Ticket #5255 (closed feature request: wontfix)
String literals cause runtime crashes when OverloadedStrings is in effect
Description (last modified by simonpj) (diff)
There has been a discussion on the web-devel list about the fate of the IsString instance for Name in the xml-types library. A Name is the name of an XML element or attribute.
That instance calls error when the string contains a certain kind of invalid domain-specific syntax. Some are even advocating expanding this behavior to any string that is syntactically invalid for XML names.
So we now have GHC as the only major compiler which can cause *runtime* crashes depending on what characters are used in a string literal.
OverloadedStrings as a more general mechanism is very convenient in many settings. One of them is XML names; another is attoparsec-text parsers. I must admit I have succumbed to the temptation of making this deal with the devil and benefiting from them.
But when used this way OverloadedStrings is really just another syntax for quasi-quotation, and that is what should have been used explicitly instead of these unsafe domain-specific IsString instances.
I propose fixing the problem in one of the following ways:
A. Make string literals syntax in fact a specialized quasi-quotation when OverloadedStrings is turned on. That way, exceptions are caught at compile time as they should be.
B. Bless Text, and possibly ByteString, as the only types that get magical behavior of string literals.
C. Remove OverloadedStrings altogether.
Option A is by far the nicest. But it requires GHC to know the type of the string literal before the cast is applied. We might also need some way to help GHC find the cast function at the right time, beyond just having an IsString instance somewhere in scope.
By submitting this bug, I am making it clear that I am opposed to Option D, leaving things the way they are and wishing everyone the best of luck. The OverloadedStrings pragma is not really optional anymore now that Text is becoming the default string type in practice for Haskell. It is not acceptable to have to wrap every string awkwardly with (T.pack "") and give up the chance of it being CAFfed. In fact, the blaze-html library relies on OverloadedStrings for its performance.
I am also opposed, though less so, to providing a deprecation route by using a new language pragma for Option A or B. The current behavior is dangerous and should be summarily removed.