pip confuses ccache

[pv]

Just a note which may be useful for people using ccache:

Using virtualenv+pip causes stuff to be built in directories whose names have random components.
As a consequence, ccache seems to produce mostly cache misses when compiling.

A possible workaround is here: https://github.com/pv/scipy-bench/blob/master/run.py#L29

# Required to make ccache work properly --- pip builds and installs stuff
# to paths with random components in directory names, which confuses ccache.
#
# To work around this, we need to encourage ccache to be sloppy, and to
# remove the -g flag from compiler flags --- it causes file names to be
# embedded to the object files, which ccache apparently does not like.

import sysconfig
os.environ['CCACHE_SLOPPINESS'] = 'file_macro,time_macros'
os.environ['CCACHE_UNIFY'] = '1'
os.environ['CFLAGS'] = drop_g_flag(sysconfig.get_config_var('CFLAGS'))
os.environ['OPT'] = drop_g_flag(sysconfig.get_config_var('OPT'))
os.environ['LDSHARED'] = drop_g_flag(sysconfig.get_config_var('LDSHARED'))

def drop_g_flag(flags):
    """
    Drop -g from command line flags
    """
    if not flags:
        return flags
    return " ".join(x for x in flags.split() if x not in ('-g', '-g1', '-g2', '-g3'))

Not sure whether all of the above are needed, but it appeared that at least UNIFY and dropping -g were necessary to make ccache work with Scipy/Numpy. Files produced by Cython however still seem to produce cache misses --- didn't manage to figure that out yet.

Making asv define the above seems very dirty, but mentioning the pitfall and workaround in docs might be useful.

[mdboom]

Thanks. An earlier version of asv used pip install --editable which builds things in the same place everytime. But that had other issues. I'll play with this suggestion when I get a chance.

[mdboom]

Fixed by #203.