, which is unsafe. Since we're changing the flags, it's a chance to issue an error when specifying plain -opt:inline without a pattern.

Could we make opt:inline:pattern enalbe all optimizations, remove the all option, and require a pattern?

I will defer to your expert opinion on those two gotchas.

The only use case for fine-grained options is, "The optimizer broke my code, so I have to turn off a feature for this file."

I agree, it was a nuisance to fiddle with -opt:all in the tests. (I intended to enable only minimal options for testing, but didn't want to comprehend every test.)

On not defaulting to double star, it would be nice if there were a convenient default, such as, inline from classes with sources in this build. Then an option to inline from dependencies; from std lib; from platform lib. But maybe real programmers don't mind saying com.example.**. Oh wait, there is <sources>, why isn't that the default? Probably in sbt that only works for a clean build? and not for subprojects.

I will follow up with the tighter screws before I forget how it works.

About defaulting to <sources>: it's already not a good idea today to enable the optimizer during development, but I wouldn't be surprised if it still happens. Defaulting to <sources> makes would make things even more random.

I think requiring people to think about inlining when enabling it is a good thing :)

Thanks lrytz, you have me singing the classic tune, "Puttin' on the Rytz". Probably I've made that pun before.

The new behavior is much better. One difference is that before, you needed -opt:inline or probably -opt:l:inline and -opt-inline-from:** to enable inlining, the way you need both keys turned simultaneously to launch the missiles.

Specifying only one option would silently not inline anything.

Now, -opt:inline requires the pattern, and -opt-inline-from:** is just an alias that works like -opt:inline:** so that by itself it enables.

I needed this the other day when I couldn't remember how to use -opt, so I used -optimise, with an ess in case it didn't accept zed. It was to check out -Wperformance.

@som-snytt do you want a second round of review from Lukas, or are you reasonably confident you've addressed his concerns and we should go ahead and merge?

yes, now it is quite gorgeous after following up his tweaks. It is tweakèd beauty.

The old settings (-opt:l:XX and -opt-inline-from:XX) continue to work and without deprecation warning, but we should still update the PR description to explain what's changed and include it the release notes.