> Hmm, ISTR we had such mitigations in place (or have) elsewhere keying
> on the most significant bit set instead of power-of-two.
fold_plusminus_mult_expr only factors out for power-of-two:
if (exact_log2 (absu_hwi (int11)) > 0 && int01 % int11 == 0
/* The remainder should not be a constant, otherwise we
end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
increased the number of multiplications necessary. */
&& TREE_CODE (arg10) != INTEGER_CST)
> But your case
> likely recurses and runs into the extract_multiv limiting to eventually
> stop, even for (N + 4) * 8, right?
Yes, it oscillates between extract_multiv and fold_plusminus_mult_expr until
reaching the maximal depth.
> If so shouldn't we prevent this even for !TYPE_OVERFLOW_WRAPS? Also
>
> + && !(tree_fits_shwi_p (c)
> + && exact_log2 (absu_hwi (tree_to_shwi (c))) > 0))
The code only distributes for TYPE_OVERFLOW_WRAPS though:
/* The last case is if we are a multiply. In that case, we can
apply the distributive law to commute the multiply and addition
if the multiplication of the constants doesn't overflow
and overflow is defined. With undefined overflow
op0 * c might overflow, while (op0 + orig_op1) * c doesn't. */
if (code == MULT_EXPR && TYPE_OVERFLOW_WRAPS (ctype))
return fold_build2 (tcode, ctype,
fold_build2 (code, ctype,
fold_convert (ctype, op0),
fold_convert (ctype, c)),
op1);
> is better written as
>
> && exact_log2 (wi::to_wide (c)) > 0
>
> not sure why the sizetype constant for you fits in a signed HWI
> or you need to compute its absolute value. Eventually you
> need to use wi::abs(wide_int::from (wi::to_wide (c), TYPE_PRECISION
> (TREE_TYPE (c)), SIGNED))
> or so.
This is just mirrored on what fold_plusminus_mult_expr does.
--
Eric Botcazou
