Since the build 154, many tests fail on the AMD64 FreeBSD CURRENT Debug 3.x buildbot slave because of float rounding errors. Failing tests:

• test_cmath
• test_float
• test_json
• test_marshal
• test_math
• test_statistics
• test_strtod

http://buildbot.python.org/all/builders/AMD64%20FreeBSD%20CURRENT%20Non-Debug%203.x/builds/154/steps/test/logs/stdio

Problem: none of build 154 changes are related to floats

• commit f9f87f0
• commit 9476299

It looks more like a libc change of the FreeBSD CURRENT slave.

Example of errors:

======================================================================
FAIL: test_specific_values (test.test_cmath.CMathTests)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "/usr/home/buildbot/python/3.x.koobs-freebsd-current.nondebug/build/Lib/test/test_cmath.py", line 420, in test_specific_values
    msg=error_message)
  File "/usr/home/buildbot/python/3.x.koobs-freebsd-current.nondebug/build/Lib/test/test_cmath.py", line 149, in rAssertAlmostEqual
    '{!r} and {!r} are not sufficiently close'.format(a, b))
AssertionError: acos0036: acos(complex(-1.0009999999999992, 0.0))
Expected: complex(3.141592653589793, -0.04471763360830684)
Received: complex(3.141592653589793, -0.04471763360829195)
Received value insufficiently close to expected value.

======================================================================
FAIL: test_floats (test.test_json.test_float.TestPyFloat)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "/usr/home/buildbot/python/3.x.koobs-freebsd-current/build/Lib/test/test_json/test_float.py", line 8, in test_floats
    self.assertEqual(float(self.dumps(num)), num)
AssertionError: 1.9275814160560202e-50 != 1.9275814160560204e-50

FAIL: test_bigcomp (test.test_strtod.StrtodTests)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "/usr/home/buildbot/python/3.x.koobs-freebsd-current.nondebug/build/Lib/test/test_strtod.py", line 213, in test_bigcomp
    self.check_strtod(s)
  File "/usr/home/buildbot/python/3.x.koobs-freebsd-current.nondebug/build/Lib/test/test_strtod.py", line 104, in check_strtod
    "expected {}, got {}".format(s, expected, got))
AssertionError: '0x1.8265ea9f864bcp+579' != '0x1.8265ea9f864bdp+579'
- 0x1.8265ea9f864bcp+579
?                 ^
+ 0x1.8265ea9f864bdp+579
?                 ^
 : Incorrectly rounded str->float conversion for 29865e170: expected 0x1.8265ea9f864bcp+579, got 0x1.8265ea9f864bdp+579

test_strtod fails on FreeBSD using clang 4.0, but pass using clang 3.8.

Links:

• https://gcc.gnu.org/wiki/FloatingPointMath
• https://gcc.gnu.org/wiki/Math_Optimization_Flags
• ICC fp-model: https://software.intel.com/en-us/node/682946

GCC options:

• -ffp-contract=off / -ffp-contract=fast
• -fp-model (not supported by clang)
• -mfpmath=sse: no effect on Clang, it's already the default and -mfpmath=387 is not supported on x86_64 ISA
• -ffast-math
• -fexcess-precision=standard: not supported by clang

I created a test case: attached dtoa2.c.

Ok with GCC 6.3.1 (on Fedora 25) and clang 3.8 (on FreeBSD 11) with -O3:

haypo@selma$ gcc -O3 dtoa2.c -o x && ./x
text: 29865e170 -> float: 2.9865e+174
0x1.8265ea9f864bcp+579
{bc 64 f8 a9 5e 26 28 64}

[haypo@freebsd ~/prog/python/master]$ clang -O3 dtoa2.c -o x && ./x
text: 29865e170 -> float: 2.9865e+174
0x1.8265ea9f864bcp+579
{bc 64 f8 a9 5e 26 28 64}

Still ok with clang 3.8 with -O1:

[haypo@freebsd ~/prog/python/master]$ clang40 -O1 dtoa2.c -o x && ./x
text: 29865e170 -> float: 2.9865e+174
0x1.8265ea9f864bcp+579
{bc 64 f8 a9 5e 26 28 64}

Error with clang 4.0 using -O2 (on FreeBSD 11):

[haypo@freebsd ~/prog/python/master]$ clang40 -O2 dtoa2.c -o x && ./x
text: 29865e170 -> float: 2.9865e+174
0x1.8265ea9f864bdp+579 <~~~~~ HERE, bd instead of bc
{bd 64 f8 a9 5e 26 28 64} <~~~~~ HERE, bd instead of bc

Still ok with clang 3.8 with -O1:

Sorry, you should read clang 4.0 here.

Much simpler test:

[haypo@freebsd ~/prog/python/master]$ clang40 -O1 dtoa5.c -o x && ./x
k=1
aadj=3.364
test ok
[haypo@freebsd ~/prog/python/master]$ clang40 -O2 dtoa5.c -o x && ./x
k=1
aadj=1.682
test FAILED: BUG!

Why is -ffast-math being used? That seems like it's asking for trouble.

Hmm. Just looked at dto5.c; nice work reducing it to something so simple.

This is looking a bit like either a strict-aliasing related problem, or a compiler bug, or both. My understanding of the "union trick" that dtoa.c uses is that it's well-defined under C99+TC3 (though that wasn't clear under the original C99), so it shouldn't be legitimate for an optimisation to cause different results to be produced.

So this is smelling like a compiler bug to me.

Why is -ffast-math being used? That seems like it's asking for trouble.

Sorry, my comment was unclear: I only tried to list all compiler options which have an impact on floating point numbers.

So this is smelling like a compiler bug to me.

Right, I will report it to LLVM (clang).

This is most likely the same issue we found in https://bugs.freebsd.org/216770, which was reported upstream to LLVM here: https://bugs.llvm.org//show_bug.cgi?id=31928, and resulted in a very long and not really productive discussion about whether type punning in this way is officially allowed, or a GNU extension. I will gladly leave it to language lawyers. :)

In our case, easy fix was to use -fno-strict-aliasing, as we did here:
https://svnweb.freebsd.org/changeset/base/313706

Dimitry: thank you! Indeed, that looks like the very same issue. It's a shame to have to use -fno-strict-aliasing here, since that might prevent legitimate optimisations elsewhere in the code. But if that's what we have to do to maintain correctness with clang, then I guess we have little choice.

I seem the same issues on OS X (10.9.5) when compiling with clang 4.0 from MacPorts. So this doesn't seem to be FreeBSD-specific.

Changing title and raising priority: failure for repr (and marshal, JSON, etc.) to roundtrip seems like a critical breakage to me.

mdickinson$ CC=/opt/local/bin/clang-mp-4.0 ./configure && make && ./python.exe -m test test_strtod
[... log from successful build omitted ...]
Run tests sequentially
0:00:00 [1/1] test_strtod
test test_strtod failed -- multiple errors occurred; run in verbose mode for details
test_strtod failed

1 test failed:
test_strtod

Total duration: 65 ms
Tests result: FAILURE

New changeset 28205b2 by Victor Stinner in branch 'master':
bpo-30104: Use -fno-strict-aliasing on clang (bpo-1221)
28205b2

Mark: "I seem the same issues on OS X (10.9.5) when compiling with clang 4.0 from MacPorts. So this doesn't seem to be FreeBSD-specific."

I confirm, I reproduced the bug on Linux using Clang 4.0.

Mark: "But if that's what we have to do to maintain correctness with clang, then I guess we have little choice."

I pushed my change adding -fno-strict-aliasing compiler option to clang. I created the issue bpo-30124 to fix the aliasing issue in dtoa.c.

I didn't check yet if we have to backport the -fno-strict-aliasing change, maybe we can wait until the aliasing issue is fixed and backport this future fix?

FWIW, I can confirm that adding -fno-strict-aliasing fixes all failing float-related tests on OS X.

Oh... I only tested my configure patch on Linux using ./configure CC=clang, but the change has no effect on FreeBSD where $CC is /usr/bin/cc and so my code doesn't detect clang correctly.

I proposed #1233 which runs "cc --version" to check if it contains clang. I tested this change on FreeBSD: it works ;-) (it adds -fno-strict-aliasing)

New changeset 35f3d24 by Victor Stinner in branch 'master':
bpo-30104: configure now detects when cc is clang (bpo-1233)
35f3d24

Ok, the FreeBSD CURRENT buildbot is repaired: "AMD64 FreeBSD CURRENT
Debug 3.x is complete: Success".

Thank you for investigating Victor.

This appears to require backporting to 3.6 and 3.5 which are also failing:

FAIL: test_repr (test.test_float.ReprTestCase)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "/usr/home/buildbot/python/3.5.koobs-freebsd-current.nondebug/build/Lib/test/test_float.py", line 674, in test_repr
    self.assertEqual(v, eval(repr(v)))
AssertionError: 5.311000000000001e+245 != 5.311000000000002e+245

Ran 42 tests in 0.242s

FAILED (failures=1, skipped=1)
7 tests failed again:
test_cmath test_float test_json test_marshal test_math
test_statistics test_strtod

This appears to require backporting to 3.6 and 3.5 which are also failing:

I'm aware of that and I plan to fix these versions as well (2.7 also, no?), but I would like to first decide if we "fix" dtoa.c aliasing, or if we restrict -fno-strict-aliasing flag on dtoa.c: see discussion on the issue bpo-30124. It seems like restricting the flag to dtoa.c is the favorite option.

2.7 branch on koobs-freebsd-current worker doesn't appear to be failing currently, but there may be differences in (or a lack of equivalent) coverage compared to 3.*.

I also note that -fno-strict-aliasing is being included in 2.7's compile arguments:

cc -pthread -c -fno-strict-aliasing -OPT:Olimit=0 -g -O2 -g -O0 -Wall -Wstrict-prototypes

I also note that -fno-strict-aliasing is being included in 2.7's compile arguments

Ah right, PyObject structures of Python 2 doesn't respect strict aliasing...

I would like to fix FreeBSD CURRENT buildbots of Python 2.7, 3.5 and 3.6, so here my attempt to restrict the -fno-strict-aliasing option to the dtoa.c file: #1340

I chose to add the flag for any C compiler. If you think that a C compiler may not support that option, we can start by only adding the option on clang, and maybe add it on other C compilers if someone complains.

New changeset 826f83f by Victor Stinner in branch 'master':
bpo-30104: Only use -fno-strict-aliasing on dtoa.c (bpo-1340)
826f83f

New changeset 809101f by Victor Stinner in branch '3.6':
bpo-30104: Use -fno-strict-aliasing on clang (bpo-1376)
809101f

New changeset 03a7ab7 by Victor Stinner in branch '3.5':
bpo-30104: Use -fno-strict-aliasing on clang (bpo-1376) (bpo-1377)
03a7ab7

dtoa.c is now compiled with -fno-string-aliasing (for any clang version, not only clang 4.0) on Python 3.5, 3.6 and 3.7.

It was decided to not touch dtoa.c to not diverge from upstream.

Thanks Dimitry Andric and Mark Dickinson for your reviews and support in this cryptic issue!

Note 1: We consider that Clang 4.0 is wrong, whereas GCC respects the C99 standard for aliasing on unions. But I don't think that it matters much who is wrong or not :-)

Note 2: Python 2.7 already used -fno-strict-aliasing on all .c files because of its design of PyObject structures, fixed in Python 3.