Two things land together: the plan and contract draft P6 needs before it starts, and the
working notes that until now existed on exactly one machine.
1. P6's plan and contract draft (neither started; not one line of code)
docs/Disaggregated/P6-SPAWN-PLAN.md the package plan
docs/Disaggregated/P6-CONTRACT-DRAFT.md the contract draft (moves to
MG_Remote/CONTRACT-P6.md when c6 lands)
The three findings that mattered, all written into the documents:
a. P6 is SMALLER than its roadmap cell suggests. The transport primitives are already in
the tree and already tested: SocketDoorbell including hangup detection and the two
lost-wakeup fences (Doorbell.cpp:211-360, four cases in FdPassingTest), SCM_RIGHTS fd
passing, ShmSegment::Adopt, and the SurfaceOp / WindowKind schema already defined in
protocol.fbs. inproc uses a real second shm mapping rather than an alias, and
SessionRings.h:22-28 says that was done so today is not the first time it runs. P6 is
assembly plus a control plane plus a process, not "write a socket transport".
b. The evidence for "P6 is only a transport swap" has EXPIRED. It was bought by P5c, whose
audit found 59 direct memory accesses in a tree that already looked clean. P5e changed
the tree again. So the first package, a6, is a read-only audit that writes no code, and
every contract row that depends on it is marked PENDING a6 rather than decided in
advance.
c. P6 does not get a game running on the phone. ANativeWindow* is meaningless in the child,
and protocol.fbs:186-196's own comment says Android transfers the window out of band -
that band is P12. P6 lands pbuffer / surfaceless / offscreen, which is what HeadlessGL,
the split-equivalent lane and trace replay need, and which is exactly what P6's own exit
gate measures. A real window reaching the server is a named refusal. This gets its own
section because "P6 is done" would otherwise read as "the game runs now".
The contract adds rule G: a role may not name a process-local handle. It extends P5c's
rule E (a role may not name the other role's memory) from memory to identity - EGLDisplay /
EGLSurface / EGLContext / ANativeWindow* / fd number / pid / mapped address either cross as
a token this contract defines or they do not cross, and a descriptor crosses only through
SCM_RIGHTS, where the kernel renumbers it.
What P5e left for P6 gets its own section. The latch semantics are already designed
(ARCHITECTURE.md:504) and are copied unchanged. What is new is that after run-ahead the
client no longer waits while the peer process can DIE: the client holds records that were
published and will never be applied, and Android's freezer suspends the child without
closing anything. So slow and dead must be distinguishable WITHOUT a timeout threshold.
That is P6's only new design; everything else is carriage. It shares a predicate with E1's
outstanding re-specification (ID-122), and the two should be designed together.
Two gates carried over from P5e's lessons: G2 requires integration-spawn's test-name set to
match integration-split's name for name, because a differing set means the lane is quietly
running less; and the arm-proof gate requires every entry to record the child pid and
transport=spawn in its OWN private log - ID-124 established that this kind of false green
really happens, and a spawn lane that actually ran monolith would pass every other gate.
S2's shape must be settled before it is written, and the document says why: E1 is the
cautionary case.
2. The working notes move into the repository (docs/Disaggregated/notes/, 415 .md files)
ROADMAP / CURRENT_STAGE_PROGRESS / MEASUREMENTS cite ruling numbers, audit line numbers and
the provenance of gate figures from ~/w7/notes/, a directory that exists on one machine and
is under no version control. A record cited hundreds of times that lives only locally is not
a record, so it moves in whole, with its directory structure intact.
Only .md moves. Scripts, run logs, census and gate result JSON, patches, APKs and nm dumps
stay where they are: they are evidence DATA rather than documents - large, regenerated per
run, and mostly carrying absolute paths and device serials that mean something only on the
machine that produced them. The mapping and this split are written in
docs/Disaggregated/notes/README.md.
Two exceptions are recorded there too: P6's BRIEF has been promoted to a first-class
document and keeps no copy under notes/, so the two cannot drift apart; and recovered/ holds
material salvaged from failed workflow runs, which may duplicate the like-named phase
directories and is kept because a few of those were the only complete version at the time.
CI's citation lint takes docs/Disaggregated/*.md only (non-recursive), and notes/ is
deliberately outside it: those notes' file:line citations were written against their own
code heads, and resolving them against today's HEAD would fail wholesale - which is not an
error, it is their correct state as a historical record. To check one, use the revision it
declares (--rev).
3. Inconsistencies fixed in passing
- README.md's status line still said P5e was in progress while the stage table already said
closed.
- CURRENT_STAGE_PROGRESS's next-step item 1 still listed the exit gate as pending, though
(a) and (b) were run in the previous commit; rewritten to closed plus the one item that
actually remains (E1).
- ROADMAP's P6 row still had the pre-phase plan in three of its cells; it now carries the
package order, the scope boundary and the two new gates.
Checked with check_doc_citations.py: 125 citations across docs/Disaggregated/*.md, 0 problems.
29 KiB
P2 package C — p2/espryt (Espryt 0b, the first Track H slice) — result v2 (rework round 2)
Tree: ~/w7/p2-espryt, branch p2/espryt, base tag p2/contract (9c6a8a25).
Build dirs now present: build-linux (pull, the G1 build), build-push, build-verify
(-DMOBILEGL_PIPE_VERIFY=ON -DMOBILEGL_ITEST_REQUIRE_GPU=ON), build-nolegacy
(-DMOBILEGL_PIPE_PUSH=ON -DMOBILEGL_PIPE_LEGACY_MEMOS=OFF) and build-bench
(-DMOBILEGL_BUILD_BENCHMARK=ON -DMOBILEGL_PIPE_PUSH=ON, DriverBench only).
The last three are new in this round and exist to answer review minors 9 and 12 and MAJOR 4.
build-nolegacy/ shows up as an untracked directory in git status; it is a build dir, not
content, and is not committed.
1. Commits
Rework round 2 added two commits on top of the v1 four:
| sha | subject |
|---|---|
dc9b7237 |
[Feat] (Espryt): give the backend a dense {slot, gen} twin table beside the address-keyed registry |
fbdaeef3 |
[Refactor] (Espryt): key every backend twin on {slot, gen} instead of the frontend object heap address |
1e0f4d6d |
[Test] (Espryt): pin the twin table identity contract - a reclaimed slot is a new handle and the stale one resolves to nothing |
d714600a |
[Fix] (Espryt): do not return a reference through a null slot pointer, and stop a comment claiming a memo that is no longer there |
d89fb684 |
[Fix] (Espryt): sweep the twin table on object churn again, refuse to run an arm the operator disabled, and give the hot lookups back their array probe |
5f245ac7 |
[Test] (Espryt): pin the churn-driven sweep cadence, the null tolerance, and the one slot per object the two Track H slices share |
Files touched across all six — still exactly package C's rows of the C.5 ownership table:
MG_Backend/DirectGLES/{SlotTables.h (new), Managers.h, Managers.cpp, DirectGLES.cpp} and
MG_Test/SanityTest.cpp. No file owned by another package was edited in this round either.
2. What the two new commits do, review item by review item
MAJOR 1 — the churn-driven sweep is back (FIXED)
BackendSlotTable now carries both sweep drivers, exactly as the map arm does:
kGCInterval = 1024 on the draw-path tick and kCreationGCInterval = 64 on first-time
insertions, swept at the top of GetOrCreate before the Entry& exists (same reason the map
arm sweeps there: EntryAt can grow m_slots and move every element). CollectGarbage{IfNeeded, Now} reset the creation tick when they sweep, so the two drivers do not double-count.
Pinned by a new case that never calls CollectGarbage* at all:
DirectGLESSlotTable.ObjectChurnAloneDrivesTheSweep churns 256 objects through GetOrCreate,
letting each die immediately, and asserts (a) at most interval + 2 live entries ever coexist,
(b) the table ends at most that big, (c) the allocator's HighWater for the kind grew by at most
interval + 2 rather than by 256.
Negative control run (ObjectChurnAloneDrivesTheSweep must be able to go red for the reason
it exists): with if (m_creationTick >= kCreationGCInterval) changed to if (false && …) and
build-push rebuilt, ctest -R DirectGLESSlotTable gives
The following tests FAILED:
88 - DirectGLESSlotTable.ObjectChurnAloneDrivesTheSweep (Failed) unit
— it is the only one of the seven that fails — and after cp /tmp/slot.memo …; cmake --build,
7/7 Passed again. (A first attempt at this control raised kCreationGCInterval to 0xFFFFFFFF
instead; that made the test itself loop 2^32 times, so the churn count is now a fixed 256
with an ASSERT_LT(interval, kChurn) guard, and the control fails fast instead of hanging.)
v1 §4.1 point 2 was wrong and is retracted. The memory behaviour of the handle arm is now the same as the map arm's, not worse; what is still missing is the improvement e2 would bring (death announced rather than discovered) — see MAJOR 3.
MAJOR 2 — Fatal{PipeLegacyMemosDisabled} now stops, and stops early (FIXED)
MGLOG_F(...) is followed by std::abort() (Managers.cpp), matching the codebase's own
Fatal convention (MG_Impl/Pipe/PipeFill.cpp's BadKnob and its verify trap). The resolution is
also forced at backend context creation (InitDisplayAndContext, first statement after
DestroyEGLContext()), so a process that never twins anything still learns its knobs left it
with no arm.
Reviewer's exact reproduction, re-run:
| command | v1 (reviewer) | v2 |
|---|---|---|
MOBILEGL_PIPE_PUSH=0 MOBILEGL_PIPE_LEGACY_MEMOS=0 ./MobileGLIntegrationTest --gtest_filter=CrossFrameBufferScenario.VertexBufferSubData |
exit=0, [ PASSED ] 1 test, Fatal at line 317/319 |
exit=0, [ PASSED ] 0 tests, [ SKIPPED ] 1, harness prints "the EGL bring-up ABORTS on this platform: a forked pre-flight child died on signal 6 (Aborted)"; the Fatal is the last line of the log (69/69), i.e. at context creation |
… ctest -L integration-gpu -R 'DirectGLES.CrossFrameBuffer' |
100% tests passed … out of 13 |
all 13 (Skipped) |
same + MOBILEGL_ITEST_REQUIRE_GPU=1 (what build-verify is configured with) |
— | 13 (Failed), Errors while running CTest |
So the arm the operator disabled no longer runs, and the A/B lever D18 depends on can no longer
be mis-set silently. The residual sharp edge is the harness's own: without
MOBILEGL_ITEST_REQUIRE_GPU=1, a whole-suite abort presents as skips, which the harness itself
warns about in the same breath ("a run that skipped everything is otherwise indistinguishable
from one that passed"). That is package E's lever and it already exists.
MAJOR 3 — step e2 is still absent (NOT FIXED — still a blocking, integrator-owned deviation)
Unchanged from v1 and for the reason the reviewer accepted: e2 has to be installed in
MG_State/GLState/{Texture,Framebuffer,Renderbuffer,Sampler,Program,VertexArray}State/*, and
C.5 gives every one of those files to package B (or, for VertexArrayObject.h, to D); C.2's own
Files line grants package C none of them. The standing instruction is to stop and record a
blocking deviation rather than cross an ownership line, so that is what this is, again.
What changed is the damage: with MAJOR 1 fixed, e3 landing without e2 is now a no-op on
twin lifetime rather than a regression. ARCHITECTURE.md:363's census still cannot count the six
registries' GC as deleted, and the seven CollectGarbageIfNeeded call sites
(DirectGLES.cpp:1273, 1660, 2025, 2026, 2027, 2857, 2858 at HEAD) still stand, now driving
ReclaimDeadSlots().
Integrator decision required, exactly as the review says: re-assign e2 to package B, or land B
before C and re-open C for one commit. Nothing in the table needs to change when it lands —
ReclaimDeadSlots() becomes the fallback path of an explicit Destroy(handle).
MAJOR 4 — the hash probe, and a cost datum that now exists (FIXED, with an honest caveat on the desktop noise floor)
Two code changes, then three measurements.
(a) EsprytSlotTablesEnabled() is now inline (SlotTables.h) over an out-of-line
ResolveEsprytSlotTablesArm() (Managers.cpp). It was a cross-TU call with no LTO, consulted on
every Find/GetOrCreate/HandleOf/ForEachLive/CollectGarbage*.
(b) HandleOf keeps a one-entry lifetimeId -> handle memo. The three per-draw resolution
paths ask the same table for the same object every draw, so the steady state is an integer
compare plus an array index again — which is what D13 traded the three TwinLookupMemos for. It
cannot serve a stale answer, by two independent arguments: a lifetime id is never handed out
twice (so a recycled heap address cannot hit it, which is what the deleted address-keyed memos
needed a WeakPtr for), and every consumer resolves through FindByHandle, which compares Gen.
It is cleared anyway when the sweep frees the memoised slot.
Measurement 1 — call counts under gdb (the reviewer's own instrument; noise-free).
CrossFrameBufferScenario.VertexBufferSubData on build-push, ctest environment, no
MOBILEGL_PIPE_PUSH override:
| breakpoint | memo OFF (negative control) | memo ON (HEAD) |
|---|---|---|
MGPipeSlotAllocator::FindByLifetimeId |
24 | 6 |
MGPipeSlotAllocator::Acquire |
5 | 5 |
MGPipeSlotAllocator::Allocate |
5 | 5 |
MGPipeSlotAllocator::Free |
0 | 0 |
| the arm selector (out-of-line body) | 1 | 1 (was 92 EsprytSlotTablesEnabled() calls in v1) |
Acquire itself calls FindByLifetimeId (SlotAllocator.cpp:107-111), so 5 of the 6 remaining
probes are inside object creation. The resolution path went from 19 hash probes to 1 (the
cold first lookup that fills the memo). Free is 0 in this scenario because it churns 5 objects
and none of them dies — the sweep cadence is pinned by the unit case above instead.
Measurement 2 — DriverBench, desktop, ~/w7/p2-espryt/build-bench (D.4.3's instrument).
run_driver_bench.sh espryt <lib>, EGL_PLATFORM=surfaceless, Mesa/llvmpipe, taskset -c 20-23,
9 repeats per arm, arms interleaved round-robin so drift is shared. Four arms: pull =
build-linux/libMobileGL.so; push_default = build-push with the default bitmask;
push_bit0 = build-push with MOBILEGL_PIPE_PUSH=0; push_no_espryt = build-push with
MOBILEGL_PIPE_PUSH=0x5f (this slice's bit clear, everything else as default).
| case | T1 min / med | T2 min / med | arm A/B (0x7f − 0x5f), same binary min / med |
|---|---|---|---|
draw_tiny |
+441 / +1294 | +498 / +1408 | −66 / +481 |
draw_multi_vao |
+506 / +1583 | +549 / +2011 | +28 / −1003 |
program_pingpong |
+637 / +1864 | +550 / +509 | +183 / −41 |
tex_pingpong |
+363 / +1849 | +358 / +1278 | +193 / −1802 |
mc_vanilla_draw |
+510 / +1813 | +569 / +801 | −17 / −354 |
mc_state_toggle |
+750 / +4489 | +1130 / +10222 | +40 / +987 |
(ns/op; op = draw for mc_vanilla_draw, 5495 per frame. Raw rows: ~/w7/e2-bench2.csv.)
What this does and does not say. The host is llvmpipe and was running at load average 29–51 on 28 cores (the other P2 packages were building and testing on it throughout), so the within-arm spread of the pull arm alone is 65–95 % of its own median. On that noise floor:
- T1 ≈ T2 on every case by the min statistic (+363…+750 vs +358…+1130 ns/op), i.e. essentially all of the pull→push cost is P1's residual fill, and P2 adds nothing measurable on top.
- The same-binary arm A/B, which differs only in this slice's arm, is −66…+193 ns/op by the min statistic and changes sign by the median. It is inside the noise; what can be said is that no several-hundred-ns per-draw regression from this slice exists on this host.
- The brief's ceiling (
T1 ≤ 45 ns/drawon Adreno 830,≤ 60on Mali) is a device number and this host cannot resolve it. D.4.2/D.4.3 on the two devices remain owed, and this slice should be measured there with the same0x7fvs0x5farm A/B, which is the cheapest attribution available (one binary, one lock hold).
Recorded honestly rather than dressed up: v1 produced no cost datum at all, v2 produces a call-count datum that is decisive about the mechanism and a timing datum that only bounds the magnitude on a contended software rasteriser.
Minors
| # | review item | disposition |
|---|---|---|
| 1 | TwoTablesOfTheSameKindAgreeOnOneObjectsHandle cannot fail |
fixed. Replaced by TwoTablesOfTheSameKindShareOneSlotAndKeepTheirOwnTwin, which asserts the allocator's LiveCount(kind) grows by exactly one however many tables hold a twin of the object (a per-table allocator passes the old pure compare and fails this), and that the shared handle still addresses each table's own twin. |
| 2 | GetOrCreate(nullptr) derefs null in release |
fixed. Returns a per-table parking BackendPtr, never live, never swept, never handed a handle — the map arm's documented tolerance. The MOBILEGL_ASSERT was removed from that path on purpose: a DEBUG build must not trap where release quietly does the defined thing. Pinned by GetOrCreateToleratesANullStateObject. |
| 3 | latent UAF in ReclaimDeadSlots |
fixed. The twin is std::moved into a local, the entry is finished with inside a scope, the slot is freed, and only then is the local released — so a twin destructor that re-entered GetOrCreate and grew m_slots cannot make the writes land in freed memory. |
| 4 | nothing returns slots to the allocator on destruction/reset | argued, not fixed. D13 requires the table to keep ScopedDirectGLESTextureBindings' copy-assign-and-restore shape, so two live table values can name the same slot (the fixture's saved and the working table do exactly that). A destructor or reset that called Free would therefore be a double free through a copy — the double free is real and reachable: Free on the same {slot, gen} twice pushes the slot on the free list twice and two later Allocates hand out the same slot. Returning slots has to be owned by the explicit-destroy path (e2), which is blocked. Recorded as debt; the new two-table case carries a comment saying why it deliberately does not sweep. Consequence the integrator inherits: MGPipeSlotAllocator::HighWater/LiveCount are still not a process-wide health signal from a test process. |
| 5 | the backend mints client handles | recorded, not fixed. A P3+ DEBT block at the top of SlotTables.h names MGPipeHandles.h:13-16, says the minting and the lifetimeId -> handle resolution both belong on the client, says the backend should receive the handle in the verb payload under split, and says check_include_closure.py does not probe MG_Backend headers so nothing catches it. It is monolith glue and is not "Track H done". |
| 6 | MOBILEGL_PIPE_PUSH=0 is no longer the faithful P1 control |
fixed. g_fbSlotCache is compiled under #if !MOBILEGL_PIPE_PUSH || MOBILEGL_PIPE_LEGACY_MEMOS and GetFramebufferBindingSlotChecked uses it when !EsprytSlotTablesEnabled(). So the poison bypass is closed on the arm that ships and MOBILEGL_PIPE_PUSH=0 reproduces P1 at those five sites too, which is what ConfigLoader.cpp:251-254 promises. build-nolegacy (no legacy arm) takes the #else, an unconditional checked read. |
| 7 | MGB_TWIN_KIND_PARAM/ARG leak into every TU |
fixed. MGB_TWIN_KIND_ARG is gone; the twelve declaration and definition sites name a TwinRegistry<S, B, kKind> alias template that swallows the kind in the pull build (an alias template may have an unused parameter and emits no symbol, so the pull mangling is unchanged — confirmed by G1 below). MGB_TWIN_KIND_PARAM is #undef'd immediately after the class. |
| 8 | shadowing at BindCurrentFBO |
fixed. The twin-table lookup is twinEntry; the framebuffer binding slot keeps the name slot. |
| 9 | the build-nolegacy row was unverifiable |
fixed. build-nolegacy exists and was built and run; see §3. |
| 10 | §5.3's cold-cache flake could not be reproduced | root-caused, and v1's explanation was wrong. See §5.2 — it is a ctest scheduling flake, not a cold shader cache. |
| 11 | §4.1 point 2 ("unchanged") is factually wrong | retracted, and the underlying defect fixed (MAJOR 1). It does not appear in this file except as a retraction. |
| 12 | no build-verify, so the verify read-hook was never run against the handle arm |
fixed. ctest --test-dir build-verify -L integration-verify -j 4 → 818/818 passed, and grep -l 'Fatal{' over the six pipe-verify-*.log lanes is empty. The five converted GetFramebufferBindingSlotChecked sites now go through the checked accessor on the handle arm and produce no divergence. |
| 13 | the reported G3 SSIM minimum is wrong | fixed and confirmed. The reviewer's figures reproduce exactly; see §3. |
3. Verification — commands run and what they printed
All from ~/w7/p2-espryt, CCACHE_BASEDIR=/home/swung/w7. Logs under ~/w7/e2-*.
| gate / check | command | result |
|---|---|---|
| build | cmake --build {build-linux,build-push,build-verify,build-nolegacy,build-bench} -j 12 |
all rc=0 |
| G1 | python3 scripts/symbol_report.py --before ~/w7/p2-before-libMobileGL.so --after build-linux/libMobileGL.so --threshold 0 |
27799 -> 27799 defined symbols: 0 added, 0 removed, 4 resized, 0 renamed; resized = RenderState::RenderState() +148, RenderState::SetCapability +77, RenderState::IsCapabilityEnabled +29, _GLOBAL__sub_I_DirectGLES.cpp −9 — all four are the contract commit's |
| G1 attribution | same script, --before ~/w7/p2-contract/build-linux/libMobileGL.so --after build-linux/libMobileGL.so |
.text +0, 0 added, 0 removed, 0 resized, 0 renamed — package C's pull-build delta is still exactly zero after this round, including the TwinRegistry alias |
| G2 | ctest -N | sed -n 's/^ *Test *#[0-9]*: //p' | sort on all four dirs, diff |
build-linux 2374, build-push 2374, build-nolegacy 2374 — both diffs empty. (build-verify lists 3192: it adds the verify-only lanes, which is by design.) |
| G5 | git show {48268068, p2/contract, HEAD}:…/DirectGLES.cpp | awk '/namespace RenderStateImpl {/,/} \/\/ namespace RenderStateImpl/' | sha256sum |
all three d8fd1c48716056c536752fde09db3a9e5aa8494b70404798bfc15d11220efe27, equal to ~/w7/p2-before-syncrenderstate.sha |
| G13 | grep -rc 'pGLContext' MobileGL/MG_Backend | grep -v ':0$' |
empty |
| G13 | python3 scripts/check_include_closure.py |
4 probes, 0 skipped, 0 problem(s), rc 0 |
| G13 | gen_pipe.py --check / --self-test |
rc 0 / rc 0 |
| G13 | stdio grep over the four touched backend files | 9 hits, all the English word "puts" in comments (the base ref has 6 of the same in DirectGLES.cpp alone); no printf/cout/cerr |
| G14 | comm -23 ~/w7/p2-before-ctest-names.txt <pull names> |
empty. Added: the contract's 4 *.PlaceholderUntilTheOwningPackageFillsThisIn and this package's 7 DirectGLESSlotTable.* (5 from v1, one of which was renamed, plus 3 new) |
| legacy residue | preprocessor-stack walker over DirectGLES.cpp for OwnerEquals|TwinLookupMemo|g_fbSlotCache|GetFramebufferBindingSlotFast |
24 hits, UNGUARDED: 0 — each is inside #if MOBILEGL_PIPE_LEGACY_MEMOS, the #else of #if MOBILEGL_PIPE_PUSH, or the new #if !MOBILEGL_PIPE_PUSH || MOBILEGL_PIPE_LEGACY_MEMOS. GetFramebufferBindingSlotFast is gone entirely |
| unit | ctest -L unit --no-tests=error -j 6 on all four dirs |
100% tests passed, 0 failed out of 1496 × 4 |
| slot-table cases | ctest --test-dir build-push -R DirectGLESSlotTable --output-on-failure |
7/7 Passed |
| slot-table cases, pull | same on build-linux |
all 7 present and (Skipped) — a visible skip, as C.4 wants |
| churn negative control | neuter the creation sweep, rebuild build-push, ctest -R DirectGLESSlotTable |
only ObjectChurnAloneDrivesTheSweep fails; 7/7 again on restore |
| integration, handle arm | ctest --test-dir build-push -L integration-gpu --no-tests=error -j 4 -R DirectGLES |
446/446 (run 1 and run 2) |
| integration, legacy arm | same with MOBILEGL_PIPE_PUSH=0 |
446/446 |
| integration, no legacy arm compiled | ctest --test-dir build-nolegacy … -R DirectGLES |
run 1 444/446 (the scheduling flake, §5.2), run 2 446/446 |
| integration, verify build | ctest --test-dir build-verify … -R DirectGLES |
run 1 853/855 (same two), run 2 855/855 |
| integration-verify (G4 half) | ctest --test-dir build-verify -L integration-verify --no-tests=error -j 4 |
818/818 passed; grep -l 'Fatal{' build-verify/…/pipe-verify-*.log empty |
| Fatal{PipeLegacyMemosDisabled} | reviewer's two commands, verbatim | see the table in §2 — the disabled arm no longer runs |
| G3 (DirectGLES half) | python3 ~/w7/retrace_gate.py --tree ~/w7/p2-espryt --lib $PWD/build-push/libMobileGL.so --out ~/w7/retrace-out/e2-espryt -j 4 --only 'DirectGLES$' |
rc 0, passed 39 / 39; failed: []. Five lowest SSIM: 0.993475 minecraft-1.21.11-main-menu, 0.995426 …iris-makeup-ultrafast-in-world, 0.995920 …iris-iterationt-nodsa-in-world, 0.996197 …iris-derivative-main-d24.4.14-in-world, 0.996301 …iris-sundial-lite-in-world. The reviewer's figures are the correct ones; v1's "lowest 0.997009" was wrong. All 39 ≥ 0.99 |
| DriverBench T1/T2 + arm A/B | §2 MAJOR 4, raw rows ~/w7/e2-bench2.csv |
produced; noise-bounded |
| gdb call counts | §2 MAJOR 4 | FindByLifetimeId 24 → 6; arm selector 92 → 1 |
4. Deviations from the brief (v2)
4.1 BLOCKING — step e2 is not implemented (unchanged from v1)
See §2 MAJOR 3. The brief contradicts itself: C.2 step e2 requires files that C.2's own
Files line does not grant and that C.5 gives to B and D. Integrator decision.
4.2 The TwinLookupMemos are replaced by a table-local one-entry memo, not by a bare array index
D13's "direct slot indexing" is true of the table (FindByHandle is a bounds check plus an
index) but not of resolving a frontend object to its handle, which goes through the allocator's
lifetimeId -> slot hash. v1 left that hash on the three hottest paths; v2 puts a one-entry
lifetimeId -> handle memo in front of it (§2 MAJOR 4). This is the mitigation the review names
("restoring the memo keyed on {slot, gen}"), except that the key is the lifetime id — which is
strictly safer than {slot, gen} would be, because a lifetime id is never reused at all while a
slot is, and the Gen compare in FindByHandle remains as the second line of defence.
4.3 UnitBindingsSnapshot holds lifetime ids, not {slot, gen} (unchanged from v1)
A texture that is bound but never synced has no twin and therefore no handle, so a handle-keyed snapshot would read two never-synced textures as equal. The reviewer independently agreed.
4.4 The four pDefaultFramebufferInfo->defaultFBO compares are not retired (unchanged from v1)
The default framebuffer is never registered in the twin table, so HandleOf answers the null
handle for it; somebody has to own minting and installing kMGPipeDefaultFramebuffer first. The
reviewer agreed this is not actionable as written.
4.5 Some deletions are arm-split rather than outright (unchanged from v1)
ScopedDetachedTextureFramebufferAttachments' walk and UnitBindingsSnapshot keep the pull text
verbatim under #else, because the shared form moved pull-build codegen and G1 admits no resize
beyond the contract's.
4.6 retrace_gate.py has no --backend or --ssim flag (unchanged from v1)
It takes --tree --lib --out -j --only; the SSIM threshold comes from the reference
build-retrace CTestTestfile.cmake it parses. C.2's and C.3's verification blocks and G3's row
in section A spell flags that do not exist. Tree wins on the fact.
4.7 New in v2: the MOBILEGL_ASSERT on a null state object was deliberately removed
On the handle arm only. Null is now a defined answer, so trapping in a DEBUG build while release quietly parks the twin would be the worst of both. The map arm keeps its assert.
4.8 New in v2: one v1 test name was changed
DirectGLESSlotTable.TwoTablesOfTheSameKindAgreeOnOneObjectsHandle →
…ShareOneSlotAndKeepTheirOwnTwin. G14 is untouched (the old name was added by this package on
this branch and is not in ~/w7/p2-before-ctest-names.txt), but the integrator should know the
name moved between review rounds.
5. Where the tree contradicted the brief
5.1 The contract commit resizes a fourth pull-build symbol (unchanged from v1, re-confirmed)
_GLOBAL__sub_I_DirectGLES.cpp is 1340 → 1331 bytes (−9) at p2/contract with zero package-C
changes; contract→HEAD is a byte-identical .text. G1's admitted set (D15 point 4, C.0's
verification block) must be widened by that symbol, and the attribution belongs to package A's
D3.1, not to C. Re-verified this round after the TwinRegistry alias landed.
5.2 §5.3 of v1 was wrong about the flake: it is a ctest scheduling flake, not a cold cache
DirectGLES.UnlocatedIoBlocks.UnlocatedIoBlockScenario.TheEmulationIsActuallyArmedWhenTheEnvironmentPinsItOn
and DirectGLES.PointSizeDemotion.PointSizeDemotionScenario.TheDemotionIsActuallyArmedWhenTheEnvironmentPinsItOn
fail on the first ctest -L integration-gpu -R DirectGLES -j 4 of a build directory and pass
on every run after. v1 guessed "cold shader/translation cache". That is refuted:
$ rm -f build-push/Testing/Temporary/CTestCostData.txt # warm binaries, unchanged library
$ ctest --test-dir build-push -L integration-gpu -j 4 -R DirectGLES --output-on-failure
99% tests passed, 2 tests failed out of 446
2336 - DirectGLES.UnlocatedIoBlocks…TheEmulationIsActuallyArmedWhenTheEnvironmentPinsItOn (Failed)
2369 - DirectGLES.PointSizeDemotion…TheDemotionIsActuallyArmedWhenTheEnvironmentPinsItOn (Failed)
Deleting only ctest's cost data — nothing else changes, same library, same build — reproduces
it exactly; the next run (cost data regenerated) is 446/446. It reproduced on three separate
build directories this round (build-push after the cost-data removal, first runs of
build-nolegacy and build-verify) and on both arms, so it is not attributable to this package:
the library is byte-identical between the failing and the passing run of the same directory.
The failure itself is the log-window assertion:
PointSizeDemotionScenario.cpp:512: Failure
Expected: (appended.find("demoted to an ordinary varying")) != (std::string::npos)
LibraryLogSize() / LibraryLogSince(before) read a byte window of
${CMAKE_CURRENT_BINARY_DIR}/point-size-demotion-gles.log, which every entry of that lane shares
(MG_IntegrationTest/CMakeLists.txt:387, 370). Running the lane alone under -j 4 is green, so
the losing interleaving needs the full 446-test schedule; ctest's cost-data ordering is what
decides whether it happens. A clean CI runner never has cost data, so D.3 will meet this
every time. Fix belongs to package E (the lane's entries need RUN_SERIAL or a RESOURCE_LOCK,
or the assertion needs a per-test log path); it is not fixable from package C's file set.
5.3 The G2 / G14 ctest -N extraction in section A silently drops most tests (unchanged from v1)
grep -E '^\s+Test #' misses Test #83: once the total is four digits. Working form used
throughout: ctest --test-dir <dir> -N | sed -n 's/^ *Test *#[0-9]*: //p' | sort. Section A
(G2, G14) and D.1/D.3 need the fix. The reviewer confirmed this independently (1373 of 2372).
5.4 git stash -u in a P2 worktree destroys the copied trace fixtures (unchanged from v1)
wsl_p2_tree.sh copies the 803 MB fixture set in and marks it --assume-unchanged; git stash
resets through that bit and leaves LFS pointer files. Not hit this round (the negative controls
used cp to /tmp, never git stash), and the fixture set is intact (du -sh = 803M).
5.5 p2/contract is an ambiguous refname in this worktree
git show p2/contract:… warns refname 'p2/contract' is ambiguous (a tag and a branch-shaped
path both match). Harmless — it resolves to the tag — but the integrator will see the warning in
every G5 run.
6. Unfinished
- Step
e2, the explicit-destroy hook — blocked on file ownership (§4.1). The one scoped deliverable of C.2 that is not present. With MAJOR 1 fixed it is no longer a regression, butARCHITECTURE.md:363's census still cannot count the six registries' GC as deleted, and the sevenCollectGarbageIfNeededcall sites still stand. - The four
pDefaultFramebufferInfo->defaultFBOcompares are not replaced bykMGPipeDefaultFramebuffer(§4.4) — needs an owner for the reserved handle. - Slots are never returned to the allocator on table destruction or reset (minor 4). Argued
rather than fixed: two table values can name one slot under the copy-and-restore fixture
shape D13 mandates, so a
Freeon discard is a reachable double free. The release belongs toe2.HighWater/LiveCounttherefore remain unusable as a process health signal in tests. - The backend still mints client handles (minor 5). Recorded as P3+ debt in
SlotTables.h; it must not be counted as "Track H done", andcheck_include_closure.pydoes not probeMG_Backendheaders so nothing will catch a second instance. - Device measurement. The desktop DriverBench numbers exist now but the host (llvmpipe,
load average 29–51) cannot resolve a 45 ns/draw ceiling. D.4.2's two-device paired A/B and
D.4.3's T1/T2 on device are still owed, and this slice should get the
MOBILEGL_PIPE_PUSH=0x7fvs0x5farm A/B there — one binary, one lock hold, exact attribution. - The
integration-gpufirst-run flake (§5.2) is root-caused but not fixed; the fix is in package E's files. D.3 will hit it on every clean runner. - No
HandleRecycleScenario(package E owns it), noPipeStatsgate for this subsystem (package A ownsPipeStats.{h,cpp}), so the integrator's A/B still has no counter that says "N twins resolved through the slot table this frame" — only the gdb counts above.