Debug code in `bentleyOttmann` writes to /tmp, prints to stdout, and races

[j-modernc-org]

Hi - thanks for canvas, we lean on it heavily.

path_intersection.go carries what looks like leftover debugging in toleranceSquares.breakupCrossingSegments (v0.0.0-20260708151538-b3beae08c578, line 1553; present in every version we have back to 2026-04):

square.Lower = prev
if square.Upper == nil {
    // TODO: this happens sporadically, add to unit tests
    w, _ := os.CreateTemp("", "canvas-testcase-*.gob")
    gob.NewEncoder(w).Encode([]any{_ps, _qs, _op, _fillRule})
    w.Close()
    fmt.Println("NOTE: new test case written to", w.Name())
    square.Upper = prev
}

and the package-level state it serialises, set on every call at line 1803:

var _ps, _qs Paths
var _op pathOp
var _fillRule FillRule

func bentleyOttmann(ps, qs Paths, op pathOp, fillRule FillRule) Paths {
    ...
    _ps, _qs, _op, _fillRule = ps, qs, op, fillRule

We understand why it is there - you want the input that triggers the case. Three consequences make it awkward for a library, and one of them is a bug independent of the debugging.

1. _ps/_qs/_op/_fillRule are a data race

bentleyOttmann writes those globals unconditionally, so any two goroutines doing boolean path operations race. This is the part we would flag even if the debug branch went away: concurrent rasterisation is a normal thing to do, and nothing in the API suggests Path.And is not safe to call from more than one goroutine.

func TestConcurrentBooleanOps(t *testing.T) {
    a := canvas.Rectangle(100, 100)
    b := canvas.Rectangle(50, 50).Translate(25, 25)
    var wg sync.WaitGroup
    for i := 0; i < 8; i++ {
        wg.Add(1)
        go func() { defer wg.Done(); for k := 0; k < 50; k++ { _ = a.And(b) } }()
    }
    wg.Wait()
}

WARNING: DATA RACE
Write at 0x000000bd6ba0 by goroutine 17:
  github.com/tdewolff/canvas.bentleyOttmann()
      path_intersection.go:1803
  github.com/tdewolff/canvas.(*Path).And()
      path_intersection.go:148
Previous write at 0x000000bd6ba0 by goroutine 10:
  ... same two frames

2. It panics when the temp file cannot be created

os.CreateTemp's error is dropped, and (*os.File).Name has no nil check - Close does, and gob only returns an error, so w.Name() is where it lands:

CreateTemp err = open /nonexistent/x-1224150888.gob: no such file or directory
panic: runtime error: invalid memory address or nil pointer dereference

A container with a read-only or absent TMPDIR turns a rendering call into a panic.

3. stdout and unbounded temp files

fmt.Println to stdout is hard for an embedder to deal with: a program whose stdout is the artifact (a PDF or PNG written to a pipe) gets corrupted output, and one that emits structured logs gets a stray line in the middle. The .gob files also accumulate with no cleanup - a long-running renderer trickles files into /tmp indefinitely.

What we see

We render HTML and use canvas for the vector back end, where boolean ops implement CSS clipping. We sometimes observe:

NOTE: new test case written to /tmp/canvas-testcase-3369210964.gob
NOTE: new test case written to /tmp/canvas-testcase-652071197.gob
... 12 more on one page load

We have not managed to reduce it to a small path pair - it fires on real page content and not on anything we have been able to trim down, which we realise is exactly the problem you left the branch there to solve. If it helps, we can send you the .gob files a real page produces; they are the inputs you are trying to capture. Say the word and we will attach them, or wire up a build that saves them somewhere you would prefer.

Suggested shape

Whatever suits you, but from an embedder's side:

• Put the capture behind an opt-in - a build tag, or an exported
  canvas.DebugPathIntersection io.Writer (nil by default) that the branch writes to.
  That also lets a caller collect the cases for you without patching the module.
• Drop the globals, or pass the inputs down to the point of use, so concurrent boolean
  ops stop racing.
• If the branch stays as is: check os.CreateTemp's error, and print to os.Stderr
  rather than stdout.

Happy to send a PR for any of these if you would like - just say which shape you prefer.

[tdewolff]

Thank you, all output is written to log which uses Stderr by default. You can now opt-in by setting canvas.DebugPathIntersection = true.

And yes please, I'm very interested in the test cases you've accumulated!

[j-modernc-org]

Gladly - except that when I went to package them up, every one of the 476 files on this
machine turned out to be 12 bytes, and byte-identical:

$ ls -l /tmp/canvas-testcase-*.gob | awk '{print $5}' | sort -n | uniq -c
    476 12
$ md5sum /tmp/canvas-testcase-*.gob | awk '{print $1}' | sort -u | wc -l
1

That is the third item in #388: Encode([]any{_ps, _qs, _op, _fillRule}) cannot work,
because gob will not encode a value through an interface whose concrete type has not been
registered. It fails with gob: type not registered for interface: canvas.Paths, its
error was dropped, and what lands on disk is the 12-byte type descriptor and nothing else.

So I have nothing to send you, and I suspect that is why you have never received a usable
capture from anyone - the channel has never worked, which would explain a TODO that
"happens sporadically" but never arrives with a reproducer attached.

Rather than send you a zip of zeros I tried to regenerate real ones. On current master
with DebugPathIntersection = true, rendering a range of real pages (okoun.cz,
pkg.go.dev, katex.org, news.ycombinator.com, sqlite.org, danluu.com, lobste.rs) and our
entire local test suite: 0 captures. We normally skip the intersection when the path's
bounds are wholly inside or outside the clip, so I disabled that too and put every
clipped draw through Path.And - 114 real intersections in one run, still 0 captures. I
also tried the April build we were pinned to when that batch appeared, on the same pages:
also 0.

So the case does not reproduce on demand for us today. It may well be that your work on
the intersection code since April closed whatever we were hitting.

What I can offer instead: we run a browser engine over real pages all day, so once #388
(or your own version of it) lands, I will leave DebugPathIntersection on in our
dogfooding runs and send you the first real captures the moment the branch fires. With
that patch they are a concrete pathIntersectionTestCase, so they read back with
gob.NewDecoder(f).Decode(&tc) from inside the package rather than being stubs.

Thanks for turning #387 around so quickly.

[tdewolff]

Thank you! Fixed now

[tdewolff]

I'm sorry, I thought to be possibly responding to AI content but figured that might not be the case. I'd be very happy to receive any bugs and test data you might find and will do my best to fix them, thank you in advance for the effort! Kind regards, Taco de Wolff

…

On Wed, Aug 19, 2026, 14:37 j-modernc-org ***@***.***> wrote: *j-modernc-org* left a comment (tdewolff/canvas#387) <#387 (comment)> Gladly - except that when I went to package them up, every one of the 476 files on this machine turned out to be 12 bytes, and byte-identical: $ ls -l /tmp/canvas-testcase-*.gob | awk '{print $5}' | sort -n | uniq -c 476 12 $ md5sum /tmp/canvas-testcase-*.gob | awk '{print $1}' | sort -u | wc -l 1 That is the third item in #388 <#388>: Encode([]any{_ps, _qs, _op, _fillRule}) cannot work, because gob will not encode a value through an interface whose concrete type has not been registered. It fails with gob: type not registered for interface: canvas.Paths, its error was dropped, and what lands on disk is the 12-byte type descriptor and nothing else. So I have nothing to send you, and I suspect that is why you have never received a usable capture from anyone - the channel has never worked, which would explain a TODO that "happens sporadically" but never arrives with a reproducer attached. Rather than send you a zip of zeros I tried to regenerate real ones. On current master with DebugPathIntersection = true, rendering a range of real pages ( okoun.cz, pkg.go.dev, katex.org, news.ycombinator.com, sqlite.org, danluu.com, lobste.rs) and our entire local test suite: *0 captures*. We normally skip the intersection when the path's bounds are wholly inside or outside the clip, so I disabled that too and put *every* clipped draw through Path.And - 114 real intersections in one run, still 0 captures. I also tried the April build we were pinned to when that batch appeared, on the same pages: also 0. So the case does not reproduce on demand for us today. It may well be that your work on the intersection code since April closed whatever we were hitting. What I can offer instead: we run a browser engine over real pages all day, so once #388 <#388> (or your own version of it) lands, I will leave DebugPathIntersection on in our dogfooding runs and send you the first real captures the moment the branch fires. With that patch they are a concrete pathIntersectionTestCase, so they read back with gob.NewDecoder(f).Decode(&tc) from inside the package rather than being stubs. Thanks for turning #387 <#387> around so quickly. — Reply to this email directly, view it on GitHub <#387?email_source=notifications&email_token=ABKOGHWIIHW6IX276YGV3OD5KXXWPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZUGY2DINZXGE3KM4TFMFZW63VMON2GC5DFL5RWQYLOM5S2KZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-5346447716>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABKOGHVT5QC2KFWRGMOIXIT5KXXWPAVCNFSNUABEKJSXA33TNF2G64TZHM4TCOJQG42TQMJ3JFZXG5LFHM2TCNRVGM2DAMRQHCQXMAQ> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/ABKOGHV62LTFF2SKD5DAIE35KXXWPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZUGY2DINZXGE3KM4TFMFZW63VMON2GC5DFL5RWQYLOM5S2KZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/ABKOGHVU3RFBZDDBXPVNTL35KXXWPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZUGY2DINZXGE3KM4TFMFZW63VMON2GC5DFL5RWQYLOM5S2KZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you modified the open/close state.Message ID: ***@***.***>

[j-modernc-org]

You are responding to often 100% AI-generated text, posted however only after I read and approve it or manually edit it on a per-case basis as I see necessary. However, I can miss errors, edit-in my own mistakes, forget to do this step even when intended etc.