Date: Mon, 15 Mar 2004 17:48:08 +0900 From: Alexander Nedotsukov <bland@FreeBSD.org> To: ticso@cicely.de Cc: alpha@FreeBSD.org Subject: Re: Is the way static data alligned correct? Message-ID: <40556DC8.2060900@FreeBSD.org> In-Reply-To: <20040315072848.GW55325@cicely12.cicely.de> References: <405554D9.4000208@FreeBSD.org> <20040315072848.GW55325@cicely12.cicely.de>
next in thread | previous in thread | raw e-mail | index | archive | help
Bernd Walter wrote:
>On Mon, Mar 15, 2004 at 04:01:45PM +0900, Alexander Nedotsukov wrote:
>
>
>>Guys,
>>
>>Recently doing some development I hit alignment problem on alpha
>>machines. Problem solved but the way I did it in fact the reason why I
>>asking pros to judge who is wrong here.
>>Inside code I have linked-in data blocks generated by external tool.
>>They looks like this:
>> const unsigned char raw_data_block[] = { bla-bla-bla };
>>Internal block content is 8 bytes aligned. Each block resides in a
>>single .c file. So after compilation I have a set of .o files linked
>>later into single shared object (library). Taking an address of such
>>data block and casting it to pointer to some real structure was a
>>problem because it is not necesary 8 byte aligned. Obvious solution was
>>to ask explicitly for such alignment:
>> const unsigned char raw_data_block[] __attribute__ ((alligned(8))) =
>>{ bla-bla-bla };
>>But I wonder is this gcc bug or space optimization for char arrays only?
>>
>>
>
>You request an array of bytes and you get the natural alignment of
>bytes - why should the compiler waste bytes that are not required for
>the variable you define?
>
>
Mmm... It was my pure speculation around the fact that Alpha can read
memory only from addreses divisible by eight or raise an alignment
error. So natural byte alignment was not so natural for Alpha in my
understanding. Btw on stack (automatic variables) where space savings
may have more sence than in .data segment char[] allocations are always
8 bytes aligned.
But I sholud admit that my expirience in alpha programming not much
bigger than beast uptime :-)
>A portable way would be to define it as a union with an int64_t.
>
>But the real point is that you shouldn't access any variable with
>stricter alignment than they are.
>If you have const data as in your example then define it as the
>correct structure.
>
>
Thanks for the portability tip. I think originally code was written this
way to avoid differences in generated binary file supposed to be shared
across different platforms and linked-in parts. I'll try to do things
right in next code review.
>If it's not constant data then bcopy it into your structure or
>directly read it in a malloc'ed buffer.
>
>
>
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?40556DC8.2060900>