https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112566
Bug ID: 112566
Summary: Some ctz/popcount/parity/ffs optimizations
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
Target Milestone: ---
I believe
ctz(ext(x)) == ctz(x) if UB on zero,
popcount(zext(x)) == popcount(x),
parity(zext(x)) == parity(x),
parity(sext(x)) == parity(x) if the extension is by even number of bits and
ffs(ext(x)) == ffs(x).
So, e.g. with x86 -O2 -m32 -mbmi2 -mlzcnt -mpopcnt
int foo (unsigned int x) { return __builtin_ctzll (x); }
int bar (unsigned int x) { return __builtin_popcountll (x); }
int baz (unsigned int x) { return __builtin_parityll (x); }
int qux (int x) { return __builtin_ffsll (x); }
int corge (int x) { return __builtin_ctzll (x); }
int garply (int x) { return __builtin_parityll (x); }
int fred (unsigned int x) { return __builtin_ffsll (x); }
shouldn't use any double-word bit query, or similarly
int foo (unsigned _BitInt(256) x) { return __builtin_ctzg ((unsigned
_BitInt(512)) x); }
int bar (unsigned _BitInt(256) x) { return __builtin_popcountg ((unsigned
_BitInt(512)) x); }
int baz (unsigned _BitInt(256) x) { return __builtin_parityg ((unsigned
_BitInt(512)) x); }
int qux (_BitInt(256) x) { return __builtin_ffsg ((_BitInt(512)) x); }
int corge (_BitInt(256) x) { return __builtin_ctzg ((unsigned _BitInt(512)) x);
}
int garply (_BitInt(256) x) { return __builtin_parityg ((unsigned _BitInt(512))
x); }
int fred (unsigned _BitInt(256) x) { return __builtin_ffsg ((_BitInt(512)) x);
}
Of course, we shouldn't do this if we deoptimize some supported precision into
an unsupported narrower one.
For clz(zext(x)) = clz(x)+difference_in_precision, but at that point it might
not be a win.
