Hookshot/Hookshot

Watch

Fork 10

perf(build): a renderer header edit compiles the files that read it, not the whole back end #2706

Open

github-sync wants to merge 2 commits from github/pr/2706 into main

github-sync commented

2026-11-12 08:30:50 +00:00

(Migrated from git.magicbean.dev)

@jordanella opened this pull request on GitHub · 2026-11-12 07:43 UTC · original

Editing any renderer header recompiles the whole back end: on Windows a comment in vk_present_rect.h, which two files include, compiled 35 files again (417 s of compiler time; a 44 s build against 26 s for a .cpp edit, and minutes on a 4–8 thread PC). This is the dev-loop proposal's A6 (#2636), validated by hckrman101's Linux numbers there (16 objects for one header).

Change

Depfiles: the back end's own files (gpu/vk/*.cpp) are compiled with -MD, and clang writes a depfile beside each object (vk_x.d).
Narrower keys: the next build keys each object by the headers its depfile lists under the same folders as before (gpu/vk, gpu/render, gpu/compat, suyu's texture decoders), by name and content (renderer_deps). A header the file doesn't read no longer changes its key; one it reads, or one that's gone, does.
The blob stays as the fallback: for a file with no depfile yet, or one whose depfile names nothing under those folders.
No forced recompiles: flag_text leaves -MD out, so no key changes and nothing compiles again for this change itself. Depfiles appear as files are compiled.
Re-keying: an object compiled now is cached and stamped under the key its new depfile gives (rekey_depfiles), so the next build doesn't compile it once more. Only an object of this build, so made from this build's source and headers, is re-keyed.
Cached by its own depfile only: an object whose planned key came from the last compile's depfile is neither cached nor stamped under that key (it hashes the new source with the old include list, so it doesn't vouch for a header read for the first time). compile_one stamps it depfile-pending, and rekey_depfiles stamps and caches it by its own depfile's key, also when that's the planned key. A build that stops before re-keying leaves it pending, and the next build compiles it again. (hckrman101's review: an include added, removed, the header edited, the include back, used to get the first object back from the cache.)
Unchanged: Eden's and sirit's files keep their own headers' blob, and the host keeps its blob (it mixes in non-header inputs such as the FFmpeg version key, so it's left for a separate change).
Same limit as make/ninja: a newly added header that shadows another earlier in the include path isn't seen until something the file reads changes.

Measured

On Windows (Ryzen 9 3900X, NVMe):

Build	Files compiled	Compiler time	Build
one comment in `vk_present_rect.h`, before	35	417 s	44 s
the same, with depfiles	4	20 s	27 s
the header reverted	0	0 s	18 s

The four are the files that read it: vk_framegen.cpp, vk_menu.cpp, vk_replay.cpp, and host/hos_main.c, through the host's own blob. Getting depfiles took one compile of each back-end file, done here by commenting them all. The re-keying then kept the reverted files from compiling a third time.

In the game

Windows 11, RTX 3080 (NVIDIA 621.14), from this branch's build.py (with the review fix) on main ee29dd55: a build with depfiles, then one with a comment added to vk_present_rect.h (5 files compiled: the four above, and vk_hangdump.cpp, which reads no repo header, so its depfile names nothing under those folders and it keeps the blob; no object left pending; with the comment taken out again, 0 files). That game, launched headless through the control bridge from a save: in play in Kokiri Forest, 27–34 fps over five samples, nothing in the log about a lost device or a crash.

Testing

tests/test_renderer_depfiles.py (new, 11 cases):
- the depfile read as clang 23 writes it on Windows (/ and \ in paths with spaces, CRLF continuations), plus $$ and \#;
- -MD is in no key;
- no depfile keeps the blob;
- an unread header or one outside the folders changes nothing, while a read header or a gone one changes the key;
- the re-keying, and that it leaves alone an object not of this build;
- the wiring;
- hckrman101's four steps, with a stand-in compiler that writes the headers it read into the object and a real depfile: on the previous commit step 4 got back step 2's object, built with the old b.h.
It fails against main's build.py.
test_compiler_failure.py and test_mem_gate_leases.py take DEPFILE_DEPS and DEPFILE_PENDING from build.py with compile_one.
test_dlss_sdk_cache.py and test_shader_stamp_cache.py run the job loop with its new helpers. Both still take an older build.py (--build-script), and test_shader_stamp_cache passes against main's.
Every check that reads build.py (39), on Windows, passes apart from test_driver_cache_folder, test_fpx_sticky and test_linux_update, which fail the same way on main here.
Not tested: Linux. gcc and clang write the same make-style depfile.

🤖 Generated with Claude Code

> [@jordanella](https://github.com/jordanella) opened this pull request on GitHub · 2026-11-12 07:43 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2706) Editing any renderer header recompiles the whole back end: on Windows a comment in `vk_present_rect.h`, which two files include, compiled 35 files again (417 s of compiler time; a 44 s build against 26 s for a `.cpp` edit, and minutes on a 4–8 thread PC). This is the dev-loop proposal's A6 (#2636), validated by hckrman101's Linux numbers there (16 objects for one header). ## Change - **Depfiles:** the back end's own files (`gpu/vk/*.cpp`) are compiled with `-MD`, and clang writes a depfile beside each object (`vk_x.d`). - **Narrower keys:** the next build keys each object by the headers its depfile lists **under the same folders as before** (`gpu/vk`, `gpu/render`, `gpu/compat`, suyu's texture decoders), by name and content (`renderer_deps`). A header the file doesn't read no longer changes its key; one it reads, or one that's gone, does. - **The blob stays as the fallback:** for a file with no depfile yet, or one whose depfile names nothing under those folders. - **No forced recompiles:** `flag_text` leaves `-MD` out, so no key changes and nothing compiles again for this change itself. Depfiles appear as files are compiled. - **Re-keying:** an object compiled now is cached and stamped under the key its new depfile gives (`rekey_depfiles`), so the next build doesn't compile it once more. Only an object of this build, so made from this build's source and headers, is re-keyed. - **Cached by its own depfile only:** an object whose planned key came from the *last* compile's depfile is neither cached nor stamped under that key (it hashes the new source with the old include list, so it doesn't vouch for a header read for the first time). `compile_one` stamps it `depfile-pending`, and `rekey_depfiles` stamps and caches it by its own depfile's key, also when that's the planned key. A build that stops before re-keying leaves it pending, and the next build compiles it again. (hckrman101's review: an include added, removed, the header edited, the include back, used to get the first object back from the cache.) - **Unchanged:** Eden's and sirit's files keep their own headers' blob, and the host keeps its blob (it mixes in non-header inputs such as the FFmpeg version key, so it's left for a separate change). - **Same limit as make/ninja:** a newly added header that shadows another earlier in the include path isn't seen until something the file reads changes. ## Measured On Windows (Ryzen 9 3900X, NVMe): | Build | Files compiled | Compiler time | Build | |---|---|---|---| | one comment in `vk_present_rect.h`, before | 35 | 417 s | 44 s | | the same, with depfiles | **4** | **20 s** | **27 s** | | the header reverted | 0 | 0 s | 18 s | The four are the files that read it: `vk_framegen.cpp`, `vk_menu.cpp`, `vk_replay.cpp`, and `host/hos_main.c`, through the host's own blob. Getting depfiles took one compile of each back-end file, done here by commenting them all. The re-keying then kept the reverted files from compiling a third time. ## In the game Windows 11, RTX 3080 (NVIDIA 621.14), from this branch's `build.py` (with the review fix) on `main` ee29dd55: a build with depfiles, then one with a comment added to `vk_present_rect.h` (5 files compiled: the four above, and `vk_hangdump.cpp`, which reads no repo header, so its depfile names nothing under those folders and it keeps the blob; no object left pending; with the comment taken out again, 0 files). That game, launched headless through the control bridge from a save: in play in Kokiri Forest, 27–34 fps over five samples, nothing in the log about a lost device or a crash. ## Testing - **`tests/test_renderer_depfiles.py` (new, 11 cases):** - the depfile read as clang 23 writes it on Windows (`/` and `\ ` in paths with spaces, CRLF continuations), plus `$$` and `\#`; - `-MD` is in no key; - no depfile keeps the blob; - an unread header or one outside the folders changes nothing, while a read header or a gone one changes the key; - the re-keying, and that it leaves alone an object not of this build; - the wiring; - hckrman101's four steps, with a stand-in compiler that writes the headers it read into the object and a real depfile: on the previous commit step 4 got back step 2's object, built with the old `b.h`. It fails against `main`'s `build.py`. - **`test_compiler_failure.py` and `test_mem_gate_leases.py`** take `DEPFILE_DEPS` and `DEPFILE_PENDING` from `build.py` with `compile_one`. - **`test_dlss_sdk_cache.py` and `test_shader_stamp_cache.py`** run the job loop with its new helpers. Both still take an older `build.py` (`--build-script`), and `test_shader_stamp_cache` passes against `main`'s. - **Every check that reads `build.py` (39), on Windows,** passes apart from `test_driver_cache_folder`, `test_fpx_sticky` and `test_linux_update`, which fail the same way on `main` here. - **Not tested:** Linux. gcc and clang write the same make-style depfile. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

github-sync commented

2026-11-12 09:33:05 +00:00

(Migrated from git.magicbean.dev)

@hckrman101 commented on nx2/build.py line 889 on GitHub · 2026-11-12 09:24 UTC · original

This adds the correct key, but compile_one has already cached the object under the planned key, which for a depfile job came from the previous compile's depfile (the new source hashed with the old include list). That entry is what a later build can hit with a stale header (steps in the review body). Doing the cache_put here even when key == old, and not under the planned key in compile_one for DEPFILE_JOBS objects, would keep only sound keys in the cache.

> [@hckrman101](https://github.com/hckrman101) commented on `nx2/build.py` line 889 on GitHub · 2026-11-12 09:24 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2706#discussion_r4548556488) This adds the correct key, but `compile_one` has already cached the object under the planned key, which for a depfile job came from the previous compile's depfile (the new source hashed with the old include list). That entry is what a later build can hit with a stale header (steps in the review body). Doing the `cache_put` here even when `key == old`, and not under the planned key in `compile_one` for `DEPFILE_JOBS` objects, would keep only sound keys in the cache.

github-sync commented

2026-11-12 09:33:06 +00:00

(Migrated from git.magicbean.dev)

@hckrman101 reviewed on GitHub · 2026-11-12 09:24 UTC · original

Thanks, this is a nice win and the numbers in the description are exactly what I wanted to see (same edit before/after, the hardware named). The depfile parser, the -MD exclusion from the key, the blob fallback and the "only re-key an object of this build" guard all read right to me, and CI is green.

One correctness problem before I can approve: a back-end object can come back from the object cache built against an old header.

compile_one still caches every compiled object under its planned key (cache_put(obj, key) at the end, not in this diff). For a depfile job that planned key comes from the previous compile's depfile, so it hashes the new source with the old include list. That key is lossy: it doesn't cover a header the new source reads for the first time. rekey_depfiles adds the correct key, but the lossy one stays in the cache, and a later build can land on it:

vk_x.cpp includes a.h; build. Its depfile says {a.h}.
Add #include "b.h" to vk_x.cpp; build. Planned key = (new source, {a.h}). Compiled with b.h v1 and cached under that key; re-keyed under (new source, {a.h, b.h}).
Take the include back out (or switch to another revision without it); build. It compiles again, and the depfile is back to {a.h}.
Edit b.h, then put the include back (undo/redo, or jj edit back to the revision that has it); build. The key is (new source, {a.h}) again, so it's a cache hit on step 2's object, still built with b.h v1. A cache hit doesn't write a depfile, so rekey_depfiles computes the same key and the stale object stays until a.h or the source changes.

Switching between revisions where one adds an include is an everyday thing, and a stale struct layout here means a crash or corruption in the game, not a build error, so I'd rather close this hole. The simplest fix I see: for a DEPFILE_JOBS object, only ever cache it under the key its own depfile gives. Skip the cache_put under the planned key in compile_one when that key came from a depfile (a blob key is fine, since it covers every header) and let rekey_depfiles do the put, also when the key didn't change. A lookup can then only hit an object whose depfile lists exactly the headers being hashed, and that's sound. A case in test_renderer_depfiles.py that walks through the steps above would pin it down.

Smaller things:

-MD (and so the depfiles) only goes to the back end's own files, but host/hos_main.c still goes through the host's blob, as the description says. That's fine for this PR.
On testing: the build measurements cover the build side. Since the risk here is a wrong object rather than a slow one, please also say in the description that the game built this way starts and runs (one launch after the header edit is enough).

> [@hckrman101](https://github.com/hckrman101) reviewed on GitHub · 2026-11-12 09:24 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2706#pullrequestreview-5872022957) Thanks, this is a nice win and the numbers in the description are exactly what I wanted to see (same edit before/after, the hardware named). The depfile parser, the `-MD` exclusion from the key, the blob fallback and the "only re-key an object of this build" guard all read right to me, and CI is green. One correctness problem before I can approve: a back-end object can come back from the object cache built against an old header. `compile_one` still caches every compiled object under its *planned* key (`cache_put(obj, key)` at the end, not in this diff). For a depfile job that planned key comes from the **previous** compile's depfile, so it hashes the new source with the old include list. That key is lossy: it doesn't cover a header the new source reads for the first time. `rekey_depfiles` adds the correct key, but the lossy one stays in the cache, and a later build can land on it: 1. `vk_x.cpp` includes `a.h`; build. Its depfile says `{a.h}`. 2. Add `#include "b.h"` to `vk_x.cpp`; build. Planned key = (new source, `{a.h}`). Compiled with `b.h` v1 and cached under that key; re-keyed under (new source, `{a.h, b.h}`). 3. Take the include back out (or switch to another revision without it); build. It compiles again, and the depfile is back to `{a.h}`. 4. Edit `b.h`, then put the include back (undo/redo, or `jj edit` back to the revision that has it); build. The key is (new source, `{a.h}`) again, so it's a cache hit on step 2's object, still built with `b.h` v1. A cache hit doesn't write a depfile, so `rekey_depfiles` computes the same key and the stale object stays until `a.h` or the source changes. Switching between revisions where one adds an include is an everyday thing, and a stale struct layout here means a crash or corruption in the game, not a build error, so I'd rather close this hole. The simplest fix I see: for a `DEPFILE_JOBS` object, only ever cache it under the key its **own** depfile gives. Skip the `cache_put` under the planned key in `compile_one` when that key came from a depfile (a blob key is fine, since it covers every header) and let `rekey_depfiles` do the put, also when the key didn't change. A lookup can then only hit an object whose depfile lists exactly the headers being hashed, and that's sound. A case in `test_renderer_depfiles.py` that walks through the steps above would pin it down. Smaller things: - `-MD` (and so the depfiles) only goes to the back end's own files, but `host/hos_main.c` still goes through the host's blob, as the description says. That's fine for this PR. - On testing: the build measurements cover the build side. Since the risk here is a wrong object rather than a slow one, please also say in the description that the game built this way starts and runs (one launch after the header edit is enough).

github-sync commented

2026-11-12 12:37:29 +00:00

(Migrated from git.magicbean.dev)

@jordanella commented on GitHub · 2026-11-12 11:41 UTC · original

Thanks: that's a real hole, and your steps reproduce it exactly. Fixed in f0f345b6:

No caching under the planned key. compile_one no longer caches, or stamps, an object under a key that came from the last compile's depfile. It stamps it depfile-pending instead.
rekey_depfiles does the keying. It stamps and caches the object by the key its own new depfile gives, also when that's the planned key. The cache now only ever holds blob keys (every header) or a compile's own depfile key.
A stamp-only variant of the same hole. The old code also stamped the object with the lossy key. A build that stopped before re-keying (some other file failing) could leave that stamp for a later build whose depfile gives that key. A pending stamp never matches, so that build compiles the file again.
test_renderer_depfiles.py walks your four steps with a stand-in compiler that writes its included headers into the "object" and a real depfile. On the previous commit it fails: step 4 got back step 2's object, built with the old b.h.

On running the game: with this branch's build.py (the fix included) on Windows 11 and an RTX 3080, I built once to get the depfiles, then built again after adding a comment to vk_present_rect.h. That compiled the four files plus vk_hangdump.cpp (no repo headers, so it keeps the blob), and left no object pending. The game from that build, launched headless from a save, plays in Kokiri Forest at 27-34 fps with nothing in the log about a lost device or a crash. With the comment taken out, the next build compiled nothing. That's in the description now.

test_compiler_failure.py and test_mem_gate_leases.py also cut compile_one out, so they now take the two new constants from build.py as well.

> [@jordanella](https://github.com/jordanella) commented on GitHub · 2026-11-12 11:41 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2706#issuecomment-6822627288) Thanks: that's a real hole, and your steps reproduce it exactly. Fixed in f0f345b6: - **No caching under the planned key.** `compile_one` no longer caches, or stamps, an object under a key that came from the last compile's depfile. It stamps it `depfile-pending` instead. - **`rekey_depfiles` does the keying.** It stamps and caches the object by the key its *own* new depfile gives, also when that's the planned key. The cache now only ever holds blob keys (every header) or a compile's own depfile key. - **A stamp-only variant of the same hole.** The old code also stamped the object with the lossy key. A build that stopped before re-keying (some other file failing) could leave that stamp for a later build whose depfile gives that key. A pending stamp never matches, so that build compiles the file again. - **`test_renderer_depfiles.py`** walks your four steps with a stand-in compiler that writes its included headers into the "object" and a real depfile. On the previous commit it fails: step 4 got back step 2's object, built with the old `b.h`. On running the game: with this branch's `build.py` (the fix included) on Windows 11 and an RTX 3080, I built once to get the depfiles, then built again after adding a comment to `vk_present_rect.h`. That compiled the four files plus `vk_hangdump.cpp` (no repo headers, so it keeps the blob), and left no object pending. The game from that build, launched headless from a save, plays in Kokiri Forest at 27-34 fps with nothing in the log about a lost device or a crash. With the comment taken out, the next build compiled nothing. That's in the description now. `test_compiler_failure.py` and `test_mem_gate_leases.py` also cut `compile_one` out, so they now take the two new constants from `build.py` as well.

github-sync commented

2026-11-12 12:37:29 +00:00

(Migrated from git.magicbean.dev)

@jordanella commented on nx2/build.py line 889 on GitHub · 2026-11-12 11:42 UTC · original

Done in f0f345b6: compile_one stamps a depfile-keyed object depfile-pending and caches nothing, and rekey_depfiles does the cache_put under the key its own depfile gives, also when key == old. Details and the four-step test are in https://git.magicbean.dev/MagicBeanDev/Hookshot/pulls/2706.

> [@jordanella](https://github.com/jordanella) commented on `nx2/build.py` line 889 on GitHub · 2026-11-12 11:42 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2706#discussion_r4549943428) Done in f0f345b6: `compile_one` stamps a depfile-keyed object `depfile-pending` and caches nothing, and `rekey_depfiles` does the `cache_put` under the key its own depfile gives, also when `key == old`. Details and the four-step test are in https://git.magicbean.dev/MagicBeanDev/Hookshot/pulls/2706.

github-sync commented

2026-11-12 14:44:25 +00:00

(Migrated from git.magicbean.dev)

@Joey-Magnus approved on GitHub · 2026-11-12 14:34 UTC · original

hckrman101's stale-object hole is closed in f0f345b6. compile_one stamps an object it compiled under a depfile key depfile-pending and doesn't cache it (nx2/build.py:952). rekey_depfiles then caches and stamps it under the key from its own new depfile, so the cache only holds blob keys or a compile's own depfile key. I traced the cache-hit path (compile_one restoring from OBJ_CACHE). A hit needs the planned key to equal a key computed from that object's own depfile, so the source and the set of hashed headers match, and leaving the old .d on disk is sound. A build that fails before rekey_depfiles leaves pending stamps, which never match, so those files compile again. I merged this into current main and ran the depfile, compile, mem_gate, dlss_sdk, shader_stamp and build checks: 31 passed. test_precompile_compute_guard and test_portable_local_builder also fail on main on this machine, so they're not from this PR. The in-game run after the header edit is what was asked for. The red appimage job timed out in the PPA apt step, which is infrastructure, so a re-run should clear it.

> [@Joey-Magnus](https://github.com/Joey-Magnus) approved on GitHub · 2026-11-12 14:34 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2706#pullrequestreview-5876236083) hckrman101's stale-object hole is closed in f0f345b6. `compile_one` stamps an object it compiled under a depfile key `depfile-pending` and doesn't cache it (`nx2/build.py:952`). `rekey_depfiles` then caches and stamps it under the key from its own new depfile, so the cache only holds blob keys or a compile's own depfile key. I traced the cache-hit path (`compile_one` restoring from `OBJ_CACHE`). A hit needs the planned key to equal a key computed from that object's own depfile, so the source and the set of hashed headers match, and leaving the old `.d` on disk is sound. A build that fails before `rekey_depfiles` leaves pending stamps, which never match, so those files compile again. I merged this into current main and ran the depfile, compile, mem_gate, dlss_sdk, shader_stamp and build checks: 31 passed. `test_precompile_compute_guard` and `test_portable_local_builder` also fail on main on this machine, so they're not from this PR. The in-game run after the header edit is what was asked for. The red appimage job timed out in the PPA apt step, which is infrastructure, so a re-run should clear it.

github-sync commented

2026-11-12 16:55:49 +00:00

(Migrated from git.magicbean.dev)

@mooki000 commented on GitHub · 2026-11-12 16:43 UTC · original

Linux result on f0f345b6, against its main base 46780503: the identical header edit recompiles 5 files instead of 35, and both resulting games load Kokiri Forest and exit cleanly.

Setup: Arch Linux, kernel 7.2.15, i9-13900K, RTX 4090, NVIDIA 615.78.08, Clang 23.1.3 release builds with the same PGO profile and compatible translated inputs. Separate checkouts and game folders. After the initial builds, I added the same harmless comment to gpu/vk/vk_present_rect.h in each checkout and rebuilt through nx2/build.py.

Build	Files actually compiled after the edit
Base `46780503`	35
PR `f0f345b6`	5

The five were vk_framegen.cpp, vk_menu.cpp, vk_replay.cpp, host/hos_main.c, and vk_hangdump.cpp. This matches your explanation: the host still uses its header group, and vk_hangdump.cpp has no repository-header dependency and keeps the fallback. Counts come from the build's own @@file ... new records, asserted by script. Other work was running, so I make no wall-time speed claim.

The stale-cache review fix also passes with real compilers. A separate check drives the actual cache/key/compile/rekey functions, adds an include, removes it, edits that header, and adds it back. It links and runs each resulting object. GCC 16.2 and Clang 23.1.3 both return the new value 50 at step 4 on f0f345b6; both return the stale value 30 on 632b55ac. This does not use the stand-in compiler from the PR's test.

Game check after the header edit: windowless, fresh copied save, 2048×1152 game output / 1920×1080 offscreen presentation, 60 cap, DLSS/frame generation off. Base: 37.9 seconds, 1,505 game-present rows, exit 0. PR: 34.3 seconds, 1,454 rows, exit 0. Both show Link, the forest and HUD; neither logs a crash, lost device or kernel driver fault. Load averages were 36.37→34.31 and 13.85→13.41, respectively: these are not frame-rate comparisons. The experimental comments were removed afterward; the retained tested binaries are explicitly the post-comment builds.

All six changed checks pass in both compiler environments. Against base source, the new depfile check reports one flag-key failure and ten missing-helper errors; the compile-failure/memory-gate adaptations reject missing constants. The existing compile-retry, DLSS-cache and shader-stamp checks pass on base too, as compatibility coverage. Full suite on the head: 473 passed, 0 failed, 20 skipped.

The hosted AppImage job separately timed out in System packages, exit 124, before compiling this change: https://github.com/NotRexed/Hookshot/actions/runs/43211349705/job/131159467435.

Not tested: Windows runtime, all possible include-path shadowing changes, other game scenes, or quiet repeated build timings. No new defect found in the exercised head.

> [@mooki000](https://github.com/mooki000) commented on GitHub · 2026-11-12 16:43 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2706#issuecomment-6829816712) Linux result on `f0f345b6`, against its main base `46780503`: the identical header edit recompiles **5 files instead of 35**, and both resulting games load Kokiri Forest and exit cleanly. **Setup:** Arch Linux, kernel 7.2.15, i9-13900K, RTX 4090, NVIDIA 615.78.08, Clang 23.1.3 release builds with the same PGO profile and compatible translated inputs. Separate checkouts and game folders. After the initial builds, I added the same harmless comment to `gpu/vk/vk_present_rect.h` in each checkout and rebuilt through `nx2/build.py`. | Build | Files actually compiled after the edit | |---|---:| | Base `46780503` | 35 | | PR `f0f345b6` | 5 | The five were `vk_framegen.cpp`, `vk_menu.cpp`, `vk_replay.cpp`, `host/hos_main.c`, and `vk_hangdump.cpp`. This matches your explanation: the host still uses its header group, and `vk_hangdump.cpp` has no repository-header dependency and keeps the fallback. Counts come from the build's own `@@file ... new` records, asserted by script. Other work was running, so I make no wall-time speed claim. **The stale-cache review fix also passes with real compilers.** A separate check drives the actual cache/key/compile/rekey functions, adds an include, removes it, edits that header, and adds it back. It links and runs each resulting object. GCC 16.2 and Clang 23.1.3 both return the new value **50** at step 4 on `f0f345b6`; both return the stale value **30** on `632b55ac`. This does not use the stand-in compiler from the PR's test. **Game check after the header edit:** windowless, fresh copied save, 2048×1152 game output / 1920×1080 offscreen presentation, 60 cap, DLSS/frame generation off. Base: 37.9 seconds, 1,505 game-present rows, exit 0. PR: 34.3 seconds, 1,454 rows, exit 0. Both show Link, the forest and HUD; neither logs a crash, lost device or kernel driver fault. Load averages were 36.37→34.31 and 13.85→13.41, respectively: these are not frame-rate comparisons. The experimental comments were removed afterward; the retained tested binaries are explicitly the post-comment builds. All six changed checks pass in both compiler environments. Against base source, the new depfile check reports one flag-key failure and ten missing-helper errors; the compile-failure/memory-gate adaptations reject missing constants. The existing compile-retry, DLSS-cache and shader-stamp checks pass on base too, as compatibility coverage. Full suite on the head: **473 passed, 0 failed, 20 skipped**. ![Kokiri Forest after the same header edit: base and this pull request](https://gist.githubusercontent.com/mooki000/b86bd3f35969803fb032f71f035844f8/raw/67d20c0bce7f3135e063d3f1c0a6651784d50843/hookshot-pr2706-linux-after-header-edit.jpg) The hosted AppImage job separately timed out in **System packages**, exit 124, before compiling this change: https://github.com/NotRexed/Hookshot/actions/runs/43211349705/job/131159467435. Not tested: Windows runtime, all possible include-path shadowing changes, other game scenes, or quiet repeated build timings. No new defect found in the exercised head.

notrexed referenced this pull request from a commit

2026-11-13 22:36:48 +00:00

perf(build): skip unchanged links, write compile_commands.json, key the hot pieces by the profile's contents (#2893)

This pull request has changes conflicting with the target branch.

nx2/build.py

View command line instructions

Manual merge helper

Use this merge commit message when completing the merge manually.

Merge commit title

Merge pull request 'perf(build): a renderer header edit compiles the files that read it, not the whole back end' (#2706) from github/pr/2706 into main

Merge commit body

Reviewed-on: https://git.notrexed.dev/Hookshot/Hookshot/pulls/2706

Merge pull request 'perf(build): a renderer header edit compiles the files that read it, not the whole back end' (#2706) from github/pr/2706 into main

Reviewed-on: https://git.notrexed.dev/Hookshot/Hookshot/pulls/2706

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u origin 'github/pr/2706:github/pr/2706'

git switch 'github/pr/2706'

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch 'main'

git merge --no-ff 'github/pr/2706'

git switch 'github/pr/2706'

git rebase 'main'

git switch 'main'

git merge --ff-only 'github/pr/2706'

git switch 'github/pr/2706'

git rebase 'main'

git switch 'main'

git merge --no-ff 'github/pr/2706'

git switch 'main'

git merge --squash 'github/pr/2706'