[PATH] have gcse consider a compare with the same register as a constant (except for floating point)

[Andrew Pinski <pinskia at physics dot uc dot edu>]

This patch fixes a problem on PPC where it moves a compare out of a 
loop but the compare is actually constant at compile time but gcc would 
not remove the branch inside the loop, the code is derived from art in 
SPEC2000 :
void temp(int winner, int limit, int *x)
{
        int i, test;
        int p;
        test = winner;
        for(i = 0; i<limit;i++)
        {
                if (test == winner)
                {
                        p = i<<2;
                }
                else
                {
                        p = i*30;
                }
                x[i] = p;
        }
}
Old gcc would produce:
_temp:
        li r9,0
        cmpw cr0,r9,r4
        bgelr- cr0
        cmpw cr7,r3,r3		<--- notice this is a compare before the loop 
and is constant.
        mtctr r4
L20:
        slwi r0,r9,2
        mr r2,r0
        beq- cr7,L16		<--- this branch will be always be taken
        mulli r0,r9,30
        slwi r2,r9,2
L16:
        stwx r0,r2,r5
        addi r9,r9,1
        bdnz L20
        blr


gcc with this patch, no extra branch in the loop:
_temp:
        li r2,0
        cmpw cr0,r2,r4
        bgelr- cr0
        mtctr r4
L20:
        slwi r3,r2,2
        addi r2,r2,1
        stwx r3,r3,r5
        bdnz L20
        blr

Another case I have to include -fno-guess-branch-probability because 
gcc gets the basic blocks wrong for some reason, I think it is because 
of gcse pulling rtl out of a loop and not putting it into its own bb 
but the bb for inside of the loop.


ChangeLog:

2003-05-02	Andrew Pinski <pinskia@physics.uc.edu>

	* gcse.c (gcse_constant_p): COMPARE of the same registers is a constant
	if they are not floating point registers.

Patch:

Attachment: gcse.diff
Description: Binary data

Bootstraped and tested on i686-unknown-openbsd3.1.
Bootstraped on powerpc-apple-darwin6.5 also,
	but not testing because I do not have a fixed kernel.


Thanks,
Andrew Pinski