diff --git a/docs/patches-v110.md b/docs/patches-v110.md index 66b20c05..06846667 100644 --- a/docs/patches-v110.md +++ b/docs/patches-v110.md @@ -13,8 +13,10 @@ The built-in table is [recomp/patches-v110.txt](../recomp/patches-v110.txt). The | pooloff | 2 | Widen texture-pool offsets unsigned, covering 2..4 GiB without changing offsets below 2 GiB. Enabled by default; `TOTK_POOL_OFFSET_PATCH=0` retains the original path. | | shadow | 1 | The sun shadow map takes its fixed size from page +92 before rounding. The shadow-resolution setting enables it; `shadow_res=game` leaves the game's sizing. | | cascade | 4 | Read the four sun shadow cascade defaults from page +64..76. Shadow range 100 leaves the original constructor. | +| climb | 1 | Climbing speed above 60 fps: the 1.2.1 climb factor in the 1.1.0 climb controller. Follows `TOTK_CLIMB_FIX` and needs the real-time step. | +| gears | 2 | The gear rotation check scales its one-frame threshold by the update's frame delta, so shrine gears read as turning above 30 fps. Follows `TOTK_HIGHFPS_FIX`. | -At the current upstream baseline `6b0974f08d580571ebf16f6f1405367c61405eec`, the table has 14 sets and 51 instruction sites before this follow-up, and 16 sets and 54 instruction sites after adding one climb site and two gears sites. Its existing 889 spin back-edges are unchanged. The new sets use the existing runtime page and settings; no new setting, allocator or renderer path is introduced. +The table has 16 sets and 54 instruction sites in total. Its 889 spin back-edges are unchanged. The sets use the existing runtime page and settings; no new setting, allocator or renderer path is introduced. The camera sites were cross-checked by normalized instruction-window matching from the local 1.0.0 main. The viewport routine has two matching copies: both are patched. Other sites were checked through their data flow: the equipment timer and its gate, the map clock, the ability clock, the NVN texture storage/package calls, the shadow size publication and the cascade parameter constructor. FOV's s8 is overwritten by the next original instruction; x10 is saved. The cursor sites use the same scratch registers as their 1.2.1 counterparts. @@ -40,34 +42,15 @@ Without ExeFS, the test uses synthetic memory and the patch table only. It compa These checks establish instruction and code-generation contracts. World-scene visual acceptance, texture packs allocating a pool above 2 GiB, physical gyro and comparisons of shadow quality still require gameplay checks. -## 0.4.0 follow-up: climbing and gears +## Climbing and gears -The historical audit below compared the final patches and runtime gating added after `3d386fe` against upstream `c89b0f3ee3b131275c5c5c26426128f16469f2cd`. The target is **1.1.0 / v131072**, not 1.1.1. The current baseline is `6b0974f08d580571ebf16f6f1405367c61405eec`: upstream now supplies the 1.1.0 fixed-dynamic-resolution exit (`fixres`) and the version-specific `sync_ahead` frame-wait address. Those are retained unchanged; this follow-up still adds only climb and gears. The table records the older audit boundary, not a claim that current upstream lacks those features. +Both sets port their 1.2.1 counterparts in [recomp/patches-v121.txt](../recomp/patches-v121.txt); the comments there explain the formulas. -| Upstream change | 1.1.0 audit | Scope of this follow-up | -| --- | --- | --- | -| climb | Local main+0x1927948 has the matching velocity compensation, r=s9 and s8=0.5 clamp, but computes the coefficients in a different order. | Port one site at +0x19279f0, using the existing TOTK_CLIMB_FIX and realtime dependency. | -| gears | main+0x213eec4 tests rotation against RotateStartDeg/30 without the update context delta. | Port two sites under TOTK_HIGHFPS_FIX, independently of realtime. | -| fixres / large scene sizes | Existing output/scene sites remain; +0xce9280 is a candidate fixed-dynamic-resolution exit. | Not ported: tiers, render scale, HeapDef and actual rendering still need acceptance checks. | -| grass | +0x100e838 and its +0x100e890 range constant are candidates. | Not ported: locating all six sites and checking textures, heaps and world scenes remain separate work. | -| optref / ability / report | The same crash root cause has not been established on v110; the report loop is not structurally equivalent on every version. | Not ported; guards must not hide an unlocated problem. | -| sync_ahead | frame_wait_lr in [nvn_api_core.c](../runtime/nvn_api_core.c) lists only 121/143 and also requires shadow storage. | Keep the version gating; GPU, readback and fence behaviour has not been accepted. | +**climb**, one site at main+0x19279f0 (1.2.1: main+0x19182cc) in the climb controller's velocity step at main+0x1927948. The site replaces `fmov s1, #1.5` after the shipping factor 1.272 - 0.362 (r - 1), r in s9, s8 = 0.5. The body between prologue and epilogue is word for word the 1.2.1 one: it reads u = max(page +0, r) and gives s0 = max(s0, h), s1 = max(1.5, h). At r >= 1 h stays at or below the shipping factor, so 30 fps and slower keep it bit for bit; at 60 fps the factor becomes 1.4.3's, about 1.8665 instead of 1.453; above 60 fps g(u) takes over. Unlike 1.2.1, the site does not rely on dead registers: a 96-byte stack frame saves x8 and full q5, q6, q7, q16 and q17 and restores them before the original code continues, so only s0 and s1 change. `TOTK_CLIMB_FIX=0` runs the original `fmov`; [frame_step_init](../runtime/frame_rate.c) switches the set off without the real-time step, where the page is not written. -### Shipping code and compensation semantics +**gears**, two sites in the ActorCheckRotateModule update at main+0x213eec4 (vtable 0x347e8e8 slot +0x110; 1.2.1: main+0x21350c0). The first, main+0x213eee4, loads the update context's frame delta (x1 + 8) into s12 before the function overwrites x1; the prologue saved d12 at main+0x213eec8, nothing writes s12 before the second site, and callees preserve the low 64 bits of v12. The second, main+0x213f0d4, turns `fdiv s12, s0, s4` into s12 = (s0 / s4) * delta, the RotateStartDeg / 30 threshold scaled by the delta, which the three `fcmp s8, s12` at main+0x213f10c, +0x213f174 and +0x213f1b0 read. It saves full q4 in a 16-byte stack slot and restores it and SP. The set does not read the patch page and works with or without the real-time step; `TOTK_HIGHFPS_FIX=0` keeps both original sites. -The climb compensation comes from this repository's [1.2.1 climb definition](../recomp/patches-v121.txt) at the fixed baseline. Local 1.1.0 first multiplies by -0.181, doubles with fadd, then adds 1.272. This is algebraically 1.272 - 0.362(r-1), but merging the multiplication changes its floating-point rounding order. This patch extends only the final clamp and preserves the preceding v110 instructions. - -The shipping factor's lower 32 bits stay unchanged at normal 30 Hz and at slower steps with r>=1. The 60 Hz change is intentional: at normal r=0.5, compensation is about 1.8665 instead of the shipping 1.453. Above 60 Hz it uses upstream g(u), with u=max(page+0,r), so reducing the time scale alone does not trigger the higher-frame-rate g branch. It reads existing page+0 without adding fields. - -climb uses a 96-byte, 16-byte-aligned stack frame to preserve x8 and full q5/q6/q7/q16/q17. Only s0/s1 are declared outputs; scalar writes clear their vector upper bits. This avoids relying on 1.2.1 dead-register assumptions. TOTK_CLIMB_FIX=0 executes only the original fmov; existing [frame_step_init](../runtime/frame_rate.c) disables the set when realtime=0. - -The gear function is reached through vtable 0x347e8e8 slot +0x110. AI dispatchers +0x1205b90 and +0x2ae29ac copy context+8 into the passed structure's +8; AI root update +0x1085c88 calls the owner's virtual +0x30 and publishes its returned frame delta there. The first site, +0x213eee4, follows the original str d12 save. This function does not write s12/d12/q12 between that site and the threshold site. Across calls it relies only on AAPCS64 preserving d12's low 64 bits: the delta occupies the low 32 bits and does not depend on the vector upper half surviving. The second site produces a fresh scalar s12, and the original epilogue restores d12; q12's upper half is not callee-saved. - -The threshold becomes (RotateStartDeg/30)*ctx.delta, shared by the three original fcmp/b.le pairs at +0x213f10c, +0x213f174 and +0x213f1b0. The second site also preserves full q4 and restores SP before continuing. gears does not read the patch page, remains active with realtime=0, and restores both shipping sites with TOTK_HIGHFPS_FIX=0. - -### Unique matches and original-word guards - -Each of these original windows occurs exactly once in the local 1.1.0 main. These are addresses and hashes, not game-byte dumps: +Each original window below occurs once in the 1.1.0 main (addresses and hashes, not game bytes): | Window | Length | SHA-256 | | --- | ---: | --- | @@ -75,27 +58,10 @@ Each of these original windows occurs exactly once in the local 1.1.0 main. Thes | +0x213eee4 | 32 bytes | 17487c0fe195655e9a1bcaf49835208bd1ebca86eeef9a2ca0059752f503e378 | | +0x213f0d0 | 28 bytes | d07279a2abea53cffbd8861e95034cd1a5c85edf90f6094c5261ba55da203577 | -The generator still checks each original instruction. Optional local tests also check the shipping constants, d12 save, vtable entry, all three compares and their b.le targets. No version-wide fixed address shift was used. - -### Added validation - -The [high-frame-rate patch test](../tests/patches_v110_highfps_check.py) reuses shared CC selection, platform suffixes and mappings, with pathlib paths. The [patch parser](../recomp/patches.py) explicitly reads UTF-8, including the Unicode encoding fixture, independently of the Windows default code page. - ```sh -env CC=clang python tests/patches_v110_highfps_check.py +python tests/patches_v110_highfps_check.py python tests/patches_v110_highfps_check.py --cflags=-march=x86-64-v2 -python tests/patches_v110_highfps_check.py --exefs --work-dir +python tests/patches_v110_highfps_check.py --exefs ``` -- Windows v3/v2, each with synthetic and local ExeFS input: 3240 native/Unicorn cases per run passed, covering on/off, zero/tiny deltas, adjacent ULPs around 0.5, 30/60/120/240 Hz, slower steps and all 16 NZCV values. -- All 31 GPRs, 32 full 128-bit SIMD registers, SP, NZCV and the complete scratch-memory CRC are compared. Every undeclared output must match its input bit for bit. Nonzero upper-64-bit sentinels in q4/q8/q9 are explicit preconditions. -- Exact save-slot contents, 16-byte stack alignment, restored SP, slot gaps, surrounding canaries and all other scratch memory are checked. No whole memory region or broad SIMD allowlist is excluded. -- Nine default/explicit setting combinations compile the real [patches.c](../runtime/patches.c), frame_step_init and step_store and pass. The existing [#1002 regression](../tests/patches_v110_check.py) retains its SIMD allowlist fix; all four runs pass 408 cases each. -- All five in-memory negative mutations are rejected: corrupt x8, corrupt q17, write outside the save area, restore only d4, and restore only s4. No mutation was applied to the patch file. Combined cases also clobber q12's upper 64 bits to check that calls need preserve only low d12. -- Native WSL Ubuntu 22.04 compilation and execution pass all nine gating combinations. The complete SIMD native replay fails to compile with the installed Clang 14 because the shared [simd_fp.h](../runtime/simd_fp.h) requires newer elementwise builtins. **There is no claim that 3240 Linux replay cases passed.** Synthetic inputs/outputs remain local for a newer-toolchain replay and are not committed. Linux Python differential dependencies were not installed, and the distribution environment was not modified. - -These establish instruction, numeric and configuration contracts, not acceptance of real climbing, shrine gears, slow motion, severe stalls or GPU scenes. This branch did not launch or deploy the game; maintainers perform full builds, installation updates and gameplay acceptance separately. - -### Regeneration requirements - -Re-lift the generated chunks containing main+0x1927948 and main+0x213eec4 and regenerate patch metadata. A normal complete generation is recommended; never hand-edit generated C. RT_PATCH_COUNT changes from 12 to 14, retaining the original 12 indices and appending climb/gears. The patch page address and layout stay unchanged: climb reuses +0 and gears adds no page fields. Do not reuse stale generated objects or lift hashes; rebuild runtime libraries against the regenerated patch macros. +The test lifts both sets and runs them against Unicorn on and off, at 30, 60, 120 and 240 Hz and slower steps, with zero and tiny deltas, adjacent ULPs around 0.5 and at each compare threshold, and all 16 NZCV values (3240 cases). It compares all general and full 128-bit SIMD registers, SP, NZCV and the whole scratch memory: only s0/s1 (climb), x1/s12 (gears) and the exact save slots may change. It also compiles [patches.c](../runtime/patches.c) with frame_step_init for the nine combinations of the three switches. With `--exefs` it checks every original word, the shipping constants, the d12 save, the vtable entry and the three compares with their branch targets. These are instruction and switch contracts; climbing and the shrine gears still need checks in the game.