Files
psxrecomp/runtime/include/dirty_ram_interp.h
T
Matthew Stanley 71c2df0f54 kernel-image bless: byte-verified static dispatch for the relocated kernel window
The generated dispatcher ran kernel-window table hits (RAM [0x500,0x8500),
the BIOS's boot-time copy of ROM 0x1FC10000+) natively with NO check that
the live RAM bytes still match the ROM image the functions were compiled
from — unsound for runtime-patched kernel code (the BIOS installs pad/SIO
handlers INSIDE compiled bodies; live diff: 15 clusters, 348 bytes). It
never misfired only because entry points happened to avoid patched keys.

A static kernel hit now requires psx_kernel_bless_dispatchable(): the
emitter emits psx_bios_kernel_bodies[] (key -> full RAM extent of the code
reachable from it; continuations carry their parent's extent, 499 entries),
and the runtime lazily byte-compares each entry's body against the ROM
source, caches the verdict, and invalidates precisely on writes into the
body (store chokepoint, DMA/EXE range marks, savestate restore). Clean ->
guarded native; patched/unverifiable -> the faithful dirty-RAM interpreter,
exactly as Rule 18 prescribes. This implements the runtime half of the
relocation-manifest contract (docs/RELOCATION_MANIFEST_FORMAT.md).

Measured on Tomba (SCUS-94236): 208-214 of 219 verified functions clean
(the 5 mismatches are the genuinely patched pad-driver family, faithfully
interpreted); millions of guarded native hits per boot; title 1.00x,
village 0.92x (parity — kernel keys were already native, now they are
PROVEN native); LLE and HLE boots + gameplay input verified. Kernel-window
interp measures 1.3% of interp burden (dirty_ram_stats), so this is a
soundness fix, not a speedup; the dominant interp cost is overlay-region
coverage (cg-tag hygiene), tracked separately.

Observability: TCP kernel_bless (entries/clean/mismatch/native_hits/
verifies/invalidations), always-on. PSX_KERNEL_BLESS=0 reverts to
unguarded-table behavior for A/B.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-10 19:12:32 -07:00

256 lines
12 KiB
C

/* dirty_ram_interp.h — interpret-on-dispatch for install-at-runtime RAM.
*
* See CLAUDE.md Rule 18 and docs/dynamic_handler_install.md for the full
* rationale. The PS1 BIOS dynamically writes 4-instruction dispatch stubs
* into kernel RAM (notably RAM 0xCF0 for the SIO data-byte handler). A
* static recompiler can't see those bytes at compile time, so a small MIPS
* interpreter here runs them at dispatch time on the same CPUState.
*
* Scope: this is NOT a fallback for code the recompiler failed to translate.
* It runs only against PCs in pages that have been written-to since boot.
* Static-recompiled code continues to handle ROM-resident code and game
* RAM. See docs/dynamic_handler_install.md for the inline note about a
* potential future migration to runtime JIT (Option B).
*/
#ifndef PSXRECOMP_DIRTY_RAM_INTERP_H
#define PSXRECOMP_DIRTY_RAM_INTERP_H
#include <stdint.h>
#include "cpu_state.h"
#ifdef __cplusplus
extern "C" {
#endif
/* Returns 1 if `addr` lies in a dirty kernel-RAM page and the interpreter
* ran a basic block at that PC. Returns 0 if `addr` is clean (caller must
* fall back to the static dispatch table). On return with 1, cpu->pc is
* either 0 (block ended on jr $ra style return) or the target of a tail
* jump that the dispatch trampoline should re-enter.
*
* `stop_addr` is the dispatch loop's return contract (psx_dispatch_call's
* return_addr; 0 for plain tail dispatches). The interpreter's overlay
* local-flow chaining MUST NOT run past it: when a transfer (or fall-
* through) reaches stop_addr, the interpreter exits with cpu->pc ==
* stop_addr so the dispatch loop returns into the suspended native
* caller. Skipping this check lets the interp execute the native
* caller's tail itself; the suspended native host frame then resumes and
* double-executes that tail against a moved guest stack, restoring
* garbage callee-saved registers (the pig-throw blue-screen/MMIO-fatal
* corruption, 2026-06-10). */
int dirty_ram_dispatch(CPUState* cpu, uint32_t addr, uint32_t stop_addr);
/* Overlay-cache windows — the address ranges eligible for capture, offline
* recompilation, and per-entry-validated native execution (Rule 18 code that
* does not exist in any compile-time image):
*
* Kernel RAM [0x00000, 0x10000): the BIOS part-2 image relocated to RAM
* plus install-at-runtime stubs (e.g. the SIO data-byte stub at 0xCF0).
* Dirty-tracked per CPU store (dirty_ram_mark_kernel_write).
* Overlay region [OVERLAY_REGION_FLOOR, RAM_SIZE): game overlays loaded by
* CD DMA (dirty_ram_mark_executable_range).
*
* Main-EXE text [0x10000, OVERLAY_REGION_FLOOR): CLEAN pages are statically
* recompiled and never captured. But games that discard their boot/init code
* load gameplay overlays OVER this region (Tomba 2: hot gameplay code at
* 0x1DA88/0x24548/0x3116C, well below its 0x38800 floor). Post-baseline
* (dirty_ram_clear_image_baseline) a DIRTY page here is by definition live
* overlay code — the compiled image is stale for it — so dirty boot-text
* pages ARE a capture/candidate window (2026-07-06; before this they could
* only interpret, which dominated wall time at ~300ns/insn).
*
* The interpreter's LOCAL-FLOW gates (is_local_dirty_target /
* phys_is_overlay_flow_region) use DIRTY_RAM_KERNEL_WINDOW_END alone: only
* the kernel window stays per-block (it runs in exception context where the
* verified dispatch cadence is delicate). Native coverage for the kernel
* window comes from the overlay loader, not from interp chaining. */
#define DIRTY_RAM_KERNEL_WINDOW_END 0x00010000u
/* The overlay-region floor is the END of THIS game's statically-recompiled
* main-EXE text (phys). Everything at/above it is runtime-loaded overlay code,
* eligible for in-interpreter local-flow chaining and overlay-cache capture.
*
* It is a RUNTIME value (g_overlay_region_floor), NOT a compile-time constant,
* because it differs per game: Tomba 1's text ends at 0x98000 (0x10000+0x88000)
* but Tomba 2's boot EXE is only 0x28800 (text ends at 0x38800). A hardcoded
* 0x98000 misclassified Tomba 2's overlays (loaded at 0x85000+) as main-EXE
* text, forcing them onto the slow block-by-block dispatch path AND through the
* bail-prone non-local-call contract — the Whoopee-Camp splash freeze. main.cpp
* sets g_overlay_region_floor = (load_address + text_size) & 0x1FFFFFFF at game
* load; OVERLAY_REGION_FLOOR_DEFAULT applies for BIOS-only runs. */
#define OVERLAY_REGION_FLOOR_DEFAULT 0x00098000u
extern uint32_t g_overlay_region_floor;
#define OVERLAY_REGION_FLOOR (g_overlay_region_floor)
/* Test whether a given physical kernel-RAM address is in a page that was
* written-to since boot. Defined in memory.c. */
int dirty_ram_is_dirty(uint32_t phys);
/* Kernel-image bless (memory.c). A dispatch key in the relocated kernel
* window (RAM [0x500,0x8500), the BIOS's boot-time copy of ROM 0x1FC10000+)
* may run its statically-recompiled native function IFF the live RAM bytes
* of the code reachable from it (emitter-supplied body extent) byte-match
* the ROM source. Lazily verified, cached per entry, invalidated precisely
* on writes into the body. Runtime-patched bodies (pad/SIO install stubs)
* never verify and keep interpreting — faithful either way. */
int psx_kernel_bless_dispatchable(uint32_t phys);
void psx_kernel_bless_note_range(uint32_t phys, uint32_t len);
void psx_kernel_bless_stats(uint64_t out[6]);
/* Capture/candidate window membership. Kernel window and overlay region are
* unconditional; main-EXE text [KERNEL_WINDOW_END, FLOOR) is included only
* when the page is dirty — i.e. a gameplay overlay overwrote the boot text
* after the game-start baseline, so the bytes there are runtime overlay code
* (see the window model note above). Clean boot text never enters the window
* (it runs compiled; capturing it would violate the static-first design). */
static inline int overlay_cache_window_contains(uint32_t phys) {
return phys < DIRTY_RAM_KERNEL_WINDOW_END
|| phys >= OVERLAY_REGION_FLOOR
|| dirty_ram_is_dirty(phys);
}
uint32_t dirty_ram_get_bitmap(void);
uint32_t dirty_ram_get_bitmap_word(uint32_t word_index);
uint32_t dirty_ram_get_bitmap_word_count(void);
void dirty_ram_mark_executable_range(uint32_t phys, uint32_t len);
void dirty_ram_register_text_image(uint32_t phys_lo, const uint8_t *bytes,
uint32_t len);
int dirty_ram_text_native_ok(uint32_t phys);
int dirty_ram_text_image_registered(void);
/* Bless an intentional runtime data patch (e.g. text_xlate string/glyph tables)
* into the text reference image so it is not mistaken for self-modifying code. */
void dirty_ram_text_bless(uint32_t phys, const uint8_t *bytes, uint32_t len);
uint64_t dirty_ram_text_native_blocked(void);
uint32_t dirty_ram_text_diverged_pages(void);
/* Counters for visibility / TCP debug. Increment in interpreter; expose
* via debug_server.c if helpful. */
extern uint64_t g_dirty_ram_blocks_run; /* basic blocks interpreted */
extern uint64_t g_dirty_ram_insns_run; /* instructions interpreted */
extern uint64_t g_dirty_window_dispatches; /* interp dispatches inside a
* capture window (autocapture
* pressure signal, step 2.8) */
extern uint64_t g_dirty_ram_aborts; /* unsupported-opcode aborts */
extern uint64_t g_dirty_ram_guard_yields; /* long dirty loops yielded */
extern uint64_t g_dirty_ram_unsupported_midblock;
extern uint32_t g_dirty_ram_last_unsupported_entry;
extern uint32_t g_dirty_ram_last_unsupported_entry_ra;
extern uint32_t g_dirty_ram_last_unsupported_entry_sp;
extern uint32_t g_dirty_ram_last_unsupported_insns;
extern uint32_t g_dirty_ram_last_unsupported_pc;
extern uint32_t g_dirty_ram_last_unsupported_insn;
extern const char *g_dirty_ram_last_unsupported_reason;
/* Per-entry-PC counters. Aggregate counters above hide which install
* stubs actually fire — a single noisy spurious-dispatch site can mask
* a legitimate handler that never runs. This open-addressed table keys
* on the entry PC of each interpreted block. */
/* 65536 -> 262144 (2026-07-03): Tomba2's Trolley attract demo interprets a
* working set past 64K distinct PCs; a full table degraded every lookup to a
* full-table scan (0.3 fps host burn in pc_table_get_or_insert_in) and
* silently truncated overlay-capture seed lists (MAX_CAPTURE_PCS aliases
* this). Probe length is bounded in pc_table_get_or_insert_in either way. */
#define DIRTY_RAM_PC_TABLE_SIZE 262144
typedef struct {
uint32_t pc; /* entry PC, 0 = empty slot */
uint64_t hits; /* number of times dispatched here */
uint64_t insns; /* total instructions executed across hits */
uint64_t entry_hits; /* dispatches that arrived from NATIVE code (call /
* dispatch loop), not from the interpreter's own
* block-to-block chaining. `hits` conflates the two:
* below the local-flow floor every taken branch
* re-dispatches, so chain blocks dominate. entry_hits
* is the evidence stream for interior-alias seeds. */
} DirtyRamPcEntry;
extern DirtyRamPcEntry g_dirty_ram_pc_table[DIRTY_RAM_PC_TABLE_SIZE];
/* Companion table: every PC the interpreter actually executes (not just block
* entries). overlay_capture uses both to report execution-verified seeds. */
extern DirtyRamPcEntry g_dirty_ram_exec_pc_table[DIRTY_RAM_PC_TABLE_SIZE];
/* Block-entry ring buffer. Records every dispatch into dirty RAM with the
* caller's RA at entry, plus argument context — answers
* "who tried to JALR into this RAM stub, with what args".
* Always-on, eviction keeps memory bounded; callers query the window of
* interest (CLAUDE.md global rule on ring buffers).
*
* Limitation: this captures dispatches into RAM-resident code only. ROM
* recompiled-C → recompiled-C calls (direct C function calls in
* generated dispatch table) are NOT captured here. For BIOS shell
* investigation that's fine — shell code lives in RAM (0x800XXXXX).
*
* `ra` is cpu->gpr[31] at dispatch time. For normal JAL-style calls,
* (ra - 8) gives the caller's PC. For J/JR/tail-calls it's only a
* heuristic — treat ra as "ra_callsite_guess", not authoritative. */
#define DIRTY_RAM_BLOCK_LOG_CAP (1u << 22) /* 4M entries (~128 MB).
* At ~580K dispatches/s during boot, this retains ~7s of history; at
* ~10K/s during modal idle, ~400s. Sized for retroactive press-window
* analysis without the prior 16K ring's 28-ms eviction problem. */
typedef struct {
uint64_t seq; /* monotonic, unique per entry */
uint32_t target; /* entry PC (RAM address) */
uint32_t ra; /* cpu->gpr[31] at dispatch — caller's return target */
uint32_t a0; /* cpu->gpr[4] at dispatch */
uint32_t a1; /* cpu->gpr[5] at dispatch */
uint32_t a2; /* cpu->gpr[6] at dispatch */
uint32_t a3; /* cpu->gpr[7] at dispatch */
uint32_t t0; /* cpu->gpr[8] at dispatch */
uint32_t t1; /* cpu->gpr[9] at dispatch */
uint32_t t2; /* cpu->gpr[10] at dispatch */
uint32_t sp; /* cpu->gpr[29] at dispatch */
uint32_t frame; /* s_frame_count at the time of dispatch */
} DirtyRamBlockLogEntry;
extern DirtyRamBlockLogEntry g_dirty_ram_block_log[DIRTY_RAM_BLOCK_LOG_CAP];
extern uint64_t g_dirty_ram_block_log_seq;
#define DIRTY_RAM_FLOW_LOG_CAP (1u << 16)
typedef struct {
uint64_t seq;
uint32_t pc;
uint32_t target;
uint32_t ra;
uint32_t a0;
uint32_t a1;
uint32_t a2;
uint32_t a3;
uint32_t sp;
uint32_t frame;
} DirtyRamFlowLogEntry;
extern DirtyRamFlowLogEntry g_dirty_ram_flow_log[DIRTY_RAM_FLOW_LOG_CAP];
extern uint64_t g_dirty_ram_flow_log_seq;
#define DIRTY_RAM_INSN_LOG_CAP (1u << 16)
typedef struct {
uint64_t seq;
uint32_t pc;
uint32_t insn;
uint32_t next_pc;
uint32_t target;
uint32_t before_s0;
uint32_t after_s0;
uint32_t sp;
uint32_t ra;
uint32_t v0;
uint32_t v1;
uint32_t a0;
uint32_t a1;
uint32_t a2;
uint32_t a3;
uint32_t t0;
uint32_t t1;
uint32_t t2;
uint32_t current_tcb;
uint32_t task_ptr;
uint32_t task_mode;
uint32_t task_submode;
uint32_t frame;
uint8_t transferred;
uint8_t pad[3];
} DirtyRamInsnLogEntry;
extern DirtyRamInsnLogEntry g_dirty_ram_insn_log[DIRTY_RAM_INSN_LOG_CAP];
extern uint64_t g_dirty_ram_insn_log_seq;
#ifdef __cplusplus
}
#endif
#endif /* PSXRECOMP_DIRTY_RAM_INTERP_H */