Re: Doloop transform resulting in extra code

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

On May 11, 2004, at 16:05, Pat Haugen wrote:

> Zdenek,
>
> Has anyone pointed out to you that the doloop transform in mainline 
> results
> in extra code when compared to the 3.4 branch?  Actually, that's not 
> true
> for the -m64 case since 3.4 didn't generate a doloop for -m64 but I've
> included it anyway since it suffers from the same problem.
>
> Following is example source (trimmed down from bzip2.c) and generated 
> code
> for PowerPC target.  The extra tests (one for -m32, two for -m64) will
> never be true since the first test guarantees the loop count is > 0.

I think the reason why nobody pointed it out is because it works 
correctly
already on the tree-ssa (which should be merged in tomorrow into the 
mainline)
so this will be magically fixed.

Thanks,
Andrew Pinski
someone who does look at code gen on PPC

For an example with your code on the tree-ssa on powerpc-apple-darwin:
_hbAssignCodes:
        mr. r0,r5
        mtctr r0
        blelr- cr0
        li r2,0
L4:
        slwi r0,r2,2
        addi r2,r2,1
        stwx r4,r3,r0
        bdnz L4
        blr


and on the mainline:
_hbAssignCodes:
        li r2,0
        cmpw cr7,r2,r5
        bgelr- cr7
        cmpwi cr7,r5,0
        mtctr r5
        ble- cr7,L13
L8:
        slwi r0,r2,2
        addi r2,r2,1
        stwx r4,r3,r0
        bdnz L8
        blr
L13:
        li r0,1
        mtctr r0
        b L8