Improve tailcallstress testing

[erozenfeld]

1. Dispatch all tail calls under TailCallStress via helpers. That
   increases our coverage since valid non-tail-prefixed calls are
   dispatched as fast calls in normal non-stress mode.

2. Don't attempt to tail call from methods that have a localloc
   unless there is an explicit tail prefix. Note that we already disallowed
   fast tail calls from such methods so this change only affects
   tailcallstress mode.

3. Fix a bug in TestInvokeDOPAndCancel. As the test was written
   this assert was firing under tailcallstress:
   

   runtime/src/libraries/System.Threading.Tasks.Parallel/tests/ParallelForTests.cs

   Line 1074 in 480c49b

   Assert.False(counter == numActions, "TestInvokeDOPAndCancel: > FAILED! Cancellation+exception not effected.");

   
   When this call to cts.Cancel:
   

   runtime/src/libraries/System.Threading.Tasks.Parallel/tests/ParallelForTests.cs

   Line 1065 in 480c49b

   if (newVal == 2) cts.Cancel();

   
   is dispatched via helpers, it takes much longer than when it's
   dispatched via a fast tail call (we have to jit a couple of IL stubs).
   Because of that, it's possible that all other actions complete on the
   other thread before cancellation is completed.

[erozenfeld]

runtime-coreclr jitstress: https://dev.azure.com/dnceng/public/_build/results?buildId=778817&view=results
runtime-coreclr libraries-jitstress: https://dev.azure.com/dnceng/public/_build/results?buildId=778819&view=results

[erozenfeld]

@AndyAyersMS @dotnet/jit-contrib PTAL
@tarekgh @danmosemsft PTAL at the test change

[tarekgh]

@erozenfeld test changes LGTM.

[AndyAyersMS]

Thanks, LGTM.

I was thinking we should make tail call stress be modal and perhaps still try and use fast tail calls in some cases (perhaps randomly under jitstress, etc). But that can wait.

[erozenfeld]

I investigated the System.Threading.Threads.Tests.ThreadTests.ConstructorTest failure in runtime-coreclr libraries-jitstress (Windows x64) under tailcallstress: https://helix.dot.net/api/2019-06-17/jobs/d017bf2e-6058-4ed6-819f-ee589065f3a5/workitems/System.Threading.Thread.Tests/files/console.8238c349.log

The stack trace is: #40698 (comment)

The test was attempting to launch a thread with a 64kb stack and it was failing with StackOverflow. On my machine it was failing with StackOverflow even without tailcallstress. I made a change to double SmallStackSizeBytes from 64k to 128k.

@kouvel @mangod9 PTAL at the test change

[erozenfeld]

Libraries Test Run release coreclr Windows_NT x86 Debug failures are an infrastructure issue. Ignoring since it passed on a previous run.

[kouvel]

Thanks!