| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| ... | |
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A TUI must not die because one background load failed, so this codebase catches
broadly -- ~50 `except Exception` sites, two dozen resolving to `pass`. Right
policy, one bad consequence: with 44 `@work(thread=True)` workers, a failure in
a background load leaves no trace whatsoever. The view stays empty and there is
nothing to read afterwards, because the app owns the screen.
kittygfx already solved this for itself with $IDATUI_KITTY_LOG. idatui/diag.py
is the same idea for everything else: $IDATUI_LOG writes every swallowed error
plus its traceback to a file, and the last 50 are kept in memory regardless so a
driver can ask a live app what went wrong. Unset, it costs an environ lookup.
Wired in where losing the error changes a DECISION rather than just a pixel:
* rename: a resolve() that throws renames as DATA instead of as a function.
* name: a function_of() that throws means we never learn the address is a
function start, so the index keeps the old name and every readback says the
rename didn't happen.
* retype: a resolve() that throws retypes the ENCLOSING function instead.
* decompile: a failed full-body fetch silently returns CLIPPED pseudocode.
* trail: a failed decomp_map stops the pseudocode being painted, silently.
Deliberately NOT wired into the query_one guards -- a modal owning the screen is
normal and constant, and logging it would bury the real entries in noise.
New RPC verb `diag {n?, clear?}`, documented in docs/RPC.md: the answer to "the
verb reported success and the pane shows nothing".
Also a flake, same shape as the others: follow_xrefs waited on the nav depth but
asserted on _cur, and a follow pushes the source entry BEFORE opening the
target -- so the check could run in between and see the function it jumped
from. About one run in ten. It waits on the postcondition it asserts now; three
clean full runs since.
833 checks; --fast is 344 in 3.5s.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
_sweep_locks removes the scratch IDA unpacks beside a .i64 (.id0/.id1/.id2/
.nam/.til) when an open fails, keyed on both the full name and the stem. It
never touched the .i64, which is the dangerous one everybody thinks of.
It did delete the input. '.til' is an unpacked-DB suffix AND the extension of an
IDA type library, so 'ida-tui mylib.til' swept its own argument out of
existence -- irreversibly, on a path that runs automatically. Same for anything
named *.id0/*.id1/*.id2/*.nam. Now the sweep skips whatever it was asked to
open, compared as an absolute path so a relative argument is covered too.
tests/test_launch.py pins the whole contract: what it takes, what it must never
take (the .i64, the input, the neighbours), and what it reports. Pure, in the
--fast tier. It is the right shape of test for code whose failure mode is
deleting the wrong file.
Also: _load_args parsed the base with bare int(), which raises on the
'0x8000000' string a project file writes. Unreachable from our own CLI (which
int()s first) but the asymmetry with project._as_addr was a trap, so both go
through the same parser now.
813 checks; --fast is 324 in 3.4s.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The layer between the app and idalib had no tests, which is awkward: it is
where failures are silent. A worker that dies during startup, a socket that
drops mid-call, two UI threads sharing one socket -- none of those look like
bugs from outside, they look like the TUI hanging or showing stale data.
None of it needs IDA. WorkerClient spawns whatever _WORKER_PY points at, so the
suite points it at a fake speaking the same length-prefixed pickle and tells it
to misbehave on demand: die at startup, never bind, drop the connection, fail a
tool, take its time. 40 checks in the --fast tier.
Two things the tests found:
call() reconnects when _sock is None, which is what makes a dropped socket
recoverable -- but it made an explicitly CLOSED client resurrect too, spawning a
whole new idalib worker to serve one stray call (verified: pid 1066961 ->
1066962). close() runs on teardown and on binary-switch while @work threads are
still in flight, so quitting during a decompile could leave a fresh process
re-opening the .i64 we had just released, which is the wedging hazard. A closed
client now refuses; connect() still revives it, which is all _reconnect needs
(it builds a new client anyway).
connect() polled on a flat 0.2s sleep, so every caller paid a fifth of a second
even when the worker was ready in milliseconds -- a seeded .i64, a small binary.
Backs off from 5ms instead.
786 checks, 144.6s; --fast is 297 in 3.3s.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
_active was a bare string with 49 comparisons across four modules and a fifth
value nobody meant to keep. "disasm" was assigned on exactly one path -- a
decompile that failed with nowhere to return to -- and named the same widget as
"listing". Four sites understood it; five compared against "listing" alone and
silently took the wrong branch:
* Tab out of a failed decompile set "listing" instead of "decomp", so the
first press appeared to do nothing.
* rpc.py carried a workaround for a mode change that never arrived, keyed on
being ALREADY in the ghost state -- so it fired in the rare case and not in
the common one. Now keyed on LISTING, which is the case that happens.
* drive.py asked the socket to show it "disasm", a value the app will now
never report, and would have toggled twice and given up.
ViewMode is a StrEnum on purpose: _active goes straight to drivers as
cursor.kind and the pilot compares it to plain strings, so members being strings
keeps every payload and comparison working. What it buys is one place that says
which modes exist, and an AttributeError instead of silence on a typo.
Read it through is_listing/is_decomp/is_hex/is_graph/in_code rather than ==.
The bare comparisons are what let the ghost hide, and they are what the next
mode would have to hunt down -- adding "graph" already cost one crash that way
(_active_code_view returning None when a prompt closed).
view_modes_all_handled walks the enum and asks the app the questions it asks
itself. Verified it bites: adding a fifth unhandled member fails it twice.
746 checks, 142.3s.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The split resync loop (f898350) had no test. Two attempts at one were worthless
and are not in this commit: a scroll-based guard passed with the bug
reintroduced, and a constructed anchor -- inside the loaded function, outside
its mapped span -- skipped, because on this target the map covers the whole
function. The real trigger is the race window while the decomp map lags the
decompiler re-pointing, which is tedious to force but wide open in split_view's
own flow.
So split_view counts lookup_funcs across its body and bounds it. Verified both
ways, which is the only reason it's worth having: 29,227 calls with the bug put
back, under 500 with the fix. The bound is loose because the bug was three
orders of magnitude out, not a near miss.
733 checks, 139s.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
all_funcs() forced a full load of the function index only when it was EMPTY, so
a partially streamed index -- non-empty but incomplete, which is exactly the
state during boot and after any bump_items() -- came back truncated. Every
fixture picked through find_func/biggest therefore depended on how far streaming
had got by the time a scenario asked.
That is the graph_minimap flake: on an unlucky run find_func(size > 0x300)
picked a much larger function than usual, whose graph never finished inside the
scenario's own 60s wait. Three failures and 65 seconds, one run in several, with
no code change to blame. Three consecutive clean runs at 1.7s since.
CORRECTION to f898350, which said a range cache for function_of 'broke
graph_minimap'. It did not. The failure happened in the run after I added the
cache and I attributed it without checking; it recurred with the cache long
gone. The cache is still not here, but for the honest reason: with the resync
loop fixed, function_of is down to 340 calls and 1.4s across the whole suite,
so caching it is not worth the invalidation surface.
Suite 195.7s -> 138.4s, 732 checks.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
test_trace_ui spent 18.8 of its 35.2 seconds in flat pilot.pause() calls placed
to let an async seek land. Two loops were most of it: 6 iterations at 0.5s and
28 at 0.3s, 11.4s of sleeping to check that a step moves the cursor.
They are condition waits now. The questions are unchanged -- does the listing
cursor reach the pc, does the pseudocode cursor follow -- but they cost what
they cost instead of a fixed budget. The second loop settles on something that
does NOT presuppose the answer (the listing cursor arriving, and the trail map
belonging to the loaded function): waiting on 'is this pc mapped' would have
burned the timeout on every unmapped instruction, about half of them, and come
out slower than the sleep it replaced.
35.2s -> 20.9s, 39 checks, stable over repeated runs.
tests/_fixtures.py collects the staging both this suite and test_scenarios need
-- scratch copy, seeded from a golden .i64 nothing writes back to -- which was
private to test_scenarios. Worth saying plainly: on targets/echo the seeding is
worth 0.19s, not the analysis time I assumed when I went looking. It is shared
for the deduplication and for whatever gets pointed at a bigger binary.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
_split_range is the min/max of the decomp_map's addresses, which does not cover
every address in the function -- Hex-Rays doesn't attribute them all. An anchor
inside the loaded function but outside that span therefore asked _sync_split for
a resync, _apply_resync found the function already decompiled, called
_sync_split again, and it asked again. One thread worker and one lookup_funcs
round trip per iteration, for as long as the cursor sat there.
Measured in the pilot: 23,888 function_of calls in one scenario across FOUR
distinct addresses, 21,156 of them for 0x2060 alone. In the live app that is an
idle split view pegging the worker.
_sync_split grows a resync flag; the one caller that is itself the resync passes
resync=False, so the branch can be entered at most once per chain.
While measuring, three scenarios waited on "fail" appearing in the status --
the app says "cannot decompile". decomp_fallback burned its full 25s timeout and
then passed a check on _active == "listing", which was already true before Tab
was pressed: it asserted nothing, slowly. Now waits for the real text and checks
that the fallback actually said something.
scenarios 115.8s -> 74.9s, suite 195.7s -> 153.3s, 732 checks green.
Not included: a range cache for function_of. It broke graph_minimap (the graph
stopped loading at all -- the 65s was that scenario's own 60s wait timing out)
and with the loop gone it buys little. Left out rather than shipped
half-understood.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Second cut at the 4000-line class, and the biggest: 702 contiguous lines of
rename/comment/retype/make-data/literal-format/define, now EditController.
What stays on IdaTui is what Textual insists on owning -- on_<Message> handlers,
which it dispatches by name on the DOMNode, and @work entry points, whose worker
machinery wants a DOMNode host. Both are one-line delegates.
Underneath them was the duplication that made this worth doing rather than just
moving lines. Opening a prompt (hide the status bar, set placeholder, can_focus,
display, value, focus) was written out five times; closing it four; and Esc was a
six-branch ladder in on_key with one copy of the same four lines per prompt. They
had drifted -- search cancelled its highlight, goto restored focus, the edit
prompts did neither consistently. Prompt/PromptBar own that discipline once, and
close() HANDS BACK the context it was holding, so it can't be read twice or go
stale: the listing's rename had to capture _rename_addr by hand before
_end_rename cleared it, or the name went to address 0.
_line_ea_for stays on the app -- rpc.py and the goto readback ask the same
question, so it was never an edit helper.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
First cut at the 4000-line class. Trace is the cleanest seam: 348 contiguous
lines, one coherent job (where we are in time and everything that moves us),
and two suites already covering it.
TraceController owns the state now -- the trace, the timestamp, the trail maps.
IdaTui keeps the keys, because Textual only merges BINDINGS from DOMNode
subclasses and a mixin's would be silently dropped, and it keeps the @work entry
points, because the worker machinery wants a DOMNode host. Both are one-line
delegates.
_trace/_t/_trail_map/_trail_map_ea/_trail_line_of stay readable on the app as
properties: the pilot suite and rpc.py read the position by those names, and a
property means one owner rather than a copy that can drift. rpc.py itself now
goes through the controller.
The parallel line-map that _apply_split_map used to poke into five attributes is
now one adopt_map() call -- same single shared index, but the sharing is stated
rather than implied by two places assigning the same fields.
731 checks, unchanged.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fourteen test files, each its own __main__, and no way to run them but from
memory -- so in practice you ran the one you were working on and hoped. Worse,
nothing said which files need a licensed IDA and a real worker (minutes) and
which are pure stdlib (milliseconds), so the cheap ones nobody ran either.
tests/run.py runs the lot and prints one table. --fast selects only the suites
that need nothing, which is 257 checks in half a second under any python3 --
that's the one you run between edits.
The classification lives in the test files, not in a table here that would rot
the first time someone adds a test: each declares NEEDS_IDA at module scope and
run.py reads it with ast (it can't import them -- they run their suite at
import). A file without the marker is a hard error rather than a silent guess.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The artwork is now a proper transparent PNG with soft edges (24% of its pixels
carry partial alpha) instead of opaque art on black with a stray full-width
scan line along the bottom. Cropped to its content and resized 1024 -> 768px,
which halves the file and costs nothing visible; logo-trans.png keeps the
master for future re-renders.
Two things the new art exposed, both wrong before it:
fit() assumed cells were 1:2. This terminal reports 9x22, i.e. 1:2.44. The old
logo was 474x516 -- close enough to the assumption that nobody noticed -- but a
square image at the hardcoded 60x33 would have been visibly stretched. The
graphics query now asks for the cell size too (CSI 16 t rides along in the same
round trip, before the DA1 that already synchronises it) and fit() uses the
answer.
The footprint was a constant. logo_cells() derives it from the artwork and the
measured cell size, so the art can be replaced without anyone remembering to
edit a number.
logo.ans was stale: the block-art fallback for terminals that can't draw an
image was still the OLD artwork, scan line included. tools/make_logo_ans.py
regenerates it from logo.png so the two cannot drift again. It understands
alpha -- a transparent cell emits no colour and lets the terminal background
through, and a cell with only one opaque half uses the matching half block so
the pixel lands on the correct side.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
F1 is swallowed before it ever reaches us on at least one setup here -- the
app's own binding fires when the key is injected directly, zellij has no F1
binding of its own, and every common F1 encoding written straight into the pane
(SS3 ESC O P, CSI ESC [11~, CSI-u ESC [1;1P) opens it. So the key is being eaten
by something upstream, which is not ours to fix, and a cheatsheet reachable only
through a function key is fragile anyway: terminals and multiplexers claim them
routinely.
H opens and closes it too. '?' stays with the incremental search, which is what
it has always done in the code views.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Clicking the minimap panned to the exact coordinate under the pointer and moved
the cursor only if a block happened to sit there. Since one minimap cell covers
many canvas cells, "there" was almost always padding: you got a jump into empty
space and the cursor stayed behind, so you had to click a block afterwards to
actually go anywhere.
Blocks cover a few percent of a laid-out graph -- 4.6% of an 87-block function,
0.8% of a 424-block one -- and the rest is the space that keeps edges apart. So
coordinates are the wrong thing to navigate by here. The minimap now snaps to
the nearest block and takes the cursor with it, and a drag scrubs from block to
block. Distance is measured with the column halved, because cells are twice as
tall as they are wide and otherwise "nearest" is not what looks nearest.
A drag-pan or ctrl+d/pageup that ends with no block on screen at all now eases
to the nearest one too, since an empty screen leaves nothing to navigate back
by. It only fires when nothing is visible, so a deliberate pan is never fought.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Click it to jump the view to that part of the graph, drag to scrub. If the
point you clicked is over a block the cursor lands in it, so the keyboard
carries on from where you pointed instead of snapping back.
This also fixes a real bug rather than only adding a feature. The minimap
FLOATS over the canvas -- it is pinned to the viewport, not drawn into the
graph -- so a click on it was being translated into canvas coordinates and
dropping the cursor into whatever block happened to lie underneath. It has to
be hit-tested before the canvas, which is what on_click now does.
_minimap_rect() is the one source of truth for where it is: the renderer and
the hit-test both take the position from it, so the two-column inset that
keeps it clear of the ScrollView's scrollbar can't drift between them.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
logo.ans is 60x33 cells of half-blocks -- a 60x66 pixel image. logo.png is
474x516. On a terminal that speaks the kitty graphics protocol we now send the
real thing, in the same cell footprint (fit() lands on exactly 60x33, so the
layout is unchanged), and fall back to the block art everywhere else.
Three findings, each of which cost a round of "it renders nothing":
Support cannot be sniffed from the environment. Under a multiplexer that passes
the protocol through, TERM is xterm-256color and KITTY_WINDOW_ID, TERM_PROGRAM
and COLORTERM are all empty while the protocol answers OK -- detection by
terminal name would disable graphics on exactly the terminal that supports them.
So we ask: a 1x1 graphics query plus a Primary Device Attributes request, with
DA1 as the sync point.
Unicode placeholders are not usable. The tidy way to put an image in a TUI is a
virtual placement plus U+10EEEE cells that the compositor clips and moves like
text -- and it is what every Textual image library builds on -- but this
terminal answers ENOTSUPPORTED for placeholders while supporting everything
else. So the image is placed directly, anchored to screen cells Textual knows
nothing about. The splash therefore owns its lifetime: place after layout,
re-anchor when the note repaints (throttled; a placement is one short escape
with no image data), delete on unmount, or a leftover would sit on top of the
disassembly forever.
The query and the upload go on OPPOSITE sides of the alternate screen. The query
must run before Textual starts, which reads stdin on its own thread and would
eat the reply. The image must be uploaded after Textual has switched to the
alternate screen: an image uploaded to the primary screen cannot be placed from
the alternate one, and the placement reports success while drawing nothing. That
silent failure is why detection lives in launch.py and upload lives in the
splash's on_mount.
$IDATUI_KITTY_LOG traces the decisions, since none of this is visible to a test
-- correct escape sequences and visible pixels are not the same thing here.
Off-tty (the pilot suite, a pipe) detection returns False and the block art is
used, so the tests are unaffected.
|
| |
|
|
| |
The PNG logo.ans was rendered from.
|
| |
|
|
|
|
|
|
| |
cfg_dump freezes real CFGs to JSON; graph_spike renders one or --stats a
whole corpus through the shipping engine; graph_smoke is the end-to-end
tool->domain->layout check; graph_shot renders the real view headless at a
chosen size, which is the only sane way to judge it (a tiled pane is far
too narrow and the minimap sits on top of the graph).
|
| |
|
|
|
|
|
|
| |
Opening, navigation and edge-following, the three zoom levels, the drawing
actually reaching the screen (a layout that is right but paints nothing
looks fine from the outside), clicking a block, renaming from inside one,
and the mode surviving a navigation. The help test now derives its group
list from _HELP instead of hardcoding it, so adding a card isn't a failure.
|
| |
|
|
|
|
|
|
| |
It reports STRUCTURE -- blocks, typed edges, ranks, box geometry, the
cursor -- and not the box-drawing characters, which is what a driver
actually wants; screen still gives you the drawing. show is a pure read.
The line-oriented verbs refuse in the graph rather than reporting a
(block, row) cursor as a line index some later edit would trust.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Space swaps the code view for the function's basic blocks, IDA's own key.
Boxes hold the SAME Head rows the listing renders, so IDA's colour tags,
the word highlight, the execution trail and every editing verb work inside
them for free. Nothing is pre-painted: each screen row is composed on
demand from the edge index plus whichever boxes cover it, so cost tracks
the viewport and not the graph.
z cycles three zoom levels, m toggles a minimap, J/K walk edges, and the
mode is sticky -- following a call lands in the callee's graph. Above 400
blocks it declines and says so, because nothing readable comes out at that
size. Edges follow IDA's colours, and the ones touching the block under
the cursor are brightened.
Adding a value to _active means every consumer has to learn it: the one
that was missed (_active_code_view returning None) crashed the app the
first time a prompt closed in graph mode. A stale async load is also
guarded now -- without it, a rename's queued rebuild landed later and
dragged the user back into a graph they had already dismissed.
|
| |
|
|
|
|
|
| |
Two calls rather than one per block: flowchart for the shape, then a
single heads walk over the function's extent, sliced up by address. A
hundred blocks would otherwise be a hundred round trips. Cached per
function and dropped on a rename, since the rows carry live names.
|
| |
|
|
|
|
|
| |
Blocks and typed edges for the function at an address -- not their text.
The block body is an address range the client already knows how to render
with heads, and keeping text out is what lets the graph reuse the listing
rows instead of growing a second disassembly renderer.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Textbook Sugiyama, the same shape IDA's own graph uses: break cycles,
longest-path layering, dummy nodes, median/transposition ordering,
priority x-coords, then port-and-channel edge routing. Pure python -- no
IDA, no Textual, no I/O -- so it is tested offline in milliseconds with no
worker, which is the whole reason the hard part is kept out of the UI.
Dummy nodes are what make routing tractable: a long edge occupies real
horizontal space, so no edge ever has to cross a box. The tests assert
exactly that over a 128-function corpus, and it holds at 0.
Two things cost real time to find. A self-loop never drains its own
in-degree, so it deadlocks the ranking and collapses the graph into three
layers, 280 columns wide -- they are dropped from the layout and drawn as
a marker. And crossing minimisation is the entire runtime: recounting
globally per candidate swap is O(n^3) and took 20.4s on a 424-block
function, against 152ms for Fenwick inversion counting plus a local
O(deg*deg) swap delta.
The result is not a painted canvas -- that function is ~13M cells. It is
an index: per-row runs, bucketed vertical intervals, and point marks,
queried one row at a time.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Pilot scenarios for the listing and the pseudocode, for the mark moving
between operands, for a refusal not being swallowed by the previous
success, and for the cursor staying on its literal across a reflow. Plus
experiments/opfmt_tools.py, which runs the real injected tool sources
against a live database with the decorators stubbed -- faster than the
pilot and the right place for the IDA-side edge cases.
Also fixes two pre-existing bugs the work surfaced, both of which made
edits happen off screen: cursor_on searched from row 0 of the whole
segment and never scrolled, so a driver's word= edit landed in an
unrelated function while reporting success; and the cursor verb didn't
scroll either. Both now go through rpc.place_cursor.
|
| |
|
|
|
|
|
|
| |
mode is cycle/back/show or an explicit format. 'show' reports the current
format and the stops on offer without editing, which is what a driver
needs: the rendered text alone can't be trusted (a listing read before an
ARM/Thumb switch shows the old decoding). 'word' puts the cursor on a
token first, so a literal can be named instead of steered to.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
IDA's own key, and that muscle memory is worth more than the opcode
column's old claim on it -- the bytes column moves to B. The literal the
cursor is on is MARKED, and that mark is what changes, because a line
usually holds more than one. The cursor follows its literal across the
edit: 48 <-> 0x30 reflows the line, and holding the column would put the
next press on a neighbour. Land on a register -- something with no format
of its own -- and it says so and names the operand that has one, rather
than quietly reformatting a different one. Works in the pseudocode too,
on Hex-Rays' separate number formats.
|
| |
|
|
|
|
|
|
| |
Head carries `ops` (where each operand sits in the text) so a view can
tell which one the cursor is on, and Program grows op_format / pc_nums /
pc_num_format over the new tools. The pseudocode literal positions are
cached with the decompilation and dropped on a rename, since a reformat
moves every literal on its line.
|
| |
|
|
|
|
|
|
|
|
| |
op_format reads and sets an operand's display format through IDA's own
op_hex/op_dec/... nibbles, and pc_nums/pc_num_format do the same for
Hex-Rays' number formats, which are a SEPARATE set the listing's format
never reaches. The candidate stops are filtered to ones that change what
you see: no char unless the value prints as one, no offset unless the
target is already named. Operand extents come from IDA's own COLOR_OPND
markers on the generated line, so they agree with print_operand exactly.
|
| |
|
|
|
|
|
|
|
| |
The multiplexer is auto-detected ($ZELLIJ then $TMUX) and every pane
command works the same under both. Pane ids are self-identifying, so a
mixed set of tmux and zellij panes can be tracked at once. zellij has no
-l, so --size is ignored there, it always focuses a new pane (--detached
is emulated), and it leaves an EXITED husk behind that stop/reap now
clear. The pane tests skip on neither multiplexer rather than on no TMUX.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Two bugs found driving a 65KB ARM firmware image (one flat 42k-line listing, no
ELF sections to break it up):
`note` did goto + `cursor line=0` before commenting. Line 0 is the top of the
function only in the DECOMPILER; in the listing it is the top of the SEGMENT, so
every note landed at address 0 -- and scrolling a 42k-line listing there took so
long the call timed out, which read as "comments are broken". goto already lands
on the function's first line, so the cursor call just goes; note now also
reports where it landed.
The client's 90s timeout was too tight for the same reason: comments on that
listing take 26-106s (the rebuild has no function boundary to stop at), so the
CLI reported "no response ... server busy or the op is hung" for edits that had
already been applied. Believing a successful edit failed is the worse error --
the driver redoes it, or "fixes" what was never broken. Default is now 300s;
IDATUI_RPC_TIMEOUT still overrides.
|
| |
|
|
|
|
|
|
|
|
| |
Hex-Rays caches per function and does not notice that a *callee* was renamed;
worse, that cache is persisted in the .i64, so a bulk import left pseudocode
calling sub_98C0 forever while the listing and every readback said memset --
the exact readback disagreement a driver cannot detect. Batch now calls
force_recompile before bumping the local caches.
Test extended: decompile, rename via rename_many, read the pseudocode back.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Opening a headerless blob was the one workflow that fell out of the driving
surface entirely, and each gap hid the next:
- `pane spawn` couldn't pass --processor/--base/--ida-args, so the pane came up
"ready" with zero functions (x86 at 0) and the only way through was to
hand-write a project file. It now forwards them to idatui.launch.
- c/p/t/T (code, function, ARM<->Thumb, vector scan) existed as listing
bindings with no verb, so a driver had to guess raw keys -- and raw keys are
swallowed by whatever modal happens to be up. `define {kind,target?}` goes
through the app's own edit worker and reports what IDA actually did.
- every name went through the typed rename prompt: a navigation (listing page +
decompile) plus two prompt round-trips each. A 427-symbol map took tens of
minutes of driving. `rename_many {items|file}` hands IDA's rename tool the
whole list in one call (371 symbols in 3s) and refreshes the caches and the
function table once.
drive gains `define <kind> [target...]` and `syms <file.json>`.
Verified live against a real pane (tests/test_rawimage_rpc.py, 13 checks:
spawn load options, define thumb/func + unknown-kind rejection, rename_many
from a file and inline, with resolve/functions readback).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The existing trace suites (test_trace.py, test_trace_ui.py,
test_trace_vs_tenet.py) cover the model layer, the UI via the Textual pilot,
and differential correctness against Tenet's reference. None of them exercise
the path an agent actually takes: the trace RPC verb driven over the unix
socket through rpcclient.
This one spawns a real tmux pane with --trace, records a trace with the QEMU
tracer (falling back to /tmp/echotrace.0.log if the tracer isn't built), and
drives every trace operation through the RPC socket, validating the JSON
responses:
seek (absolute, percentage, string, edge-clamping)
step (forward, backward, multi-step, clamp at bounds)
step over (finds a call via SP drop, verifies it lands after the return)
goto (by name, by hex address, error on unexecuted)
changed registers in the response
cursor tracking (ea follows the trace pc)
response shape (trace key is a superset of snapshot)
interaction with non-trace verbs (pseudocode, state, view, goto)
trace position independence from navigation
All 45 checks pass against the echo binary with a 226-instruction trace.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Two fixes to the comment verb:
Drop the per-character typing delay. Rename/retype/goto use a 35 ms delay for
the visual effect (the agent's keystrokes appear one by one on the livestream),
but comments can be long — a 200-char annotation blocked the driver for 7 s of
pure animation. Comments now type instantly (delay=0); the aesthetic delay is
kept for rename, retype and goto where values are short.
Escape literal newlines before injecting into the prompt. The Input widget is
single-line, so a real 0x0a sent as a keystroke was silently swallowed. The
app's _do_comment already converts the two-char sequence '\\n' into a real
newline for IDA, so the RPC layer now does text.replace('\\n', '\\\\n')
before typing — both literal newlines from the caller and explicit \\n in the
text reach IDA as multi-line comments (each line gets its own // prefix in the
decompiler).
Verified on a live pane: a comment with an embedded newline now renders as two
// lines in the pseudocode, and long comments appear without the multi-second
typing pause.
tests/test_scenarios.py: 212 passed, 0 failed.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
goto/open ran `settle(app, pred, timeout)` and threw the result away. On a large
database the listing build routinely outruns the default 20 s, so the verb
returned a normal snapshot while the view had not moved. Every subsequent
rename/comment then applied to wherever the caller *used* to be.
Reproduced on a live pane against a 4 MB Go binary: `goto 0x1002019b0`
returned ok with the view still at 0x100001000, and the following
`rename main_inflate_zlib` renamed internal/abi.BoundsDecode instead — then the
rename's own snapshot showed `main_inflate @ 0x1002019b0`, because by the time
it was taken the goto had finally landed. Success reported, right-looking
readback, wrong function edited, and it survived a save.
This is what made an agent session stamp net_writeFull onto main_usage and
conclude the tooling was flaky.
goto/open now raise TimeoutError naming the target and where we actually are,
suggesting a larger timeout=. _press() does the same for follow/toggle_view/
hex/xrefs/structs, which had the identical "predicate ignored" shape.
tests/test_scenarios.py: 212 passed, 0 failed.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Five defects found while an agent drove a long RE session over the socket.
Each one was reproduced on a live spawned pane first (an in-process pilot
would not have shown any of them), then fixed:
pane stop truncated the save. `stop` asked the app to quit, slept 400 ms, then
unconditionally killed the pane. Quitting runs App.on_unmount, which writes
every dirty database; a 90 MB .i64 takes tens of seconds, so the kill landed
mid-write and a whole session's annotations went to /dev/null with a cheerful
{"stopped": [...]} on stdout. Now it waits for the pane to actually exit
(--timeout, default 600 s) and only force-kills on timeout, saying so.
The quit verb bypassed the dirty check. It called app.exit() directly rather
than the path a human gets, so the "unsaved changes" logic never ran. It now
routes through _on_quit_choice and reports {saving, dirty}.
Naming a function start from the listing never reached the function index.
`goto <addr>` puts the cursor on the address token, so `n` takes the
name-an-address path, which called bump_items() but left FunctionIndex holding
the old name. Result: the rename response snapshot showed the section label,
and functions()/names()/the palette all reported the rename had not happened —
so a driver that trusts its readbacks redoes work it already did. Twice, in
the session that prompted this. _do_name_addr now updates the index, the nav
stack and the table cell when the address is a function start.
A stripped binary with no entry function started up *inside a modal*.
_auto_land pushed the symbol palette when main() was missing, while ping still
answered ready:true. Every keystroke an RPC driver injected went into the
palette's search box and was silently swallowed. It now lands on the first
function instead and hints at Ctrl+N.
Verbs that inject keystrokes now refuse when a modal is on top, naming it,
instead of failing with "'goto' prompt did not open (word under cursor?)" —
a message that blamed the cursor for what was always a focus problem.
Also: `drive raw` passes k=v values through as strings, so `view lines=8` died
with "'<' not supported between instances of 'int' and 'str'". Numeric params
are now coerced centrally rather than at each call site.
tests/test_scenarios.py: 212 passed, 0 failed.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fixes the bug found while building M3. Navigations run in workers and are applied
when they land; the trace's OPENING seek goes to t=0, which for a normal binary
is _start, and that navigation is slow. It arrived after later seeks and won,
leaving the cursor and _cur on _start while the trace's pc was elsewhere — and it
never settled, measured stable for 3+ seconds. Anything cursor-based done just
after a seek (`>` asks about the address under the cursor) then acted on the
wrong address.
The decompiler path has had a staleness guard since 756589a; the listing path
never got one. It has one now (_open_at_if_current), and a seek bumps _nav_seq so
older in-flight navigations are dropped.
Verified both directions on the exact reproduction: seek to the first execution
of a repeated instruction, seek to the second, wait — cursor stays put with the
guard, and with the guard removed it drifts to 0x34d0 (_start) exactly as
reported.
Scope, deliberately narrow. I first bumped _nav_seq in _goto_ea for EVERY
navigation, which is the more general rule, and a full run then failed
follow_xrefs — a follow can be dropped by whatever navigates next. That check has
flaked before so it is not proof, but the mechanism is real and my evidence is
only about seeks, so the bump lives in _seek. TODO records what would justify the
general version and what test it needs.
tests: +1 trace UI (39) — seek, seek again, wait 3s, and the cursor is still on
the instruction the trace is at. Two consecutive full runs 212/0 after narrowing.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
M3. Stepping walks time; seeking jumps to the next time THIS thing was touched,
which is what makes a trace more than a very long single-step log.
`>` / `<` — next/previous execution of whatever the focused view addresses. One
pair of keys, two questions, because what's on screen already says which:
* listing: the instruction under the cursor. "When else did this run?"
* pseudocode: the whole C line, as the union of its instructions' executions. A
line is not one address, and falling back to its single /*ea*/ marker would
answer a narrower question — usually none at all, since most lines have no
marker.
* hex: the byte under the cursor, via memory_accesses.
It says where you landed ("execution of 0x3160: 2 of 2 @ t=320") and, at either
end, that you're AT the end rather than silently doing nothing — a key that does
nothing is indistinguishable from a broken one.
`W` — the registers with the instruction that set each to its current value, and
the distance back. Enter seeks to that write, f seeks forward. Backward is the
direction people want: you notice a bad value after it has been used. This is
the question a trace exists to answer and it was already in the model
(last_write/next_write), untested in anger until now.
tests: +13 trace UI (38) — > and < move between the two executions of a
repeated instruction, the status names which execution it is, both edges report
instead of moving, W opens, and choosing a register lands on an instruction that
REALLY wrote it (checked against the trace's own changed-set, not just the
timestamp matching).
Two things the tests taught me, both recorded:
* focus() does not make a view active outside split mode — Tab does. My first
seek test pressed > while _active was still "decomp", so it asked the
pseudocode about a line with no instructions.
* TODO gets a new entry: a stray late navigation to the entry function arrives
after a seek and wins, leaving the cursor on 'start' while the pc is
elsewhere. Same shape as the stale-decomp-result bug fixed in 756589a, which
got a sequence guard the listing path never did.
212/0 scenarios, 35/0 model, 12/0 differential.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
M2. Trace.memory(addr, length, idx) reconstructs what memory held at a moment,
returning the bytes AND a per-byte "known" mask. The mask is the point: a trace
knows what it observed and nothing else, so a byte nobody read or wrote is
genuinely unknown and must not be drawn as zero. That distinction is the whole
reason to read memory from a trace instead of the database — the database has
the file's bytes, the trace has what was actually there.
Reads count as evidence, not just writes: an instruction reading a byte reveals
what it held then.
Indexed by ADDRESS (sorted once, bisect per query), because the question is
"what was in this window at time t" and the accesses that matter are the few
touching that window, not the tens of thousands in the trace.
Where the memory actually is: measured, 0% of accesses in either real trace fall
inside the image — every one is stack or heap. So the primary view is the STACK,
in the dock, anchored at SP:
stack (rsp)
▸7ffff6f99470 ????????????????
7ffff6f99478 00007ffff6fb0b00
7ffff6f99488 00007ffff6fa94e5
The hex view overlays trace bytes on the file's contents (green = the trace saw
this byte at this timestamp, grey = still the file's idea). Correct, and it will
matter for a program that writes globals, but on these traces it shows nothing —
which is why the stack pane is the deliverable and not a nice-to-have.
One bug the work surfaced: MemOp.addr was having the image slide applied to it,
which is nonsense for a stack address — it produced -0xc838. The slide relocates
the IMAGE; stack and heap have no database counterpart. Memory op addresses now
stay in trace space, and memory_raw() queries there, while memory() takes
database addresses for the hex view.
tests: +9 model (35) covering the known-mask, reads-as-evidence, partial
coverage and the writers/accessors queries; +1 differential (12) checking
reconstructed memory state against Tenet's own get_memory at sampled timestamps;
+5 UI (30) for the stack pane — present, anchored at SP, marks unseen bytes,
follows time. 212/0 scenarios.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Finishing the thing the last commit only worked around. The split view and the
trace path each kept their own per-line instruction map of the same
pseudocode — fetched separately, indexed separately, and keyed differently: the
split one on _cur (the function the CURSOR is in), the trace one on the function
the DECOMPILER has loaded. Those are not the same thing, which is precisely how
the two ended up describing different functions and why I spent a commit
chasing a "sparse decomp_map" that was never sparse.
_apply_split_map now indexes once and both read it. Keyed on the decompiler's
loaded function, and no longer conditional on split being on — the old guard
dropped the result whenever _cur had moved while the fetch was in flight, which
during stepping is almost always.
Measured after: three decomp_map fetches across 28 steps (two for main, one for
the function stepped into), the split map and the trace map are literally the
same object, and both describe what is on screen.
Process note, because this is the second time: my first attempt at this edit
SILENTLY DID NOTHING — the pattern didn't match (a duplicated comment line I'd
mangled), the old method stayed, and the new caller hit its `not self._split`
guard, so the painting tests went from passing to "0 lines". Same failure mode
as the key bindings that never got added. Structural edits now assert that the
anchor was found and that the replacement is present before writing.
212/0 scenarios, 26/0 trace UI, 30/0 project UI, 20/0 split view.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I blamed decomp_map in the last commit. It was innocent: called directly it
returns 769 lines, 475 with addresses, for exactly the function I said it
returned four for. The four-line map belonged to a PLT stub the decompiler had
momentarily switched to, and I sampled mid-bounce.
The actual fault: _seek_split decided "has execution left the decompiled
function?" from _split_range, which is maintained by a guarded async path
(_apply_split_map drops its result if _cur moved while in flight) and therefore
lags during stepping. A stale range made every step look like a function change,
so the decompiler bounced main -> stub -> main, each bounce paying a synchronous
769-line map fetch on the UI thread.
Now the decision comes from the map the trail painting already holds, keyed to
what the decompiler currently HAS loaded. The bouncing is gone — three map
fetches across twelve steps instead of one per step — and the pseudocode cursor
follows every instruction the decompiler attributes to a line, including across
a call into another function.
What it does NOT do: guess. Roughly half of a function's instructions have no
line attributed, and the obvious fallback (nearest mapped address at or before
the pc) is unsound — C lines are not monotonic in address, and it put an
instruction early in main on line 708, "sub_2040();", near the end. The cursor
waits instead; the trail still marks where you are.
tests: +1 trace UI (26) — over ~28 steps, every instruction that IS mapped is
followed by the pseudocode cursor. 212/0 scenarios.
TODO corrected: the entry blaming decomp_map now says what actually happened,
including that _split_ea2line/_split_range are still fed by the laggy path and
remain a latent issue for the split view's own sync.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Normal navigation moves one pane and gives the companion a band, never a cursor
— that rule exists so the two can't chase each other. A trace step isn't
navigation: time is a single global position and both panes are showing the same
instant, so the cursor belongs on it in both.
_seek_split places the listing cursor on the current instruction, then hands off
to the existing _sync_split so the companion still gets its band and align() at
the driver's screen row. The anchoring machinery is used, not bypassed.
PARTIAL, and the shortfall is worth stating plainly: the LISTING cursor tracks
the pc reliably (tested over consecutive steps). The PSEUDOCODE cursor only
follows when decomp_map covers that address, and for cat's main it covers almost
nothing — four entries for a 700-line function. That is not something this
commit introduced and not something I could fix responsibly without
understanding it; TODO has what I measured, including that dec.goto(96) left the
cursor at 0 in the same run, which may or may not be the same bug.
One real fix along the way: _place_decomp_at prefers the map the trail painting
keeps (keyed to the decompiler's currently loaded function) over the split
view's _split_ea2line. The latter is refreshed by a guarded async path that
drops its result if _cur moved while in flight, and a burst of steps moves _cur
constantly — so during stepping it is frequently a map of the function you just
left.
tests: +2 trace UI (25) — stepping in split moves the listing cursor onto the pc
for six consecutive steps, and the trail marks it 'now' in both panes. 212/0
scenarios.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Caught while demoing this in a live pane, not by a test: press Tab to read the
pseudocode, press ] once, and you're back in the disassembly.
_seek() follows the trace by navigating to the new PC, and navigating to an
ADDRESS opens the listing unless the decompiler is explicitly preferred. So
every step out of C dropped you out of C — the painting work of the last commit
was unusable in the view it was built for, from the first keypress.
_seek now passes prefer_decomp=(self._active == "decomp"), the same thing the
xref handler already does for the same reason.
Worth noting what it looks like when it works: stepping in pseudocode follows
execution INTO a callee and the view switches to that function's C, which is
what you want and what makes the decompiler painting worth having.
tests: +2 trace UI (23) asserting the view survives a step in both directions.
212/0 scenarios.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The status bar said "0x2490" without saying what it belonged to. Obvious once
there are two panes open, or after switching binaries in a project — which
already prefixed its label, so single-binary sessions were the odd ones out.
[cat] .text @ 0x472b [listing] (c code · p func · u undefine · Enter follow)
Uses the opened file's basename, not _module(): that one asks the worker over
RPC and this runs on every status write. Kept in step when the path changes
(project switch, reload).
Three messages already carried the module name themselves and would have read
"[echo] echo — 128 functions"; they don't say it twice now.
tests: +3 scenarios (212) — the bar names the file, keeps naming it as you move
(the idle status is not the only writer), and doesn't say it twice.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Driving the trace viewer needed the same treatment as everything else: seeking by
hand through a few hundred keypresses to reach an interesting timestamp is not a
way to test it.
trace {seek: 120} absolute timestamp
trace {seek: "!50"} halfway through, like Tenet's timestamp shell
trace {goto: "main"} first execution of a name or 0xADDR
trace {step: 20} relative, negative goes back
trace {step: 5, over: true}
Returns the usual state snapshot plus {idx, length, pc, changed} so a driver can
see where it landed and what that instruction wrote.
pane spawn --trace FILE passes it through to the launcher, so a trace pane is one
command.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Both code views now show where you came from and where you're going: the
instruction you're on ('now'), the ~96 steps behind it ('past', warm) and the
~96 ahead ('future', cool).
A trail, not all of history. Painting every address the trace ever touched says
almost nothing on a loop-heavy program; the last and next few dozen steps say
how you GOT here. Where an address appears on both sides — a loop body, which is
most of them — the nearer side wins, because that's the one explaining the step
you just took or are about to.
**The pseudocode is painted too**, which is the reason to build this here rather
than use Tenet. A trace records instructions, so that's what Tenet paints. We
already have decomp_map from the split-view work, saying which instructions each
C line covers, so the same trail lands on the decompilation:
line 46 now | v3 = getenv("POSIXLY_CORRECT");
line 47 future | v4 = (__int64)*a2;
line 49 future | if ( v3 )
A C line covers many instructions, so it takes the strongest kind present: now
beats past beats future — if the instruction you're standing on belongs to this
line, this line is where you are.
Two things kept cheap: the trail is recomputed per SEEK rather than per repaint
(~200 lookups, and repaints vastly outnumber steps), and decomp_map is cached
per function because it's an RPC and stepping is interactive.
The colours sit deliberately under the code palette — the trail says "you came
through here", the text still has to read as code.
tests: +8 UI (21) — the listing carries now/past/future and it reaches the
screen; pseudocode is painted; exactly ONE C line is 'now' and it is the line
covering the current instruction (not merely some executed line, which is the
mistake this check exists to catch). 209/0 scenarios, 27/0 model, 10/0 diff.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
M0 of the trace viewer. --trace FILE loads a Tenet trace beside the binary and
docks a pane on the right: where you are in time, the register state there, and
a timeline.
Docked rather than modal on purpose. A trace turns every other view into "state
at time T", so time and registers are context you read WHILE looking at code,
not something you open and dismiss.
The registers the current instruction WROTE are highlighted. That difference is
the entire reason a delta trace is readable, and it's free — the trace already
says which registers each line changed.
] / [ step one instruction. } / { step over, by following the stack pointer: a
call pushes, so the callee runs with SP below where we started, and stepping
until SP comes back up lands after the return. That's cheaper and more portable
than recognising call instructions per architecture, and it degrades correctly —
on an instruction that calls nothing, SP is already >= the start and it's one
step. Verified on a real call: t=13 -> 18, past 5 instructions, where a plain
step gives 14.
The load waits for the function index because rebasing needs the database's
addresses: our echo trace runs at 0x7ffff6faa000 and the same code sits at
0x2000 in the database. Rebased -0x7ffff6fa8000, 12 functions touched.
Register values stay as the trace recorded them (they're machine state) while
everything else on screen is in database addresses, so the header shows both —
"pc 0x2aed (trace 0x7ffff6faaaed)" — rather than leaving the two to be puzzled
over side by side.
tests: test_trace_ui.py (13) records its own trace with the QEMU tracer and
drives the real UI — loads, rebases onto real functions, the dock renders, ] and
[ step and the code view follows, and } steps OVER a call found in that trace
rather than at a hardcoded index. Skips with a message if the tracer isn't
built. 209/0 scenarios.
One thing worth recording: my first attempt to add the key bindings SILENTLY did
nothing — the pattern contained a literal \\u2026 where the file has a real
ellipsis, so the replace matched nothing and the bindings never appeared. The
action worked when called directly, which made it look like a key-routing
problem. Assert on the replacement, not on the diff looking plausible.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
First slice of the execution-trace viewer: parse, index and query a Tenet trace.
No UI yet — this is the layer everything else stands on, and its shape decides
whether decompiler painting is cheap later.
A Tenet trace is a line-per-instruction delta log: registers that changed, the
PC every line, and each memory access WITH its bytes. That's enough to
reconstruct any register or address at any point in time, in either direction.
Our own reader, not a port. The reference (tenet-original) packs traces into
segments with compressed address/mask tables, which earns its keep for its Qt
timeline; we need different queries and would rather own ~400 lines than inherit
3700.
Indexed around the query the UI actually asks, which the reference answers one
address at a time: WHICH TIMESTAMPS EXECUTED THIS SET OF ADDRESSES. A listing
row is one address, but a pseudocode line covers many (we already have
decomp_map for that), so by_ip maps address -> timestamps and hits() takes a
set. Painting a pseudocode line will be one call, not one per instruction.
Registers are stored as per-register change points, so a value at time t is a
bisect, and "which instruction set this register?" (last_write) is the same
lookup — that being the question a trace explorer exists to answer.
Rebasing is not optional: our echo trace runs at 0x7ffff6faa000 while the
database has that code at 0x2490. Page offsets survive relocation, so the low 12
bits of an instruction address are invariant; bucket the database's addresses by
those bits and take the slide the most trace addresses agree on. Verified
against a real IDB: slide -0x7ffff6fa8000, and it picks out the 12 functions the
trace actually entered (main, start, ...) from 128.
Performance: 176k instructions parse in 350ms (~500k lines/s), so a 10M-line
trace is ~20s and wants a progress callback, which load() takes.
FOUND A BUG IN THE REFERENCE while building the differential test. A register
written on the LAST line of a 65535-line segment is missing from the next
segment's base state, so Tenet returns a stale value until that register is
written again — measured: wrong for all 179 timestamps of one such window. It
survives in the reference because it only shows when the register isn't
rewritten immediately.
That changed how the test works. Rather than "must agree with the reference",
it ARBITRATES with the raw text when they differ: if the text backs us it's
reported and allowed, if the text backs them it fails. Blanket agreement would
have made us copy their bug to stay green.
tests: test_trace.py (27, pure stdlib) covers reconstruction, the set queries,
rebasing (including that a lone agreeing address is not enough, and that
matching is on page offsets rather than addresses looking plausible) and
malformed input; test_trace_vs_tenet.py (10) diffs against the reference on real
traces.
|
