Pasting a posting from Walter in digitalmars.d: Hmm, I just noticed that the code generator should use FPREM1 instead to get IEEE conformance. Darn. http://www.sesp.cse.clrc.ac.uk/html/SoftwareTools/vtune/users_guide/mergedProjects/analyzer_ec/mergedProjects/reference_olh/mergedProjects/instructions/instruct32_hh/vc108.htm http://www.sesp.cse.clrc.ac.uk/html/SoftwareTools/vtune/users_guide/mergedProjects/analyzer_ec/mergedProjects/reference_olh/mergedProjects/instructions/instruct32_hh/vc109.htm
As to why the code generator doesn't use FPREM1 instead of FPREM, there's the following comment: "We don't use fprem1 because for some inexplicable reason we get -5 when we do _modulo(15, 10)" This could be a bug in older CPUs.
(In reply to comment #1) > As to why the code generator doesn't use FPREM1 instead of FPREM, there's the > following comment: "We don't use fprem1 because for some inexplicable > reason we get -5 when we do _modulo(15, 10)" > > This could be a bug in older CPUs. It isn't a bug. That's what the IEEE remainder specifies. Note that C's fmod is NOT the same as IEEE remainder. 15/10 = 1.5, so there's a choice of n == 1 or n==2. The standard specifies even n in such cases, so r == a - b*n == 15 - 2*10 == -5. That's kind of... weird, highly non-intuitive, and not terribly useful. I'm pretty sure that that behaviour would be unpopular.
Thanks for the explanation. At least I know why that happens, now. What do you suggest, then? Staying with FPREM or going with FPREM1 ?
(In reply to comment #3) > Thanks for the explanation. At least I know why that happens, now. What do you > suggest, then? Staying with FPREM or going with FPREM1 ? It's hard to justify including a primitive built-in operator that differs from IEEE. But it may be justifiable when it's the only way to avoid a major break from C and intuition. int x = 15 % 10; int y = cast(int)((cast(float)15) % 10); // Are we really comfortable with these being completely different? You know, all this time I was thinking that the behaviour of % for negative integers was because it needed to be consistent with floating-point modulus... Now it just seems to be wrong. But I think I have the answer. In IEEE, the preferred conversion from float to int uses round-to-nearest. IEEE remainder makes sense in that context. Since in cast(int), D has inherited 'chop' rounding from C, D needs to also inherit C's fmod behaviour. So D should stay with FPREM. But we need to document it properly.
We're not breaking with C because C has no % operator for floats. But I agree we should match C99's fmod behavior, which is its current behavior.
Fixed dmd 1.053 and 2.037