[uae] Re: faster blitfunc
- From: Mustafa TUFAN <mustafa.tufan@xxxxxxxxx>
- To: uae@xxxxxxxxxxxxx
- Date: Fri, 30 Mar 2007 22:02:10 +0300
Hi,
i don't know about other compilers but gcc (for arm and mips) does very
good job optimizing (when -O3 or -Os is used)..
but the patch i've done here is very simple but effective one..
for (j = b->vblitsize; j--; ) { ... }
here the test contition is j-- and it's just a; is j non-zero, decrement
and loop..
for (j = 0; j < b->vblitsize; j++) { ... }
but here it has to subtract j from b->vblitsize and check if it's not
zero, increment and loop..
more like;
AMIGA:
move.l b->vblitsize,d0
.mylop:
....
dbf d0,.mylop
PC:
mov ecx, b->vblitsize
.mylop:
...
loop .mylop
VS.
AMIGA:
move.l #0,d0
.mylop:
....
add.l #1,d0
cmp b->vblitsize,d0
bne .mylop
PC:
mov ecx,0
.mylop:
...
add ecx,1
cmp ecx,b->vblitsize
bne .mylop
my other patch was to remove unused srcc (uae_u32 srcc = b->bltcdat),
normally compiler warns about unused variables
but the thing is e.g:
void blitdofast_0 {
uae_u32 srcc = b->bltcdat; // srcc is loaded
..
b->bltcdat = srcc; //and back..
}
the srcc value does not change here.. and because of the (b->bltcdat =
srcc) line, compiler thinks this variable is used and leaves it there..
these may not make a speed difference on high end pcs,
but it makes a lot of difference for 266 mhz psp and 200 mhz gp2x.
-GnoStiC/bronx.
Toni Wilen wrote:
genblitter diff for wip4.. this should be even faster..
Shouldn't it be C-compiler's job to do these kinds of optimizations?
(I thought all modern compilers handle these easily)
Other related posts:
