https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128468
Bug ID: 128468
Summary: Missed phi-opt optimization
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: law at gcc dot gnu.org
Blocks: 26163
Target Milestone: ---
This is teeny-tiny in x264 and while the motivating case is supposed to be
pretty standard, I don't see that it'd likely be that common in encoders.
Perhaps this would make a good easyhack as a learning experience?
Consider:
int
f (int v)
{
int t = v * 2;
return t > 1 ? t : 1 - t;
}
That's equivalent to max (t, 1-t)
Right now we have this in gimple:
t_3 = v_2(D) * 2;
if (t_3 <= 1)
goto <bb 3>; [41.00%]
else
goto <bb 4>; [59.00%]
;; succ: 3
;; 4
;; basic block 3, loop depth 0
;; pred: 2
_4 = 1 - t_3;
;; succ: 4
;; basic block 4, loop depth 0
;; pred: 2
;; 3
# _1 = PHI <t_3(2), _4(3)>
So looks like a fairly straightforward phi-opt transformation.
Currently we get something like this on rv64gcbv_zicond:
slliw a4,a0,1
li a5,1
subw a0,a5,a4
sgt t0,a4,a5
czero.nez t1,a0,t0
czero.eqz t2,a4,t0
add a0,t2,t1
ret
Clearly the RTL if-converter caught this case. We'd like to get this:
slliw a4,a0,1
li a5,1
subw a0,a5,a4
max a0,a4,a0
ret
While I think we could possibly to it in ifcvt.cc, phiopt seems like a better
place given how simple the case looks in gimple.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=26163
[Bug 26163] [meta-bug] missed optimization in SPEC (2026, 2k17, 2k and 2k6 and
95)
