On Thu, Jul 20, 2017 at 11:22:10 -1000, Richard Henderson wrote:
> >Perhaps we should then enlarge both the first and last regions so that we
> >fully use the buffer.
>
> I really like the idea. That's a lot of space recovered for 64k page hosts.
>
> I do think we can make the computation clearer. How about
(snip)
>
> static inline void tcg_region_bounds(TCGContext *s, size_t curr_region,
> void **pstart, void **pend)
> {
> void *start, *end;
>
> /* ??? Maybe store "aligned" precomputed. */
> start = QEMU_ALIGN_PTR_UP(region.start, qemu_real_host_page_size);
> /* ??? Maybe store "stride" precomputed. */
> start += curr_region * (region.size + qemu_real_host_page_size);
> end = start + region.size;
>
> if (curr_region == 0) {
> start = region.start;
> }
> if (curr_region == region.n - 1) {
> end = region.end;
> }
>
> *pstart = start;
> *pend = end;
> }
That's a nice helper -- will do it this way.
For v4, should I send all patches again, or just the handful of patches
that are changing from v3?
E.
