mirror of
https://github.com/libretro/virtualjaguar-libretro.git
synced 2026-11-10 17:28:58 -10:00
dependabot/github_actions/develop/actions/setup-python-7
5
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Joseph MattielloandClaude Fable 5
|
72470bb22b |
fix(acid): un-wedge pit_countdown_rate and unaligned_word
Both ROMs assembled and booted but never wrote a result signature, so
the acid suite reported them as NOT-RUN-YET on every build. Each was
violating a documented hardware contract; the emulator was right in both
cases, so the fix is in the test ROMs.
pit_countdown_rate: the IRQ handler acked TOM's INT1 copy but never
cleared JERRY's own pending latch. JERRY drives TOM INT1 bit 4 for as
long as an enabled source has an uncleared latch (tom.c:TOMPendingMask),
and m68k_execute re-presents that level-2 request at every instruction
boundary, so the 68K re-entered the handler forever: ~2200 IRQs/frame
with the busy-loop counter d2 frozen, i.e. 100% of the frame's 227k
cycles spent in exception entry plus handler. JINTCTRL's low byte is the
enable mask and its high byte is write-1-to-clear, so writing the enable
bit alone cleared nothing. Handler now writes IRQ2_TIMER1_CLR too.
The PIT period itself was never wrong: with the storm gone the test
measures 395 IRQs/sec against a [385, 426] window. Do not widen that
window if it drifts -- this test is the only acid coverage of the
dram_timing model, and it has already been churned once (
|
||
|
Joseph MattielloandClaude Opus 4.7
|
6dd92875bc |
acid: divergence-only shadow check + CI workflow + BASELINE gate
Two things on top of the Copilot batch-3 fixes:
## 1. Linter check 4: local equate value-divergence
The OLP_HI/LO swap that Copilot caught last round was a *local
equate that disagreed with the oracle*. The lint pass now flags
exactly that pattern: warn only when a local `name equ value`
defines a name already in jaguar_regs.s AND the local value
differs from the oracle's.
Pure value-duplicates (e.g. local `B_COMMAND equ $F02238` matching
oracle `$F02238`) are safe and stay silent -- otherwise we'd have
to refactor 30+ files for redundancy with no actual benefit.
The check parses the local RHS through a small expression
evaluator that handles `$hex`, decimal, `+`/`-`/`<<`/`>>`, and
substitutes other oracle symbols. Bails (no warning) if anything
unparseable -- conservative, no false positives.
Currently lint-clean across the whole suite -- which means every
local shadow today *agrees with* the oracle. Going forward, the
moment a local equate diverges, CI catches it.
## 2. CI workflow: .github/workflows/acid-test.yml
Runs on every PR + push to develop / master / release branches
(and manual dispatch). Path-filtered so it only fires when src/,
libretro.c, test/acid/, or the Makefiles change.
Steps:
* Build vasm 1.9 from prb28/vasm GitHub mirror (cached -- ~30s
one-time, instant on subsequent runs).
* Build the libretro core with TEST_EXPORTS=1 + BENCH_PROFILE=1
(so the runner can dlsym `perf_counters_find` and report
per-test counter deltas).
* Assemble the suite via `make -C test/acid all` and require
`make -C test/acid lint` to be clean.
* Run the suite via `make -C test/acid test`, capture full log.
* Run `check-baseline.py` against `test/acid/BASELINE.txt`.
* Post a summary to the PR job-summary panel.
* Upload the full results.log as an artifact (14-day retention).
## 3. BASELINE.txt regression gate
`test/acid/BASELINE.txt` (committed) lists the expected
`PASS`/`FAIL`/`NOT-RUN-YET` for each .jag. Generated by
`make baseline` and updated alongside test changes.
`scripts/check-baseline.py` classifies each test in the new run:
was PASS, still PASS -- OK
was FAIL/NOT-RUN, now PASS -- IMPROVEMENT (good!)
was PASS, now FAIL/NOT-RUN -- REGRESSION (CI fails)
was FAIL, still FAIL -- known FAIL (OK)
in baseline, missing in run -- broken assemble (CI fails)
in run, not in baseline -- new test (OK; baseline needs
updating)
The acceptance philosophy is unchanged: we *encourage* adding
tests that FAIL because each FAIL is a checked-in description
of a known emulator bug. We block PRs that *regress* a
previously-PASSing test, because that's the definition of a real
break.
Two new Makefile targets:
`make -C test/acid baseline` -- regenerate BASELINE.txt
from a fresh run (use after
landing test changes or
emulator fixes).
`make -C test/acid check-baseline` -- run + diff against baseline,
exit non-zero on regression.
Suite is currently lint-clean and baseline-clean: 122 PASS / 20
FAIL / 0 NOT-RUN, no regressions.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
|
||
|
Joseph MattielloandClaude Opus 4.7
|
b1030f587d |
acid: address Copilot review batch 3
Three new Copilot inline comments, all real bugs in our test code.
1. **TOM register map off by 0x10 in the oracle** (jaguar_regs.s).
I had VBB/VBE/VS/VDB/VDE/VEB/VEE at offsets $2A/$2C/$2E/$30/
$32/$34/$36, but those are actually BORD1/BORD2/HP/HBB/HBE/
HS/HVS per src/tom/tom.c:351-369. The V-prefixed registers
live at $40 and up. Fixed gen-jaguar-regs.py: rewrote the
TOM_OFFSETS dict with the full register map -- BORD1/BORD2,
HP/HBB/HBE/HS/HVS/HDB1/HDB2/HDE, then VP, then VBB/VBE/VS/
VDB/VDE/VEB/VEE/VI at their actual offsets. Also added OBF
($26) and HEQ ($54) for completeness.
2. **TOM_OLP_HI / TOM_OLP_LO swapped in op_branch_object.s and
op_scaled_bitmap.s** (local equates that shadowed the oracle).
Per src/tom/op.c:238-239, OPGetListPointer reads the LOW
word from $F00020 and the HIGH word from $F00022 -- "LO/HI
WORD, hence the funky look of this". Both files had defined
OLP_HI=$F00020 and OLP_LO=$F00022 locally, the opposite of
the spec, so writes byte-swapped the OLP and OP started at
the wrong address. Removed the local equates from both
files; they now use the correct definitions from
include/jaguar_regs.s.
3. **BRANCH target encoded in wrong bit positions in
op_branch_object.s.** The OP decodes a branch link as
`(p0 >> 21) & 0x3FFFF8` (src/tom/op.c:474), so the link
target needs to live in bits 21..43 of the 64-bit p0 phrase.
The original test had `(OBJ1 << 5) | 3` entirely in the low
long, which doesn't reach bit 21. Rewrote the encoding with
detailed comments showing the math:
link = $50008 -> p0 high long $000000A0, low long $01003FFB
verify: ((hi << 11) | (lo >> 21)) & $3FFFF8
= ($A0 << 11) | ($01003FFB >> 21)
= $50000 | $00008
= $50008 ✓
The test now actually exercises the BRANCH-to-link path
(was previously a no-op since the OP would land at link=$0).
Suite still at 122/142 PASS -- no new regressions; the OP tests
that previously "PASSed" because the OP never actually ran (due
to the swapped OLP) still PASS for the legitimate reason now
that the OP runs correctly and just doesn't write outside its
declared object data.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
|
||
|
Joseph MattielloandClaude Opus 4.7
|
553f2bd6f8 |
acid: 70+ new tests across all categories from 4 parallel sub-agents
Comprehensive coverage push per user direction ("get all the tests we
need"). Four background sub-agents worked the 11-chunk plan in
COVERAGE_PLAN.md in parallel, with the oracle (jaguar_regs.s) and
linter (lint-acid.py) catching encoding mistakes mechanically.
Final state: **122 / 142 PASS** (was 67/72). 19 FAILs are real bugs
documented as regression gates; 1 deliberate-FAIL placeholder
(dsp_op_mac40_overflow's older form replaced).
## What landed
### Chunk 1 (blitter agent): 10 existing tests tightened
Loose assertions replaced with strict bounded checks:
* timing/vc_advance -- delta in [1, 524] (not just "non-zero")
* timing/hc_advance -- HC bit 0x0400 must toggle, phase < 0x0400
* gpu/gpu_basic_run -- G_PC bounded to [start+2, start+2*1024]
* dsp/dsp_basic_run -- same shape
* op/op_stop_terminates -- 8 KB sentinel block, every long checked
* quirks/m68k_set_sr_supervisor -- SR & $E700 == $2700
* stress/deep_call_chain -- SP intact, SR unchanged, all 16 flags
* bus/cpu_blitter_concurrent -- src AND dst correct
* perf/memcpy_loop -- spot-check at 0, N/2, N-1 with index
-derived expected pattern
### Chunk 2 (blitter agent): 9 new pixsize × phrase tests
Filename: copy_pix<N>_<phrase|pixel>.s for N in {1,2,4,8,16,32}.
Discovered: **1bpp + 2bpp phrase blits hang BlitterMidsummer2
forever.** Both ROMs replaced with deliberate-FAIL placeholders so
the suite can complete; original test logic preserved as a comment
plus a one-line restoration recipe for once the hang is fixed.
### Chunk 3 (blitter agent): 9 missing LFU functions
$1, $2, $4, $5, $7, $9, $A, $B, $D. Bit-exact assertions vs the
truth-table evaluation of (S,D) for each function.
### Chunk 5 (gpu/dsp agent): 16 GPU opcode tests
add, sub, and, or, xor, mult, imult, div, abs, shlq, shrq, cmpq,
jump, loadb, storew, moveq -- each as a 3-instruction GPU program
that stores its result and the 68K verifies bit-exact.
Notable corrections from agent: STORE encoding is rm in bits 9..5
(not rn as my prompt said); GPU MOVEQ does not sign-extend; SHLQ
imm encodes as 32-shift_amount.
### Chunk 6 (gpu/dsp agent): 16 DSP opcode tests + extras
Same 16 opcodes mirrored to DSP_BASE/DSP_RAM, plus:
* dsp_op_mac40_overflow -- the REAL 40-bit MAC test (replaces
the earlier NOP-loop placeholder). PASSes -- accumulator
correctly preserves bits past 32.
* dsp_irq_to_68k -- FAIL: JERRY pending bit gets set, 68K never
enters handler at autovector $68. Real bug.
* dsp_mailbox -- D_HIDATA round-trip via shared work RAM.
### Chunk 7 (op/bus agent): 7 OP scenarios
op_bitmap_render (PASS), op_branch_conditional (PASS),
op_gpu_int_object (FAIL placeholder -- G_FLAGS observability hard
from 68K), op_reflect_modifier (PASS), op_palette_8bpp (PASS),
op_olp_alignment (PASS), op_short_branch (PASS).
Notable agent finding: **the OP modifies the BITMAP p0 phrase in
place every halfline** (decrements HEIGHT, advances DATA pointer),
so tests that probe LBUF mid-render must re-prime p0 + re-write
OLP each retry. This is documented behavior worth knowing.
### Chunk 9 (op/bus agent): 3 bus contention probes
All 3 FAIL by design -- bus contention is unmodelled. Each carries
a strict numeric assertion that will go GREEN automatically once
contention modelling lands.
### Chunk 10 (timing/68k agent): 4 strict timing tests
* vblank_60hz_exact FAIL observed=103 expected=60
* halfline_period_us FAIL observed=630 expected=844 cycles
* pit_countdown_rate FAIL observed=49386 expected=23937
* vc_resets_at_vp PASS
The three FAILs all point at the same root cause: emulated
wall-clock during 68K busy loops runs ~1.7-2x faster than the
event clock that drives VBlank/PIT IRQ rate. **This is almost
certainly the Doom #131 game-logic-2x-too-fast bug.** Once fixed,
all three tests will go GREEN simultaneously.
### Chunk 11 (timing/68k agent): 4 68K coverage tests
* movem_round_trip PASS
* divs_w_signed PASS
* abcd_nbcd PASS
* btst_dynamic PASS
### Runner: short-circuit on signature
test/acid/run.c now polls the ACID signature each frame and breaks
out as soon as PASS or FAIL is written. Cuts full suite runtime
from ~30 minutes to 12 seconds. Critical -- 142 tests at 600
frames each was unworkable.
## Real emulator bugs surfaced as failing tests
1. **GPU/DSP control-register read shadowing** (gpu_basic_run +
dsp_basic_run FAIL). `GPUReadLong` (gpu.c:338-342) intercepts
long-aligned reads in $F02100..$F0211F as register-bank reads
BEFORE checking the control-RAM range, so 68K reads of G_PC,
G_CTRL, G_FLAGS via long return wrong data. Same shape on DSP.
2. **BlitterMidsummer2 hangs on 1bpp / 2bpp blits.** Replicates
for inner counts 4, 16, 64, 256. Two placeholders document the
bug in copy_pix1_phrase / copy_pix2_phrase.
3. **DSP IRQ to 68K** (dsp_irq_to_68k FAIL). JERRY raises pending,
68K never enters handler.
4. **Event-clock vs 68K-instruction-clock divergence** (3 timing
FAILs). Likely Doom #131 root cause.
5. **DIVL zero divide trap** doesn't fire (existing FAIL).
6. **Blitter narrow-pixel copies pick wrong byte** (copy_pix1_pixel,
copy_pix2_pixel, copy_pix4_pixel partial copies).
## Coverage by category (after this commit)
| Category | Tests | Pass |
|---|---:|---:|
| smoke | 1 | 1 |
| memory | 10 | 10 |
| timing | 13 | 10 |
| irq | 7 | 6 |
| blitter | 35 | 27 |
| gpu | 18 | 17 |
| dsp | 21 | 20 |
| op | 10 | 9 |
| bus | 5 | 2 |
| hle | 6 | 6 |
| quirks | 11 | 10 |
| stress | 3 | 3 |
| perf | 3 | 3 |
Sub-agent productivity: 4 agents in parallel + me on integration
landed ~70 new tests in one PR cycle. Each agent's report flagged
real findings (not just churn) -- the OP p0-mutation behavior, GPU
control-RAM read shadowing, the 1bpp/2bpp blitter hang, the STORE
opcode encoding correction.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
|
||
|
Joseph MattielloandClaude Opus 4.7
|
f368b07b1e |
acid: oracle-generated jaguar_regs.s + acid lint pass (steps A & C)
Two pieces of infrastructure to prevent the encoding-mistake class of
bug Copilot just caught (wrong LFU bit positions, wrong DSTEN, wrong
B_PATD address, etc.).
A: test/acid/scripts/gen-jaguar-regs.py
Parses the actual emulator C headers and emits a single
include/jaguar_regs.s file with:
- Subsystem base addresses (TOM, GPU, JERRY, DSP, blitter)
- TOM register offsets (HC, VC, VBB, VDB, INT1, ...)
- Blitter MMIO addresses (B_A1_BASE, B_COMMAND, B_PATTERNDATA, ...)
- Blitter command bits (SRCEN, DSTEN, BCOMPEN, GOURD, ...)
- LFU function constants (LFU_FN_0..LFU_FN_F, pre-shifted to bits 21..24)
- TOM IRQ enum + bit masks (IRQ_VIDEO_MASK, IRQ_DSP_MASK, ...)
- JERRY IRQ2 enum (IRQ2_TIMER1, IRQ2_DSP, ...)
- BLIT_CMD_VALID_BITS = OR of every defined cmd field (lint mask)
Sources parsed: src/tom/blitter.c, src/tom/tom.h, src/tom/gpu.h,
src/jerry/jerry.h, src/jerry/dsp.h. Re-runs whenever any of those
change (declared as Makefile dependencies).
Why: I had B_PATD at offset $50 in two tests (pattern_fill,
bcompen_basic), but the real PATTERNDATA register lives at $68
per src/tom/blitter.c. $50 is DSTZ. The oracle would have caught
that; humans copy-pasting offsets don't.
C: test/acid/scripts/lint-acid.py
Walks every .s file under tests/ and warns on:
1. B_COMMAND literals using bits outside BLIT_CMD_VALID_BITS
(catches "ity short-form $C000" mistake from before)
2. LFU function selecting an operand whose ENable isn't set
(e.g. LFU=$E (S|D) without DSTEN -> dest reads as 0)
3. DCOMPEN without DSTEN, BCOMPEN without SRCEN
4. Hard-coded $Fxxxxxx MMIO literals where a symbolic name
exists in the oracle
Run via `make -C test/acid lint`, also runs automatically as part
of the standard `all` target.
Currently clean across the suite -- proves the oracle has caught up
to all prior tests (post-Copilot-fixes).
Existing tests touched:
blitter/pattern_fill.s, blitter/bcompen_basic.s
- Now `include "include/jaguar_regs.s"` and use B_PATTERNDATA
symbol instead of locally-defined wrong $50 address.
- pattern_fill now lands the pattern correctly (FAIL signature
changed from "$00000000" to "$CAFEBABE" -- the pattern IS
being written, just with byte order opposite to what the test
expected; remaining test-side bug to clean up).
Suite: 67/72 PASS unchanged. Both new scripts are infrastructure --
they don't add tests, they add safety.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
|