http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51310
--- Comment #3 from Toon Moene <toon at moene dot org> 2011-12-05 20:04:21 UTC
---
At first I thought that gfortran would initialize small local arrays to
whatever -finit-real indicated by making them static, instead of stack based.
However, perusing the assembler output of
subroutine sub
real a(3)
print*,a
end
(as a diff between gfortran -S and gfortran -S -finit-real=snan):
17c17,27
< movq $.LC0, -536(%rbp)
---
> movl $1, %eax
> .L3:
> cmpq $3, %rax
> jg .L2
> leaq -1(%rax), %rdx
> movl $0x7fa00000, %ecx
> movl %ecx, -16(%rbp,%rdx,4)
> addq $1, %rax
> jmp .L3
> .L2:
and checking with a small main program
call sub
end
that the output is really different:
$ ./a.out
1.12103877E-44 0.0000000 9.80908925E-45
vs:
$ ./a.out
NaN NaN NaN
convinced me otherwise.
So one wonders why that doesn't work with variable length automatic arrays (or
why it would even be hard to initialize malloc'd automatic arrays /
allocatables).
This quest to be continued ("do you think coconuts migrate ? No - but they
could be carried.").
