housekeeping

  and gets close to gcc's. Too short (256 B) is useless, too long (4 kB) worse

  again.

- On the same machine with the `powersave` profile, a prefetch 1–4 kB ahead gave

  +20–30% (review of 2026-10-02). There the core slows down its clock while it

  +20–30% . There the core slows down its clock while it

  waits for memory, and a loop that issues its loads late loses more. Always

  report the power profile with such numbers.

prefetchers do the rest. Explicit prefetching is for irregular patterns, such

as `prefetching/ex_1` (binary search) and `prefetching/ex_3` (linked list).

### What changed in the code (2026-10-04)

- Fixed (all the versions printed wrong sums for some N):

  - v2 lost the last group of 4;

  - v3 and v4 started from the second group (elements 1–4 were never summed),

    and read past the end;

  - v5 and v6 read past the end, added an extra vector when N/4 was odd, and

    lost the final N % 8 elements;

  - v1 had an unused load.

- v1b was a standalone program with the old PAPI API (removed in PAPI 6); it is

  now a kernel that keeps the unused pre-load on purpose.

- New v7: a real software prefetch, with the distance as a parameter, and its

  control without prefetch.

- `prefetching2.vall.c`:

  - allocation sizes rounded to the alignment;

  - the kernel file is given by the Makefile (clang could not build the old

    `#include` trick);

  - each version prints its own label;

  - the transfer rate counts the bytes read once, in GB (10^9), and prints the %

    of a theoretical peak only with `-D_MEM_CLOCK=...` (it compared with

    hard-coded DDR3 numbers);

  - the cache wipe reads instead of writing, so no dirty lines are written back

    while the next timed loop runs.

- The same fixes in `prefetching2.v0.c`.

- `register` arrays removed (clang refuses them). New `Makefile`; `compile` calls

  it.

## Exercises