http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48377
--- Comment #26 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-04-07
12:48:12 UTC ---
int a = __alignof__ (double);
struct A { int b; double c; int d; } e;
int f = __alignof__ (e.c);
double *p;
int g = __alignof__ (*p);
int a2 = __alignof__ (long long);
struct A2 { int b; long long c; int d; } e2;
int f2 = __alignof__ (e2.c);
long long *p2;
int g2 = __alignof__ (*p2);
on powerpc64-linux unfortunately shows that the target hook is needed, at least
with -m64 -malign-power. Though, what the target hook does is definitely
weird:
if (TARGET_32BIT)
{
if (rs6000_alignment_flags == MASK_ALIGN_NATURAL)
return true;
if (rs6000_alignment_flags == MASK_ALIGN_POWER)
return true;
return false;
}
else
{
if (TARGET_MACHO)
return false;
/* Assuming that all other types are naturally aligned. CHECKME! */
return true;
}
I believe rs6000_alignment_flags can be either MASK_ALIGN_NATURAL, or
MASK_ALIGN_POWER, nothing else, so wonder why it just doesn't return true for
TARGET_32BIT. And, for TARGET_64BIT && !TARGET_MACHO, clearly DFmode types
aren't naturally aligned with rs6000_alignment_flags == MASK_ALIGN_POWER.
Therefore I'd say it should return false for
TYPE_MODE (strip_array_types (type)) == DFmode && !TARGET_ALIGN_NATURAL, at
least for Linux (rs6000 has unfortunately way too many ABIs and even more
target switches). For TARGET_MACHO, ADJUST_FIELD_ALIGN is
darwin.h:#define ADJUST_FIELD_ALIGN(FIELD, COMPUTED) \
darwin.h- (TARGET_ALIGN_NATURAL ? (COMPUTED) \
darwin.h- : (COMPUTED) == 128 ? 128 \
darwin.h- : MIN ((COMPUTED), 32))
thus I think it could use ADJUST_FIELD_ALIGN (NULL, TYPE_ALIGN (type)) ==
TYPE_ALIGN (type). And AIX is the same as Linux TARGET_64BIT, probably even
for 32-bit AIX though.
So, I think you want the compare_tree_int check in generic code (not sure if
is_packed isn't redundant with that check), plus a hook which will sometimes
tell you peeling won't necessarily align either.