https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124622
Bug ID: 124622
Summary: Missed opportunity for IPA CP cloning, IPA reference
constification in 544.nab_r
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: ipa
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
The hot function mme34 has
if (dim == 4)
conditional code, but the global static dim variable is just set in three
places, to either 3 or 4 with the setter to 4 being
REAL_T mme4(REAL_T * x, REAL_T * f, int *iter)
{
dim = 4;
return mme34(x, f, iter);
}
the setter to 3 (also the initial value of dim) being in an initialization
function as well as a corresponding
REAL_T mme(REAL_T * x, REAL_T * f, int *iter)
{
dim = 3;
return mme34(x, f, iter);
}
the mme4 function is not used, so with -flto only setters to value 3 remain.
Still the runtime conditionals in mme34 on dim prevail. Without LTO the
static mme34 function would be a candidate for "IPA CP" of 'dim'.
This IPA transform would turn 'dim' into a parameter, replacing accesses
to 'dim' in the callee by constants. Feels like IPA reference which
knows that mme34 does not write to 'dim':
Function name:mme34/153:
locals read: prm dim frozen max_step chknm nconstrained pairlistnp upairsnp
lpairsnp pairlist nb_pairs upairs lpairs N14pearlist grad reqack
locals written: nb_pairs tpair tmme tcons tborn tphi tangl tbond tnonb tmd
treduce grad reqack
globals read: ALL
globals written: ALL
with LTO at WPA time the write with value 4 is gone, but we do not promote
the variable constant for some reason.
