6.3.0 Tests fail on Alpine Linux x86 (musl libc)

[russkel]

Example of problem

I am in the process of fixing the Alpine Linux proj package and have found that is solely fails on the x86 platform.

The following tests FAILED:
	  2 - Builtins (Failed)
	  3 - Builtins2 (Failed)
	  6 - Ellipsoid (Failed)
	 31 - testvarious (Failed)
Errors while running CTest

test 2
      Start  2: Builtins

2: Test command: /builds/russkel/aports/community/proj/src/proj-6.3.0/bin/gie "/builds/russkel/aports/community/proj/src/proj-6.3.0/test/gie/builtins.gie"
2: Environment variables: 
2:  PROJ_LIB=/builds/russkel/aports/community/proj/src/proj-6.3.0/data
2: Test timeout computed to be: 10000000
2: -------------------------------------------------------------------------------
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -21: conic lat_1 = -lat_2
2: proj_create: Error -6: effective eccentricity < 0 or >= 1.
2: proj_create: Error -21: conic lat_1 = -lat_2
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -21: conic lat_1 = -lat_2
2: proj_create: Error -30: h <= 0 or h > 1e10 * a
2: proj_create: Error -30: h <= 0 or h > 1e10 * a
2: proj_create: Error -58: argument not numerical or out of range
2: proj_create: Error -55: lat_0 = 0
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -27: W <= 0 or M <= 0
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -6: effective eccentricity < 0 or >= 1.
2: proj_create: Error -6: effective eccentricity < 0 or >= 1.
2: proj_create: Error -32: lat_1=lat_2 or lat_1=0 or lat_2=90
2: proj_create: Error -32: lat_1=lat_2 or lat_1=0 or lat_2=90
2: proj_create: Error -32: lat_1=lat_2 or lat_1=0 or lat_2=90
2: proj_create: Error -32: lat_1=lat_2 or lat_1=0 or lat_2=90
2: proj_create: Error -32: lat_1=lat_2 or lat_1=0 or lat_2=90
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -30: h <= 0 or h > 1e10 * a
2: proj_create: Error -30: h <= 0 or h > 1e10 * a
2: proj_create: Error -37: failed to find projection to be rotated
2: proj_create: Error -33: lat_0 = 0 or 90 or alpha = 90
2: proj_create: Error -6: effective eccentricity < 0 or >= 1.
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
2: proj_create: Error 12: Out of memory
2: proj_create: Error -32: lat_1=lat_2 or lat_1=0 or lat_2=90
2: proj_create: Error -33: lat_0 = 0 or 90 or alpha = 90
2: Reading file '/builds/russkel/aports/community/proj/src/proj-6.3.0/test/gie/builtins.gie'
2: errno=-14, expected=-6
2:      -----
2:      FAILURE in builtins.gie(2489):
2:      expected: 0 4886594.2207
2:      got:      -nan   -nan
2:      deviation:  nan mm,  expected:  0.100000 mm
2: errno=0, expected=-6
2:      -----
2:      FAILURE in builtins.gie(4805):
2:      expected: 0.000000000000 8654720.000000000000
2:      got:      0.000000000000   8654720.465063288808
2:      deviation:  465.063289 mm,  expected:  0.100000 mm
2:      -----
2:      FAILURE in builtins.gie(4811):
2:      expected: 0.000000000000 -8654720.000000000000
2:      got:      0.000000000000   -8654720.465063288808
2:      deviation:  465.063289 mm,  expected:  0.100000 mm
2:      -----
2:      FAILURE in builtins.gie(5686):
2:      expected: 223395.249543407 111704.596633675
2:      got:      223395.249543474842   111704.596082000367
2:      deviation:  0.551675 mm,  expected:  0.250000 mm
2:      -----
2:      FAILURE in builtins.gie(5689):
2:      expected: 223395.249543407 -111704.596633675
2:      got:      223395.249543474842   -111704.596082000367
2:      deviation:  0.551675 mm,  expected:  0.250000 mm
2:      -----
2:      FAILURE in builtins.gie(5691):
2:      expected: -223395.249543407 111704.596633675
2:      got:      -223395.249543474842   111704.596082000367
2:      deviation:  0.551675 mm,  expected:  0.250000 mm
2:      -----
2:      FAILURE in builtins.gie(5693):
2:      expected: -223395.249543407 -111704.596633675
2:      got:      -223395.249543474842   -111704.596082000367
2:      deviation:  0.551675 mm,  expected:  0.250000 mm
2: errno=0, expected=-20
2: -------------------------------------------------------------------------------
2: total: 1728 tests succeeded,  0 tests skipped, 10 tests FAILED!
2: -------------------------------------------------------------------------------
 2/41 Test  #2: Builtins .........................***Failed    0.05 sec
test 3
      Start  3: Builtins2

3: Test command: /builds/russkel/aports/community/proj/src/proj-6.3.0/bin/gie "/builds/russkel/aports/community/proj/src/proj-6.3.0/test/gie/more_builtins.gie"
3: Environment variables: 
3:  PROJ_LIB=/builds/russkel/aports/community/proj/src/proj-6.3.0/data
3: Test timeout computed to be: 10000000
3: -------------------------------------------------------------------------------
3: proj_create: Error -1: no arguments in initialization list
3: proj_create: Error -54: missing required arguments
3: proj_create: Error -50: malformed pipeline
3: proj_create: Error -50: malformed pipeline
3: proj_create: nested pipeline not supported
3: proj_create: +step found outside pipeline
3: proj_create: Error -50: malformed pipeline
3: proj_create: Error -1: no arguments in initialization list
3: proj_create: Error -38: failed to load datum shift file
3: proj_create: Error -38: failed to load datum shift file
3: proj_create: Error -1: no arguments in initialization list
3: proj_create: Error -54: missing required arguments
3: proj_create: Error -58: argument not numerical or out of range
3: proj_create: Error -58: argument not numerical or out of range
3: proj_create: Error -58: argument not numerical or out of range
3: proj_create: Error -58: argument not numerical or out of range
3: proj_create: Error -54: missing required arguments
3: proj_create: Error -51: unit conversion factor must be > 0
3: proj_create: Error -51: unit conversion factor must be > 0
3: proj_create: Error -22: lat_0, lat_1 or lat_2 >= 90
3: Reading file '/builds/russkel/aports/community/proj/src/proj-6.3.0/test/gie/more_builtins.gie'
3:      -----
3:      FAILURE in more_builtins.gie(838):
3:      expected: 0 90 0
3:      got:      0.000000000000   0.000000000000   -6378137.000000000
3:      deviation:  999999999.999000 mm,  expected:  1.000000 mm
3:      -----
3:      FAILURE in more_builtins.gie(841):
3:      expected: 0 -90 0
3:      got:      0.000000000000   0.000000000000   -6378137.000000000
3:      deviation:  999999999.999000 mm,  expected:  1.000000 mm
3: -------------------------------------------------------------------------------
3: total: 157 tests succeeded,  0 tests skipped,  2 tests FAILED!
3: -------------------------------------------------------------------------------
 3/41 Test  #3: Builtins2 ........................***Failed    0.04 sec

test 6
      Start  6: Ellipsoid

6: Test command: /builds/russkel/aports/community/proj/src/proj-6.3.0/bin/gie "/builds/russkel/aports/community/proj/src/proj-6.3.0/test/gie/ellipsoid.gie"
6: Environment variables: 
6:  PROJ_LIB=/builds/russkel/aports/community/proj/src/proj-6.3.0/data
6: Test timeout computed to be: 10000000
6: -------------------------------------------------------------------------------
6: proj_create: Error -9: unknown elliptical parameter name
6: proj_create: Error -13: major axis or radius = 0 or not given
6: proj_create: Error -13: major axis or radius = 0 or not given
6: proj_create: Error -13: major axis or radius = 0 or not given
6: proj_create: Error -13: major axis or radius = 0 or not given
6: proj_create: unrecognized format / unknown name
6: proj_create: unrecognized format / unknown name
6: proj_create: Error -11: |radius reference latitude| > 90
6: proj_create: Error -10: reciprocal flattening (1/f) = 0
6: proj_create: Error -6: effective eccentricity < 0 or >= 1.
6: proj_create: Error -6: effective eccentricity < 0 or >= 1.
6: proj_create: Error -6: effective eccentricity < 0 or >= 1.
6: proj_create: Error -6: effective eccentricity < 0 or >= 1.
6: proj_create: Error -6: effective eccentricity < 0 or >= 1.
6: proj_create: Error -6: effective eccentricity < 0 or >= 1.
6: Reading file '/builds/russkel/aports/community/proj/src/proj-6.3.0/test/gie/ellipsoid.gie'
6:      -----
6:      FAILURE in ellipsoid.gie(48):
6:      expected: 1335833.8895192828 7326837.7148738774
6:      got:      1335833.889519282850   -7326837.714873872697
6:      deviation:  999999999.999000 mm,  expected:  0.000010 mm
6:      -----
6:      FAILURE in ellipsoid.gie(120):
6:      got errno ref_rad_larger_than_90 (-11): |radius reference latitude| > 90
6:      expected invalid_eccentricity (-6):  effective eccentricity < 0 or >= 1.
6: -------------------------------------------------------------------------------
6: total: 35 tests succeeded,  0 tests skipped,  2 tests FAILED!
6: -------------------------------------------------------------------------------
 6/41 Test  #6: Ellipsoid ........................***Failed    0.10 sec


31: Test command: /bin/bash "/builds/russkel/aports/community/proj/src/proj-6.3.0/test/cli/testvarious" "/builds/russkel/aports/community/proj/src/proj-6.3.0/bin/cs2cs"
31: Environment variables: 
31:  PROJ_LIB=/builds/russkel/aports/community/proj/src/proj-6.3.0/data
31: Test timeout computed to be: 10000000
31: ============================================
31: Running /builds/russkel/aports/community/proj/src/proj-6.3.0/test/cli/testvarious using /builds/russkel/aports/community/proj/src/proj-6.3.0/bin/cs2cs:
31: ============================================
31: doing tests into file tv_out, please wait
31: diff tv_out with tv_out.dist
31: --- tv_out
31: +++ /builds/russkel/aports/community/proj/src/proj-6.3.0/test/cli/tv_out.dist
31: @@ -41,7 +41,7 @@
31:  6378137.00      -0.00 0.00	0dE	0dN 0.000
31:  6378147.00      -0.00 0.00	0dE	0dN 10.000
31:  861996.98       -4434590.01 4487348.41	79dW	45dN 0.001
31: -0.00    -0.00 6356752.31	0dE	0dN -6378137.000
31: +0.00    -0.00 6356752.31	0dE	90dN -0.004
31:  #############################################################
31:  Test conversion from geocentric latlong to geodetic latlong
31:  0d00'00.000"W 0d00'00.000"N 0.0	0dE	0dN 0.000
31: @@ -97,9 +97,9 @@
31:  15d4'42.357"E	14d48'56.372"N 0.000	3999967.33	1999855.31 0.00
31:  ##############################################################
31:  Test robinson projection (#113)
31: --30 40	-2612096.11	4276349.63 0.00
31: --35 45	-2963455.37	4805075.13 0.00
31: -20 40	1741397.41	4276349.63 0.00
31: +-30 40	-2612095.95	4276351.58 0.00
31: +-35 45	-2963455.42	4805073.65 0.00
31: +20 40	1741397.30	4276351.58 0.00
31:  -2612095.95     4276351.58 0.00	30d0'0.004"W	40d0'0.066"N 0.000
31:  -2963455.42     4805073.65 0.00	35dW	45dN 0.000
31:  1741397.30      4276351.58 0.00	20d0'0.002"E	40d0'0.066"N 0.000
31: 
31: PROBLEMS HAVE OCCURRED
31: test file tv_out saved

I have uploaded the entire build log from the CI system.

job.log

Environment Information

• PROJ version (proj) 6.3.0
• Alpine Linux Edge on x86

Installation method

• Compiling from source

[rouault]

This is most certainly a defect in the musl libc implementation of some of the math functions that PROJ uses

[russkel]

I thought we'd see the defects extend to other archs as well. Out of the 7 or so architectures Alpine compiles for, it's only x86 that it fails on. I guess I'll post this onto the musl libc mailing list.

[richfelker]

By x86 do you mean 32-bit or x86_64?

[richfelker]

Also are you sure it's math? Naively looking at the error messages, it looks more like somehow the input/arguments didn't get processed, at least for tests 3 and 6. But I haven't read any of the relevant source yet.

[nsz-arm]

on musl non-nearest rounding mode sin/cos/tan are known to have issues and most of complex math in all rounding modes (it's a naive implementation that works for normal range inputs i think but not for difficult cases).

other musl math stuff should work (guaranteed low ulp error bounds).

[russkel]

By x86 do you mean 32-bit or x86_64?

Yes, x86 32bit. 64b has no issues with these tests.

[russkel]

Also are you sure it's math? Naively looking at the error messages, it looks more like somehow the input/arguments didn't get processed, at least for tests 3 and 6. But I haven't read any of the relevant source yet.

Good point. I will post the output of the x86_64 tests and see if the errors are different. Keep in mind the errors are only showing since I have verbosity turned on for ctest. It's possible some tests were meant to error?

[kbevers]

Ignore the proj_create: Error ... lines, they are supposed to be there. They are issued when we are testing if failures are correctly captured and handled by the software.

[richfelker]

@kbevers: From a standpoint of being familiar with the code, can you offer any insight on what math library functions might be called in the code paths that are failing? Unless it's an issue with excess precision (unique to i386 and m68k), if it's a bug in musl it pretty much has to be one of the functions with i386-specific versions, listed here: https://git.musl-libc.org/cgit/musl/tree/src/math/i386?id=v1.1.24

[kbevers]

@richfelker I can try. I haven't looked much at this, but here's a few places to start.

For the "3" failures, the test is exercising this function, so a likely candidate for the problem:

PROJ/src/conversions/cart.cpp

Lines 149 to 188 in c3e7876

	/*********************************************************************/
	static PJ_LPZ geodetic (PJ_XYZ cart, PJ *P) {
	/*********************************************************************/
	double N, p, theta, c, s;
	PJ_LPZ lpz;
	/* Perpendicular distance from point to Z-axis (HM eq. 5-28) */
	p = hypot (cart.x, cart.y);
	/* HM eq. (5-37) */
	theta = atan2 (cart.z * P->a, p * P->b);
	/* HM eq. (5-36) (from BB, 1976) */
	c = cos(theta);
	s = sin(theta);
	lpz.phi = atan2 (cart.z + P->e2s*P->b*s*s*s, p - P->es*P->a*c*c*c);
	if( fabs(lpz.phi) > M_HALFPI ) {
	// this happen on non-sphere ellipsoid when x,y,z is very close to 0
	// there is no single solution to the cart->geodetic conversion in
	// that case, so arbitrarily pickup phi = 0.
	lpz.phi = 0;
	}
	lpz.lam = atan2 (cart.y, cart.x);
	N = normal_radius_of_curvature (P->a, P->es, lpz.phi);
	c = cos(lpz.phi);
	if (fabs(c) < 1e-6) {
	/* poleward of 89.99994 deg, we avoid division by zero */
	/* by computing the height as the cartesian z value */
	/* minus the geocentric radius of the Earth at the given */
	/* latitude */
	double r = geocentric_radius (P->a, P->b, lpz.phi);
	lpz.z = fabs (cart.z) - r;
	}
	else
	lpz.z = p / c - N;
	return lpz;
	}

For the "6" test the problem is likely somewhere in https://github.com/OSGeo/PROJ/blob/master/src/ell_set.cpp. At least for the last failure. The first one may be caused by the "55s" in this gie test:

PROJ/test/gie/ellipsoid.gie

Lines 38 to 49 in c3e7876

	-------------------------------------------------------------------------------
	Then try using a built in ellipsoid
	-------------------------------------------------------------------------------
	operation proj=merc ellps=GRS80
	-------------------------------------------------------------------------------
	tolerance 10 nm
	accept 1 2
	expect 111319.4907932736 221194.0771604237
	accept 12 55s
	expect 1335833.8895192828 7326837.7148738774
	-------------------------------------------------------------------------------

Those are roughly the places that the tests are meant to check but there may be some other thing going when setting up transformations etc.

[richfelker]

Thanks, this is helpful. I'll setup a build and see if I can catch it doing something weird under debugger.

[russkel]

@richfelker https://gitlab.alpinelinux.org/alpine/aports/merge_requests/3634 if you would like an Alpine Linux APKBUILD. Thanks a ton for looking into this.

@kbevers Helpful, thank you.

[richfelker]

In the test that fails, at line 165 it takes the branch and sets lpz.phi to 0. I'm not sure yet if this is a bug on our side or on yours.

[richfelker]

OK, the cause of atan2 not truncating excess precision. I forget the history of the requirements, but historically it was underspecified whether the standard library functions returning float or double were required to do so, but my current understanding is that current or future standards require it, so we should fix this.

In theory it could be worked around by explicitly dropping the excess precision with a (double) cast in the caller, but this doesn't seem to work, probably because GCC does not use or support -fexcess-precision=standard in C++. Of course that means your C++ code is going to be subject to nasty compiler bugs around numerical stability; for example on targets without -fmath-errno or with explicit -fno-math-errno, I think GCC could use a builtin for atan2 and end up having the excess precision, breaking things again even if we fix it in musl. So you really probably need to do something to make this code safe against the badness of GCC.

I don't know the right fix here, but the current code is numerically unstable.