https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97282
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
We have the divmod discovery in tree-ssa-math-opts.c and it will handle this
case if the division/modulo is by the same variable. But for constants it
punts:
/* Disable the transform if either is a constant, since division-by-constant
may have specialized expansion. */
if (CONSTANT_CLASS_P (op1) || CONSTANT_CLASS_P (op2))
return false;
So, perhaps we want to get it through if CONSTANT_CLASS_P (op2) (but not op1)
and
some further conditions are met, e.g. that op2 is not a power of two, and
either the type's precision is > HOST_BITS_PER_WIDE_INT (then expand_divmod
punts on it), as choose_multiplier etc. work on HOST_WIDE_INTs), or compute
choose_multiplier for the constant divisor and if pre or post shift needs to be
BITS_PER_WORD or more bits.
Or optimize into DIVMOD always even for constant op2 and when trying to expand
DIVMOD ifn with constant op2, see if the target can expand division by that
constant using just shifts and +/- and if so, emit it as that division +
subtraction, otherwise throw away the division expansion and emit a divmod.
