Another note from the "big files" department:

What would *really* do wonders for the support of the images I'm working on
(multi-Gb TIFFs) is some kind of "progressive" image file load behaviour,
i.e. the ability to suppress the read of data blocks from file until it is
actually needed. This would mean transferring data from file to swap using
a similar algorithm to the one used today to copy data from swap to tile
cache. Actually, when viewing and anlysing files, but not changing their
data, the swap wouldn't even be necessary; the data could be read directly
from the file into the tile cache.

Of course, the total time taken to complete real image analysis and
transform operations might be unchanged by this, but I think I would be
able to work more efficiently because I wouldn't have to wait for several
minutes for the image window to open (when startining up in 1:1 zoom mode -
se bug 106960); I could have a quick glance at the data and start the
desired operation (and have the opportunity to verify that the correct file
was opened etc.) right away.

Also, implementing this kind of operation in a general way might be quite
hard, but it would be trivial for formats like tiled TIFF (which is also
what I'm mostly working on) where the data is already spit into smaller
sections. It would probably be a good idea to keep the impelementation of
the partial file read in the image format plugins, and have some kind of
interaction when the file is opened to decide if it is supported at all,
and for what block sizes.