std.internal.math.gammafunction: loosen some tests for 64-bit reals.

[kinke]

As supplement for JohanEngelen@2ac2581 (1st commit of #3045; that commit alone doesn't depend on the other PR).

[WalterBright]

Why? I like to have them tight, because they let me know when the algorithm changes in a way that makes the result less accurate.

[kinke]

Simple - these tests assume 80-bit reals, and it's time to support other targets too, don't you think? We're talking about 3 relaxed assertions and 9 coverage tests now only enabled for x87 reals...

[WalterBright]

Use version blocks for tests for targets that have reduced precision. The tests for all platforms should not be the worst precision on any platform. Using max precision for each target is necessary so we know when that precision decays.

[kinke]

Summary:

• line 526: previous input results in overflow for 64-bit real => extended by a dedicated assert for 64-bit real
• line 902: allow 14 instead of only 13 least-significant mantissa bits to diverge so that no additional modification is required for 64-bit reals
• disable existing extended coverage tests for non-80-bit reals as, iirc, most of those yield NaN with 64-bit reals and the coverage has most likely been measured for the x87 inline asm code only
• line 1421: relax the upper limit for the absolute difference; the new limit is used in all other asserts of that test

Using version blocks results in way clumsier code, we have enough of those PPC, PPC64, ARM, ARM64, MIPS, MIPS64 etc. stuff already in there (in LDC), so I'll stick with the static if to check the precision of the real type.

[WalterBright]

Again, please do not reduce the precision of 80 bit tests. Add a separate test for 64 bit reals.

[kinke]

Done.

[WalterBright]

thanks!

[WalterBright]

Auto-merge toggled on