Re: [sdcc-devel] About peephole rules of stm8

[Philipp K. K. <pk...@sp...>]

Am 12.02.2014 10:57, schrieb Ben Shi:
> Hello,
> 
> I find a redundant stm8 code serial, that with the following c code
> 
> int ss(int s)
> {
>         s += 1;
>         return s + 0x2030;
> }
> 
> the following asms are generated
> 
> ;       a.c: 2: int ss(int s)
> ;       -----------------------------------------
> ;        function ss
> ;       -----------------------------------------
> _ss:
> ;       a.c: 4: s += 1;
>         ldw     x, (0x03, sp)
>         incw    x
>         ldw     (0x03, sp), x
> ;       a.c: 5: return s + 0x2030;
>         ldw     x, (0x03, sp)
>         addw    x, #0x2030
>         ret
> 
> I thougth the "ldw x, (0x03, sp)" is unnecessary here, can there be
> adding a rule to the peeph.def? like
> 
> replace restart {
>     ldw %1, %2
>     ldw %2, %1
> } by {
>     ldw %1, %2
> }
> 
> Ben
> 

That rules eliminates a write of %2 and a read of %1. We may only do
this if neither is volatile. To be correct, your peephole needs a condition:

if notVolatile(%1), notVolatile(%2)

Unfortunately, notVolatile() is currently a very rough guess. This will
result in the rule probably not being applied in your case (but the rule
would help if s was global instead of local, since notVolatile() seems
to work better for global than for local variables). See RFE #407 for a
suggestion on how to improve notVolatile().

Philipp

P.S.: Why does the stm8 port even use s again for the result of s += 1,
instead of a iTemp. The z80 port does the same as the stm8, but the hc08
port seems to be more efficient.