aermod.f90 Polyhedron 2005 benchmark miscompiled with -msse4 -ffast-math -funroll-loops -O3 -fplugin-arg-dragonegg-enable-gcc-optzns

[llvmbot]

Bugzilla Link	10893
Resolution	FIXED
Resolved on	Sep 24, 2011 06:39
Version	trunk
OS	MacOS X
Reporter	LLVM Bugzilla Contributor

Extended Description

At r139290, when the aermod.f90 Polyhedron 2005 benchmark is compiled with...

/sw/lib/gcc4.6/bin/gfortran -fplugin=/sw/lib/gcc4.6/lib/dragonegg.so -msse4 -ffast-math -funroll-loops -O3 -fplugin-arg-dragonegg-enable-gcc-optzns aermod.f90 -o aermod

the benchmark shows a runtime speed regression...

Benchmark Compile Executable Ave Run Number Estim
Name (secs) (bytes) (secs) Repeats Err %

---

  aermod     24.28     1095232     29.27       1100.0000

compared to without -fplugin-arg-dragonegg-enable-gcc-optzns

  aermod     20.14     1039824     16.54       1100.0000

and produces failures during the benchmark run...

Value= 2195.3913600 Target= 2191.1145000 Tolerance=0.10000000000E-02
FAIL <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
Value= 34310.417970 Target= 34310.421880 Tolerance=0.30000000000E-01
Value= 4260.9482400 Target= 4260.9716800 Tolerance=0.50000000000E-01
Value= 37060.945310 Target= 37094.375000 Tolerance=0.20000000000E-01
FAIL <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
Value= 7937.5170900 Target= 7924.8842800 Tolerance=0.30000000000E-01
FAIL <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
Value= 37060.945310 Target= 37094.375000 Tolerance= 1.0000000000
FAIL <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<

aermod FAILED 4 fails and 2 passes

[llvmbot]

Use of -fno-tree-vectorize...

8 Sep 2011 12:16:32 de-gfortran_lin_O3_deopts_fastmath - Compile aermod
command=de-gfortran46 -msse4 -ffast-math -funroll-loops -O3 -fplugin-arg-dragonegg-enable-gcc-optzns -fno-tree-vectorize aermod.f90 -o aermod

8 Sep 2011 12:16:55 de-gfortran_lin_O3_deopts_fastmath - Execute aermod
aermod Run # 1 27.91100 27.91100 - Error=100.0000%

doesn't eliminate the run time speed regression in the case of aermod.f90. However -fno-tree-vectorize does eliminate the run-time errors in aermod when using -fplugin-arg-dragonegg-enable-gcc-optzns...

Value= 2191.1142600 Target= 2191.1145000 Tolerance=0.10000000000E-02
Value= 34310.417970 Target= 34310.421880 Tolerance=0.30000000000E-01
Value= 4260.9594700 Target= 4260.9716800 Tolerance=0.50000000000E-01
Value= 37094.367190 Target= 37094.375000 Tolerance=0.20000000000E-01
Value= 7924.8808600 Target= 7924.8842800 Tolerance=0.30000000000E-01
Value= 37094.367190 Target= 37094.375000 Tolerance= 1.0000000000

aermod PASSED 6 tests - no failures

[llvmbot]

The run-time failures in aermod are triggered with...

de-gfortran46 -msse4 -ffast-math -O3 -fplugin-arg-dragonegg-enable-gcc-optzns aermod.f90 -o aermod

but not with...

de-gfortran46 -msse4 -ffast-math -O2 -fplugin-arg-dragonegg-enable-gcc-optzns -ftree-vectorize aermod.f90 -o aermod

or

de-gfortran46 -msse4 -O3 -fplugin-arg-dragonegg-enable-gcc-optzns aermod.f90 -o aermod

so "msse4 -ffast-math -O3 -fplugin-arg-dragonegg-enable-gcc-optzns" appears to be the minimal flags to trigger the aermod run-time failures.

[llvmbot]

This is on x86_64-apple-darwin11.

[llvmbot]

The failures in aermod are gone at r139660 but the performance is still lower than with -fno-tree-vectorize...

13 Sep 2011 19:31:50 de-gfortran_lin_O3_deopts_fastmath - Compile aermod
command=de-gfortran46 -msse4 -ffast-math -funroll-loops -O3 -fplugin-arg-dragonegg-enable-gcc-optzns aermod.f90 -o aermod

13 Sep 2011 19:32:14 de-gfortran_lin_O3_deopts_fastmath - Execute aermod
aermod Run # 1 28.52800 28.52800 - Error=100.0000%

[llvmbot]

The run time of aermod is reduced from 28.32s to 14.87s at r140387 when using -msse4 -ffast-math -funroll-loops -O3 -fplugin-arg-dragonegg-enable-gcc-optzns with FSF gcc 4.6.2svn. Nice.

[llvmbot]

Closing because this is no longer miscompiled and the performance seems good.