JIT: Assertion failed 'src1->OperIs(GT_LONG)'

[mazong1123]

Run Windows_NT x86 Checked Build and Test (Jit - CoreFx) CI tests will cause a failure on current master branch code:

20:10:24   Generated serialization assembly for assembly D:\j\workspace\x86_checked_w---92c0aac3\_\fx\bin\AnyOS.AnyCPU.Release\Microsoft.XmlSerializer.Generator.Tests\netstandard\Microsoft.XmlSerializer.Generator.Tests.dll --> 'D:\j\workspace\x86_checked_w---92c0aac3\_\fx\bin\AnyOS.AnyCPU.Release\Microsoft.XmlSerializer.Generator.Tests\netstandard\Microsoft.XmlSerializer.Generator.Tests.XmlSerializers.cs'.
20:10:25   Discovered:  Microsoft.Win32.Registry.Tests
20:10:32   Starting:    Microsoft.Win32.Registry.Tests
20:10:42   
20:10:42 
  Assert failure(PID 788 [0x00000314], Thread: 6912 [0x1b00]): Assertion failed 'src1->OperIs(GT_LONG)' in 'Microsoft.Xml.Serialization.GeneratedAssembly.XmlSerializationWriter1:Write65_LongEnum(long):ref:this' (IL size 90)
20:10:42   
20:10:42       File: d:\j\workspace\x86_checked_w---92c0aac3\src\jit\lower.cpp Line: 2167
20:10:42       Image: D:\j\workspace\x86_checked_w---92c0aac3\_\fx\bin\testhost\netcoreapp-Windows_NT-Release-x86\dotnet.exe
20:10:42   
20:10:43   Finished running tests.  End time=20:10:43.22, Exit code = 123456789
20:10:44 D:\j\workspace\x86_checked_w---92c0aac3\_\fx\Tools\tests.targets(445,5): warning MSB3073: The command "D:\j\workspace\x86_checked_w---92c0aac3\_\fx\bin/AnyOS.AnyCPU.Release/Microsoft.XmlSerializer.Generator.Tests/netstandard//RunTests.cmd D:\j\workspace\x86_checked_w---92c0aac3\_\fx\bin/testhost/netcoreapp-Windows_NT-Release-x86/" exited with code 123456789. [D:\j\workspace\x86_checked_w---92c0aac3\_\fx\src\Microsoft.XmlSerializer.Generator\tests\Microsoft.XmlSerializer.Generator.Tests.csproj]
20:10:44 

It should be introduced 3 days ago.

Please see the CI failure in this PR: dotnet/coreclr#13480

[mikedn]

The relevant assert code is https://github.com/dotnet/coreclr/blob/3dd1c074f5d89533d8a8e13ea6143a4850752d3e/src/jit/lower.cpp#L2162-L2169

So we have a compare with long operands but the first hasn't been decomposed yet. Hmm...

cc @dotnet/jit-contrib

[mikedn]

Self contained repro:

enum LongEnum : long
{
    Option0, Option1, Option2
}

[MethodImpl(MethodImplOptions.NoInlining)]
static string Test(LongEnum v)
{
    string s;
    switch (v)
    {
    case LongEnum.Option0: s = "Option0"; break;
    case LongEnum.Option1: s = "Option1"; break;
    case LongEnum.Option2: s = "Option2"; break;
    default: throw new Exception();
    }
    return s;
}

When compiled with optimizations disabled and run with RyuJIT-x86-Checked the assert fires.

[mikedn]

So we have

               [000028] ------------             *  SWITCH    void  
               [000027] ------------             \--*  CAST      int <- uint <- long
               [000025] ------------                \--*  LCL_VAR   long   V05 tmp1         

and after morph we have

               [000028] -----+------             *  SWITCH    void  
               [000025] C----+------             \--*  LCL_VAR   int    V05 tmp1

Morph has retyped the lclvar node to TYP_INT and dropped the cast. I don't know why morph does this when optimizations are disabled but that's not the problem.
In LowerSwitch we have:
https://github.com/dotnet/coreclr/blob/3dd1c074f5d89533d8a8e13ea6143a4850752d3e/src/jit/lower.cpp#L519-L521

So it gets the type from the lclvar desc. Morph has retyped the lclvar node but the lclvar itself is still TYP_LONG and now switch lowering is generating various new nodes using TYP_LONG, after decomposition was done.

We should probably get the type from the lclvar node instead:

var_types  tempLclType = genActualType(temp->TypeGet());

[danmoseley]

I repro this also.

[mikedn]

I repro this also.

In the corefx tests as well or somewhere else?

[danmoseley]

The CoreFX tests. I misread the report above - it's merely the same case.

[pgavlin]

This also repros on JIT\Regression\CLR-x86-JIT\V2.0-Beta2\b425314\b425314 under MinOpts.