Filed, not started. Proton already carries Linux players and the Windows
build reaches a running mission under it; this is the answer to the
separate question of what one source tree building natively on both would
take, so that the number exists before anyone needs it.
Three sweeps of the tree, and the measurements are the part worth
keeping. Two of them make the job smaller than the earlier assessment
feared, and one makes it different.
/Zp1 was called an entanglement with no manifest of which structs mattered.
There is a manifest now: 31 distinct whole-object stream types, of which
29 are enums, scalar typedefs or all-float aggregates. Modelled at both
packings and compared by sizeof, exactly one struct changes - the
anonymous .met reader in L4D3D.cpp, 14 bytes packed against 16 default.
Dropping /Zp1 costs one pragma. That also retires the packing divergence
between RPL4TOOL and the game that FORCE:MULTIPLE has been hiding.
The include mess is not the blocker. 172 of 172 core translation units
fail under GCC, and 171 still fail with every 64-bit error removed,
because MUNGA/SCHAIN.h uses a friend-declared name as a type and sits in
the precompiled block. One line breaks the entire core, and 155 friend
sites over 117 names share the shape. The whole conformance list is about
thirty sites across twenty headers - small, and it gates everything.
And the windows.h leak into engine core is NETWORK.h, not MATRIX.h.
Winsock2.h reaches APP.h and 157 translation units, which is how
APPMGR.h's bare 'extern HWND ghWnd' has been compiling all along. The
D3DX9 include in MATRIX.h is real but narrow: ten files, two members.
docs/RP412-LINUX.md said otherwise and now says so.
What the sweeps found in our favour: RP is clean outright, 190 of 241
units touch no platform symbol, the renderer's distinct surface is
nineteen render states and one mesh draw call with no shaders, all 224 .X
files are text using eight templates, and the entire gauge production path
- including four thousand seven hundred lines of L4VB16 - is already
portable C++. RIOBase, the whole input seam, is two pure virtuals and five
scalars. DivLoader is referenced by nothing and linked by nothing; the
lobby room screen has no callers at all.
The plan is phased so that almost everything happens on Windows first,
verified by the existing build and the two-pod harness, with Linux arriving
only at phase four. A port that spends six months unable to run is a port
that gets abandoned.
The document ends by saying when this would not be worth doing, which is
most of the time: Proton already delivers the player experience, and only
the first three phases have standalone value. It becomes worth starting
when the goal changes to retiring the 2010 SDK dependency or removing Wine
as a support variable.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CydandClaude Fable 5
29ad567ffb