# New Ticket Created by Andy Dougherty
# Please include the string: [perl #42772]
# in the subject line of all future correspondence about this issue.
# <URL: http://rt.perl.org/rt3/Ticket/Display.html?id=42772 >
On Thu, 26 Apr 2007, Leopold Toetsch wrote:
> Am Donnerstag, 26. April 2007 21:44 schrieb Andy Dougherty:
> > Does anyone understand the 'dummy' element in
> > include/parrot/stacks.h? Here is the relevant snippet:
> >
> > typedef struct Stack_Chunk {
> > pobj_t obj;
> > int size;
> > const char * name;
> > struct Stack_Chunk *prev;
> > #if ! DISABLE_GC_DEBUG && defined(I386)
> > void * dummy; /* force 8 byte align for mmx and sse moves */
> > #endif
> > union { /* force appropriate alignment of 'data' */
> > void *data;
> > #ifndef I386
> > double d_dummy; /* align double values on stack */
> > #endif
> > } u;
> > } Stack_Chunk_t;
> At some earlier time in parrot history, there was a (32 bit, pointer-sized)
> 'version' [1] structure item in pobj_t, which was active
> with !DISABLE_GC_DEBUG. The dummy was needed for i386 (when x86_64 didn't
> even exist) to align the data pointer at an 8-byte boundary.
Ah, yes. Thanks for the detailed reply. At this point, removing the
dummy will put u.data back on an 8-byte boundary (for i386), so that's
probably a good thing all around.
I'd suggest the following patch:
diff -ru parrot-current/include/parrot/stacks.h
parrot-andy/include/parrot/stacks.h
--- parrot-current/include/parrot/stacks.h 2006-11-12 03:15:19.000000000
-0500
+++ parrot-andy/include/parrot/stacks.h 2007-04-27 09:24:04.000000000 -0400
@@ -29,13 +29,11 @@
int size;
const char * name;
struct Stack_Chunk *prev;
-#if ! DISABLE_GC_DEBUG && defined(I386)
- void * dummy; /* force 8 byte align for mmx and sse moves */
-#endif
- union { /* force appropriate alignment of 'data' */
+ union { /* force appropriate alignment of 'data'. If alignment
+ is necessary, assume double is good enough. 27-04-2007. */
void *data;
-#ifndef I386
- double d_dummy; /* align double values on stack */
+#if PARROT_PTR_ALIGNMENT > 1
+ double d_dummy;
#endif
} u;
} Stack_Chunk_t;
--- parrot-current/src/stack_common.c Thu Apr 26 19:15:14 2007
+++ parrot-andy/src/stack_common.c Fri Apr 27 11:49:53 2007
@@ -59,9 +59,6 @@
Stack_Chunk_t *chunk;
item_size += offsetof(Stack_Chunk_t, u.data);
- item_size += 7;
- item_size &= ~7; /* round up to 8 so that the chunk is aligned at
- the same size - the aligned MMX memcpy needs it */
make_bufferlike_pool(interp, item_size);
chunk = (Stack_Chunk_t *)new_bufferlike_header(interp, item_size);
chunk->prev = chunk; /* mark the top of the stack */
> As the past tense is indicating, above dummy is totally obsolete and the copy
> functions[2] are unused. The latter could be removed too, but might be a
> nice-to-have, when something faster than memcpy(3) is wanted or needed,
> albeit only on restricted (but common) platforms and for aligned memory only.
> Some recent CPUs do have special instructions for copying unaligned memory
> reegions too.
Yes, indeed. I could see a future optimization pairing memalign(3C) with
specialized copy functions.
Thanks again,
--
Andy Dougherty [EMAIL PROTECTED]
