Re: [Chicken-hackers] Only specialize string-length if type is definitely known

[Felix]

From: Christian Kellermann <address@hidden>
Subject: [Chicken-hackers] Only specialize string-length if type is definitely 
known
Date: Sun, 27 May 2012 20:29:47 +0200

> Hi Hackers,
> 
> as Mario pointed out the previously commited fix for #855 was only the
> sympthom not the cause of the troubles. The real cause has been the
> unconditionally specialization of string-length to ##sys#size which
> will segfault if given immediate types (since it is only a wrapper for
> C_block_size).
> 
> So the attached patch removes #:enforce from its types.db entry.
> 
> You can test it for yourself with the following example code:
> 
> (define (t l)
>   (let ((s1 (car l)))
>     (zero? (string-length s1))))
> 
> (t '(1))
> 
> I am not fully sure whether I have understood the purpose of #:enforce
> correctly. As I understood it signals the compiler to carry on even if
> it does not know for sure which type it is, as the specialization code
> will handle type errors by itself. Is that correct? Who can enlighten
> me?
> 

Enforce merely specifies that the argument is of the declared type
_in_case_the_procedure_call_returns_without_error_. Apparantly here is
a bug in the flow-analysis that causes the enforcement property of
"string-length" to be applied to early: it assumes "s1" is a string
because it is passed to "string-length", but it does so too early.
Removing the "#:enforce" declaration is not a solution to this. Please
hold, this needs scrutinizer hacking and needs to be addressed


cheers,
felix