Using four abilities instead of two means the C++ checking_tag
mechanism needs to handle multiple values, or needs to be supported
by recursion counting.
The branching test lets one level of recursion finish, and then tries
to go deeper. This tests for bugs where a recursion detection tool in
the engine gets its count reset when exiting one level of recursion.
This adds a recursion counter which runs on a per-weapon basis; if the
recursion limit is reached then the last level of recursion is assumed
to have returned false. A warning is printed (with error severity).
At the user-visible layer, that gives the following logic:
* Abilities that depend on another ability being active will be
inactive.
* Pairs of abilities, where X depends on Y being inactive and Y
depends on X being inactive, will end up with them both inactive.
The limit is defined by ATTACK_RECURSION_LIMIT in attack_type.cpp.
The minimum for passing all the unit tests is 2, but I've left some
headroom by setting it to 4.
With the recursion limit set to 1, the following tests fail, which
seems reasonable because they're checking that the special is on
a particular type of weapon.
* event_test_filter_attack_specials
* event_test_filter_attack_opponent_weapon_condition
* event_test_filter_attack_student_weapon_condition
Output from the four_cycle_recursion_branching test:
```
error unit: Recursion limit reached for weapon 'melee_attack' while checking filter 'special_type_active = parry'
error unit: Recursion limit reached for weapon 'melee_attack' while checking filter 'special_type_active = poison'
error unit: Recursion limit reached for weapon 'melee_attack' while checking filter 'special_type_active = damage'
```
There's deduplication to prevent the logfiles getting spammed during
AI turns. As implemented, the two tests mentioned below print 3
messages each; these tests are added in separate commits because the
tests are shared between PRs while we discuss the right mechanism for
guarding against recursion.
With no rate limit:
* four_cycle_recursion_branching prints 1280 logs
* event_test_filter_attack_student_weapon_condition prints 320
With a rate limit via a flag that's reset in recursion_guard's destructor:
* four_cycle_recursion_branching prints 80 logs
* event_test_filter_attack_student_weapon_condition prints 160
Resetting the flag in specials_context_t's destructor gets better numbers,
but splits the logic across .cpp files. I think the trade-off isn't worth it:
* four_cycle_recursion_branching prints 40 logs
* event_test_filter_attack_student_weapon_condition prints 132
Co-authored-by: newfrenchy83 <eric.belleux@gmail.com>
* Update xBRZ to v1.8 (resolves#8307)
This replaces our implementation with stock xBRZ sans any of our custom code. Made these changes from plain source:
- Renaming .h -> .hpp
- Conditionally use [[likely]]
- Added a trailing newline
- Comment out unused parameters
- Disable `-Wunused-function` warning on macOS
There are no callers in mainline. The functions are not meant to be called directly at all – for each, there are wrapper functions that are intended to be called instead, which are still exposed.
