------- Additional Comments From amylaar at gcc dot gnu dot org 2005-04-08
17:04 -------
(In reply to comment #1)
:2005-03-09 J"orn Rennecke <[EMAIL PROTECTED]>
: Richard Shann <[EMAIL PROTECTED]>
:
: * emit-rtl.c (gen_lowpart_common): Handle TRUNCATE. Generate
: TRUNCATE for non-noop truncations.
: * expr.c (force_operand): Handle TRUNCATE.
: * rtlanal.c (reg_overlap_mentioned_p): Likewise.
: * loop.c (combine_movables): Take TRULY_NOOP_TRUNCATION into account.
: * optabs.c (expand_binop): Likewise.
: * expmed.c (extract_fixed_bit_field): Check before assuming that
: op is a MEM.
Note that this a rather radical approach, in that it changes the behaviour
of gen_lowpart_common to generate TRUNCATE instead of an unsuitable
no-op truncation. This is efficient to safeguard against wrong code from the
numerous places that call gen_lowpart, but the downside is that then TRUNCATE
appears in interesting new places. This requires handling both in
machine-independent code and in every port that doesn't set
TRULY_NOOP_TRUNCATION to be always true. As far as I can see, this affects
exactly two ports: SH64 and mips. I will pursue this patch further if the
mips maintainers agree that this is a sensible approach to pursue, but not
otherwise.
In the meantime, I'll post a patch to gcc-patches that follows the more
'traditional' approach of fixing the callers of gen_lowpart, where these
problems have been detected. This has very little risk of breaking an existing
port, but a high probability to be incomplete.
--
What |Removed |Added
----------------------------------------------------------------------------
CC| |rsandifo at redhat dot com,
| |echristo at redhat dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20396
