Assertion failed 'data->GetRegNum() != REG_WRITE_BARRIER_DST'

[markples]

runtime-coreclr libraries-pgo
Libraries Test Run checked coreclr Linux x64 Release

This has hit at least twice in System.Runtime.Serialization.Xml.ReflectionOnly.Tests, but that test run usually hits #75827.

https://dev.azure.com/dnceng-public/public/_build/results?buildId=52661&view=results
https://helix.dot.net/api/2019-06-17/jobs/6e3b9174-8eec-4bbf-8302-0033b34d4120/workitems/System.Runtime.Serialization.Xml.ReflectionOnly.Tests/console

https://dev.azure.com/dnceng-public/public/_build/results?buildId=52628&view=results
https://helix.dot.net/api/2019-06-17/jobs/fd3be114-dde0-4513-92ce-def4be1c2535/workitems/System.Runtime.Serialization.Xml.ReflectionOnly.Tests/console

+ grep COMPlus
+ printenv
COMPlus_JitRandomlyCollect64BitCounts=1
COMPlus_ReadyToRun=0
COMPlus_TC_QuickJitForLoops=1
COMPlus_JitRandomGuardedDevirtualization=1
COMPlus_TieredCompilation=1
COMPlus_JitRandomEdgeCounts=1
COMPlus_DbgMiniDumpName=/home/helixbot/dotnetbuild/dumps/coredump.%d.dmp
COMPlus_TieredPGO=1
COMPlus_DbgEnableMiniDump=1
...
  Starting:    System.Runtime.Serialization.Xml.ReflectionOnly.Tests (parallel test collections = on, max threads = 2)

Assert failure(PID 11707 [0x00002dbb], Thread: 11713 [0x2dc1]): Assertion failed 'data->GetRegNum() != REG_WRITE_BARRIER_DST' in 'System.Runtime.Serialization.XmlObjectSerializerWriteContext:SerializeWithoutXsiType(System.Runtime.Serialization.DataContracts.DataContract,System.Runtime.Serialization.XmlWriterDelegator,System.Object,System.RuntimeTypeHandle):this' during 'Generate code' (IL size 87; hash 0x9f1fa2ca; Tier1)

    File: /__w/1/s/src/coreclr/jit/codegenxarch.cpp Line: 5144
    Image: /datadisks/disk1/work/ADF20904/p/dotnet

cc @dotnet/jit-contrib

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

Issue Details

---

runtime-coreclr libraries-pgo
Libraries Test Run checked coreclr Linux x64 Release

This has hit at least twice in System.Runtime.Serialization.Xml.ReflectionOnly.Tests, but that test run usually hits #75827.

https://dev.azure.com/dnceng-public/public/_build/results?buildId=52661&view=results
https://helix.dot.net/api/2019-06-17/jobs/6e3b9174-8eec-4bbf-8302-0033b34d4120/workitems/System.Runtime.Serialization.Xml.ReflectionOnly.Tests/console

https://dev.azure.com/dnceng-public/public/_build/results?buildId=52628&view=results
https://helix.dot.net/api/2019-06-17/jobs/fd3be114-dde0-4513-92ce-def4be1c2535/workitems/System.Runtime.Serialization.Xml.ReflectionOnly.Tests/console

+ grep COMPlus
+ printenv
COMPlus_JitRandomlyCollect64BitCounts=1
COMPlus_ReadyToRun=0
COMPlus_TC_QuickJitForLoops=1
COMPlus_JitRandomGuardedDevirtualization=1
COMPlus_TieredCompilation=1
COMPlus_JitRandomEdgeCounts=1
COMPlus_DbgMiniDumpName=/home/helixbot/dotnetbuild/dumps/coredump.%d.dmp
COMPlus_TieredPGO=1
COMPlus_DbgEnableMiniDump=1
...
  Starting:    System.Runtime.Serialization.Xml.ReflectionOnly.Tests (parallel test collections = on, max threads = 2)

Assert failure(PID 11707 [0x00002dbb], Thread: 11713 [0x2dc1]): Assertion failed 'data->GetRegNum() != REG_WRITE_BARRIER_DST' in 'System.Runtime.Serialization.XmlObjectSerializerWriteContext:SerializeWithoutXsiType(System.Runtime.Serialization.DataContracts.DataContract,System.Runtime.Serialization.XmlWriterDelegator,System.Object,System.RuntimeTypeHandle):this' during 'Generate code' (IL size 87; hash 0x9f1fa2ca; Tier1)

    File: /__w/1/s/src/coreclr/jit/codegenxarch.cpp Line: 5144
    Image: /datadisks/disk1/work/ADF20904/p/dotnet

cc @dotnet/jit-contrib

Author:	markples
Assignees:	-
Labels:	area-CodeGen-coreclr, blocking-clean-ci-optional
Milestone:	8.0.0

[jakobbotsch]

Haven't seen this recently. This and #77141 might have been fixed by something. Will removing blocking label for now and keep monitoring.

[jakobbotsch]

This reproed again this Sunday:
https://dev.azure.com/dnceng-public/public/_build/results?buildId=152015&view=ms.vss-test-web.build-test-results-tab&runId=3204063&resultId=196116&paneView=debug

Unfortunately I still haven't been able to get a repro of this.

[jakobbotsch]

I was finally able to get a repro by setting DOTNET_TC_AggressiveTiering=1 in addition to all the other environment variables, and then running this in a loop for a bit.

Attached an SPMI context that repros it with unix-x64 JIT on commit dfe1076.

repro-27528.zip

[jakobbotsch]

Seems like LSRA gives an illegal register assignment:

N3111 (  3, 10) [003011] H---------Z                 t3011 =    CNS_INT(h) ref    REG rsi $b16
                                                            ┌──▌  t3011  ref    
               [004437] -----------                 t4437 = ▌  RELOAD    ref    REG rdi
N3113 (  3, 10) [001640] H----------                 t1640 =    CNS_INT(h) long   0x7f7ee0008440 static Fseq[<unknown field>] REG rsi $b14
                                                            ┌──▌  t1640  long   
                                                            ├──▌  t4437  ref    
N3115 (???,???) [003877] -A--G------                         ▌  STOREIND  ref    REG NA

The address is t1640, i.e. a constant. It is put into rsi, while the data, t4437, is put into rdi. But the write needs a write barrier, and the write barrier expects the address in rdi and the data in rsi.

[jakobbotsch]

The problem is that gcIsWriteBarrierCandidate is expected to return the same results in LSRA and codegen, yet the checks on the data do not skip LSRA inserted COPY/RELOAD nodes. Will submit a fix.
As a side note the register allocation is very odd here (why is LSRA even inserting a RELOAD on top of a constant, right after its def?). After the fix we end up with this very subpar codegen:

       48BEE8086F5BBF7F0000 mov      rsi, 0x7FBF5B6F08E8
       4889B570FEFFFF       mov      gword ptr [rbp-190H], rsi
       48BE408400E07E7F0000 mov      rsi, 0x7F7EE0008440      ; data for <unknown class>:<unknown field>
       488BBD70FEFFFF       mov      rdi, gword ptr [rbp-190H]
       48893E               mov      gword ptr [rsi], rdi

cc @kunalspathak, the SPMI context above provides the repro case. There's some reuse of constants, maybe that's related.

[003011] 3112.#2455 C501  Def    Alloc    rsi  │     │V116a│     │V4  a│     │     │     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│
                                 Spill    rsi  │     │V116a│     │V4  a│     │     │     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│
[001640] 3114.#2456 C502  Def    Alloc    rsi  │     │V116a│     │V4  a│C502a│     │     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│
[003877] 3115.#2457 C502  Use *  Keep     rsi  │     │V116a│     │V4  a│C502i│     │     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│
         3115.#2458 C501  Use *  ReLod    rdi  │     │V116a│     │V4  a│     │C501a│     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│
                                 Keep     rdi  │     │V116a│     │V4  a│     │C501i│     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│
[003012] 3120.#2459 C503  Def    Reuse    rdi  │     │V116a│     │V4  a│     │C503a│     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│
[001644] 3121.#2460 C503  Use *  Keep     rdi  │     │V116a│     │V4  a│     │C503i│     │     │     │V117a│V2  a│V34 a│V3  a│V1  a│