Many ocean.math unittests fail when built with LDC

[joseph-wakeling-frequenz]

I've given a go to building ocean v5.0.x tests with ldmd2 instead of dmd (v5.0.x because it includes the getIeeeFlags fix that was a blocker to using compilers other than DMD). The command used was:

DC=ldmd2 make V=1 test ALLOW_DEPRECATIONS=1

... with LDC 1.16.0 being used to test (matching DMD 2.086.1).

My suspicion (not yet confirmed) is that all of these reflect LDC using different core math routines to DMD (e.g. implementation of cos, sin, sqrt, etc.). I suspect that similar results might be obtained for GDC, for similar reasons.

No worries if this is a WONTFIX (or at least very low priority) from the maintainers' side, but since I went through this all systematically today I thought I'd share the results.

The following unittests all fail when LDC is used:

ocean.math.Elliptic:

ocean/src/ocean/math/Elliptic.d

Line 349 in 63ef0e8

test(ellipticEComplete(1-x) == ellipticE(PI_2, x) );

ocean.math.Probability:

ocean/src/ocean/math/Probability.d

Lines 261 to 271 in 63ef0e8

	test(isfeqabs(studentsTDistributionInv(10, 0.995), 3.169_272_67L, 0.000_000_005L));
	test(isfeqabs(studentsTDistributionInv(8, 0.6), 0.261_921_096_769_043L,0.000_000_000_05L));
	// -TINV(2*0.4, 18) == -0.257123042655869
	test(isfeqabs(studentsTDistributionInv(18, 0.4), -0.257_123_042_655_869L, 0.000_000_000_05L));
	test( feqrel(studentsTDistribution(18, studentsTDistributionInv(18, 0.4L)),0.4L)
	> real.mant_dig-5 );
	test( feqrel(studentsTDistribution(11, studentsTDistributionInv(11, 0.9L)),0.9L)
	> real.mant_dig-2);

ocean/src/ocean/math/Probability.d

Line 368 in 63ef0e8

test(fabs((1-fDistribution(12, 23, 0.1)) - 0.99990562845505L)< 0.0000000000005L);

ocean.math.Math:

ocean/src/ocean/math/Math.d

Line 422 in 63ef0e8

if (!isIdentical(r, t)) test(fabs(r-t) <= .0000001);

ocean/src/ocean/math/Math.d

Lines 428 to 429 in 63ef0e8

	if (!isIdentical(r, t) && !(isNaN(r) && isNaN(t)))
	test(fabs(r-t) <= .0000001);

ocean/src/ocean/math/Math.d

Line 435 in 63ef0e8

test(isIdentical( tan(NaN(0x0123L)), NaN(0x0123L) ));

ocean/src/ocean/math/Math.d

Line 490 in 63ef0e8

test(sin(2.0+0.0i) == sin(2.0L) );

ocean/src/ocean/math/Math.d

Lines 513 to 514 in 63ef0e8

	test(cos(1.3L+0.0i)==cos(1.3L));
	test(cos(5.2Li)== cosh(5.2L));

ocean/src/ocean/math/Math.d

Line 632 in 63ef0e8

test(asin(sin(0+0i)) == 0 + 0i);

ocean/src/ocean/math/Math.d

Line 740 in 63ef0e8

test(sinh(4.2L + 0i)==sinh(4.2L));

ocean/src/ocean/math/Math.d

Line 762 in 63ef0e8

test(cosh(8.3L + 0i)==cosh(8.3L));

ocean/src/ocean/math/Math.d

Line 976 in 63ef0e8

test(cfeqrel(sqrt(4+16i)*sqrt(4+16i), 4+16i)>=real.mant_dig-2);

ocean/src/ocean/math/Math.d

Line 1504 in 63ef0e8

test(exp(1.3e5Li)==cos(1.3e5L)+sin(1.3e5L)*1i);

ocean/src/ocean/math/Math.d

Line 1506 in 63ef0e8

test(exp(7.2 + 0.0i) == exp(7.2L));

[don-clugston-sociomantic]

Some of these tests seems extremely reasonable, and shouldn't be failing.
test(asin(sin(0+0i)) == 0 + 0i);
But, the code for almost all of this should be the same as Phobos (the Phobos code was taken from Tango and improved, that code was copied from Phobos1 and improved)
It should be just copy-and-paste to make it work.

[joseph-wakeling-frequenz]

TBH I wonder if we shouldn't just import/alias from phobos try to drop as much as possible of ocean.math. But it would be nice to work out why we get these failures, given (as you say) that they should not reasonably be happening.

[Geod24]

At the moment dmd-transitional does not include Phobos (on purpose) so a simple alias isn't possible.

[don-clugston-sociomantic]

Honestly I'm horrified at some of the changes that were made to this module in Phobos. There's some good fixes which we want, that's all.

[joseph-wakeling-frequenz]

Could we have a chat about that some time so that I get a good list of what those horror changes are? I remember approxEqual but that's the only one.