Repository navigation
[libraries-jitstress-random] Assertion failed '(emitThisGCrefRegs & regMask) == 0' during 'Emit code' #83963
Description
Several tests are failing with this assert, in both win-x86 and linux-arm32.
Test run: https://dev.azure.com/dnceng-public/public/_build/results?buildId=201954&view=results
Example console log: https://helixre107v0xdcypoyl9e7f.blob.core.windows.net/dotnet-runtime-refs-heads-main-9423de4c7e154d2480/System.Private.Xml.Tests/1/console.d22bcbec.log?helixlogtype=result
Assert failure(PID 6288 [0x00001890], Thread: 8356 [0x20a4]): Assertion failed '(emitThisGCrefRegs & regMask) == 0' in 'System.String:Concat[ushort](System.Collections.Generic.IEnumerable`1[ushort]):System.String' during 'Emit code' (IL size 417; hash 0x33a0d879; FullOpts)
File: D:\a\_work\1\s\src\coreclr\jit\emitxarch.cpp Line: 13883
Image: C:\h\w\A6C2090E\p\dotnet.exe
Activity
JulieLeeMSFT commentedon Mar 27, 2023
@TIHan PTAL.
TIHan commentedon Mar 28, 2023
I have managed to reproduce this locally, so I can iterate on this.
TIHan commentedon Mar 28, 2023
I think I know the underlying problem.
A register,
edx, is assigned for a ref null node:CNS_INT ref null REG edx $VN.NullBut later we decide to re-use
edxfor a constant zero node:CNS_INT int 0 reuse reg val REG edx $40This occurs because the two nodes are seen as equivalent, but we do not update the
emitThisGCrefRegsto removeedxas GC tracked whenedxis re-used in a non-GC way.The decision to re-use the register is handled in LSRA, though the updating of the
emitThisGCrefRegsappears to only be in the emitter. A possible fix might be to updateemitThisGCrefRegsif we see that we are going to emit a constant int 0 that is re-using a register that was previously GC tracked.// cc @kunalspathak - not sure if there needs to be something done in LSRA in this case
full dump: dump.txt