Fix type layout diagnostics in the CoreCLR runtime #60305
Conversation
In my change from the summer I added provisions to the type layout check to support printing the differences as an aid for investigation of this class of failures; I however put the diff display after the assertion so that it actually doesn't get hit as the assertion tears down the process. This change fixes the ordering and should let us review the particular mismatch that occurs in the arm64 runs (dotnet#60036). Thanks Tomas
src/coreclr/vm/jitinterface.cpp
Outdated
| @@ -13640,10 +13640,10 @@ BOOL LoadDynamicInfoEntry(Module *currentModule, | |||
|
|
|||
| #ifdef _DEBUG | |||
| { | |||
| TypeLayoutCheck(pMT, pBlob, /* printDiff */ TRUE); | |||
| StackScratchBuffer buf; | |||
| _ASSERTE_MSG(false, fatalErrorString.GetUTF8(buf)); | |||
| // Run through the type layout logic again, after the assert, makes debugging easy | |||
There was a problem hiding this comment.
Thanks Jan for pointing that out. I have moved and rephrased the comment to mention it is usable not only for local debugging but for lab diagnostics too. That led me to realize this is actually not tied to the debug-specific instrumentation for local scenarios so I moved the check a few lines up to fire even in release builds because I believe it doesn't hurt anything.
| @@ -13632,6 +13632,9 @@ BOOL LoadDynamicInfoEntry(Module *currentModule, | |||
| } | |||
| else | |||
| { | |||
| // Run through the type layout logic again to report the mismatch in lab logs and make debugging easy | |||
| TypeLayoutCheck(pMT, pBlob, /* printDiff */ TRUE); | |||
There was a problem hiding this comment.
This was previously within a _DEBUG block. Should this be also?
There was a problem hiding this comment.
As I described in my response to Jan Kotas' PR feedback, I originally put it under the _DEBUG block (admittedly in the wrong place after the assertion) as I viewed it as a tool specific to local debugging; the fix and the recent lab failure made me realize there's actually no reason to have it under the _DEBUG block - it doesn't incur any perf penalty (in the sense that it only happens when the app is about to crash anyway) and having it outside of the _DEBUG block makes us receive the same diagnostics for release runs too - I believe that to be a win-win situation, please let me know if you think otherwise.
There was a problem hiding this comment.
The path only runs for ENCODE_VERIFY_TYPE_LAYOUT. The first call of TypeLayoutCheck runs with printDiff=TRUE for ENCODE_VERIFY_TYPE_LAYOUT.
I think the net effect of this change is going to be that the diff is going to be printed twice in the log if there is a mismatch.
There was a problem hiding this comment.
OK, my bad, I was under the impression I don't see the mismatch in the log but apparently I just overlooked it. In such case this change is indeed superfluous as you say, I'll just close the PR. Thanks for pointing that out!
ghost
locked as resolved and limited conversation to collaborators
In my change from the summer I added provisions to the type layout
check to support printing the differences as an aid for investigation
of this class of failures; I however put the diff display after the
assertion so that it actually doesn't get hit as the assertion tears
down the process. This change fixes the ordering and should let us
review the particular mismatch that occurs in the arm64 runs
(#60036).
Thanks
Tomas
/cc @dotnet/crossgen-contrib