Compiling with fast-math flag options (-ffast-math, -Ofast, etc.) may result in execution errors in some or many test programs.
Details
Many test programs in this test suite output computation results to the standard output and they are compared with the reference output files (*.reference_output files in the same directory) by lit and fpcmp. Compiling with fast-math flag options may change computation result of floating-point number. While fpcmp may be able to absorb the difference, it may not be sufficient.
Many other test programs compares the computation result and the expected value in the test programs themselves (without using fpcmp).
These may result in execution errors in some or many test programs.
Workaround
Don't compile with fast-math options.
Future direction
We want to support fast-math flag options but we haven't decided how to handle this problem yet.
This is a known issue.
Problem
Compiling with fast-math flag options (
-ffast-math,-Ofast, etc.) may result in execution errors in some or many test programs.Details
Many test programs in this test suite output computation results to the standard output and they are compared with the reference output files (
*.reference_outputfiles in the same directory) bylitandfpcmp. Compiling with fast-math flag options may change computation result of floating-point number. Whilefpcmpmay be able to absorb the difference, it may not be sufficient.Many other test programs compares the computation result and the expected value in the test programs themselves (without using
fpcmp).These may result in execution errors in some or many test programs.
Workaround
Don't compile with fast-math options.
Future direction
We want to support fast-math flag options but we haven't decided how to handle this problem yet.