search

manufactured-solutions / MASA

Method of Manufactured Solutions Repository
Other
47 stars 27 forks source link

Compiler flags identifying floating divide by zero runtime error #55

Open davidcraigpenner opened 2 years ago

davidcraigpenner commented 2 years ago

The version number of the MASA library (versioning information can be obtained by running the masa_version binary located in the bin/ directory of a local MASA installation) the hardware and operating system

-------------------------------------------------------------------------------
MASA Library: Version = 0.51.0 (5100)

Development Build

Build Date   = 2019-10-09 14:05
Build Host   = nia-login05.scinet.local
Build User   = dcp
Build Arch   = x86_64-pc-linux-gnu
Build Rev    = 7351ff92

C++ Config   = mpicxx -O0
F90 Config   = mpifc-O0

Optional Features:
   Python support enabled = no
-------------------------------------------------------------------------------

A description of the bug behavior I am using MASA via the Fortran90 interface and compiling my source code with the following Intel Fortran compiler:

mpif90 for the Intel(R) MPI Library 2018 Update 3 for Linux*
Copyright(C) 2003-2018, Intel Corporation.  All rights reserved.
__INTEL_PRE_FFLAGS options:  -gcc-name=core-gcc -gxx-name=core-g++
ifort version 18.0.3

When I use the following flags (among others) to compile my source code, specifically

-check all,noarg_temp_created -init=snan -init=zero -init=arrays

I obtain the following runtime error when running my executable:

forrtl: error (73): floating divide by zero
Image              PC                Routine            Line        Source
jetstream_x86_64   000000000043E8CE  Unknown               Unknown  Unknown
libpthread-2.17.s  00002B34AD110630  Unknown               Unknown  Unknown
jetstream_x86_64   00000000004FEEA1  Unknown               Unknown  Unknown
libfmasa-0.51.so   00002B34AB54832D  _ZN4MASA27fans_sa     Unknown  Unknown
libfmasa-0.51.so   00002B34AB547F7B  _ZN4MASA27fans_sa     Unknown  Unknown
libfmasa-0.51.so   00002B34AB546D84  _ZN4MASA27fans_sa     Unknown  Unknown
libfmasa-0.51.so   00002B34AB482130  Unknown               Unknown  Unknown
libfmasa-0.51.so   00002B34AB480A9A  Unknown               Unknown  Unknown
libfmasa-0.51.so   00002B34AB48C254  _ZN4MASA9masa_ini     Unknown  Unknown
libfmasa-0.51.so   00002B34AB5147C0  masa_init             Unknown  Unknown
libfmasa-0.51.so   00002B34AB4804DB  masa_mp_masa_init     Unknown  Unknown
libdiablo.so       00002B34A05933CE  verify_mod_mp_gen       10911  Verify_Mod.f90

For reference, Line 10911 in my Fortran90 source code is a call to masa_init:

!-- initialize the problem
call masa_init('test',trim(adjustl(desired_mms_function)))

where desired_mms_function = 'euler_2d'. Note that when I compile without the aforementioned flags (i.e., without -check all,noarg_temp_created -init=snan -init=zero -init=arrays), MASA runs correctly. For reference, the functions of the flags of interest are as follows (more info here):

My guess is that there are some unintialized variables in MASA, potentially in fans_sa.cpp (based on the above error message that I am getting), that are being caught by these flags, however I have been unable to identify specifically what part of the code is leading to the runtime error. Does anyone know how to address this issue?

roystgnr commented 2 years ago

Do you get any runtime errors from make check?

I tried running all our test applications with valgrind, but no reported errors there. (Which doesn't mean no uninitialized values, IIRC, because valgrind only catches those when they affect a branch condition)

I also tried a build with

../configure --enable-fortran-interfaces CXXFLAGS='-Wuninitialized -Wmaybe-uninitialized -Werror -O0 -fno-unsafe-math-optimizations -Dgnu_compiler -Wno-literal-suffix' ac_test_CXXFLAGS=set CFLAGS='-Wuninitialized -Wmaybe-uninitialized -Werror -O0 -fno-unsafe-math-optimizations -Dgnu_compiler -Wno-builtin-declaration-mismatch -Wno-implicit-function-declaration' ac_test_CFLAGS=set FCFLAGS='-Wuninitialized -Wmaybe-uninitialized -Werror -O0'

but I don't see any uninitialized variable warnings from the compiler and I still pass make check.

davidcraigpenner commented 2 years ago

I do not get any runtime errors with make check. Thanks for running those tests. I tried a build with the additional flags you mention, i.e.,

./configure --prefix=$PREFIX --enable-fortran-interfaces CXXFLAGS='-Wuninitialized -Wmaybe-uninitialized -Werror -O0 -fno-unsafe-math-optimizations -Dgnu_compiler -Wno-literal-suffix' ac_test_CXXFLAGS=set CFLAGS='-Wuninitialized -Wmaybe-uninitialized -Werror -O0 -fno-unsafe-math-optimizations -Dgnu_compiler -Wno-builtin-declaration-mismatch -Wno-implicit-function-declaration' ac_test_CFLAGS=set FCFLAGS='-Wuninitialized -Wmaybe-uninitialized -Werror -O0'

however when I look at the summary of the build, it seems like some of the specified flags are not being used:

----------------------------------- SUMMARY -----------------------------------

Package version................. : masa-0.51.0

C++ compiler.................... : mpicxx
C++ compiler flags.............. : -O0
C compiler...................... : mpicc
C compiler flags................ : -O0 -fp-model precise -Dintel_compiler
Install dir..................... : /home/z/zingg/dcp/jetstream/masa
Build user...................... : dcp
Build host...................... : nia-login05.scinet.local
Configure date.................. : 2021-11-11 09:24
Build architecture.............. : x86_64-pc-linux-gnu
Source control revision......... : 48d2494bf

Optional Features:
   Enable gcov code coverage.... : no
   MetaPhysicL.................. : no
   Enable absolute error tests.. : no
   Enable all warnings.......... : no
   Enable python interfaces..... : no
   Enable fortran interfaces.... : yes
Fortran compiler................ : mpifc
Fortran compiler flags.......... : -O0

-------------------------------------------------------------------------------

Perhaps some of the flags are skipped because I am using MPI compilers? Or am I interpreting the summary incorrectly?

Also, are there some flags I could build MASA with that would provide more detailed error information? For example, the current error message I get when I run my executable does not indicate any Line or Source information for MASA, however having this additional information would be helpful for identifying the problem: 

forrtl: error (73): floating divide by zero
Image              PC                Routine            Line        Source
jetstream_x86_64   000000000043E8CE  Unknown               Unknown  Unknown
libpthread-2.17.s  00002B34AD110630  Unknown               Unknown  Unknown
jetstream_x86_64   00000000004FEEA1  Unknown               Unknown  Unknown
libfmasa-0.51.so   00002B34AB54832D  _ZN4MASA27fans_sa     Unknown  Unknown
libfmasa-0.51.so   00002B34AB547F7B  _ZN4MASA27fans_sa     Unknown  Unknown
libfmasa-0.51.so   00002B34AB546D84  _ZN4MASA27fans_sa     Unknown  Unknown
libfmasa-0.51.so   00002B34AB482130  Unknown               Unknown  Unknown
libfmasa-0.51.so   00002B34AB480A9A  Unknown               Unknown  Unknown
libfmasa-0.51.so   00002B34AB48C254  _ZN4MASA9masa_ini     Unknown  Unknown
libfmasa-0.51.so   00002B34AB5147C0  masa_init             Unknown  Unknown
libfmasa-0.51.so   00002B34AB4804DB  masa_mp_masa_init     Unknown  Unknown
libdiablo.so       00002B34A05933CE  verify_mod_mp_gen       10911  Verify_Mod.f90