https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107591
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
To answer my own question:
int
foo (int x)
{
return x + x;
}
int
bar (int x)
{
return x * x * x * x * x * x;
}
float
baz (float x)
{
return x + x;
}
float
qux (float x)
{
return x * x * x * x * x * x;
}
float
corge (float x)
{
return x * x;
}
match.pd or fold-const.cc canonicalizes x + x to x * 2, so no need to special
case that, x * x isn't canonicalized for integral types nor without -ffast-math
for floats either, with -ffast-math reassoc1 canonicalizes the multiplications
into __builtin_powif (x, 6) resp. __builtin_powif (x, 2) until powcabs pass
later on splits those into multiplications again.
So, for frange, if we implement good range op for CASE_CFN_POWI, we could just
use that, perhaps for VREL_EQ on mult.
