in the function F1 below, gfortran generates calls to _gfortran_size1 which (I
believe) makes that function about a factor 2-3 slower than should be. In
principle the optimizer should/could realize that size(a) is invariant, and at
least move it out of the loop.
> cat test.f90
INTEGER FUNCTION F1(a)
integer :: a(:,:)
F1=0
DO k=1,SIZE(a,2)
DO j=1,SIZE(a,1)
F1=F1+a(j,k)
ENDDO
ENDDO
END FUNCTION F1
[EMAIL PROTECTED]:~> cat main.f90
INTERFACE
INTEGER FUNCTION F1(a)
integer :: a(:,:)
END FUNCTION
END INTERFACE
INTEGER :: a(2,2),i,s
a=0
DO i=1,100000000
s=s+F1(a)
ENDDO
END
[EMAIL PROTECTED]:~> ifort -O3 -xW test.f90 main.f90 ; time ./a.out
main.f90(9): (col. 1) remark: (irrelevant) LOOP WAS VECTORIZED.
real 0m1.381s
user 0m1.376s
sys 0m0.004s
[EMAIL PROTECTED]:~> gfortran -O3 -march=native test.f90 main.f90 ; time ./a.out
real 0m3.578s
user 0m3.576s
sys 0m0.000s
--
Summary: calls to SIZE not optimized out of loops
Product: gcc
Version: 4.4.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: fortran
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: jv244 at cam dot ac dot uk
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38181
