On Mon, 21 Mar 2022 at 15:05, Fred J. Tydeman <tydeman@tybor.com> wrote:
On Mon, 21 Mar 2022 10:35:25 +0000 Jonathan Wakely via Liaison wrote:
>
>There is some overlap with parts of
>http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p2551r0.pdf
>regarding what "has_denorm" means for the new C++ traits. But the answer to
>"what does the C macro mean?" and "what does the C++ trait mean?" should
>not be the same, as they're actually asking different questions, and so the
>proposed changes in this paper do not impact C++.

I have looked at p2551r0 and p1841r2 and I do not find "has_denorm" in either.

The new denorm_min introduced by P1841 is supposed to make it unnecessary, but doesn't really do so, which is the subject of question 1 in P2551. Question 4 in P2551 is also related.
 

The only version of C++ that I have is WG21 N4860.  In it, I do find "has_denorm"
and it appears to be equivalent to C's *_HAS_SUBNORM.

Yes, and P1841 adds a new way to get that same info, but requires you to check whether std::denorm_min<T>::value == std::norm_min<T>::value, which is awkward. That's what P2251 Q1 is about.


It appears that neither standard can give those a useful value if the treatment
of subnormals can be changed at runtime (as can be done on ARM chips).
They also do not cover the two case where:
  operands are flushed to zero, but results are not flushed.
  results are flushed, but operands are not flushed.
C implementations should define the macros as: -1 indeterminable.
C++ implementations should define as: denorm_indeterminate
While that is the correct value, it is not useful.

Aside:  References to 559 should be to 60559.

That's https://cplusplus.github.io/LWG/issue2248

 
There should be "norm_max" for the maximum normalized number.

That's std::numeric_limits<T>::max() in all published standards, and std::finite_max<T>::value with the P1841 additions.

 

The definition of "epsilon" in C++ differs from that in C.  C  added
"normalized" to the definition.  It mattes for the case where
long double is implemented as a pair of doubles.

Does "round_error" cover subnormals numbers flushed to zero?

---
Fred J. Tydeman        Tydeman Consulting
tydeman@tybor.com      Testing, numerics, programming
+1 (702) 608-6093      Vice-chair of PL22.11 (ANSI "C")
Sample C99+FPCE tests: http://www.tybor.com
Savers sleep well, investors eat well, spenders work forever.