Jeffrey A Law law@cygnus.com
Wed Oct 28 00:10:00 GMT 1998

  In message <Pine.GSO.4.02A.9810270957060.8596-100000@colt.informatik.rwth-aac
hen.de>you write:
  > The patch below undoes H.J. Lu's Jul 26 change in i386.h.  It should no lon
  > ger
  > be necessary now that PREFERRED_RELOAD_CLASS has been fixed to work with
  > standard i386 constants.  It is also buggy, because it can cause aborts whe
  > n
  > trying to constrain_operands valid insns during reload.
  > HJ's original test case failed because during reload_as_needed, the variabl
  > e
  > current_function_uses_pic_offset_table was set, changing the elimination
  > offsets.  I suggest we perform some sanity checks that verify that nothing
  > vital changes during reload_as_needed, and abort if something is wrong.  Th
  > e
  > patch below contains such a sanity check, it works by making sure that neit
  > her
  > the initial elimination offsets nor the frame size change during
  > reload_as_needed.
  > Bernd
  > 	* reload1.c (verify_initial_offsets): New function.
  > 	(reload): Call it after reload_as_needed.  Also verify that the frame
  > 	size stays constant during reload_as_needed.
  > 	* i386.h (CONST_DOUBLE_OK_FOR_LETTER_P): Undo Jul 26 change.
I updated the reload changes to match the current sources and installed
this patch.


More information about the Gcc-patches mailing list