RFC - COST of const_double for x86 prevents constant copy propagation in cse

[Richard Henderson]

On Thu, Aug 25, 2005 at 12:37:32PM -0700, Ian Lance Taylor wrote:
> For what it's worth, as I told Fariborz, I suspect that returning 0 is
> correct for SFmode, but I'm somewhat doubtful for DFmode.

Indeed.

> And his test case is odd since the resulting code has more instructions
> and is larger.

Yes, and by more than double -- 21 vs 46 bytes.  I refuse to
believe this is better in the general case.

The only thing I can think is that in the specific case of eon,
either the integer memory ports are faster, or the xmm memory
ports are overcommitted or something.  Stuff that we don't really
have good modeling for in the backend, because we have no insight
into register allocation at the point at which this cse decision
must be made.


r~