Launch Crossgen2 via explicitly specified .NET host

[trylek]

/cc @dotnet/crossgen-contrib

Figuring out right now how to inject this into the runtime repo in order to actually test it, sending out in hope of early PR feedback.

Thanks

Tomas

[AntonLapounov]

Crossgen2 is published as a self-contained app. Why do we want to launch it using a host? Is that to support scenarios when crossgen2 is not self-contained?

[trylek]

I believe I have addressed the gist of David's feedback. For Anton's suggestion, as discussed offline this is non-trivial to do as publishing of the "self-contained Crossgen2 app" by itself needs the crossgenned framework as a prerequisite. Having said that, I don't think we can afford to wait until this fix propagates to the runtime repo by means of rolling forward the SDK version. The SDK upgrade is scheduled to be merged in as part of the March infra roll-out, please see

dotnet/runtime#47245

so that one more upgrade (to include changes in this PR) will likely happen in April if not later and that's far too late for our other CG2 efforts. Apart from finalizing this change I plan to try to address or mitigate reviewer feedback on my other PR

dotnet/runtime#47349

so that we can start testing CG2-compiled SDK ASAP. One way or another, please take another look at this change when you have a chance.

Thanks

Tomas

My concerns have been addressed

[davidwrighton]

@trylek, My concerns have been addressed, but I'd rather somebody else walk through and verify that the behavior is correct as I'm generally not familiar with this codebase. I'd like to see @AntonLapounov provide a review.

[AntonLapounov]

I think all error messages should go to resource files to allow localization.

[trylek]

@AntonLapounov - sigh, apparently I need to add some provisions to PrepareReadyToRunCompilation to cater for the fact that Crossgen2 until now hasn't been generating the MVID into the PerfMap name. I'll send out a PR for fixing that but that will only make it to Preview 2 - am I right to presume that we need something like a minor version check in the code?

[AntonLapounov]

I don't think you need a minor version check. Make the fix in Runtime repo and wait a bit until it is picked up by the SDK repo. SDK will use the latest Crossgen2 package.

[trylek]

Well, if I don't merge the SDK change in this week, it's going to slip to Preview 3 and its back-propagation to runtime will get delayed until Preview 3 ships (becomes generally available); I'm worried that pushes us uncomfortably close to the general .NET 6 shipping date.

[AntonLapounov]

@trylek You may temporarily disable emitting perf maps in R2R SDK tests for .NET 6 to unblock this PR. Do I understand correctly that you want to place the MVID in the file name for non-composite mode only? Is having MVID in the file name important? Below are the errors I see in the log:

error MSB3030: Could not copy the file "obj/Debug/net6.0/ubuntu.16.04-x64/R2R/Crossgen2TestApp.ni.{4e16f2c1-0ed5-4bc0-88f3-2b978f1a98e5}.map" because it was not found.
error MSB3030: Could not copy the file "obj/Debug/net6.0/ubuntu.16.04-x64/R2R/ClassLib.ni.{233d699a-af57-4160-a632-0ad7769be9a8}.map" because it was not found.

error MSB3030: Could not copy the file "obj/Debug/net6.0/ubuntu.16.04-x64/R2R/Crossgen2TestApp.r2r.ni.map" because it was not found. 

[trylek]

I guess that the purpose of the MVID in the perfmap name is similar to some checksum that's in the PDB to let diagnostic software verify that it matches, @hoyosjs or @brianrob might know. I guess you're right that I can disable the perfmap generation for now, it's just unfortunate to accrue debt.

[AntonLapounov]

@trylek Is that a problem that for the composite perf map you are not using any MVID? If not, maybe omitting MVID in non-composite mode would be OK as well.

[brianrob]

Right, the MVID is used to make sure that we only use a "matching" symbol file. Sounds like for composite mode, we'd need to do something else, but perhaps we'd just depend on a recent enough version of the perf tools to be available on the box, and then it shouldn't matter, since we'd use the built-in Linux perf tool's jitdump functionality.