Limit reload creativity on load 1 into xmm register

[Uros Bizjak]

Hello Jan!

>> On Fri, May 13, 2005 at 11:54:48PM +0200, Jan Hubicka wrote:
>> > While fixing the constraints I also noticed one missing #rf in movsf_1
>> > so fixed it one the way.
>> 
>> I'm largely convinced that these # constraints *are* the bug that you're
>> trying to work around.  When I was re-doing the SSE stuff a couple of 
>> months ago, I got better results by removing them.
>> 
>> I don't see how adding "!", meaning "last resort", is appropriate.
>
>Hi,
>apparently I suceeded with my homework and created self contained
>testcase for some reason working on i386 only, but it gives the idea.
>(it is tricky to do so as the wrong reload happens only when we have
>reg-reg copy where source is REG_EQUIV of 1 and it is reloaded out.  For
>example replacing the test2 asm in the testcase makes caller save code
>to save&load the constant on stack that is wrong too.)
>

I think that the testcase in PR 19653: "x87 reg allocated for constants 
for -mfpmath=sse" shows the same problem. I'm sorry that I'm a bit late, 
but I hope that it is of some help anyway. Please note, that sometimes 
loading of constant 0 could also confuse gcc to use fldz -> mem -> SSE 
reg sequence.

A realted problem could be observed when zero is produced in XMM 
register (or FP reg!) and then pushed to memory. A couple instructions 
later, this zero is loaded back from memory into XMM register. There is 
no reason not to calculate zero directly in XMM eg by using xor insn..

Uros.