Bug 25189

Summary: division by zero in Gfx::doAxialShFill
Product: poppler Reporter: M.Drochner
Component: cairo backendAssignee: poppler-bugs <poppler-bugs>
Status: NEW --- QA Contact:
Severity: major    
Priority: medium    
Version: unspecified   
Hardware: All   
OS: NetBSD   
Whiteboard:
i915 platform: i915 features:
Attachments: patch to avoid division-by-zero
new patch
better patch

Description M.Drochner 2009-11-19 10:15:09 UTC
Created attachment 31318 [details] [review]
patch to avoid division-by-zero

A PDF document (which I can make available but which is too large to post here)
could be viewed in "evince" but when I tried to print it the resulting PDF
was invalid. It turned out that it contained "nan" literally in linear shade
objects where numbers are expected.
The NaNs were originating from poppler's Gfx::doAxialShFill function. Both
tMin and tMax are zero in that situation (as was dy; dx was non-zero).
For now, I've using the appended patch to work around the problem.
(It seems to me that problem is in the intersection code but I didn't
look deeper yet.)
Comment 1 M.Drochner 2009-11-19 10:37:23 UTC
after a second look at my patch: substituting by tMin doesn't look correct --
if at all, a zero should be passed. But cairo clips the value anyway
(just NaN slips through), so this is not a major issue.
Comment 2 Albert Astals Cid 2009-11-19 11:35:36 UTC
Please send me or add an URL to the document
Comment 3 M.Drochner 2009-11-19 11:49:38 UTC
> Please send me or add an URL to the document
just did so per PM, please tell me if it doesn't arrive
Comment 4 Albert Astals Cid 2009-11-19 14:25:16 UTC
CairoOutputDev problem as it's the only one using useFillColorStop. Carlos?
Comment 5 M.Drochner 2009-11-20 06:51:23 UTC
The problem is not the intersection code but a somewhat arbitrary test that
prevents it from being used: dx is non-zero but less than 0.01 where the
problem occurs. So dxZero gets set unnecessarily.
My impression is that a axial shading with both "Coords" set to exactly
the same value is invalid, so why not bail out completely in that case?
I'll append a new patch which removes the problematic test, but does not
reject the "dxZero && dyZero" case yet -- perhaps I've missed some clause
in the PDF spec. With that patch, everything works well for me.
Comment 6 M.Drochner 2009-11-20 06:52:04 UTC
Created attachment 31342 [details] [review]
new patch
Comment 7 M.Drochner 2009-11-20 08:04:14 UTC
Created attachment 31344 [details] [review]
better patch

I've redone the patch slightly, to take into account the possibility that
a multiplication can hit the numerical limits. Even if it can't happen
in practice it looks more logical to me.