qi-hardware/wernermisc
1
0
Fork 0
mirror of git://projects.qi-hardware.com/wernermisc.git synced 2024-11-15 11:20:17 +02:00
Code

m1/perf/sched.c: tried and rejected more accurate LCPF distance heuristics

Browse Source 
They turned out to produce the same or even slightly worse results.
This commit is contained in:
Werner Almesberger 2011-09-24 20:20:04 -03:00
parent 9db83ae72c
commit bb8c1ce741
1 changed files with 22 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
m1/perf/sched.c
View File

@ -423,9 +423,31 @@ static void init_scheduler(struct fpvm_fragment *frag)
for (i = frag->ninstructions-1; i >= 0; i--) {
insn = sc->insns+i;
#if 0
/*
* Theoretically, we should consider the distance through
* write-write dependencies too. In practice, this would
* mainly matter if we had operations whose result is ignored.
* This is a degenerate case that's probably not worth
* spending much effort on.
*/
if (insn->next_setter) {
insn->distance =
insn->next_setter->distance-insn->distance+1;
if (insn->distance < 1)
insn->distance = 1;
}
#endif
foreach (dep, &insn->dependants)
if (dep->insn->distance > insn->distance)
insn->distance = dep->insn->distance;
/*
* While it wold be more correct to add one for the cycle
* following the write cycle, this also has the effect of
* producing slighly worse results on the example set of
* patches. Let's thus keep this "bug" for now.
*/
// insn->distance += insn->latency+1;
insn->distance += insn->latency;
}
#endif