29/11/13 13:47, David Matthews wrote:
The ML standard is not specific about what the "program context" should be when disambiguating flexible records and overloading. Poly/ML uses as wide a context as possible but that can mean that a program that compiles with Poly/ML won't necessarily compile with other implementations. I've added a switch PolyML.Compiler.narrowOverloadFlexRecord that defaults to false. If set to true the compiler uses a much narrower context; essentially the val binding.
I've looked at a few other implementations to see what they do. The most restrictive is Hamlet which corresponds to what Poly/ML now does with narrowOverloadFlexRecord set to true. It rejects let fun f x y = x + y in f 1.0 end; which SML/NJ (110.74) accepts. Neither accept let fun f {x, ...} = x in f {x=1} end; Mlton seems to use the whole program context as Poly/ML does.
There was a discussion about this on the MLton mailing list while ago and MLton now has an annotation to control the context used: http://thread.gmane.org/gmane.comp.lang.ml.mlton.user/1335/focus=6086 http://sourceforge.net/p/mlton/code/7618/
I thought MLton's default value for this annotation (equivalent to its previous behaviour) caused a narrower context than PolyML, so I have been using -default-ann 'resolveScope topdec' to get (nearly) equivalent behaviour between the compilers.
Phil