------- Additional Comments From amodra at bigpond dot net dot au 2004-12-06
11:26 -------
I think this particular problem was introduced with one of Geoff's patches,
http://gcc.gnu.org/ml/gcc-patches/2004-08/msg00022.html
rs6000_mode_dependent_address returns true for LO_SUM addresses, which in turn
makes them not offsettable. Thus movdi_internal32 won't match DImode gpr
stores.
It's clear that LO_SUM addresses can't be offset (since an offset may bump the
final address over a 64k+32k boundary, and the compiler can't know when this
might happen). Also, the rtl generated by -fprofile-generate looks OK. I don't
think there's anything wrong with Geoff's patch either, but it's boxed us into a
corner. A fix that will work (I think) is to limit legitimate_lo_sum_address_p
to 32 bit modes, but that could pessimize all loads and stores of 64 bit values.
--
What |Removed |Added
----------------------------------------------------------------------------
CC| |geoffk at geoffk dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18751
