steam-frame/Cesarus85--mjolnir-vr
1
0 Fork 0
mirror of https://github.com/Cesarus85/ringfall-mjolnir-vr.git synced 2026-10-23 05:00:32 +02:00
Code Issues Packages Projects Releases 18 Wiki Activity
Files
8fed3d94de39af96077fc8621d2f6f20d8ade9a2
Cesarus85--mjolnir-vr/docs/object_matching_logs/tif_getimage_obj.md
T

6.5 KiB
Raw Blame History

tif_getimage.obj exact-match ledger

Scope and invariant

  • Translation unit: source/bitmaps/libtiff/tif_getimage.c
  • Target: January 2002 tif_getimage.obj
  • Starting commit: 64f068c27cc56bcc55d58ceb5382863087082f3f
  • Toolchain: XDK 3911 CL 13.00.9254.1
  • Flags: /O2 /Oy- /DDEBUG /Dxbox
  • Acceptance: every function and every target-owned data section must pass the hardened COFF comparator. No inline assembly, volatile codegen coercion, undefined behavior, binary patches, or compiler-flag changes.

The starting state was 24/27 strict-exact functions. All 745 target-owned data bytes were already exact. The three residual functions were makecmap (code_00058c40), initYCbCrConversion (code_00059a60), and putRGBContigYCbCrClump (code_00059ab0).

initYCbCrConversion

Target: 80 bytes, 10 relocations.

Experiment Result
Initialize reciprocal = 1.0f/LumaGreen before D1 80 bytes, 9 relocations; rejected
Compute D1, then reciprocal, with direct three-factor products 96 bytes, 10 relocations; rejected
Compute D1, then reciprocal, and stage the reused multiply through product STRICT EXACT

Accepted source shape keeps 1/LumaGreen live on the x87 stack and reuses it for both coefficient paths:

D1 = 2 - 2*LumaRed;
one_over_luma_green = 1.0f / LumaGreen;
product = one_over_luma_green*LumaRed;
D2 = product*D1;
D3 = 2 - 2*LumaBlue;
product = one_over_luma_green*LumaBlue;
D4 = product*D2;

Final strict evidence:

  • size: 80/80
  • relocations: 10/10
  • normalized SHA-256: 4c8558ad61a2a1d2cf69dc295bd14364eb926b8349e85e66ba7bc09faca8d50a

putRGBContigYCbCrClump

Target: 464 bytes, 27 relocations.

Experiment Result
Explicit invariants before an ordinary for loop 464/27; zero-trip scheduling wrong
Guarded if (cw > 0) around invariant setup and for 480/27; redundant precheck
Direct authentic expressions in ordinary for 464/27; frame 0x1c vs target 0x20, broad drift
Guard + invariant setup + k=0; do/while 464/27; k initialization scheduled after test
Move k=0 before guard best: 464/27, all relocations exact, four x87 operand bytes differ
Reverse commutative spelling (Y + D1Cr, etc.) no change; rejected as a non-control
Split R/B accumulation into assignment plus += 480/27; rejected
Change invariant temporaries from double to float no change; rejected as a non-control

Current best residual is confined to two x87 additions:

  • target R: fld st(3); fadd st(1)
  • candidate R: fld st(0); fadd st(4)
  • target B: fld st(2); fadd st(1)
  • candidate B: fld st(0); fadd st(3)

Everything else, including all relocation addresses, the zero-trip guard, and the three consecutive x87 pops, is exact. Do not repeat the commutative spelling, split-accumulation, or float/double local-type experiments.

makecmap

Target: 832 bytes, 18 relocations.

Experiment Result
Baseline local palmap, global PALmap[i] loop store 832/18; non-exact register schedule
Exact-sibling-style local palmap[i] store 832/17; loses required global relocation; rejected
Global-only/no local palmap 816/18; rejected

The target retains the allocation result in EDX for the first iteration, keeps the output pointer in ECX and the loop index in EAX, then reloads global PALmap on later backedges. The baseline instead uses EAX for the output pointer, ECX for the loop index, and reloads the global before the first store.

High-evidence untried controls are:

  1. declaration order matching exact sibling makebwmap: i, nsamples, p, then palmap, while retaining PALmap[i] = p;
  2. explicitly peeled first-iteration topology using local palmap[i] = p, then palmap = PALmap on the loop backedge.

Provenance breakthrough and final closeout

The remaining source topology was recovered from the authentic libtiff RCS revision 1.8 source. The original translation unit used separate file-scope statics, not the reconstructed aggregate that had coupled alias analysis and register allocation across the three residual functions. Restoring the separate objects made the authentic makecmap global accesses and the direct nested putRGBContigYCbCrClump loop strict-exact.

Uninitialized file-scope statics were allocated in an unstable order by this compiler. Explicit zero initializers made their source order deterministic and recovered the January .bss layout. The placeholder _bss_0031be54 symbol was then replaced in config/symbols.json by the recovered local symbols _D4, _D3, _D2, _D1, _PALmap, _BWmap, _refBlackWhite, _YCbCrCoeffs, _YCbCrVertSampling, _YCbCrHorizSampling, _filename, _stoponerr, _bluecmap, _greencmap, _redcmap, _orientation, _photometric, _samplesperpixel, and _bitspersample, at the exact January offsets. Each recovered entry is marked static in the csplit manifest; llvm-nm reports matching lowercase b ownership in both target and candidate.

This file retains its authentic 1992 libtiff K&R declaration style rather than mixing modern prototypes into a legacy third-party translation unit. The two modified void functions nevertheless end in explicit return; statements per the project's house rule; VC7 emits no additional instructions for them.

Rejected closeout experiments:

  • one externally linked BSS anchor plus a D4 macro: exact makecmap and initYCbCrConversion, but changed clump scheduling through the linkage/alias model;
  • raw uninitialized separate statics: correct code, nondeterministic/wrong BSS allocation order;
  • preserving the aggregate while copying authentic expressions: retained the aggregate-induced scheduling mismatches.

Final strict evidence for the former residuals:

  • makecmap: 832/832 bytes, 18/18 relocations, normalized SHA-256 bf6832724e2051eaa90147ed6b5db734189bea61696903dfbc39cb03e37ff7af;
  • initYCbCrConversion: 80/80 bytes, 10/10 relocations, normalized SHA-256 4c8558ad61a2a1d2cf69dc295bd14364eb926b8349e85e66ba7bc09faca8d50a;
  • putRGBContigYCbCrClump: 464/464 bytes, 27/27 relocations, normalized SHA-256 61f4a8ab80dc48d34197c99801ed6cfcdf0e55acbfe86d21c7d0ebdd6df55828.

Final admission state

  • Hardened function sweep: 27/27 strict exact
  • Target-owned data: 745/745 bytes strict exact
  • .bss: 74/74 bytes
  • .data: 91/91 bytes
  • .rdata: 580/580 bytes
  • Consolidated ninja progress build: successful
  • Progress admission: complete unit retained; no premature-completion revocation
  • Object status: Matching
Reference in new issue View Git Blame Copy Permalink