Very odd behaviour with return value.

[fcard]

Sorry for the undescriptive title but I don't understand what's happening here at all.

I was trying to do the llvm Kaleidoscope tutorial in julia and in the middle of writing the parser I was met with a very confusing behaviour. Here is the smallest code I could reproduce the problem with (sorry if it's still too big). What happens seems to change, sometimes it crashes, sometimes things like this happen:

julia> include("c:/github/kaleidoscope/src/old/error_test.jl")
WARNING: replacing module ErrorTest
expr
primary 1
primary 4
identifier 1
args
expr
primary 1
primary 2
Should return: ErrorTest.AST.Numeric{Int64}(1)
EX: args
TYPE: args
ERROR: LoadError: MethodError: `convert` has no method matching convert(::Type{ErrorTest.AST.Expression}, ::ASCIIString)
This may have arisen from a call to the constructor ErrorTest.AST.Expression(...),
since type constructors fall back to convert methods.
Closest candidates are:
  call{T}(::Type{T}, ::Any)
  convert{T}(::Type{T}, ::T)
 in push! at array.jl:419
 in parse_args! at c:\github\kaleidoscope\src\old\error_test.jl:98
 in parse_identifier! at c:\github\kaleidoscope\src\old\error_test.jl:82
 in parse_primary! at c:\github\kaleidoscope\src\old\error_test.jl:72
 in parse_expr! at c:\github\kaleidoscope\src\old\error_test.jl:53
 in include at boot.jl:259
 in include_from_node1 at loading.jl:266
while loading c:\github\kaleidoscope\src\old\error_test.jl, in expression starting on line 115

expr through primary 2 are the functions being called, Should return is the value that should be returned to here, EX is the value actually returned, and TYPE is the type of the value, and somehow they're the same...? A few times EX was blank and julia crashes before TYPE is printed, but I couldn't reproduce that anymore.

Other inputs to parse_expr! seemed to work fine. "f()" works, "1" works, "x" works, only when it's called from within parse_args! does this happen.

I tried this both on linux and windows, and on 0.4 and 0.3, and the same thing happened.

For 0.4 I am using:

Version 0.4.0-dev+6859 (2015-08-20 19:38 UTC)
Commit 92ddae7 (0 days old master)

Am I missing something obvious? I really don't understand this.

[vtjnash]

from the code_typed, this looks a lot like #9770

[ViralBShah]

Is this a bug we can fix for 0.4 and perhaps even backport?

[vtjnash]

fixed by #15300

[fcard]

Yes!! Thank you very much for fixing this. You people are great! \o/ \o/ \o/