This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: rtx_equal_function_value_matters


   Date: Wed, 25 Nov 1998 00:26:45 -0700
   From: Jeffrey A Law <law@hurl.cygnus.com>

   What cases are you thinking of?

None that exist, you see, I'm thinking ahead :-)

If someone finds a useful place, within the window you describe
(between cse_not_expected being set and life analysis pass 1), to emit
a constant load via gen_move(), I'd like to generate pseudos for the
benefit of scheduling.

Let's plan ahead, thats all I'm saying.

(BTW another incentive is that the constant formation code on sparc64
 can output very suboptimal things when new pseudos cannot be created,
 depending upon the properties the 64-bit constant has.

 For address formation in certain code models, it gets even worse,
 and tends to stress reload very heavily with secondary and tertiary
 reloads for such objects.

 So this is another incentive to allow new pseudo generation as long
 as possible, at least for sparc64.)

Later,
David S. Miller
davem@dm.cobaltmicro.com



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]