Hamfisted "fix" for pinterest blur crash

The Android perf tests are crashing (https://code.google.com/p/chromium/issues/detail?id=282418) when rendering pinterest on a Galaxy Nexus after r10936 (Blur refactoring - https://codereview.chromium.org/21835004/). I was not able to reproduce the exact crash they were seeing but did find one that may be the same one.

In the bug I found, zooming out on the pinterest page eventually causes the filterRectMask/BlurRect code to yield a 2x1 smallR[0]. The downstream 9-patch code doesn't deal with this properly and accesses invalid memory. 

I tried fixing this by rounding out the destination in BlurRect (since the src rect was being rounded out the dest rect's round was causing the dest rect to be smaller then the src rect). That did "fix" the crash but negatively impacted the GM images.

I'm not claiming this fix is right but it does stop the bleeding.

[robertphillips, 2013-09-03 14:33:04]

The reason this is happening is:

Back in the old days only the radius was passed through the CTM in
computeXformedRadius, then all the callers would (roughly) multiple the result
by 2 and add 1 (guaranteeing a lower bound on the blur radius).

Nowadays sigma is passed through the CTM (in computeXformedSigma) and then used
raw (i.e., 3*sigma out on each side), so the clamp on the minimum blur size has
been removed.

My guess for the pinterest case (since I haven't been able to directly reproduce
it) is that the zoom out required on the Galaxy Nexus is sufficient to get into
the case I  was able to observe.

[humper, 2013-09-03 14:57:23]

On 2013/09/03 14:33:04, robertphillips wrote:
> The reason this is happening is:
> 
> Back in the old days only the radius was passed through the CTM in
> computeXformedRadius, then all the callers would (roughly) multiple the result
> by 2 and add 1 (guaranteeing a lower bound on the blur radius).
> 
> Nowadays sigma is passed through the CTM (in computeXformedSigma) and then
used
> raw (i.e., 3*sigma out on each side), so the clamp on the minimum blur size
has
> been removed.
> 
> My guess for the pinterest case (since I haven't been able to directly
reproduce
> it) is that the zoom out required on the Galaxy Nexus is sufficient to get
into
> the case I  was able to observe.

LGTM.

I can see how this would happen, and could be that this might happen on the
desktop as well under certain circumstances.

[robertphillips, 2013-09-03 17:17:56]

committed as r11057