On 2/26/19 3:38 AM, David Hildenbrand wrote:
> We sometimes want to work on a temporary vector register instead of the
> actual destination, because source and destination might overlap. An
> alternative would be loading the vector into two i64 variables, but than
> separate handling for accessing the vector elements would be needed.
> This is easier. Add one for now as that seems to be enough.
Hmm, I'll reserve judgment until I see how this is used.
For ARM SVE, I would allocate this temporary on the stack within the helper,
and move one of the operands out of the way. E.g.
void helper(foo)(void *vd, void *vx, *void *vy
{
VectorReg tmp;
TYPE *d = vd, *x = vx, *y = vy;
if (vx == vd || vy == vd) {
tmp = *(VectorReg *)vd;
if (vx == vd) {
vx = &tmp;
}
if (vy == vd) {
vy = &tmp;
}
}
process d, x, y as normal.
}
This minimized the amount of code inline. However, SVE vectors are quite a bit
larger, at 256 bytes, so the copy itself was out of line most of the time
anyway.
Provisionally,
Reviewed-by: Richard Henderson <richard.hender...@linaro.org>
r~
