[Bug middle-end/30201] gcc doesn't unroll nested loops

[rguenth at gcc dot gnu dot org]

------- Comment #10 from rguenth at gcc dot gnu dot org  2006-12-14 15:46 
-------
I get on the trunk with -O2 -funroll-loops

main:
.LFB2:
        xorl    %eax, %eax
        .p2align 4,,7
.L2:
        cvtsi2sd        %eax, %xmm0
        addl    $1, %eax
        cmpl    $1000000000, %eax
        movq    $0, data+24(%rip)
        movq    $0, data+48(%rip)
        movq    $0, data+8(%rip)
        movq    $0, data+56(%rip)
        movq    $0, data+16(%rip)
        movq    $0, data+40(%rip)
        movsd   %xmm0, data(%rip)
        movsd   %xmm0, data+32(%rip)
        movsd   %xmm0, data+64(%rip)
        jne     .L2
        xorl    %eax, %eax
        ret

it doesn't look like we can do better w/o cheating and moving the
benchmark-loop-invariant movq $0 out of the loop ;)

The UNROLL variant looks like

.L2:
        leal    8(%rdx), %ecx
        addl    $9, %edx
        movq    $0, data+8(%rip)
        cmpl    $1000000000, %edx
        movq    $0, data+16(%rip)
        movq    $0, data+24(%rip)
        cvtsi2sd        %ecx, %xmm0
        movq    $0, data+40(%rip)
        movq    $0, data+48(%rip)
        movq    $0, data+56(%rip)
        movsd   %xmm0, data(%rip)
        movsd   %xmm0, data+32(%rip)
        movsd   %xmm0, data+64(%rip)
        jne     .L2


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30201