ARM memcpy on unaligned pointers and -O2

[Shaun Jackman <sjackman at pathwayconnect dot com>]

I believe I've found a bug in GCC 3.0.2 wrt the ARM target and -O2 
optimisation. I found it when I wrote the following functions to load and 
store unaligned pointers. This is the objdump -S output.

/**
 * unaligned load
 */
uint32
ld32( const uint32* p)
{
        uint32 ret;
        memcpy( &ret, p, sizeof( ret));
  90:   e5900000        ldr     r0, [r0]
        return ret;
  94:   e1a0f00e        mov     pc, lr
}

/**
 * unaligned store
 */
void
str32( uint32* p, uint32 val)
{
        memcpy( p, &val, sizeof( *p));
  98:   e5801000        str     r1, [r0]
  9c:   e1a0f00e        mov     pc, lr
}


As you can see this did not have the intended effect. GCC recognised the 
length of the memcpy as being constant and optimised it away to a simple ldr 
or str instruction. What it doesn't seem to have taken into account though, 
is the fact that the pointer it's passed may not be aligned. I believe it's 
allowed to pass unaligned pointers to memcpy and the compiler should 
handle it correctly.

Please cc me as I'm not a member of this mailing list,
Cheers,
Shaun