Support ICC in configure

[zware]

BPO	25827
Nosy	@brettcannon, @doko42, @bitdancer, @skrah, @zware
Files	configure_icc.diff configure_icc.v2.diff

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2015-12-22.03:06:32.810>
created_at = <Date 2015-12-09.07:45:54.623>
labels = ['type-feature', 'build']
title = 'Support ICC in configure'
updated_at = <Date 2015-12-22.03:06:32.809>
user = 'https://github.com/zware'

bugs.python.org fields:

activity = <Date 2015-12-22.03:06:32.809>
actor = 'zach.ware'
assignee = 'none'
closed = True
closed_date = <Date 2015-12-22.03:06:32.810>
closer = 'zach.ware'
components = ['Build']
creation = <Date 2015-12-09.07:45:54.623>
creator = 'zach.ware'
dependencies = []
files = ['41272', '41359']
hgrepos = []
issue_num = 25827
keywords = ['patch']
message_count = 6.0
messages = ['256141', '256183', '256282', '256708', '256807', '256822']
nosy_count = 7.0
nosy_names = ['brett.cannon', 'doko', 'r.david.murray', 'skrah', 'python-dev', 'zach.ware', 'alecsandru.patrascu']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue25827'
versions = ['Python 2.7', 'Python 3.5', 'Python 3.6']

[zware]

Attached is a patch that adds support for ICC to configure, making it easier to build with ICC on Unix and adjusting arguments to better fit ICC.

• Adds '--with-icc' argument to configure, which sets 'CC=icc' and 'CXX=icpc'
• Adds support for ICC PGO
• Prevents '-Wno-unused-result' from being added to CFLAGS if CC=icc (bpo-24709)
• Adds '-fp-model strict' to BASECFLAGS (I'm not 100% sure that's the right place, perhaps CFLAGS_NODIST would be better?). Adding '-fp-model strict' clears up all the failures on the Ubuntu ICC Non-Debug buildbot, and most of the failures I get on OSX with -O3. From a quick run of the 'math' benchmarks (perf.py -b math), adding '-fp-model strict' does not hurt performance (with the flag was always faster or negligibly slower).

Be sure to run autoreconf before testing.

[bitdancer]

This patch looks correct to me, but I don't have much experience with configure. Are the .dyn files icc's profiling output?

I'm adding Alecsandru and Brett as nosy since they worked on the PGO stuff. I'm also adding Stefan since he's shown interest in icc.

[alecsandrupatrascu]

Thank you David for including me in this issue.

On ICC, when executing the instrumented applications, it will generate dynamic information file that has a unique name and .dyn suffix. From my previous experience with GCC and CLANG PGO patch (bpo-24915), also the output of the autoconf will have to be included in the final diff.

Regarding the place for adding the '-fp-model strict' flags, I had the same dillema when working on the LTO patch (bpo-25702), because I needed the flags to be propagated to _all_ compilation and linking phases, and finally decided to add them to the CONFIGURE_CFLAGS, CONFIGURE_CFLAGS_NODIST, CONFIGURE_CPPFLAGS, CONFIGURE_LDFLAGS variables in the Makefile.pre.in file. You can try there also and keep the BASECFLAGS clean .

[zware]

After testing, I think I like '-fp-model strict' better in CFLAGS_NODIST. On Ubuntu, CC=gcc ./python -m test.test_distutils passes when the flag is in CFLAGS_NODIST, but fails when it's in BASECFLAGS. On OSX, it fails both ways, due to deployment target issues; I suspect that's probably solvable by setting some other environment variable correctly (I don't know enough about OSX to say for sure, though). I don't think this patch needs to go as far as the LTO patch does to add the argument to CFLAGS; most third party extensions probably won't necessitate '-fp-model strict', those that do can add it themselves.

On the other hand, 2.7 doesn't have CFLAGS_NODIST (and I don't want to add it), so BASECFLAGS is about the only option there.

Barring objections, I'll commit this new patch in a few days.

[python-dev]

New changeset d7b5c2f99a99 by Zachary Ware in branch '2.7':
Issue bpo-25827: Add support for ICC to configure
https://hg.python.org/cpython/rev/d7b5c2f99a99

New changeset c5e419464585 by Zachary Ware in branch '3.5':
Issue bpo-25827: Add support for ICC to configure
https://hg.python.org/cpython/rev/c5e419464585

New changeset 29ea3827cfaa by Zachary Ware in branch 'default':
Issue bpo-25827: Merge with 3.5
https://hg.python.org/cpython/rev/29ea3827cfaa

[zware]

The buildbots appear to be happy with this, so I'm closing the issue. The intel-ubuntu-icc Non-Debug builders each had their first green build after this changeset. The OSX ICC builder still has some issues, but they seem to be OSX specific (see bpo-25589).