TotK rewrites per-object uniform blocks on other threads soon after it submits the frame that reads them, and the
replay records for milliseconds after the submit call, so late draws could read another object's matrix (phantom
rocks, a flat far field). The guard in nvn_api_core.c obs_release_read only knows 1.2.1's producer. runtime/vk_usnap.c
copies every uniform range of imported pools that the submission binds into an arena, after MacNativeGames'
TOTK_UBO_SNAPSHOT, and ubo_info binds the copy; ranges a submission still in flight writes are left out, and GPU writes
within the submission drop the copies they touch. TOTK_UBO_SNAPSHOT=0 binds live memory; =ab alternates.
Town on an RX 9070 (in-launch A/B, 70 s): 2235 ranges, 1.9 MiB per submission in 0.49 ms; interval +0.02 ms, replay
+0.19 ms. The sort of the ranges was half of the cost with qsort (0.53 ms for about 10,000 bindings) and is a radix
sort now. fight2 interval within noise. The four probes passed on 1.4.3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The generator emits a guest poll such as `ldr w9, [x19]; cmp w9, #1; b.eq` as a plain C load, and clang hoists it out
of the loop: the lock wait in f_7001fa0e2c compiled to `cmp; jne; jmp .`, so a thread that found the lock held spun
forever after its holder released it. A fight run hung this way in 3 of 13 launches on an RX 9070, with the UBO
snapshot on and off; gdb showed the thread on that `jmp` in every sample. The gate is an opaque call on every
back-edge of the 868 loops tools/spin_waits.py lists, so each pass reloads memory: the exe goes from 182 self-jumps to
148, and 144 of those are the game's own `b .` instructions. 8 of 8 fight runs then passed; tests/diff/func.py with
--patches passed 5470 cases on 878 functions, and the four probes passed on 1.4.3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The lifted decoder ran its copy loops as chains of small functions. On the island and fight routes TOTK_ZSTD_CHECK=1 decoded every call both ways with no difference; the host took 0.72 s where the guest decoder took 1.99 s. Load time and fps on an RX 9070 and a 7900X did not change: the gain is CPU time on the loading threads, for machines with fewer cores. The caller at main+0x547a04 stays on the guest decoder: it resets a static DCtx with which the game's zstd rejects some frames libzstd decodes, and the game relies on that result.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The FP helpers compared c->fpcr with a cached copy before every operation to resync MXCSR lazily. MXCSR now
follows FPCR.FZ at once: MSR FPCR calls rt_fpcr_sync, and so do rt_call (thread, fiber and host entry into guest
code), fiber resume (FIBER_FLAG_FLOAT_SWITCH fibers keep their own MXCSR) and the differential test harness per
case. op and vop only test FPCR.RMode for the slow path, and their redo check is one ordered-not-equal compare of
|r| with the smallest normal, the same NaN-or-smallest-normal set as before.
Hot guest code shrinks by about a quarter (132 functions holding half the guest cycles, 1.4.3); fight frame time
moves 1-3%, inside the run-to-run spread. lift, simd, func, atomic and flags_check pass on 1.4.3; simd.py fails
without the harness sync.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With TOTK_TRACE=sched, TOTK_CRIT=<lo>-<hi> records every state change and wake of the guest threads and every
present over those frames (runtime/kernel.c crit_add from acct() and wake(), crit_frame at the present) and writes
them to TOTK_CRIT_OUT at the last frame. tools/crit_path.py walks back from each present through the waits to the
thread whose signal ended them and splits the frame interval by thread and state; the sums equal the interval.
A 1.4.3 fight: with the replay on the submitting thread the path is 7.3 ms of replay on the render thread and 4.3 ms
of guest code; with TOTK_REPLAY_ASYNC=1 it is 10.1 ms of guest code through the fork-join of the main thread and five
workers, at the same frame interval.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every lifted instruction read and wrote c->x[n] and c->sp, so each register passed from one guest instruction to the
next went through a store and a load, and clang could not keep it in a host register across a guest store.
recomp/regs.py now copies the six most used X registers and SP of a function into locals at its entry. A forward pass
over the function's flow tracks which locals changed since the last sync, and only those are stored back before a
call, an SVC, a trap, a return or a tail exit. After an SVC every local reloads; after a call only X0-X18 and X30,
since AAPCS64 has the callee preserve X19-X29 and SP. TotK does not break that: main and subsdk0 import neither
longjmp nor the unwinder, sdk's have no direct callers, and the runtime writes no register above X3 except X30.
gen.py --no-callee-saved-trust reloads them too, --reg-locals sets the count. Functions with an unresolved computed br
keep their registers in the Cpu, as they keep NZCV.
All registers in locals was slower than none: clang spills them to the stack and the syncs stay. Over 300 hot
functions, 4-6 locals gave the fewest instructions and memory operands.
1.4.3 .text 299 -> 266 MiB. Fight bench (bench-fight2-2026-10-27b, frames 5400-7500, uncapped): 11.30 and 11.92 ms vs
13.5-13.9 on main; frames 2400-5300 11.24 and 11.36 vs 13.25. Without the callee-saved trust 12.58 ms.
tests/diff/func.py: 0 failing on 1.4.3 (889 functions, 583 with calls) and 1.2.1 (810); a plan with no spills fails
at once. Probes kernel, vkexec, nvn-tex and shader pass in both versions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every flag-setting instruction packed N, Z, C and V into c->nzcv and every reader unpacked them. recomp/flags.py
now runs a reaching-definitions pass per function: a compare stores its operands in a local Flags F, and a reader
whose definitions are all of one kind computes its condition from them, so cmp and b.cond become one host compare.
CCMP, FCMP, MSR NZCV and ADCS write the packed form, which readers reached by several kinds also use. Calls and
returns publish nothing (AAPCS64 neither passes nor preserves NZCV); tail branches, falls into the next function and
br through rt_dispatch publish c->nzcv, since discovery can start a function inside another one. Functions with an
unresolved computed br keep flags in c->nzcv.
1.4.3 main: 492k direct and 89k packed flag reads, .text 342 -> 299 MiB. Fight bench (bench-fight2-2026-10-27b,
frames 5400-7500, uncapped) 13.15 and 13.63 ms vs 13.9 before. tests/flags_check.py checks the new helpers against
the packed ones; tests/diff/func.py no longer compares NZCV, which generated code leaves untouched.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
memo_match compared the vertex and fragment stages' 64 texture and 64 image handles on every draw. Every OP_BIND_TEX and now every OP_BIND_IMG bumps tex_serial; a key whose serial is current has the same handles, and a key that matched by the compare takes the current serial. memo_match went from 0.95% to 0.27% of the fight replay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A fight frame of about 2,000 draws made 15 parts on the 7900X at 64 draws a part, each split costing the render thread a submit, a new command buffer and a mirror prologue (vk_sync_split, 7.6% of the replay). maybe_split now also waits until TOTK_SPLIT_US (1000, 0 off) has passed since the part began, so a frame makes 7-9 parts there. A millisecond of recording is well under a millisecond of GPU work, so the GPU still starts early in the frame.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
vk_exec_pass_begin created every attachment view again in each submission and freed them at its fence: 2% of the fight replay on the 7900X, more on drivers with a slower vkCreateImageView. view_of now takes the view from vk_exec_target_view, which keeps it in the sampled-view tables under the texture lock, and vk_exec_image_gone releases it with the image. Images that are not registered textures keep the per-submission views.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The left stick now moves the menu cursor as the D-pad does (ImGui only scrolled with it). The page's rows are a
direct NavFlattened child of the menu window: behind a second child level ImGui did not clip them, so up from
Continue went to the tabs and down from the tabs to Continue.
The full overlay reads temperature, board power, clock and GPU load from ADL's PM log on AMD GPUs (matched by PCI
bus), and shows the window, game image and 3D scene sizes, the scaling to the window (FSR and its sharpness), the
frame limit and the presentation mode in use.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.2.1 renders the scene size it reads, while 1.4.3 scales it by output width / 1920: on 1.2.1 the scene set is on
whenever the output grows (2560x1440 gave a 1600x900 scene before, 2133x1199 now). An output at or below the shipped
size needs no grown HeapDef. The release deps carry the static libzstd and its header, totk.ini.example the four new
keys, THIRD_PARTY_NOTICES.md the MacNativeGames sites, FSR 1 and zstd.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The output and scene patch sets make the game read its output, swapchain and scene sizes from the patch page
(MacNativeGames' sites for 1.2.1 and 1.4.3); resolution.c picks them from TOTK_OUTPUT_RES and TOTK_RENDER_SCALE and
grows the HeapDef texture heap, its parent and the application pool to match. vk_fsr.c runs AMD FSR 1 EASU and RCAS
on the present when the window is larger than the game's output (TOTK_UPSCALER, TOTK_SHARPNESS).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The MacNativeGames BKTR and title-key changes and our UTF-8 manifest lived only in the maintainer's toolchain folder, so
nobody else could build the hactool that extraction and the installer need. The script's exe matches the one in use
byte for byte apart from its build time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/building.md merges the two toolchain docs and the README build steps, and lists the developer checks. The
Maxwell reference moves to docs/maxwell.md. architecture.md loses the milestone table, controls.md describes timed
script lines and input recording, and the tester guide includes shader preparation.
The agent skills, hooks, playbook, backlog and the study notes on upstream projects are the maintainer's local setup:
.claude/, notes/ and CLAUDE.local.md are ignored now, and tools/worktree.py links them into task worktrees. AGENTS.md
keeps the project rules.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comments named the agent task (T1-T7, M2-M5, Task 2.x) that once owned each file or block, and pointed at local
notes the repository does not carry (backlog items, milestone plans, research notes). They now say what the code
does. gfx_stubs.c held weak stand-ins for functions that every build has defined for a long time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>