consider the following source and timings, were a natural form of a subroutine S1, and two hand optimized forms are timed: > cat test.f90 SUBROUTINE S1(N,A) REAL :: A(3) DO I=1,N CALL S2(-A) ENDDO END SUBROUTINE SUBROUTINE S1_opt1(N,A) REAL :: A(3) REAL, ALLOCATABLE :: B(:) ALLOCATE(B(SIZE(A,1))) DO I=1,N B=-A CALL S2(B) ENDDO END SUBROUTINE SUBROUTINE S1_opt2(N,A) REAL :: A(3),B(3) DO I=1,N B=-A CALL S2(B) ENDDO END SUBROUTINE > cat main.f90 SUBROUTINE S2(A) REAL :: A(*),D COMMON /F/D D=D+A(1)+A(2)+A(3) END SUBROUTINE INTEGER, PARAMETER :: N=100000 REAL :: A(3),T1,T2,T3,T4,D COMMON /F/D D=0.0 A=0.0 CALL CPU_TIME(T1) DO I=1,10000 CALL S1(N,A) ENDDO CALL CPU_TIME(T2) DO I=1,10000 CALL S1_opt1(N,A) ENDDO CALL CPU_TIME(T3) DO I=1,10000 CALL S1_opt2(N,A) ENDDO CALL CPU_TIME(T4) write(6,*) "Default [s]:",T2-T1 write(6,*) "OPT1 [s]:",T3-T2 write(6,*) "OPT2 [s]:",T4-T3 write(6,*) D END gfortran-4.4 -O3 test.f90 main.f90 Default [s]: 18.293142 OPT1 [s]: 6.2603912 OPT2 [s]: 6.2563915 ifort -O3 test.f90 main.f90 Default [s]: 6.256391 OPT1 [s]: 6.252390 OPT2 [s]: 6.256390 so, gfortran by default is about 3x slower than ifort, which by default moves the generation of the temporaries out of the loop. FYI, allowing for multi file IPO, I hope LTO gets that far... ifort -O3 -fast test.f90 main.f90 (includes ipo) Default [s]: 3.752234 OPT1 [s]: 1.276080 OPT2 [s]: 3.752234
This could also be useful when done in the middle-end, see PR 21046.
seemingly being discussed, since useful for tonto http://gcc.gnu.org/ml/fortran/2010-02/msg00157.html
(In reply to comment #2) > seemingly being discussed, since useful for tonto > > http://gcc.gnu.org/ml/fortran/2010-02/msg00157.html > But there: "it's unfortunately not possible to avoid the temporary creation without serious data-flow analysis work - too late for the frontend" Thus, this seems to be more a middle-end item. Regarding the current timing, I get with ifort -O3 -fast (v 11.1) vs. gfortran -flto -fwhole-program -O3 --fast-math -march=native (today's 4.5) It is also interesting that gfortran is much faster for the optimized version than ifort. gfortran ifort Default [s]: 24.881554 Default [s]: 5.108319 OPT1 [s]: 1.6641045 OPT1 [s]: 3.280205 OPT2 [s]: 1.6641045 OPT2 [s]: 4.988311 0.0000000 0.0000000E+00 real 0m28.420s real 0m13.400s
(In reply to comment #3) > (In reply to comment #2) > > seemingly being discussed, since useful for tonto > > > > http://gcc.gnu.org/ml/fortran/2010-02/msg00157.html > > > > But there: "it's unfortunately not possible to avoid the temporary creation > without serious data-flow analysis work - too late for the frontend" > > Thus, this seems to be more a middle-end item. right, changing component as such. This would actually be much more powerful as a middle-end thing, since it would also capture the case where a programmer would explicitly allocate/deallocate stuff in a loop.
Another case of interest is "automatic arrays". An interesting example is the polyhedron test nf.f90. On Core2 Duo and Darwin the following patch --- nf.f90 2005-10-11 22:53:32.000000000 +0200 +++ nf_v2.f90 2010-10-07 16:49:38.000000000 +0200 @@ -153,7 +153,7 @@ integer :: nx , nxy , nxyz , maxiter real(dpkind),dimension(nxyz):: ad,au1,au2,au3,x,b real(dpkind)::targrms -real(dpkind),allocatable,dimension(:) :: r,q,p,z,g,gi +real(dpkind),allocatable,dimension(:) :: r,q,p,z,g,gi,t,u real(dpkind):: alpha,beta,qr,qrp,rmserr integer :: iter , tbase , tgi , tcg , tickspersec , maxticks @@ -163,7 +163,7 @@ call GetGI3D(1,nxyz) ! c call system_clock(tgi,tickspersec,maxticks) deallocate(g) -allocate (r(nxyz),q(nxyz),p(nxyz),z(nxyz)) +allocate (r(nxyz),q(nxyz),p(nxyz),z(nxyz),t(nxyz),u(nxyz)) CALL SPMMULT(x,r) ; r = b - r ! compute initial residual vector write(*,'(A)') ' Iter Alpha Beta RMS Residual Sum of Residuals' @@ -171,12 +171,12 @@ write(*,'(I4,24X,2G18.7)') 0,sqrt(DOT_PR ! Do a single iteration with alpha =1 ! to reduce sum of residuals to 0 -p = r ; CALL NF3DPrecon(p,1,nxyz) ; CALL SPMMULT(p,z) +p = r ; CALL NF3DPrecon(p,t,u,1,nxyz) ; CALL SPMMULT(p,z) x = x + p ; r = r - z write(*,'(I4,F12.5,12X,2G18.7)') 0,1.0,sqrt(DOT_PRODUCT(r,r)/nxyz),sum(r) do iter = 1 , maxiter - q = r ; CALL NF3DPrecon(q,1,nxyz) + q = r ; CALL NF3DPrecon(q,t,u,1,nxyz) qr = DOT_PRODUCT(q,r) if ( iter==1 ) then beta = 0.0 @@ -197,7 +197,7 @@ call system_clock(tcg,tickspersec,maxtic write(*,'(/A,F10.3/A,F10.3/A,F10.3)') ' Time for setup ',REAL(tgi-tbase)/REAL(tickspersec) , & ' Time per iteration ',REAL(tcg-tgi)/REAL(tickspersec*min(iter,maxiter)) , & ' Total Time ',REAL(tcg-tbase)/REAL(tickspersec) -deallocate(r,q,p,z,gi) +deallocate(r,q,p,z,gi,t,u) contains !========================================= ! Banded matrix multiply b = A.x ========= @@ -253,7 +253,7 @@ end subroutine GetGI2D !== !========================================= ! solve for a plane of cells using ====== -subroutine NF2DPrecon(x,i1,i2) ! 2D NF Preconditioning matrix +subroutine NF2DPrecon(x,t,i1,i2) ! 2D NF Preconditioning matrix integer :: i1 , i2 real(dpkind),dimension(i2)::x,t integer :: i @@ -272,11 +272,12 @@ end subroutine NF2DPrecon !== subroutine GetGI3D(i1,i2) ! compute gi for a 3D block of cells ===== integer :: i1 , i2 integer :: i +real(dpkind),dimension(nxyz)::t g = ad do i = i1 , i2 , nxy ! advance one plane at a time if ( i>i1 ) then ! get contribution from previous plane g(i-nxy:i-1) = au3(i-nxy:i-1) - call NF2DPrecon(g,i-nxy,i-1) + call NF2DPrecon(g,t,i-nxy,i-1) g(i:i+nxy-1) = g(i:i+nxy-1) - au3(i-nxy:i-1)*g(i-nxy:i-1) endif call GetGI2D(i,i+nxy-1) ! get contribution from this plane @@ -285,17 +286,17 @@ end subroutine GetGI3D !== !========================================= ! solve for a 3D block of cells using -subroutine NF3DPrecon(x,i1,i2) ! 3D Preconditioning matrix +subroutine NF3DPrecon(x,t,u,i1,i2) ! 3D Preconditioning matrix integer :: i1 , i2 -real(dpkind),dimension(i2)::x,t +real(dpkind),dimension(i2)::x,t,u integer :: i do i = i1 , i2 , nxy if ( i>i1 ) x(i:i+nxy-1) = x(i:i+nxy-1) - au3(i-nxy:i-1)*x(i-nxy:i-1) - call NF2DPrecon(x,i,i+nxy-1) + call NF2DPrecon(x,u,i,i+nxy-1) enddo do i = i2-2*nxy+1 , i1 , -nxy t(i:i+nxy-1) = au3(i:i+nxy-1)*x(i+nxy:i+2*nxy-1) - call NF2DPrecon(t,i,i+nxy-1) + call NF2DPrecon(t,u,i,i+nxy-1) x(i:i+nxy-1) = x(i:i+nxy-1) - t(i:i+nxy-1) enddo end subroutine NF3DPrecon !========================================= cuts the execution time from ~28s to ~20s (Note that with the options I use all the procs are inlined).
Marc, I think your recently posted patch: http://gcc.gnu.org/ml/gcc-patches/2013-11/msg01049.html could fix the problem with the testcase subroutine S1, even though 'moving allocations out of loops' is more or less a side effect.
(In reply to Joost VandeVondele from comment #6) > Marc, I think your recently posted patch: > http://gcc.gnu.org/ml/gcc-patches/2013-11/msg01049.html > could fix the problem with the testcase subroutine S1, even though 'moving > allocations out of loops' is more or less a side effect. I don't speak fortran fluently so I tried compiling S1 with an unpatched compiler and -O2 -fdump-tree-optimized, but I don't see any call to malloc in there. Could you explain, with references to a dump, what the internal functions mean and where my patch might help?
(In reply to Marc Glisse from comment #7) > (In reply to Joost VandeVondele from comment #6) > > Marc, I think your recently posted patch: > > http://gcc.gnu.org/ml/gcc-patches/2013-11/msg01049.html > > could fix the problem with the testcase subroutine S1, even though 'moving > > allocations out of loops' is more or less a side effect. > > I don't speak fortran fluently so I tried compiling S1 with an unpatched > compiler and -O2 -fdump-tree-optimized, but I don't see any call to malloc > in there. Could you explain, with references to a dump, what the internal > functions mean and where my patch might help? Marc, looks like the fortran FE changed a lot since this bug was filed, and there is no explicit allocate anymore, in fact the variable is created on stack by the frontend... this is controlled by -fmax-stack-var-size=0 (putting it to zero, will yield your __builtin_malloc() that I recalled, in the PR38318.f90.003t.original dump). You have a precedent for getting the a reasonable size (32768 for fortran). The _gfortran_internal_(un)pack is a fortran FE thing, that guarantees that memory is contiguous... clearly a missed frontend optimization in this case. So now, the proper testcase would be: > cat PR38318-3.f90 SUBROUTINE S1(N,A) REAL :: A(3) REAL, DIMENSION(:), ALLOCATABLE :: B DO I=1,N ALLOCATE(B(3)) B=-A CALL S2(B) DEALLOCATE(B) ENDDO END SUBROUTINE which really should contain any call to _gfortran_runtime_error_at, _gfortran_os_error, __builtin_malloc, __builtin_free if all were perfect, and certainly not in the loop
(In reply to Joost VandeVondele from comment #8) > Marc, looks like the fortran FE changed a lot since this bug was filed, and > there is no explicit allocate anymore, in fact the variable is created on > stack by the frontend... Cool, the best optimizations are those you don't need to do ;-) > So now, the proper testcase would be: > > cat PR38318-3.f90 > SUBROUTINE S1(N,A) > REAL :: A(3) > REAL, DIMENSION(:), ALLOCATABLE :: B > DO I=1,N > ALLOCATE(B(3)) > B=-A > CALL S2(B) > DEALLOCATE(B) > ENDDO > END SUBROUTINE > > which really should contain any call to _gfortran_runtime_error_at, > _gfortran_os_error, __builtin_malloc, __builtin_free if all were perfect, > and certainly not in the loop Ok. If you used __builtin_abort instead of _gfortran_os_error, I think my current patch would handle it. It is hard for gcc to guess that _gfortran_os_error is safe. On the other hand, if I special case the test if(VAR==0) as mentioned in a comment in my patch, it won't look at that branch anymore and the optimization should apply. Er, no, I missed the call to s2. I would also need some attribute on s2 so the compiler knows that s2 doesn't do anything too weird. Hopefully, when the compiler has the sources for s2, we could later let it guess those attributes...
(In reply to Marc Glisse from comment #9) > > Ok. If you used __builtin_abort instead of _gfortran_os_error, I think my > current patch would handle it. It is hard for gcc to guess that > _gfortran_os_error is safe. For the Fortran FE people (not me, I'm a user), but _gfortran_os_error should have an attribute like 'abort' or 'noreturn'. However, the compiler should also be able to figure out this can never be called (if B is 'allocated on the stack') in this subroutine. > Er, no, I missed the call to s2. I would also need some attribute on s2 so > the compiler knows that s2 doesn't do anything too weird. Actually, in Fortran, S2 can't do anything 'weird' with B, in the sense that your optimization should certainly apply. Not so sure about the correct terms here, but in approximate C-speak, B 'as a pointer' is guaranteed to be pointing to exactly the same address, nothing has happened to its target, and no pointer can be pointing to whatever B was pointing to....
(In reply to Joost VandeVondele from comment #10) > (In reply to Marc Glisse from comment #9) > > Ok. If you used __builtin_abort instead of _gfortran_os_error, I think my > > current patch would handle it. It is hard for gcc to guess that > > _gfortran_os_error is safe. > > For the Fortran FE people (not me, I'm a user), but _gfortran_os_error > should have an attribute like 'abort' or 'noreturn'. abort doesn't exist, and noreturn is not sufficient, as a function that calls free on the pointer then exits is noreturn but unsafe. > However, the compiler > should also be able to figure out this can never be called (if B is > 'allocated on the stack') in this subroutine. Yes. > > Er, no, I missed the call to s2. I would also need some attribute on s2 so > > the compiler knows that s2 doesn't do anything too weird. > > Actually, in Fortran, S2 can't do anything 'weird' with B, in the sense that > your optimization should certainly apply. Not so sure about the correct > terms here, but in approximate C-speak, B 'as a pointer' is guaranteed to be > pointing to exactly the same address, nothing has happened to its target, > and no pointer can be pointing to whatever B was pointing to.... So S2 cannot call free (or realloc) on the pointer and then exit or call longjmp or do an infinite loop or anything like that in fortran? Maybe we'll need a flag set by the front-end that says whether (all) functions are safe.
(In reply to Marc Glisse from comment #11) > So S2 cannot call free (or realloc) on the pointer and then exit or call > longjmp or do an infinite loop or anything like that in fortran? Maybe we'll > need a flag set by the front-end that says whether (all) functions are safe. well, not free or realloc or longjmp, but infinite loops are allowed in S2. The point is, one is really not passing a pointer to S2 (from a Fortran point of view).