Re: PATCH: Support SHF_GNU_COMPRESSED

[Alan Modra <amodra at gmail dot com>]

On Tue, Jul 20, 2010 at 04:38:56PM -0700, H.J. Lu wrote:
> SHF_GNU_COMPRESSED isn't strictly required. I can peek into
> .debug_* sections like gdb does. On the other hand, SHF_GNU_COMPRESSED
> can avoid the peek. I can prepare a patch without SHF_GNU_COMPRESSED.
> Any comments on that?

A linker that doesn't know about compressed sections can't possibly
apply relocations for them correctly, so I think we should be setting
SHF_OS_NONCONFORMING on these sections.  I didn't have a preference
for SHF_GNU_COMPRESSED or SHT_GNU_COMPRESSED until I noticed that GNU
ld doesn't emit a diagnostic for SHF_OS_NONCONFORMING with unknown
flags (which is a bug), but does do so for unknown types.  So I favour
SHT_GNU_COMPRESSED.  That way an older GNU ld will sensibly give up if
presented a new compressed debug object file.

Now, some comments on the patch.

> 	* bfd.c (BFD_COMPRESS): New.
> 	(BFD_DECOMPRESS): Likewise.

Is there a reason why we shouldn't always decompress?  I think that
would make handling these sections so much easier.  Oh, I see you are
making objdump -s dump out the compressed contents.  I wonder if it
wouldn't be more useful to dump out the contents after decompression
anyway?

> +  unsigned int compress_status : 2;

If you always decompress, then you probably don't need this.

> +  /* The compressed size of the section in octets.  */
> +  bfd_size_type compressed_size;

Seems to me this is unnecessary.  For an input section, "rawsize"
should be the on-disk size, with "size" that after decompression.

-- 
Alan Modra
Australia Development Lab, IBM