[regression] dynamic dispatch on multiple arguments crashes compiler

[skilchen]

This program can't be compiled with the latest devel version of Nim:

type
  Foo = object of RootObj
  Bar = object of Foo
    x: int

method mm(a: Foo, b: Foo) {.base.} =
  echo "mm: Foo, Foo"    

method mm(a: Bar, b: Bar)  =
  echo "mm: Bar, Bar"
  
method mm(a: Foo, b: Bar)  =
  echo "mm: Foo, Bar"
  
method mm(a: Bar, b: Foo)  =
  echo "mm: Bar, Foo"

proc test() =
  var foo: Foo
  var bar: Bar

  mm(foo, foo)
  mm(foo, bar)
  mm(bar, foo)
  mm(bar, bar)

test()

This issue is related to the fact that currently the CI tests fail.

The very surprising and slightly disturbing fact is: this issue is caused by the line {.deprecated: [SomeReal: SomeFloat].} in system.nim.

Without this deprecation, the test-program compiles and runs correctly and so do the currently failing tests method\tmultim2 and method\tmultim6

[timotheecour]

@LemonBoy indeed, I was seeing unrelated failures :

https://travis-ci.org/nim-lang/Nim/jobs/401284582
#8213

FAIL: tmultim2.nim C
Test "tests/method/tmultim2.nim" in category "method"
Failure: reNimcCrash
Expected:
Gotten:
internal error: genMagicExpr: mTypeTrait


FAIL: tmultim6.nim C
Test "tests/method/tmultim6.nim" in category "method"
Failure: reNimcCrash
Expected:
Gotten:
internal error: genMagicExpr: mTypeTrait

hopefully the PR fixed it, but, curious how this regression was even possible; wouldn't have that failing test prevented the PR that introduced the bug from being green? or was it a flaky failure?

[LemonBoy]

curious how this regression was even possible

Luck. Every time the system.nim was modified the compiler would pick one of the three and overloads (mAnd, mBitandI, mTypeTrait) but it was lucky enough to pick one of the first two that work fine with bools.