Make JitPrintInlinedMethods less verbose

[EgorBo]

Was:

Now:

We can also introduce a verbosity level e.g. to only print successful inlinees after [guarded] devirtualizations. The goal is to allow developers to easily find out which call-sites were devirtualized (and inlined because devirtualization without inlining is not interesting) with PGO and which were not.

cc @dotnet/jit-contrib @AndyAyersMS

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Was:

Now:

We can also introduce a verbosity level e.g. to only print successful inlinees after [guarded] devirtualizations. The goal is to allow developers to easily find out which call-sites were devirtualized with PGO and which were not.

cc @dotnet/jit-contrib @AndyAyersMS

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[jakobbotsch]

I'm not a fan of removing the ordinal and IL offset, I've been using those a lot in my debug info work. Can we keep those?

[EgorBo]

I'm not a fan of removing the ordinal and IL offset, I've been using those a lot in my debug info work. Can we keep those?

I can make the "verbose mode" to be default. If make Checked jits available for external developers I'd love it to be more friendly and less verbose for them at least with an additional config switch. Especially when ="*" mode is used.

[jakobbotsch]

I'm not a fan of removing the ordinal and IL offset, I've been using those a lot in my debug info work. Can we keep those?

I can make the "verbose mode" to be default. Since we're going to publish a Checked jit for external developers I'd love it to be more friendly and less verbose for them at least with an additional config switch. Especially when ="*" mode is used.

Do we expect external developers to use the full JitDump mode? If we had a separate variable to report the things that may be interesting for external devs then we could have the less verbose output for this mode, and keep the more verbose mode for JitDump.

[EgorBo]

I'm not a fan of removing the ordinal and IL offset, I've been using those a lot in my debug info work. Can we keep those?

I can make the "verbose mode" to be default. Since we're going to publish a Checked jit for external developers I'd love it to be more friendly and less verbose for them at least with an additional config switch. Especially when ="*" mode is used.

Do we expect external developers to use the full JitDump mode? If we had a separate variable to report the things that may be interesting for external devs then we could have the less verbose output for this mode, and keep the more verbose mode for JitDump.

It's not part of JitDump, it's printed for DOTNET_JitPrintInlinedMethods=Test (or DOTNET_JitPrintInlinedMethods=*).

[jakobbotsch]

It's not part of JitDump, it's printed for DOTNET_JitPrintInlinedMethods=Test (or DOTNET_JitPrintInlinedMethods=*).

The inline tree is also printed in JitDump using this code.

[jakobbotsch]

Maybe pass verbose parameter to InlineStrategy::Dump and InlineContext::Dump for JitDump case? Then we can get rid of the new COMPlus variable too, I think the less verbose output makes sense if one is only looking at the inline tree without full jit dump.

[EgorBo]

@jakobbotsch good point! addressed in the latest commit.

[jakobbotsch]

I would personally prefer that "leaf" methods like these do not change behavior based on environment variables, e.g. using a verbose parameter instead (similar to what is already happening in InlineStrategy::Dump with its showBudget parameter). But that's not terribly important, if you prefer it this way that's fine with me.