Looking at the asm for the program below, there plenty of loops left after compiling with > gfortran -S -march=native -O3 -funroll-loops -funroll-all-loops -fpeel-loops test.f90 or any combination of these options. A full unrolling (and in that case a return of the value 3) would be possible and much faster. > cat test.f90 INTEGER FUNCTION lxy() lxy=0 DO lxa=0,1 DO lxb=0,0 DO lya=0,1-lxa DO lyb=0,0-lxb lxy=lxy+1 ENDDO ENDDO ENDDO ENDDO END FUNCTION write(6,*) lxy() END
We don't unroll non-innermost loops at the moment. I don't know if sccp can be taught to handle this case (and if it's worth it).
(In reply to comment #1) > We don't unroll non-innermost loops at the moment. I don't know if sccp can > be taught to handle this case (and if it's worth it). such small loops are quite typical for some quantum chemistry integral routines. I'm just experimenting rewriting the kernel mentioned in PR 31021. If I do this unrolling by hand I get quite a speedup on the full kernel: hand unrolled: # best time 5.260329 loops: # best time 6.616413 which is quite impressive because these loops take at most 30% of the kernel total time: The actual code in question is: coef(:,:)=0.0_wp lxy=0 ; lx=0 DO lxa=0,1 DO lxb=0,1 lx = lx + 1 g1=0.0_wp g2=0.0_wp g1k=0.0_wp g2k=0.0_wp DO lya=0,1-lxa DO lyb=0,1-lxb lxy=lxy+1 g1=g1+pyx(1,lxy)*dpy(lyb,lya,jg) g2=g2+pyx(1,lxy)*dpy(lyb,lya,jg2) g1k=g1k+pyx(2,lxy)*dpy(lyb,lya,jg) g2k=g2k+pyx(2,lxy)*dpy(lyb,lya,jg2) ENDDO ENDDO DO icoef=1,3 coef(icoef,1)=coef(icoef,1)+alpha(icoef,lx)*g1 coef(icoef,2)=coef(icoef,2)+alpha(icoef,lx)*g2 coef(icoef,3)=coef(icoef,3)+alpha(icoef,lx)*g1k coef(icoef,4)=coef(icoef,4)+alpha(icoef,lx)*g2k ENDDO ENDDO ENDDO and the hand-unrolling just explicitly expands all loops to the loop free version of exactly the same statements: coef(:,:)=0.0_wp g1=0.0_wp g2=0.0_wp g1k=0.0_wp g2k=0.0_wp g1=g1+pyx(1,1)*dpy(0,0,jg) g2=g2+pyx(1,1)*dpy(0,0,jg2) g1k=g1k+pyx(2,1)*dpy(0,0,jg) g2k=g2k+pyx(2,1)*dpy(0,0,jg2) g1=g1+pyx(1,2)*dpy(1,0,jg) g2=g2+pyx(1,2)*dpy(1,0,jg2) g1k=g1k+pyx(2,2)*dpy(1,0,jg) g2k=g2k+pyx(2,2)*dpy(1,0,jg2) g1=g1+pyx(1,3)*dpy(0,1,jg) g2=g2+pyx(1,3)*dpy(0,1,jg2) g1k=g1k+pyx(2,3)*dpy(0,1,jg) g2k=g2k+pyx(2,3)*dpy(0,1,jg2) g1=g1+pyx(1,4)*dpy(1,1,jg) g2=g2+pyx(1,4)*dpy(1,1,jg2) g1k=g1k+pyx(2,4)*dpy(1,1,jg) g2k=g2k+pyx(2,4)*dpy(1,1,jg2) coef(01,01)=coef(01,01)+alpha(1,1)*g1 coef(01,02)=coef(01,02)+alpha(1,1)*g2 coef(01,03)=coef(01,03)+alpha(1,1)*g1k coef(01,04)=coef(01,04)+alpha(1,1)*g2k coef(02,01)=coef(02,01)+alpha(2,1)*g1 coef(02,02)=coef(02,02)+alpha(2,1)*g2 coef(02,03)=coef(02,03)+alpha(2,1)*g1k coef(02,04)=coef(02,04)+alpha(2,1)*g2k coef(03,01)=coef(03,01)+alpha(3,1)*g1 coef(03,02)=coef(03,02)+alpha(3,1)*g2 coef(03,03)=coef(03,03)+alpha(3,1)*g1k coef(03,04)=coef(03,04)+alpha(3,1)*g2k g1=0.0_wp g2=0.0_wp g1k=0.0_wp g2k=0.0_wp g1=g1+pyx(1,5)*dpy(0,0,jg) g2=g2+pyx(1,5)*dpy(0,0,jg2) g1k=g1k+pyx(2,5)*dpy(0,0,jg) g2k=g2k+pyx(2,5)*dpy(0,0,jg2) g1=g1+pyx(1,6)*dpy(0,1,jg) g2=g2+pyx(1,6)*dpy(0,1,jg2) g1k=g1k+pyx(2,6)*dpy(0,1,jg) g2k=g2k+pyx(2,6)*dpy(0,1,jg2) coef(01,01)=coef(01,01)+alpha(1,2)*g1 coef(01,02)=coef(01,02)+alpha(1,2)*g2 coef(01,03)=coef(01,03)+alpha(1,2)*g1k coef(01,04)=coef(01,04)+alpha(1,2)*g2k coef(02,01)=coef(02,01)+alpha(2,2)*g1 coef(02,02)=coef(02,02)+alpha(2,2)*g2 coef(02,03)=coef(02,03)+alpha(2,2)*g1k coef(02,04)=coef(02,04)+alpha(2,2)*g2k coef(03,01)=coef(03,01)+alpha(3,2)*g1 coef(03,02)=coef(03,02)+alpha(3,2)*g2 coef(03,03)=coef(03,03)+alpha(3,2)*g1k coef(03,04)=coef(03,04)+alpha(3,2)*g2k g1=0.0_wp g2=0.0_wp g1k=0.0_wp g2k=0.0_wp g1=g1+pyx(1,7)*dpy(0,0,jg) g2=g2+pyx(1,7)*dpy(0,0,jg2) g1k=g1k+pyx(2,7)*dpy(0,0,jg) g2k=g2k+pyx(2,7)*dpy(0,0,jg2) g1=g1+pyx(1,8)*dpy(1,0,jg) g2=g2+pyx(1,8)*dpy(1,0,jg2) g1k=g1k+pyx(2,8)*dpy(1,0,jg) g2k=g2k+pyx(2,8)*dpy(1,0,jg2) coef(01,01)=coef(01,01)+alpha(1,3)*g1 coef(01,02)=coef(01,02)+alpha(1,3)*g2 coef(01,03)=coef(01,03)+alpha(1,3)*g1k coef(01,04)=coef(01,04)+alpha(1,3)*g2k coef(02,01)=coef(02,01)+alpha(2,3)*g1 coef(02,02)=coef(02,02)+alpha(2,3)*g2 coef(02,03)=coef(02,03)+alpha(2,3)*g1k coef(02,04)=coef(02,04)+alpha(2,3)*g2k coef(03,01)=coef(03,01)+alpha(3,3)*g1 coef(03,02)=coef(03,02)+alpha(3,3)*g2 coef(03,03)=coef(03,03)+alpha(3,3)*g1k coef(03,04)=coef(03,04)+alpha(3,3)*g2k g1=0.0_wp g2=0.0_wp g1k=0.0_wp g2k=0.0_wp g1=g1+pyx(1,9)*dpy(0,0,jg) g2=g2+pyx(1,9)*dpy(0,0,jg2) g1k=g1k+pyx(2,9)*dpy(0,0,jg) g2k=g2k+pyx(2,9)*dpy(0,0,jg2) coef(01,01)=coef(01,01)+alpha(1,4)*g1 coef(01,02)=coef(01,02)+alpha(1,4)*g2 coef(01,03)=coef(01,03)+alpha(1,4)*g1k coef(01,04)=coef(01,04)+alpha(1,4)*g2k coef(02,01)=coef(02,01)+alpha(2,4)*g1 coef(02,02)=coef(02,02)+alpha(2,4)*g2 coef(02,03)=coef(02,03)+alpha(2,4)*g1k coef(02,04)=coef(02,04)+alpha(2,4)*g2k coef(03,01)=coef(03,01)+alpha(3,4)*g1 coef(03,02)=coef(03,02)+alpha(3,4)*g2 coef(03,03)=coef(03,03)+alpha(3,4)*g1k coef(03,04)=coef(03,04)+alpha(3,4)*g2k
Subject: Re: unroll/peel loops not aggressive enough > We don't unroll non-innermost loops at the moment. I don't know if sccp can > be taught to handle this case (and if it's worth it). It is fairly easy to make gcc completely unroll non-innermost loops, I am working on that.
Note that in addition to unrolling the outermost loop you can experiment with adjusting the --param max-completely-peeled-insns param. Also I wonder if DO lxb=0,0 is really common (if so, the frontend might want to lower this differently).
The optimization asked for in this PR is now being performed: > gfortran -O3 -funroll-loops -S test.f90 yields globl lxy_ .type lxy_, @function lxy_: .LFB2: movl $3, %eax ret .LFE2: .size lxy_, .-lxy_ .section .eh_frame,"a",@progbits .Lframe1: