There are several problems with a Form XObject's /Matrix entry 
created by pdfmark_BP() and with the cm command written to PDF by 
pdfmark_SP(), which lead to Form XObjects being scaled and placed 
incorrectly. For a sample file, see the one attached to bug 687429 
"Form XObject's /BBox may contain reals in exponential format".

(1)
The parameters for the cm command written by pdfmark_SP() are 
wrong. With the sample file, even if the user space is not modified 
at /SP time relative to the user space at /BP time (as is the case 
for the first /SP), the Form XObjects are mirrored, scaled and 
placed incorrectly.

(2)
Currently, pdfmark_SP() should output a cm command that corresponds 
to the transformation between the user space in effect when the 
Form XObject is created (/BP) and the user space in effect when 
the Form XObject is rendered (/SP). For example, if the user space 
did not change the corect command is "1 0 0 1 0 0 cm"; if between 
/BP and /SP there is a "2 2 scale", the command is "2 0 0 2 0 0 cm" 
(to display the Form XObject twice as big). However, pdfmark_SP() 
does not have enough information to compute this transformation, 
because at /SP time there is no trace of the user space that was 
in effect at /BP time.

(3)
Previous problem could be solved by pdfmark_BP() storing the 
inverse of its CTM. However, the Form XObject itself would still 
be incorrect. Per the PDF Reference, its /Matrix entry must 
transform the form space to user space. With GhostScript, the 
form space == device space, and /Matrix is always the identity, 
so it transforms to device space, not user space. Even if 
pdfmark_SP() could compensate for this by concatenating the 
inverse CTM (eventualy) stored by pdfmark_BP() with its own CTM, 
the Form XObject will still display incorrectly if used in places 
other than /SP pdfmarks, for example as an annotation appearance.

The proposed patch solves these problems by changing pdfmark_BP() 
to set /Matrix to the inverse of CTM, and pdfmark_SP() to use its 
CTM for generating the cm command.

Note:
A better solution would be to set form space == user space and 
/Matrix to the identity transform. This (a) eliminates some 
rounding errors and (b) works even if the CTM at /BP time cannot 
be inverted (not really usefull...), but I suppose this requires 
extensive changes to current code.