Bug#642762: xulrunner-6.0: console flooded with unaligned access messages on ia64 (IA-64/IPF/Itanium) platform

[Mike Hommey]

On Wed, Oct 05, 2011 at 10:50:56AM +0200, Émeric Maschino wrote:
> 2011/10/4 Mike Hommey <mh at glandium.org>:
> > On Tue, Oct 04, 2011 at 10:20:16PM +0200, >
> > m0 is a 32-bit value, it needs 32-bits alignment. m1 is 16-bits, 16-bits
> > alignment, etc. Overall, the struct only requires 32-bits alignment.
> 
> Reading http://en.wikipedia.org/wiki/Data_structure_alignment, I
> thought that gcc was padding the internal data of a structure so that
> they are aligned. Furthermore, isn't nsID internal data nevertheless
> packed on 16 bytes after compilation (it is 4+2+2+8 = 16 bytes before
> compilation)?
> 
> If "explicit" 64-bit alignment is still required, do we need to add
> dummy padding data to align nsID internal data? Such as (with
> alignment recommendations taken from
> http://software.intel.com/en-us/articles/data-alignment-when-migrating-to-64-bit-intel-architecture/):
> 
> struct nsID {
>   PRUint32 m0;
>   PRUint16 m1;
>   PRUint16 padding[1];  // 2 bytes for m2 to be aligned on a 4-byte boundary (*)
>   PRUint16 m2;
>   PRUint8 m3[8];
>   PRUint8 padding[6];  // 6 bytes to make total size of nsID 24 bytes
> };

The problem is not the alignment of m2, the problem is the alignment of
the whole struct, which has a requirement of 32-bits. Which means a
struct nsID can end up at 0x0, 0x4, 0x8, or 0xc. When it's at 0x4 or
0xc, it can't be casted to a 64-bits word, because that's not 64-bits
aligned. There are two solutions: make sure struct nsIDs are 64-bits
aligned, or change the Equals function to use 2 32 bits words
comparisons.

Mike