Less conservative gt obj

[CarolEidt]

Don't mark GT_OBJ with GTF_EXCEPT or GTF_GLOB_REF if it is a local struct.
Also, fix changes needed in arm register allocation and code generation to handle the case where a struct argument is promoted, which appears to be code that was never exercised.
These changes are in preparation for further changes for First Class Structs. I am working to replace block ops with assignments, and was attempting to get to zero diffs, but it is quite difficult to do so given some inconsistencies in the treatment of GT_OBJs of locals.

[CarolEidt]

Please do not merge this PR yet. It requires additional validation.
@myungjoo, @chunseoklee, @jyoungyun, @leemgs, @prajwal-aithal, @parjong - these changes exercise arm codegen paths in codegenlegacy.cpp that were never executed before, because the struct arguments were always treated as global references and therefore those locals were never promoted. Now they are getting promoted, and the code to move the arguments from the registers to the argument locations had some bugs. If you can do some validation of this, it would be great.

[parjong]

@CarolEidt Is there any testcase related with this issue? It takes about 7 hours to run the whole unittests even in Release build (and over 1 day in Debug build). It would be very helpful if we could know which unittests are relevant to this issue.

[CarolEidt]

@parjong - I have been looking at diffs in the frameworks; I will have a look at the tests to see if I can find some that show the difference.

[parjong]

@CarolEidt Thanks 👍 Anyway, we will evaluate this patch for the whole unittest (it takes some time to get the result)

[CarolEidt]

@parjong - I am not able to duplicate the diffs I've seen in any of the unit tests that I've looked at so far. The diffs in the desktop build of the libraries had diffs for uses of DateTime and Decimal, as well as a number of struct types in the non-core libraries.

[CarolEidt]

It looks like I have missed a case in x86 ryujit, so I will have to track that one down.

[jyoungyun]

I'm trying to run unit tests on arm device. I will post the results soon.

[jyoungyun]

Um, I got the below exception error.
The test case is RegionInfoCurrentRegion.exe.
And I couldn't get any log because this exception only occurs Release mode.

Beginning test case RegionInfoCurrentRegion at 01/01/2015 11:08:34
Random seed: 20010415; set environment variable CORECLR_SEED to this value to repro

[Positive]
Beginning scenario: PosTest1:Return the CurrentRegion property in RegionInfo class

Unhandled Exception: 
   Cannot print exception string because Exception.ToString() failed.
Aborted

[CarolEidt]

@dotnet/jit-contrib PTAL - I am attempting an arm cross build, so that I can look into the RegionInfoCurrentRegion.exe failure. In the meantime, I would like to start the review process.

[BruceForstall]

@CarolEidt LGTM. However, I don't understand why the changes to codegenlegacy.cpp for reconstructing struct args from promoted structs is required, or related. That code is unfortunately complex...

[CarolEidt]

@BruceForstall - That code pre-existed, but was apparently never executed because when structs were passed as arguments, they were not being promoted. An alternative would be to prevent structs from being promoted on arm, but I'm not sure that's even desirable, as the code seems to be working correctly, modulo one possible test failure that I'm attempting to investigate - and IMO it serves as a proof point for doing this - which I hope we will be doing before too long for other targets as well.

[CarolEidt]

I have not yet been able to run the failing test on the emulator, but I ran the armaltjit on desktop for this test case, and the code produced is identical.

[CarolEidt]

@dotnet/jit-contrib ping - I would love to get another review on this.

[briansull]

Looks Good