On Tue, Dec 13, 2011 at 01:26:42PM +0100, Tom de Vries wrote:
> 2011-12-13 Tom de Vries <[email protected]>
>
> PR tree-optimization/51491
> * tree-ssa-ccp.c (insert_clobber_before_stack_restore): New function.
> (ccp_fold_stmt): Use insert_clobber_before_stack_restore after a
> successful fold_builtin_alloca_with_align.
I don't think this is safe. You don't want to look for any following
__builtin_stack_restore, but for the matching one. Consider:
int g (int *);
int
f (int n)
{
int tt = 0;
int t = 4;
{
int a[t
#ifdef DIFFERENT_BB2
+ (tt != 0 ? 6 : 0)
#endif
];
tt = g (a);
{
int b[n];
tt += g (b);
#ifdef DIFFERENT_BB
if (n > 20)
tt += 148 * g (b);
#endif
tt += b[0];
}
tt += a[0];
}
{
int a[4];
tt += g (a);
tt += a[0];
}
return tt;
}
Without any defines, this shows that looking for the first
BUILT_IN_STACK_RESTORE is wrong if you ignore BUILT_IN_STACK_SAVE calls.
And with the various defines it shows that neither the corresponding
__builtin_stack_save nor __builtin_stack_restore have to be in the same
bb as __builtin_alloca_with_align that is being folded.
Perhaps you want to look for the closest enclosing __builtin_stack_save
(search backwards in current basic block, its immediate dominator etc.?),
remember what SSA_NAME it stores its result into and then just look at
where is the (single?) user of that, which ought to be
__builtin_stack_restore.
Jakub
