topic Data alignment in common blocks in Intel® Fortran Compiler

Data alignment in common blocks

rfvujm — Fri, 12 Mar 2010 15:29:57 GMT

Hi,

I'm currently developing a static library in Fortran containing common blocks. These COMMONswill be accessed from C++ application, that links the fortran library. As the C++ app requirement is that the data from extern "C" is aligned by 1 byte, I was hoping to set this in VF11 compiler settings, however I have only found /allign: commons (4 bytes) and/allign: dcommons (8 bytes). Is there a different way of solving this problem (e.g. in-line allignment macro)? I would be most grateful for your help.

Regards,

Mario

Data alignment in common blocks

TimP — Fri, 12 Mar 2010 15:54:15 GMT

Generally speaking, it's unreliable to depend on schemes for including padding between variables in COMMON, particularly when interoperating between languages. Perhaps that is the point you are making.

The traditional recommendation to declare all data in a COMMON in decreasing order of size is still good, to avoid requirements for padding to produce correct alignments. This is satisfactory, in the absence of requirement for efficient vectorization.

In order to promote vectorization without inclusion of padding, we set arrays to specific sizes, such as multiples of 16. That may not be among your concerns.

Data alignment in common blocks

Steven_L_Intel1 — Fri, 12 Mar 2010 16:40:08 GMT

Declare your COMMON with the BIND(C) attribute, like this:

BIND(C) :: /COMMON_NAME/

This will make Fortran lay out the variables exactly as C would in a struct. Fortran does not pad COMMONs by default, which is why you don't see an option for it.

Data alignment in common blocks

rfvujm — Fri, 12 Mar 2010 17:06:13 GMT

My C++ appneeds to support the dataalignment by 1 for legacy reasons (porting from Alpha)and a chain structureof the set of software modules that passes this data. The fortran code is mid-way in this chain, it gets HUGE amount of REAL*4, INTEGER*2 and LOGICAL*1 variables and arrays and collects it in COMMONs in various order (so no natural alignment possible). Next, the C++ gets itby extern "C", wraps it into structures and passes it further on. At each stage, the offsets between each vaiable must be kept constant. I'd like to avoid low-level data manipulation and COMMONs redefinition, as the whole chain would be affected.

Any way to 'hack' :) the IVF11 compiler?

Data alignment in common blocks

TimP — Fri, 12 Mar 2010 17:29:32 GMT

As Steve pointed out, the iso_c_binding declaration guarantees consistency (almost certainly with no padding) between Fortran and extern "C"

The mis-alignment would have been problematical on Alpha, and the code was intentionally non-portable, but Steve's suggestion seems best.

Data alignment in common blocks

Steven_L_Intel1 — Fri, 12 Mar 2010 18:20:36 GMT

Even without BIND(C), Intel Fortran does not pad COMMONs unless you ask it to.

Data alignment in common blocks

rfvujm — Mon, 15 Mar 2010 14:12:12 GMT

Thanks guys!

Mario