Include StaticRecompCore.h where the interpreter uses it by dougchansan · Pull Request #7 · ExpansionPak/RecompCore

dougchansan · 2026-08-04T23:17:32Z

core does not compile at the moderngekko-vendor tip (7275711) with MSVC:

Interpreter.cpp(294): error C2065: 'g_static_recomp_core': undeclared identifier
Interpreter.cpp(299): error C2065: 'ppc_state': undeclared identifier

unknown_instruction reads both, but the translation unit includes neither the header declaring g_static_recomp_core (Core/PowerPC/StaticRecomp/StaticRecompCore.h) nor a binding for ppc_state. It is not reached through the precompiled header either â€” pch.h does not pull in StaticRecompCore.h.

The ppc_state binding already existed further down the same function, past the first use. This moves it to the top so it serves both uses; adding a second one is a redefinition (C2374).

Found while building moderngekko-launcher from ModernGekko master. It is the second of two breaks in that path â€” the first is the unguarded PCH flags, already fixed on this branch, which is why this one was not visible until now.

Two lines, no behaviour change.

How this fits

These land together as a set: savestates working end to end, the Windows build
and test suite being usable at all, and CI so none of it regresses unnoticed.
This PR is RC #7.


MG #18	Bump vendored RecompCore	prerequisite - without it the launcher cannot compile on Windows (C sources get a C++ PCH)
RC #7	`Interpreter.cpp` include	prerequisite - `core` does not compile without it
RC #8	In-game File/View menu and hotkeys	introduces `Core/SavestateLayout.h`, the one definition of where savestates live, what they are called and how they are ordered
MG #21	Launcher savestate picker, `--load-state`	consumes that header, so the launcher and the in-game menu cannot disagree
MG #22	Make the test suite pass on Windows	land before CI, or the first run is red on day one
MG #23	Build and test on PRs, three platforms	the reason the rest stayed broken unnoticed
RC #10	CI on PRs, the vendored branch, and Windows	same gap, other repo; its Windows job fails until RC #7 lands
MG #19	`--opt-level`, default `-O2`	independent
MG #20	Cache-domain affinity	independent

Suggested order: #18 and RC #7, then RC #8, then MG #21; MG #22 before MG
#23. The rest are independent.

Verified on three platforms: Windows (MSVC + clang), Linux (g++ 15.2, x86_64) and
macOS 26.1 (clang, arm64). The portable pieces - the savestate layout and its
tests, frontend_config, dol_patch - build and pass on all three. Windows-only
pieces are guarded and their tests registered behind if(WIN32).

unknown_instruction reads g_static_recomp_core and ppc_state, but the translation unit includes neither the header that declares the former nor a binding for the latter, so core does not compile: Interpreter.cpp(294): error C2065: 'g_static_recomp_core': undeclared identifier Interpreter.cpp(299): error C2065: 'ppc_state': undeclared identifier It is not reached through the PCH either -- pch.h does not pull in StaticRecompCore.h. The ppc_state binding already existed further down the same function, past the first use. Moving it to the top serves both uses rather than adding a second one, which would be a redefinition.

dougchansan · 2026-08-07T02:28:11Z

Since there are now 17 open PRs across RecompCore, ModernGekko and DolRecomp, here is a suggested merge order. Posting it here because this one has to land first regardless.

Merge these three first, in this order

RecompCore Include StaticRecompCore.h where the interpreter uses it #7 (this PR) — moderngekko-vendor HEAD does not compile without it, so every other RecompCore PR is untestable until it lands.
RecompCore JitArm64: implement the StaticRecomp fallback contract #6 — the JitArm64 fallback contract. Without it the recompiled module is entered once at boot and never again on arm64, so anything measured as "static recomp" on Apple Silicon is really JitArm64.
RecompCore GXRuntime: inline the hot paths of the FP gate and paired-single access #9 — the inline FP gate and paired-single path. DolRecomp #12 does not compile without it: the emitted code references runtime names this PR introduces.

Everything else can follow in any order, subject to the two rules below.

The two ordering rules that matter

DolRecomp #12 must come after RecompCore GXRuntime: inline the hot paths of the FP gate and paired-single access #9, and after GXRuntime: inline the hot paths of the FP gate and paired-single access #9 is vendored. This is the only hard cross-repo dependency in the set.
ModernGekko #18 should come last. It bumps the vendored pin specifically to pick up merged RecompCore work; merged early it pins a commit predating all eight RecompCore PRs and needs bumping again immediately.

Rebase clusters — annoyance, not blockers

Whichever lands first, the others need a trivial rebase:

file	PRs
`Source/Core/Core/PowerPC/Interpreter/Interpreter.cpp`	RecompCore #7, #8
`CMakeLists.txt`	ModernGekko #17, #20, #21, #22
`tools/moderngekko_run.cpp`	ModernGekko #20, #21
`src/backend/emitter.c`	DolRecomp #12, #13

All 17 report MERGEABLE against their bases as of today, so these are the only collisions.

Worth knowing before reviewing

RecompCore module-template: fix the IPO check, and default IPO off #11 needs a decision, not just review. It fixes a support check that failed silently on macOS — no module was ever built with -flto=thin despite every manifest recording it — and separately flips IPO off by default because LTO measured −0.8% on arm64. On Linux the check passes, so those builds presumably have had LTO all along and the default change would remove it. The two halves are independent; taking only the check fix is reasonable.
ModernGekko #17 is worth more than its four lines suggest. It fixes a stale-recompiler-copy trap that has now caused a wrong result twice, on two platforms.
Absolute fps figures in RecompCore JitArm64: implement the StaticRecomp fallback contract #6 and GXRuntime: inline the hot paths of the FP gate and paired-single access #9 were measured under a 2.02× CPU throttle (macOS Low Power Mode, found yesterday). Correcting comments are on both PRs. The relative deltas in every PR are unaffected — they were alternating A/B arms inside a single session, so a constant throttle scales both arms and cancels — as is JitArm64: implement the StaticRecomp fallback contract #6's JIT ratio, which re-measures at 5.58× against the 5.9× claimed.
DolRecomp #13 has been re-measured since it was opened: the original "+2.9%, ranges overlap, I would understand a decline" is now +10.6% with non-overlapping ranges on an unthrottled machine, with the artifact verified before measuring. Its description is updated.

Navigation Menu

Uh oh!

Include StaticRecompCore.h where the interpreter uses it - #7

Include StaticRecompCore.h where the interpreter uses it#7
dougchansan wants to merge 1 commit into
ExpansionPak:moderngekko-vendorfrom
dougchansan:interpreter-staticrecomp-include

dougchansan commented Aug 4, 2026 •

edited

Loading

Uh oh!

dougchansan commented Aug 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

dougchansan commented Aug 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

How this fits

Uh oh!

dougchansan commented Aug 7, 2026

Merge these three first, in this order

The two ordering rules that matter

Rebase clusters — annoyance, not blockers

Worth knowing before reviewing

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

dougchansan commented Aug 4, 2026 •

edited

Loading