tailcalls_do / tailcalls_d fail on Windows ARM64

[trylek]

Example repro:

https://dev.azure.com/dnceng/public/_build/results?buildId=494779&view=ms.vss-test-web.build-test-results-tab&runId=15897688&resultId=100793&paneView=debug

Diagnostic info:

Return code: 1
Raw output file: C:\dotnetbuild\work\98AC07F7\w\9293083E\e\JIT\Stress\Reports\JIT.Stress\ABI\tailcalls_do\tailcalls_do.output.txt
Raw output:
BEGIN EXECUTION
 "C:\dotnetbuild\work\98AC07F7\p\corerun.exe" tailcalls_do.dll --tailcalls --num-calls 1000 --no-ctrlc-summary 
Stressing tailcalls
OSVersion: Microsoft Windows NT 6.2.9200.0
OSArchitecture: Arm64
ProcessArchitecture: Arm64
Selecting win64 ABI
Mismatch in tailcall: expected -1792697222, got -241256253
Int32 ABIStress_TailCaller4(S2U, Double, Int16, Int32, Int16, S1P, S8P, S2U, S8P, S4U, S8P, Int32, S2U, Byte, S2U, S4U, Single, S2U, Byte, S8P, Int16, S1P, S4P, S4U)
Int32 ABIStress_TailCallee5178(Int64, S4P, S4U, S8P, Double, S2P, S4P, S2U, Single, Double, S8U, S8P)
Mismatch in tailcall: expected 329656105, got -1461745871
Int32 ABIStress_TailCaller9(S8U, Byte, S2U, Single, S1P, Int32, Int32, S2U, Int16, Int16, Double, Byte, S4P, S4U, S2U)
Int32 ABIStress_TailCallee8146(Int32, S1P, Int64, S8P, Int64, Int16, S4U, S4U, Single, S4P, S4P, S1P, S4P, Single)
Mismatch in tailcall: expected 1872306653, got -280200393
Int32 ABIStress_TailCaller10(Int32, S8U, S4U, Int64, S8U, S4P, Double, S2U, S4U, S4U, Double, S2U, Int64, S4U, S8U, S8P, S2P, S4U, Byte, S2P, Double, S8P)
Int32 ABIStress_TailCallee6312(Double, Double, S4P, S8U, S1P, S2P, Single, S4U, Byte, Int64, Byte, Double, S4U, S4P, Int64, S8U, S2P, S8P, Single)
Mismatch in tailcall: expected -958760121, got 1383693481
Int32 ABIStress_TailCaller31(Int16, S8P, Single, Int16, Int16, S4U, S2P, S4P, Byte, S8P, S8U, Single, S1P, Int32, Int16, S4U)
Int32 ABIStress_TailCallee9934(Int64, Int16, Double, S2U, Int64, Byte, S4P, S8P, S4U, S8U, Single, S1P, Int16, S2P, Byte)
50 callers done (45 successful tailcalls tested)
Mismatch in tailcall: expected -1032050283, got -1862624250
Int32 ABIStress_TailCaller50(S4U, S2P, S4P, S2P, Int32, S8P, S2P, Int64, S1P, Byte, S8U, S2U, S8U, S1P, Int16, S8P, Byte, Double)
Int32 ABIStress_TailCallee9994(Int64, Single, S8P, Int16, Int32, Int64, Int64, S8U, Int32, Int64, Single, Int16, S8U, S2P, Double)
Mismatch in tailcall: expected 1332452070, got 285299523
Int32 ABIStress_TailCaller52(Int64, Int32, Double, Int16, Single, S4P, Byte, S8U, S2U, S4U, S8U, S4U, S2P, S8P, S4U, Int32, Byte, Int32, Byte, S8P, Int64, S2U, Single, Int64)
Int32 ABIStress_TailCallee1312(S8P, S2P, Int16, Byte, Double, Int16, Int32, Int16, S2U, S4P, Byte, Int32, Byte, Double, Double, S8P, Byte, S1P, S4U, Single, Byte, S8U)
Mismatch in tailcall: expected -454712802, got 1026721613
Int32 ABIStress_TailCaller58(S8U, Byte, Int16, Single, S4U, Single, S4P, S2P, S8P, S8P, Byte, Int64, S1P, S2U, Single, Double, S8P, Double)
Int32 ABIStress_TailCallee9374(Int64, S8U, S4U, S4U, Int64, Single, S4U, S4P, S4P, Int64, S8U, Byte, S4U, S8U)
Mismatch in tailcall: expected 1977502728, got -1485647357
Int32 ABIStress_TailCaller63(S2P, Int16, Int64, Int32, Byte, Byte, Byte, S2U, S8U, Double, S2U, S8U, S4P, S8P, Double, Int16, S2P, S4U, S1P, S8U, Byte, S2U, S4U)
Int32 ABIStress_TailCallee2379(S4U, S8P, S2P, S4P, Single, S4U, Int16, Single, Byte, Int64, S2U, Int16, S8U)
Mismatch in tailcall: expected 1587307949, got 202074404
Int32 ABIStress_TailCaller64(Int32, S4U, S8U, S2P, S8P, S1P, S2U, Byte, S4U, Byte, S2P, S2P, S4P, S4P, S4U, Double, Int64, S4U, Double)
Int32 ABIStress_TailCallee2937(S2U, Int16, Int16, S4P, Single, S4P, S2U, S8P, S8U, S8P, Int64, Int32, S2P, S4U, S4P, Int64, Double)
Mismatch in tailcall: expected 1855809914, got -1738803405
Int32 ABIStress_TailCaller73(Double, S2P, Int16, Int16, S2P, S8U, Int16, S8P, Int16, Single, S4U, Int16, S4P, S1P, S8U, S4P, S8U, Byte, S2P, S8P, Byte, S2P, Single, Int16)
Int32 ABIStress_TailCallee1009(S2U, S4U, Int32, S2P, S1P, S4P, S8P, Int32, Double, Int16, Int64, S2P, Single, Int16)
Mismatch in tailcall: expected -1351429783, got -1492495110

The issues.targets already disables some JIT tests with the issue set to "Requires helper-based tailcalls"; not sure whether this is related. CC @janvorli who was involved in the tailcall design and parts of its implementation AFAIK.

category:correctness
theme:testing
skill-level:expert
cost:small

[BruceForstall]

@jashook @erozenfeld @jakobbotsch @dotnet/jit-contrib

I think these need to be disabled for Windows/arm64.

[janvorli]

CC @janvorli who was involved in the tailcall design and parts of its implementation AFAIK.

The new tailcall implementation based on my original design doc is still not in, see #341.

[jakobbotsch]

The tailcall stress tests do not rely on helper-based tailcalls working. This is probably an actual bug in the implementation of fast tailcalls in Windows arm64 or a bug in the implementation of the ABI on that platform.

To get a better test case on windows arm64 I would rerun the test with smaller and smaller --max-params (args) until I found the smallest example of mismatching behavior, then proceed from there.

[AndyAyersMS]

Let's look at this for .NET 5.