Re: [PATCH] Repair __builtin_setjmp/__builtin_longjmp

[Mark Mitchell <mark at codesourcery dot com>]

Eric Botcazou wrote:

> Proposed changes attached.  Tested by building an --enable-sjlj-exceptions 
> compiler on AMD64/Linux.
>
> 	* unwind-sjlj.c: Poison setjmp.
> 	* doc/install.texi (enable-sjlj-exceptions): Use more general wording.
> 	* doc/tm.texi (DWARF2_UNWIND_INFO): Likewise.
> 	(TARGET_UNWIND_TABLES_DEFAULT): Fix typo.
> 	(DONT_USE_BUILTIN_SETJMP): Document it.

I apologize for disappearing for a while here; I've been traveling, and 
am digging out of the email pile now.

> +/* The setjmp side is dealt with in the except.c file.  */
> +#define setjmp setjmp_should_not_be_used_in_this_file

Since <setjmp.h> is sometimes included in this file, you should do 
"#undef setjmp" before (re)defining setjmp.  Please make that change.

With that change, the patch is OK.

As to the issue of documenting built-ins, let's leave it aside for now; 
you're not making things any worse than they were before.

However, my position remains that we should either define the semantics 
for builtins or explicitly prohibit their use.  I would also accept 
Kenner's suggestion that we prohibit their use, but then provide a 
"-fexpose-internal-builtins" option to allow their use in test-cases. 
But, as Paul Brook and others said, I think that silently allowing the 
use of magic builtins whose semantics we don't intend to guarantee over 
time is a bad idea.

Thanks,

--
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713