Re: Switch to disable initialization of local structures in subprogram.

[Tobias Burnus <burnus at net-b dot de>]

On 09/28/2011 05:34 PM, Paul van Delst wrote:
> These structures are defined like so:
>
>    TYPE :: CSVar_type
>      TYPE(CSinterp_type) :: csi(MAX_N_LAYERS, MAX_N_CLOUDS)

How fancy is the type CSinterp_type?

Otherwise, I do not see why gfortran should do anything fancy. As 
variable with compile-time known size, it should be created on either 
the stack on in static memory, depending on the size of 
-fmax-stack-var-size= and whether -frecursive/-fopenmp/RECURSIVE is used.

Especially, if the data is in static memory, there should be no 
performance penalty, but also stack-allocated memory is fast.

As a test, you could try to use the "SAVE" attribute on the variables 
and check whether it makes a difference.

Note that by commenting code, you also change the function size - and 
thus possibly allowing to inline code which was before kept as a 
separate code.

You could use, e.g., -finline-limit=<n> (n could be for instance 600) 
for that file to see whether it makes a difference.*

You could also try to create a smaller example and use 
"-fdump-tree-original" which creates a file which contains the generated 
code with a C-like syntax. That should allow you to see what 
GCC/gfortran does internally. (Note: The dump does not contain all 
information.)

Tobias

* In GCC 4.7, the inliner has seen some modifications which should help 
Fortran. Additionally, there is the flag -fstack-arrays, which generates 
automatic arrays on the stack rather than allocating it on the heap, 
which significantly improves the performance of some programs (and 
requires a lot of stack memory).