Patch to improve mov?f patterns

[Jan Hubicka]

> On Thu, Apr 15, 1999 at 10:43:55AM +0200, Jan Hubicka wrote:
> > This is simple patch to improve mov?f patterns.  They now use new constraint
> > that refuse all constants (even 0 and 1).  This allows combine to be more
> > smart and in operations like a+1 read 1 from memory as add operand.
> 
> Do you consider this an improvement simply because isolated
> code produces one insn instead of two?  What about cases like
> 
> 	for (i = 0; i < n; ++i)
> 	  flt[i] += 1.0f;
> 
> which will now do 
> 
> 	flds	.LC0
> loop:
> 	flds	(mem)
> 	fadd	%st(1)
> 	fstsp	(mem)
> 	...
> 
> instead of using `fld1' to begin.  Also, fld1 is theoritically
> faster due to memory load latencies.  This is probably hard to
> measure in practice though.
I believe, that fld1 will be still used in case combine fails to put
constant to some other operation. (at least it was used in my test)
thats why I allow immediate constants in reload.
(I my tests reload was replacing memory references to standard constant
by const doubles nicely)
I will do some more testing.

Honza