This repository has been archived on 2026-09-16. You can view files and clone it. You cannot open issues or pull requests or push a commit.
Files
Syplheed-Reborn/docs/re/structures/hud-glyph-quad.md
Sylpheed RE agent 6a8c366a11 formats: correct the glyph class - a SECONDARY vtable of adjustor thunks
The first write-up called 0x820B2A64's 32 entries "methods" and the three xrefs
to it "construction sites". The DB's disassembly refutes both: every slot points
at a 16-byte adjustor thunk - lwz r11,-4(r3); subf r3,r11,r3; b <method> - which
is the PPC/MSVC thunk for a SECONDARY base subobject. So this is a secondary
vtable and the class uses multiple inheritance, and sub_823C3148, listed as a
constructor because it references the table, is actually the branch target of one
of the table's own thunks.

Following the 32 branch targets instead: 32 distinct addresses, 18 of them known
function starts, in two clusters (0x823c3xxx beside the thunks and 0x823e3xxx),
and NONE of them appears in any of the 1150 vtables the DB classified - so
nothing places this class in a named hierarchy, the disc's RTTI having no class
names. The two-cluster split is noted as a reading rather than a measurement.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PMRJjbxLqZtsb5Vb7KunPE
2026-08-24 10:16:12 +00:00

3.8 KiB
Raw Blame History

The HUD's glyph quad — vtable 0x820B2A64

Status: CONFIRMED for the object layout and the atlas size, read live off a Stage 02 run and checked against itself. 🟡 the class has no name (RTTI carries none disc-wide).

How it was reached

mission-objective-counter.md found that the words moving with REMAINING OB are four ASCII digits and four guest pointers. Following the pointers lands on four objects with an identical shape, whose first word is 0x820B2A64.

It is a SECONDARY vtable — 32 adjustor thunks

Read out of guest memory: 32 consecutive code pointers in 0x823c43b0 … 0x823c45a0, terminated by 0xfffffffc.

🔴 Corrected 2026-08-24, same session. The first write-up called these "32 methods" and the three xrefs to the table "three construction sites". Both were wrong, and sylpheed.db's disassembly says why — every slot points at a 16-byte adjustor thunk:

lwz   r11, -4(r3)     ; the this-adjustment stored just before the object
subf  r3, r11, r3     ; fix up `this`
b     <the real method>
.long 0

That is the classic PPC/MSVC thunk for a secondary base subobject, so 0x820B2A64 is a secondary vtable and the class uses multiple inheritance. sub_823C3148, which the first write-up listed as a constructor because it references the table, is in fact the branch target of one of the table's own thunks.

Where the real methods live. Taking the 32 branch targets: 32 distinct addresses, 18 of them known function starts in sylpheed.db, in two clusters — 0x823c3xxx (beside the thunk block itself) and 0x823e3xxx. The rest land on further stubs immediately after the block. None of the 32 appears in any of the 1 150 vtables the DB classified, so nothing places this class in a named hierarchy; the disc's RTTI carries no class names.

The two clusters are what a derived class looks like — its own overrides next to its own thunks, and inherited implementations elsewhere — but that is a reading, not a measurement.

The instance

 -0x0c  f32   1280.0        ) the design space, immediately before the object
 -0x08  f32    720.0        )
 +0x00  u32   0x820B2A64    vtable
 +0x1c  f32     34.0        width  in pixels
 +0x20  f32     42.0        height in pixels
 +0x30  u32   0xffffffff    vertex 0 colour
 +0x34  f32   u                     u
 +0x38  f32   v                     v
 +0x48 … vertex 1   (+0x18 stride)
 +0x60 … vertex 2
 +0x78 … vertex 3

Measured on one digit glyph:

vertex colour u v
0 0xffffffff 0.5477 0.5013
1 0xffffffff 0.5742 0.5013
2 0xffffffff 0.5742 0.5560
3 0xffffffff 0.5477 0.5560

Four corners of an axis-aligned rectangle, in order, all untinted — a textured quad.

The atlas is 1280 × 768, and that is a check rather than a guess

The UV rectangle is 0.0265 × 0.0547. Multiply by 1280 × 768:

0.0265 × 1280 = 33.9   ->  the width  stored at +0x1c is 34.0
0.0547 ×  768 = 42.0   ->  the height stored at +0x20 is 42.0

Two independent fields of the same object agree to within a rounding step, which is what makes the atlas size a measurement instead of a plausible number.

⚠️ These objects churn

The four pointers next to the counter swap as the digit changes, and re-reading one of the old targets a minute later returned unrelated data — the object had been recycled. Anything read here is a snapshot of a live pool, not a stable address.

What this is for

It gives the port a decoded HUD text primitive: a glyph is a quad with a pixel size and four UVs into a 1280×768 atlas, drawn by a class with 32 virtual methods and three constructors. And it explains why the objective counter was so easy to find in RAM — the value sits inside the widget that renders it.