Usually addresses for variables are only known after the linking stage. - Too late for the compiler to notice and generate more efficient code for aligned access.

If instead the address of a variable is known at compile time (here by the use of the 'at' keyword) the compiler is given a chance to obmit testing the 'low address overflow':

#include <8052.h>

unsigned char xdata at 0xfe00 buffer[200];

void f(void)
{
  unsigned char i;
  for(i=0; i!=sizeof buffer; i++)
     P3 = buffer[i];
}

should compile to something like:

   0000                     338 _f:
                            347 ;       buffer.c:8: for(i=0; i!=sizeof buffer; i++)
   0000 7A 00               348         mov     r2,#0x00
   0002                     349 00101$:
   0002 BA C8 01            350         cjne    r2,#0xC8,00110$
   0005 22                  351         ret
   0006                     352 00110$:
                            353 ;       buffer.c:9: P3 = buffer[i];
   0006 8A 82               354         mov     dpl,r2
   0008 75 83 FE            355         mov     dph,#(_buffer >> 8)
   000B E0                  356         movx    a,@dptr
   000C F5 B0               357         mov     _P3,a
   000E 0A                  359         inc     r2
   000F 80 F1               360         sjmp    00101$

with SDCC more recent than december 2003.