MechaCat02 e3cc6d11d0 [iterate-4A] jit: per-block GPR register caching — bodies now beat interp
Guest GPRs were reloaded from PpcContext memory on every use and stored back
after every write, forcing memory traffic that held the JIT bodies at mere
interpreter parity. Now each GPR is loaded ONCE per block into a Cranelift SSA
value (kept in a host register by the register allocator); writes update an
in-EmitCtx cache (RegCache: [Option<Value>;32] + dirty bits) without storing;
dirty GPRs are flushed to memory only at the block boundary (cache_flush at the
end of emit_node_body).

Sound because nothing a block calls back into touches guest GPRs: the memory
trampolines take addr/val as args (I pass the cached SSA values), and the
FP-punt runs interpreter FP ops that touch fpr/fpscr/cr only. So the cache never
goes stale mid-block; flushing at the boundary makes gpr[] interpreter-identical
for the runner and the next block. FPR/CR still go through memory (the FP-punt
writes them, which would need flush-around-punt — deferred).

emit_op / ea_d / ea_x / gpr32 / store_gpr32z rewired to cache_read_gpr /
cache_write_gpr; cache_read_gpr copies the Option out before re-borrowing (a
RefCell double-borrow would panic).

Validated: 7/7 jit unit tests; diff MISMATCHES=0 over 47.9M blocks (foreground).
Perf (foreground, RUST_LOG=warn, -n 3e9 to amortize compile cost — -n<=1.5e9 is
compile-dominated and misleads): interp 95.8 MIPS -> jit-no-chain 100.2 MIPS
(bodies now BEAT interp, were parity) -> chain 100.4 MIPS / 57.95s = -15.6% wall
vs interp (was ~-6% at B2), dispatch 18.9%, 110M calls (1.95x merge). Same-batch
-n 2e9: interp 44.8s -> chain 41.0s = -8.6%. Register caching is what makes the
bodies fast enough that chaining's dispatch cut converts to real wall time
(B2's bodies had slowed to ~82 MIPS and cancelled the win).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-08 06:15:16 +02:00

xenia-rs

Rust reimplementation of the Xbox 360 emulator xenia, focused on reverse-engineering and preservation rather than full-speed play. The initial target is Project Sylpheed — Arc of Deception; getting the title disassembled, traced, and far enough into its init path to understand its engine.

Heavy cross-reference to xenia-canary for CPU context setup, kernel export behavior, and XEX loading semantics.

Status

  • XEX loader — XEX2 header parsing, LZX decompression, AES decryption, PE section parsing.
  • VFS / XISO — XGD2 dual-layer disc images (with the 0x0FD90000 partition offset).
  • PPC interpreter — 200+ opcodes, PowerPC 32/64-bit GPR/FPR, VMX128 decoding.
  • Static analyzer — function discovery (prolog/epilog heuristics), cross-references, labels, save/restore helper detection, assembly text + SQLite database output.
  • Kernel HLE — minimal subset driving Project Sylpheed: ~170 xboxkrnl + xam exports (critical sections, events, TLS, virtual memory, Vd stubs, XAM input/user/content).
  • Debugger — in-memory step/break, SQLite execution + import-call + branch tracing.

Not yet: GPU (xenos/xe-shader), APU audio, HID, kernel scheduler, full threading, exception delivery.

Workspace

crates/
  xenia-types       # shared primitive types, bitflags
  xenia-memory      # guest memory, paged allocator, page table
  xenia-cpu         # PPC decoder, interpreter, context
  xenia-xex         # XEX2 loader, PE parser, LZX, AES
  xenia-vfs         # XISO / disc-image reader
  xenia-kernel      # HLE kernel state, exports, XAM
  xenia-gpu         # (stub) Xenos command processor
  xenia-apu         # (stub) XAudio
  xenia-hid         # (stub) XInput
  xenia-debugger    # in-memory trace, breakpoints, step modes
  xenia-analysis    # function/xref analysis, assembly formatter, SQLite DbWriter
  xenia-app         # `xenia-rs` CLI binary

CLI

Build:

cargo build --release

The binary xenia-rs accepts XEX2 files or ISO / XISO disc images as input (the loader auto-detects discs and extracts default.xex).

info / browse / disasm

Quick header / disc / first-N-instructions inspection. See --help.

extract — unpack PE + metadata

xenia-rs extract <xex-or-iso> [-o <out-dir>] [--db <sqlite-path>]

Writes <name>.pe (decompressed/decrypted PE image) and <name>.xex.json (header metadata). With --db, also emits a SQLite database containing the base tables: metadata, sections, imports.

dis — full disassembly

xenia-rs dis <xex-or-iso> [-o <asm-file>] [--db <sqlite-path>] [--quiet]

Runs function + cross-reference analysis and produces:

  • assembly text to stdout or -o <file> (unless --quiet)
  • optional SQLite DB with the base tables + disasm tables: functions, labels, instructions, xrefs

exec — interpret with tracing

xenia-rs exec <xex-or-iso> [-n <max-instrs>] [--db <sqlite-path>]
             [--trace-instructions] [--trace-imports] [--trace-branches]

Loads the title, initializes CPU state per xenia-canary, intercepts import thunks with HLE kernel calls, and interprets from the entry point. Without -n, runs until halt/fault. With --db, produces a DB that is a superset of dis --db plus opt-in trace tables:

flag table rows
--trace-instructions exec_trace one row per interpreted instruction (PC, r3/r4, LR, SP)
--trace-imports import_calls one row per kernel/XAM call (module, ordinal, args)
--trace-branches branch_trace taken branches classified as call/return/jump/branch

Cumulative DB layering

Each command's DB is a superset of the previous. A single xenia-rs exec <iso> --db full.db --trace-instructions --trace-imports --trace-branches produces the full picture in one pass — base tables, complete static disassembly, and runtime traces correlatable by address/cycle.

Performance knobs

  • XENIA_DB_BATCH_SIZE — rows per streaming commit / trace-buffer flush (default 100_000). Lower values reduce memory use; higher values reduce fsync overhead on slow disks.

The DB writer uses journal_mode=OFF, synchronous=OFF, locking_mode=EXCLUSIVE and commits in batches; no ANALYZE is run at finalize. Indices are created after bulk insertion with progress messages.

Example queries

-- Top 20 kernel functions called during early init
SELECT name, COUNT(*) FROM import_calls GROUP BY name ORDER BY 2 DESC LIMIT 20;

-- All basic-block leaders (targets of taken branches) not already labelled
SELECT DISTINCT bt.target
FROM branch_trace bt LEFT JOIN labels l ON l.address = bt.target
WHERE l.address IS NULL;

-- Correlate a traced call site with its static disassembly
SELECT et.cycle, i.disasm, i.ext_disasm
FROM exec_trace et JOIN instructions i ON i.address = et.address
WHERE et.address = 0x824AB748 ORDER BY et.cycle;

License

BSD-3-Clause, matching upstream xenia.

Description
No description provided
Readme 9 MiB
Languages
Rust 98%
WGSL 1.7%
Python 0.3%