• A new subclass of Mapping, a SplineMap, has been added, which
  represents a 2-in, 2-out Mapping defined by a pair of two-dimensional
  B-spline surfaces.

• A bug that could cause astMapSplit to enter an infinite loop has been
  fixed.
• A memory leak in the Moc class has been fixed.
• A PolyMap that is equivalent to a simple shift of origin and/or a linear
  scaling on each axis is now simplified to a combination of equivalent
  simpler classes..
• If a PolyMap contains multiple coefficients that relate to the same
  input axis powers, they are now combined into a single coefficient.
• Zero-valued coefficients are now ignored by the PolyMap constructor,
  rather than being stored within the PolyMap.
• The FitsChan class can now create FITS headers containing SIP
  distortion keywords from a wider range of FrameSets. A PolyMap must
  still be present in the FrameSet but it is now allowed to deviate from
  the strict requirements of the SIP documentation in various ways (but
  only if it can be transformed into a PolyMap that does meet the
  requirements).

A bug has been fixed in the astRate method of the Mapping class, which
could cause numerical instability in the algorithm and thus return
incorrect values.

• A new attribute called IgnoreBadAlt has been added to the FitsChan
  class. If set, it allows unreadable alternate axis descriptions within FITS-WCS
  headers to be ignored. By default, such headers would cause an error to
  be reported.
• A new warning called BadAlt has been added to the FitsChan class. If
  included in the list of active warnings (see attribute "Warnings"), a
  warning will be created if any alternate axes are ignored when reading a
  FITS-WCS header.
• The FitsChan class now supports 64 bit integer keyword values.
• The Table class now supports 64 bit integer columns.
• The FitsTable class now supports 64 bit integer columns.

A bug in the flagging of permanent memory blocks when AST is configured --with-memdebug has been fixed.

• A bug has been fixed in the astRebinSeq4 wrapper where lbnd and ubnd
  had dimensions ndim_out rather than ndim_in. Additionally the dimension
  check error message was corrected to refer to input dimensions. These
  and other recent fixes to astRebinSeq were applied also to astRebin.

• Some memory allocations in PolyMap were corrected.

A bug has been fixed in fitschan.c that could cause an error when compiling using some versions of clang.
ast-9.2.9.tar.gz

Main Changes in this Version

• The Fortran interface has been changed to fix a bug caused by a change in the way that gfortran passes character arguments. As of gfortran V8 the length of each character argument passed to a subroutine is stored in a size_t value rather than an int value. AST now tests the version of gfortran at configure-time, and uses int or size_t within the Fortran API as appropriate. Prior to this change int was always used, which could cause undefined behaviour and segmentation faults when calling AST routines under gfortran v8 and later.

• A bug has been fixed in astRebinSeq that could cause NaN values to appear in the output array in cases where the overlap between two input arrays is very small. The fix corrects the way in which the mean weight per input pixel is calculated. It may cause some general minor changes to the decision about which output pixels pass the "wlim" criterion.

The FitsChan class has a new attribute called FitsRounding, which controls rounding applied when floating point values are formatted as FITS keyword values. Note, this changes the default rounding behaviour. Previously, sequences of four or more 9's or 0's were always rounded. Now, such sequences are only rounded if they extend beyond the tenth significant digit (the default value for FitsRounding being 10). To restore the original behaviour, set FitsRounding to 1.