[patch] fix sync builtin expansion and i386 pattern

[Michael Matz]

Hi,

On Fri, 16 Mar 2007, Ian Lance Taylor wrote:

> >         * builtins.c (expand_builtin_sync_operation,
> >         expand_builtin_compare_and_swap,
> >         expand_builtin_lock_test_and_set): Care for extending CONST_INTs
> >         correctly.
> > 
> >         * config/i386/sync.md (sync_double_compare_and_swapdi_pic,
> >         sync_double_compare_and_swap_ccdi_pic): Use "SD" as constraint
> >         for operand 3.
> 
> This is OK for mainline.

committed (after testing on x86-64-linux succeeded).

> It's OK for 4.1 and 4.2 if it is a regression (I've lost track).  If 
> it's not a regression, check with Mark on the backport.

Mark: I don't know if it's a regression, I believe the incorrect code was 
there since the introduction of __sync_* builtins, which would make it a 
non-regression.  The failure mode is, that calls like

unsigned long long *ptr;
__sync_fetch_and_add (ptr, -1LL);
__sync_compare_and_swap (ptr, a, -1LL);

would be miscompiled, the 64 bit -1 would be represented as 0xffffffff.  I 
have no feeling how often that is actually used in real code (or how often 
the __sync_* builtins are used at all), especially how often any constant 
arguments would be -1LL instead of e.g. just 0 or 1 (which both would be 
represented correctly).

There's another miscompilation on x86, also only involving 64bit atomic 
primitives, which also leads to misrepresented 64bit constants.


Ciao,
Michael.