Multiplying by a constant

[David Brown]

This is just an idea for an improvement in the ColdFire backend.

When multiplying by a constant, the compiler generates a series of 
shifts and adds (or subs) instead of using the appropriate "mul" 
instruction.  While this is optimal when the shift and add combination 
is small, it is sub-optimal if more than 3 instructions are needed 
(multiplies take 4 clocks on a v2 Coldfire).  This is a hang-over from 
the pre-Coldfire m68k cores which had very slow multiplies.

Test code:

extern unsigned int x, y;
void foo(void) {
   x = y * 11;
}

Generated code (-O2 -Wa,-ahlsd -m528x):

   92               	foo:
   93 008e 4E56 0000 		link.w %fp,#0
   94 0092 2079 0000 		move.l y,%a0	| y, y.6
   94      0000
   95 0098 2008      		move.l %a0,%d0	| y.6, y.6
   96 009a 2200      		move.l %d0,%d1	| y.6, tmp33
   97 009c E589      		lsl.l #2,%d1	|, tmp33
   98 009e 2001      		move.l %d1,%d0	| tmp33, tmp34
   99 00a0 E588      		lsl.l #2,%d0	|, tmp34
  100 00a2 9081      		sub.l %d1,%d0	| tmp33, tmp34
  101 00a4 9088      		sub.l %a0,%d0	| y.6, D.1159
  102 00a6 23C0 0000 		move.l %d0,x	| D.1159, x
  102      0000
  103 00ac 4E5E      		unlk %fp
  104 00ae 4E75      		rts


mvh.,

David