Pointer typecasts may write to incorrect addresses on ARM?

[John Breitenbach]

Subramanyam Badri wrote:
> Hi,
>  
> I am using a GCC compiler for ARM and see an alignment problem with pointer typecasting.  We're porting code to ARM and found some writes to incorrect addresses.  Can you help?  For example:
>  
> char * pChar = (char *)0x100000;  /* hardcoded for example purposes */
> pChar[0] = 0x0;
> *(int *)&pChar[1] = 0x11223344;
>  
> /* We really want pChar[0] to equal 0x00, but it equals 0x44 on ARM!!!! */
> printf ("pChar[0] = 0x%02x\n", pChar[0]);
>
> My understanding is that non-aligned memory access is allowed on ARM, so it's not clear why typecasting and writing data adjusts the destination address. 
>
>   
Your understanding isn't the same as mine - I would expect the cpu to 
either trap on the unaligned access, or possibly just ignore the 
misalignment.

If you're *lucky* enough to be running under linux, you can echo 3 to 
/proc/cpu/alignment to enable the alignment fault handler.  IIRC, its 
default is to do the wrong thing, leaving partial memory updates like 
you are seeing.

John