1517551 - Extremely wide "view-source" for New Yorker article shows the wrong text (doesn't match what you copypaste)

Status: NEW

(Core :: Layout: Text and Fonts, defect, P3)

Description

[Daniel Holbert [:dholbert]]

Attachment: testcase 1

STR:
 1. View the source of attachment (which is slightly-reduced HTML from https://www.newyorker.com/magazine/2019/01/07/father-time )

 2. Do find-in-page (in the view-source tab) for the string "historical data"


EXPECTED RESULTS:
There should be 2 highlighted matches, for instances of the string "historical data".

ACTUAL RESULTS:
The highlighted matches show different characters -- it's as if we didn't find the string correctly. The first match looks like
> 1970s," and that
And the second match looks like:
>  the distant ma

But in fact, if you select the "matched" text and copy it to clipboard and then paste it somewhere else, you'll find that it was the correct characters -- but we were just painting the wrong thing.


Regression range:
https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=64bab5cbb9b63808d04babfbcfba3175fd99f69d&tochange=aa958b29c149a67fce772f8473e9586e71fbdb46

Comment 1

[Daniel Holbert [:dholbert]]

Attachment: screenshot showing the incorrect text painted for the find-in-page matches

Comment 2

[Daniel Holbert [:dholbert]]

In the regression pushlog, this looks likely to be a regression from:
	d1f4b0f6182e	Jonathan Kew — Bug 1408612 - patch 1 - Use single-precision floats more consistently for glyph position computations when drawing text. r=jrmuizel

...which seems to have reduced our precision in some text positioning calculations when drawing text.  (And that breaks this scenario which has text placed at extremely large positions.)

Comment 3

[Xidorn Quan [:xidorn] UTC+10]

FWIW, I cannot reproduce this on macOS.

Comment 5

[Mike Kaply [:mkaply]]

Why would a regression be a P5?

Comment 6

[Daniel Holbert [:dholbert]]

(In reply to Mike Kaply [:mkaply] from comment #5)

Why would a regression be a P5?

Yeah, sorry - this should really be P3. I think our P1/P3/P5 classification has been formalized a bit better since this was triaged (at least in the layout team's triage protocol) -- and under our current interpretation, P3 feels like the right bucket.

(Having said that -- RE this being a regression: note that we shipped with this bug for 2 years without any reports, & without anybody noticing. So it seems not to be a particularly prominent regression, and it's perhaps not one that many users are likely to encounter. So, I think it's reasonable to consider it somewhat lower-priority than more user-visible regressions & regressions that we catch before or soon-after shipping them. Anyway - still a valid bug, though!)

Comment 7

[Mike Kaply [:mkaply]]

Agreed, but I wonder if there are other things that regressed because of this that we haven't caught?