Use a pointer for PC instead of an instruction counter

[pguyot]

Also mark code in modules as const.
This optimization brings an additional 20% speed upgrade to the Sudoku
benchmark.

These changes are made under both the "Apache 2.0" and the "GNU Lesser General
Public License 2.1 or later" license terms (dual license).

SPDX-License-Identifier: Apache-2.0 OR LGPL-2.1-or-later

[bettio]

I found out that for some reason this floating point test is way slower:

-module(pi).
-export([run/0, pi/4, ok/0]).

ok() ->
    ok.

run() ->
    pi(4, 3, -1, 1000000).

pi(Value, _N, _Sign, 0) ->
    Value;
pi(Value, N, Sign, Iterations) ->
    pi(Value + Sign * (4 / N), N + 2, -Sign, Iterations - 1).

[pguyot]

I found out that for some reason this floating point test is way slower:

I added the test to atomvm_benchmark in pguyot/atomvm_benchmark#3.

However, I don't get your result on macOS with OTP26.

paul@yuzu ~/P/A/A/build (master)> ./src/AtomVM ../../atomvm_benchmark/_build/default/lib/benchmark.avm libs/atomvmlib.avm
Running tests:
pingpong_speed_test: 2298446
prime_speed_test: 106290
prng_test: 624476
pi_test: 681115
sudoku_solution_test: 167670
sudoku_puzzle_test: 6324736
pingpong_speed_test [schedulers=1]: 175687
prime_speed_test [schedulers=1]: 373233
Return value: ok

paul@yuzu ~/P/A/A/build (w34/use-pc-pointer-instead-of-i)> ./src/AtomVM ../../atomvm_benchmark/_build/default/lib/benchmark.avm libs/atomvmlib.avm
Running tests:
pingpong_speed_test: 2260228
prime_speed_test: 90409
prng_test: 608688
pi_test: 627082
sudoku_solution_test: 156944
sudoku_puzzle_test: 6008711
pingpong_speed_test [schedulers=1]: 174490
prime_speed_test [schedulers=1]: 296484
Return value: ok

Everything was compiled with OTP26. Which platform and which OTP version did you test it with?

[pguyot]

The results depend on the SDK/gcc versions.

sdk 5.1

with PR:
pi_test: 119369657

with master:
pi_test: 199435054

sdk 4.4

with PR:
pi_test: 129949250

with master:
pi_test: 82786098

[pguyot]

On top of #805, this brings a significant speed increase (38%) to pi_test benchmark on ESP32 with sdk 4.4

[bettio]

For future reference I'm posting some benchmarks and some further information that can be useful for further optimizations.

The following are normalized scores (lower is better here too):

So there is a general improvement, even if the issue with prime test (here attached) should be investigated.

prime.zip

The .S file has been generated using OTP 24.