Changes 1.3.23:

Special Issues:
* Due to GCC bug 53967, several key agorithms (e.g. convolution) may execute much faster (e.g. 2-3X) for x86-64 and/or when SSE is enabled for floating point math (-mfpmath=sse) if the GCC option -frename-registers is used. Default 32-bit builds do not experience the problem since they use '387 math. It is not clear in what version of GCC this problem started but it was not noticed by the developers until the GCC 4.6 timeframe. Other compilers do not suffer from this bug. Please lobby the GCC project to fix this embarrassing performance bug.

Security Fixes:
* ScaleImage(): While not strictly a security issue, requesting to scale an image while retaining the original number of rows will lead to a program crash or memory corruption due to double-free.

Bug fixes:
* ScaleImage(): Fix problem with new width/height match original (regression added by 1.3.22).
* ScaleImage(): Fix double-free when new rows matches original rows (regression added by 1.3.22).
* MinGW build fix related to eliminating a sleep() macro which conflicts with a MinGW-provided inline sleep() function.
* PNG: Issue a warning instead of an error when attempting to read a PNG file containing a zero-length profile. This allows the file to be read.
* identify: Fix problem in that identify -format "%A" (to test if transparency is supported in image) does not always produce the correct results.


(adam)