Factor text size device mapping in SkScalerContext.

Factor text size device mapping in SkScalerContext.

All of our font back-ends use the text size in some way other than
simple concatentation with the current matrix. The code here finds
the full device matrix and then decomposes it to extract the text
size.

FreeType, GDI, and DirectWrite use the text size as the pre-hint scale.
CoreText will not scale color emoji above the requested text size,
and certain features like 'trak' are performed on the text size.

Committed: https://skia.googlesource.com/skia/+/5f14c5e038a16c80d50c2f87ae7c8775f977456c

[bungeman-skia, 2014-12-04 20:02:32]

bungeman@google.com changed reviewers:
+ mtklein@google.com, reed@google.com

[bungeman-skia, 2014-12-04 20:02:33]

I think I've got this doing more or less the same thing it used to do, except
now the code is all in one place and better documented. I'll run this through
our bots and the blink layout bots before committing.

I think there may be one extra copy of a matrix on the FreeType path in the fast
case, not sure that it matters. The ways I've tried to remove it make things
rather awkward.

[reed1, 2014-12-04 20:11:53]

1. how about you define a struct with all of the decomposed answers, and just
return that.

2. how about we always perform the decomposition in scalarcontext's constructor
(baseclass) and just make the answer struct (see #1) be available to the
subclasses.

[bungeman-skia, 2014-12-04 20:50:23]

On 2014/12/04 20:11:53, reed1 wrote:
> 1. how about you define a struct with all of the decomposed answers, and just
> return that.

In some sense

PreMatrixScale preMatrixScale, SkVector* scale, SkMatrix* remaining

go together (as required parameters).

SkMatrix* remainingWithoutRotation, SkMatrix* remainingRotation = NULL

also go together, as it makes no sense to have one without the other.

GDI and Mac (will) both need the total matrix, but they both will pre-scale it
with the EM size, because they're using it to work around bugs by just skipping
the underlying library altogether.

I'll see if I can get these related things together better.

> 
> 2. how about we always perform the decomposition in scalarcontext's
constructor
> (baseclass) and just make the answer struct (see #1) be available to the
> subclasses.

The issue is that an indeterminate amount of what is computed here is never
stored long term, but instead converted into the underlying platform's types,
like a DWRITE_MATRIX, MAT2, FT_Matrix, and the like. Computing all of this in
the base class constructor implies that we're going to keep the storage around.

[bungeman-skia, 2014-12-05 19:51:29]

For the time being, can I just add a big comment describing what it does and how
the results are used?

[reed1, 2014-12-05 20:12:20]

lgtm

[bungeman-skia, 2014-12-05 20:16:29]

The CQ bit was checked by bungeman@google.com

[commit-bot: I haz the power, 2014-12-05 20:17:09]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/748883005/120001

[commit-bot: I haz the power, 2014-12-05 20:26:50]

Committed patchset #7 (id:120001) as
https://skia.googlesource.com/skia/+/5f14c5e038a16c80d50c2f87ae7c8775f977456c

[bungeman-skia, 2015-03-17 21:15:19]

https://codereview.chromium.org/748883005/diff/120001/src/ports/SkFontHost_wi...
File src/ports/SkFontHost_win.cpp (right):

https://codereview.chromium.org/748883005/diff/120001/src/ports/SkFontHost_wi...
src/ports/SkFontHost_win.cpp:637: SkScalar gdiTextSize = scale.fY;
Here we need to call SkScalarRoundToScalar(). The one thing this calculation
does backwards from the old way is that the old way calculates the scale from
the text size, but here we're getting the text size from the scale (which is
arguably the right way to think about it).

In the 'no hinting' case, this was already rounded to an integer (so that the
rest of the matrix is minimized), so there is no change there. In the 'hinted'
case, we've removed the entire scale from the matrix, so the scale value is the
'real' residual size. We used to round this size, and we should continue to do
so.