Menu
▾
▴
#985 Z80 an optimization of some simple returns
The following code:
unsigned char t, x, y, z;
unsigned char f( void );
unsigned char g( unsigned char c );
unsigned char my( void ) {
t = 4;
return f()
&& g( x )
&& g( y )
&& g( z );
}
produces using:
SDCC : mcs51/z80/z180/r2k/r2ka/r3ka/sm83/tlcs90/ez80_z80/z80n/r800/ds390/pic16/pic14/TININative/ds400/hc08/s08/stm8/pdk13/pdk14/pdk15/mos6502/mos65c02/f8 TD- 4.5.0 #15240 (MINGW64)
sdcc -mz80 --opt-code-speed --no-c-code-in-asm --max-allocs-per-node20000000 -c -o zout/check-speed.rel check.c
the following code:
_myf::
call _f
or a, a
jr Z, 00103$
ld a, (_x)
call _g
or a, a
jr Z, 00103$
ld a, (_y)
call _g
or a, a
jr Z, 00103$
ld a, (_z)
call _g
or a, a
jr NZ, 00104$
00103$:
xor a, a
ret
00104$:
ld a, #0x01
ret
but the functionally equivalent, but clumsier:
unsigned char myf2( void ) {
if ( !f() )
return 0;
if ( !g( x ) )
return 0;
if ( !g( y ) )
return 0;
if ( !g( z ) )
return 0;
return 1;
}
surprisingly, already knows how to do ret Z when A is already zero. (!)
_myf2::
call _f
or a,a
ret Z
ld a, (_x)
call _g
or a,a
ret Z
ld a, (_y)
call _g
or a,a
ret Z
ld a, (_z)
call _g
or a,a
ret Z
ld a, #0x01
ret
In the starting myf the only used peephole rule in the conditional jumps was just:
; common peephole 163 changed absolute to relative conditional jump.
and in the myf2 there were four (!):
or a,a
; common peephole 154a removed redundant zeroing of a (which has just been tested to be #0x00).
;ic:7: __iffalse_0($2) :
; genLabel
ret Z
; common peephole 161 replaced jump by return.
; common peephole 81 removed jp by using inverse jump logic
; common peephole 159 removed unused label 00102$.
And how it looks like with --no-peep:
call _g
; genMove_o size 1 result type 2 source type 2 hl_dead 1
;ic:10: if iTemp1 [k6 lr8:9 so:0]{ ia0 a2p0 re0 rm0 nos0 ru0 dp0}{unsigned-char fixed}[a ] != 0 goto __iffalse_1($4)
; genIfx
or a, a
jp NZ, 00104$
;check.c:40: return 0;
;ic:13: ret 0x0 {unsigned-char literal}
; genRet
; genMove_o size 1 result type 2 source type 1 hl_dead 1
xor a, a
jp 00109$
;ic:14: __iffalse_1($4) :
; genLabel
00104$:
where behind 00109$ is just a ret.
My guess is that for such an optimization to always happen, it would have to happen in some earlier stages.
