Re: [Pnet-developers] ILTypeIdentical

[Rhys Weatherley]

On Monday 25 October 2004 05:47 pm, Simon Posnjak wrote:

> The problem I am seeing is that the memory allocated by malloc is
> aligned, but the layout of the struct in the allocated memory is not
> aligned. The structs are *packed* on CRIS and using union and struct
> makes it impossible to make all fields of that struct/union to align to
> 4 (or 2^n). So to fix this it will not help if I change the malloc,
> because this problem is related to how CRIS gcc does struct/union
> layout.

It should only be necessary for the start of a malloc'ed region to be aligned.  
The ILType pointers refer to the start of a structure, not something in the 
middle.  The alignment of other fields shouldn't matter.

I think however I am beginning to see the true problem.  ILType structures are 
normally allocated from a memory pool, which allocates structures on a 
multiple of IL_BEST_ALIGNMENT (defined in "include/il_align.h").

Pnet tries to automagically determine the best alignment by measuring 
structure padding.  If, as you say, gcc packs everything with a sub-4 
alignment value, then this could cause IL_BEST_ALIGNMENT to end up being 1 or 
2 or something similarly small.

This is easy to fix.  Add the following to the end of "il_align.h":

#ifdef cris
#undef IL_BEST_ALIGNMENT
#define IL_BEST_ALIGNMENT 4
#endif

where "cris" is whatever macro you need to use to detect your CPU.

Cheers,

Rhys.