GFORTRAN Optimization Question

[jelanier]

TL;DR Summary

compiling with O level parameters individually does not equal O level

I am trying to troubleshoot why GFORTRAN versions beyond 5.4 will not compile with optimization on some of my .f source.
You can request options included in each level by:
Gfortran -Q -O1 --help=optimizers > listO1.txt (as an example)
When I enter the enabled flags individually and compile, they do not equal the -O1 composite. I know this by compiling and running a program using each example. Huge difference. What am I missing? Is this a major bug?

Thanks on advance

Jim

 

[jim mcnamara]

Do you know how to compile to turn on profiling? You can run gprof for each compiled version.

Why am I suggesting this. Are you opening files - Linux puts recently opened files into memory buffers so version 1 will run more slowly when compared against version 2 (assuming it ran second from the same parent process - e.g., keyboard command line) - because 2 reads from memory. Other OSes do this, too. This is an example of how your timing test is likely to be confounded. Use profiling. And AFAIK the two should be very much the same.

See:: https://docs.oracle.com/cd/E19957-01/805-4940/6j4m1u7q6/index.html

 

[jelanier]

Do you know how to compile to turn on profiling? You can run gprof for each compiled version.

Why am I suggesting this. Are you opening files - Linux puts recently opened files into memory buffers so version 1 will run more slowly when compared against version 2 (assuming it ran second from the same parent process - e.g., keyboard command line) - because 2 reads from memory. Other OSes do this, too. This is an example of how your timing test is likely to be confounded. Use profiling. And AFAIK the two should be very much the same.

See:: https://docs.oracle.com/cd/E19957-01/805-4940/6j4m1u7q6/index.html

That isn't it. I run this on Windows minGW and MacOS. Any compiled executable regardless order of execution has huge difference in speed and behavior. The composite Options -O -O2 -O3 will not work the same as when flags are individually added to compile batch.

 

[jim mcnamara]

http://www.mingw.org/Reporting_Bugs

Try reporting, you can search the bug list first. Especially since you have a cross-platform problem - Mac & Windows, they should respond quickly.

Helpful hint:
We would have been here a lot sooner here if you had provided those last bits of information. Testing & platform. You will have to give EXACT, complete information when reporting a bug. Or it will be ignored. That will include your compile command(s) and OS versions (and windows 10 current update), and library versions. Especially if you are linking against things like LAPACK. That broke optimizations a while back.

 

[jelanier]

i have since found the culprit and I will report. I took the diff between -Og and -O1 . I then added those flags to -Og to make an equivalent -O1 and it works. So basically calling -O in the batch adds something screwy that kills it. This has been in Gfortran code on all versions starting with v6.x.

Later my friend,

Jim