This brings down the reload time from 22 seconds to 4.3 seconds with
my test document.  This is what I've been looking for!

These are deprecated and not used in LO.

that commit should have been cleaned up before it went upstream

it uses GOT instead of PLT on i386 with -02

White-space changes - code was in parts unreadable after tab/spaces
conversion. While at it, removed a few lines of commented-out code.

Did I mention I hate boilerplate copied all over the place? ;)

The clear() previously offsetted the topleft rect corner for sprites
by _the sprite topleft_ position, which is rather unhelpful but for
sprites located at (0,0).

Another fix for a rendering glitch mentioned in fdo#45219 -
vclcanvas sadly does not properly implement compositing, so all-
black background yields dark shadows on semi-transparent objects.

Ultimately though, we should retire vclcanvas instead.

Use clearAll() here, to really clear the whole sprite area (clear()
does respect potential clip polygons)

vclcanvas has 1bpp alpha - so cut-off transparency at 98%
This looks for all practical cases close enough to full alpha.

Rather unexpectedly, calling setupOutDevState() with IGNORE_COLOR
does not return a proper transparency, but null. Fixed now.

So, the Impress transparence default is 50% - we'd want that to
*not* be fully transparent on vclcanvas, so shift the cut-off by one

This nowadays really is a pessimization.

Make sure transparent polygons really only cover the polygonal area,
not the whole bounding box for sprites.

Use GetIndex() instead of GetBlue() (does the same, but more
self-documenting); use opaque when no alpha channel consistently.

We're creating the bitmap to exactly match our output size -
so why rounding off total transformation again, incurring another
expensive bitmap scale down in the canvas?