Add a Canvas::SaveAndUnscale() method for pixel-accurate drawing.

ui/gfx/canvas.cc

Index: ui/gfx/canvas.cc
diff --git a/ui/gfx/canvas.cc b/ui/gfx/canvas.cc
index 49e32350a42e45c8c498aa7cd00534f6c578b4fb..503cee00e9a7b8ef12c29d0e8653dbacfd4ec106 100644
--- a/ui/gfx/canvas.cc
+++ b/ui/gfx/canvas.cc
@@ -180,6 +180,13 @@ void Canvas::Save() {
   canvas_->save();
 }
 
+float Canvas::SaveAndUnscale() {

[sky, 2015/10/13 22:23:11]

How expensive is save? The common case is scale=1. Is it worth optimizing that?

[Peter Kasting, 2015/10/13 22:34:42]

Save() is very cheap.  Also we can't get rid of it when the scale factor is 1
because callers are going to unconditionally Restore().

scale(1, 1) is pretty cheap as well, so I didn't bother to try to optimize it
away.

+  Save();
+  SkScalar scale_factor = 1.0f / image_scale_;

[sky, 2015/10/13 22:23:11]

I'm not familiar enough with what we want to happen, but did you consider
unscaling by the current transform?

[Peter Kasting, 2015/10/13 22:34:42]

I did, but I'm not convinced that's what we actually want.  We want to undo the
"device pixel -> DIP" transformation, but imagine if some other caller, e.g. a
"magnifying glass effect" implementation, had added an additional scale atop
this.  We'd still want that scale to apply to the current canvas.

+  canvas_->scale(scale_factor, scale_factor);
+  return image_scale_;
+}
+
 void Canvas::SaveLayerAlpha(uint8 alpha) {
   canvas_->saveLayerAlpha(NULL, alpha);
 }