test_lwgeom_split fails on various architectures

[sebastic]

The Debian package builds for postgis 2.2.1 failed on various architectures due to test failure of test_lwgeom_split introduced for #3401 in r14549:

Suite: split
  Test: test_lwline_split_by_point_to ...passed
  Test: test_lwgeom_split ...[cu_split.c:267]
 Expected: -20
 Obtained: -20
[cu_split.c:272]
 Expected: -20
 Obtained: -20
FAILED
    1. cu_split.c:267  - CU_ASSERT_EQUAL(pt.x,(double)-20)
    2. cu_split.c:272  - CU_ASSERT_EQUAL(pt.x,(double)-20)

Build logs: ​arm64, ​ppc64el, ​s390x

Not all build results are available yet, the others will become available via the ​buildd status page

[sebastic]

The ​powerpc build also failed due to the same test failure.

amd64, i386 & mipsel are currently the only architectures on which the test succeeds.

[robe]

I'm a bit puzzled why expected and obtained values look the same.

[sebastic]

I must admit that my knowledge of these ports is not sufficient to explain this test failure. I do have access to porterboxes for these architectures, so I'm happy to help troubleshoot this issue on the architectures in question.

[gdt]

Are we sure the tests are looking for a close value as a double, vs exact bits, and that we aren't comparing the value obtained with extended precision doubles (which are more than IEEE754, common on x86) vs standard precision doubles (typical on everything that isn't x86 and isn't vax).

[robe]

Uses ASSERT_DOUBLE_EQUAL

   ASSERT_DOUBLE_EQUAL(pt.x, -20);

defined here:

​http://postgis.net/docs/doxygen/2.2/d5/d35/liblwgeom_2cunit_2cu__tester_8h_a27f9113cc18b9d98d86709a2c76f9fab.html#a27f9113cc18b9d98d86709a2c76f9fab

[strk]

Sebastic: could you give me access to one of those machines to take a look ? It sounds like closest_point_on_segment is not returning the input point even when it is already _on_ the input segment.

[strk]

It could also be tested that the following returns TRUE:

WITH inp AS ( SELECT 
 'LINESTRING(-180 0,0 0)'::geometry l, 
 'POINT(-20 0)'::geometry p 
) SELECT ST_Equals(ST_ClosestPoint(l,p), p) FROM inp;

[sebastic]

I cannot provide access to the Debian porterboxes, if you really need that access we can request a ​guest account.

Since that's quite an involved process, it's probably better for me to just run the tests. I'll do that after work tonight.

[pramsey]

Added tolerance to the double tests at r14589 in 2.2 and r14590 in trunk

[sebastic]

With the changes from r14589 added to the Debian package, regress/st_modedgesplit now fails:

--- regress/st_modedgesplit_expected    2016-01-11 19:02:03.825936435 +0000
+++ /tmp/pgis_reg/test_21_out   2016-01-11 19:06:18.937928368 +0000
@@ -47,4 +47,6 @@
 t3401.N2|2
 t3401.L1|1
 t3401.split|3
+t3401.valid_after|edge start node geometry mis-match|2|3
+t3401.valid_after|edge end node geometry mis-match|1|3
 t3401.end|Topology 'bug3401' dropped

[pramsey]

You're certain? The changes in r14589 are completely contained in the library test module, the online tests you see failing are completely unrelated.

[sebastic]

Are you sure it's completely unrelated?

The test is for the same issue (#3401), I think the problem just got moved.

As I don't have time to endlessly debate this, I'll close this issue again because test_lwgeom_split doesn't fail any more. I'll keep the tests disabled for the problematic architectures.

[strk]

The tests _are_ related. Robustness issues in lwgeom_split were found to be the cause for corrupting topology, so rather than hiding the failure in the C library, we should keep it exposed and try to fix it instead.

The correct solution would be to keep the split point as a vertex, rather than re-projecting it, when found to be already ON the segment.

[pramsey]

Related insofar that you have to pass the C tests in order to even run the online tests. I see.

[strk]

I've a temptative fix locally, about to push it.

[strk]

(In [14594]) Avoid any drift of cutter point on lines split

Should fix splitting operations on at least arm64, ppc64el and s390x. See #3422 and #3401

Also turn ASSERT_DOUBLE_EQUAL back to a tolerance-free check (better use a different name for tolerance-aware check, so caller can decide)

[strk]

Sebastic, could you please try testing the trunk branch, as of r14594 ?

[strk]

(In [14595]) Avoid any drift of cutter point on lines split

Should fix splitting operations on at least arm64, ppc64el and s390x. See #3422 and #3401 (for 2.2 branch)

Also turn ASSERT_DOUBLE_EQUAL back to a tolerance-free check (better use a different name for tolerance-aware check, so caller can decide)

[strk]

Actually, also backported to 2.2 branch with r14595

[sebastic]

The root disk in my workstation is dying with bad sectors, so I have to fix that first to be fully functional again. I can probably test the changes from r14595 this weekend.

[sebastic]

With the changes from r14595 the tests succeed on arm64, powerpc, ppc64el, s390x.

Looks like this is fixed properly now, thanks!

[strk]

thanks for testing