On 8/1/25 23:30, Richard Henderson wrote:
On 1/8/25 14:14, Philippe Mathieu-Daudé wrote:
static void gen_bsl1n_vec(unsigned vece, TCGv_vec d, TCGv_vec n,
TCGv_vec m, TCGv_vec k)
{
- if (TCG_TARGET_HAS_bitsel_vec) {
- tcg_gen_not_vec(vece, n, n);
- tcg_gen_bitsel_vec(vece, d, k, n, m);
- } else {
Why aren't we doing the NOT n operation here?
- tcg_gen_andc_vec(vece, n, k, n);
[*]
- tcg_gen_andc_vec(vece, m, m, k);
- tcg_gen_or_vec(vece, d, n, m);
- }
+ tcg_gen_not_vec(vece, n, n);
+ tcg_gen_bitsel_vec(vece, d, k, n, m);
Pardon? It's right there, unindented.
Sorry I'm not clear. Previous to your change, in the
TCG_TARGET_HAS_bitsel_vec side we use the NOT opcode,
but not in the other side where we expand, why?
Are you asking about the code being removed?
Recall that bitsel = (n & k) | (m & ~k).
Passing n = ~n' we get (~n & k) | (m & ~k),
= (k & ~n) | (m & ~k).
which is the two andc + or operations above.
Sorry, I misread the first ANDC [*] as AND...
Reviewed-by: Philippe Mathieu-Daudé <phi...@linaro.org>
