On Fri, Sep 16, 2022 at 10:33 AM Richard Sandiford wrote: > > Aldy Hernandez via Gcc-patches writes: > > On Thu, Sep 15, 2022 at 9:06 AM Richard Biener > > wrote: > >> > >> On Thu, Sep 15, 2022 at 7:41 AM Aldy Hernandez wrote: > >> > > >> > Hi Richard. Hi all. > >> > > >> > The attatched patch rewrites the NAN and sign handling, dropping both > >> > tristates in favor of a pair of boolean flags for NANs, and nothing at > >> > all for signs. The signs are tracked in the range itself, so now it's > >> > possible to describe things like [-0.0, +0.0] +NAN, [+0, +0], [-5, +0], > >> > [+0, 3] -NAN, etc. > >> > > >> > There are a lot of changes, as the tristate was quite pervasive. I > >> > could use another pair of eyes. The code IMO is cleaner and handles > >> > all the cases we discussed. > >> > > >> > Here is an example of the various ranges and how they are displayed: > >> > > >> > [frange] float VARYING NAN ;; Varying includes NAN > >> > [frange] UNDEFINED ;; Empty set as always > >> > [frange] float [] NAN ;; Unknown sign NAN > >> > [frange] float [] -NAN ;; -NAN > >> > [frange] float [] +NAN ;; +NAN > >> > [frange] float [-0.0, 0.0] ;; All zeros. > >> > [frange] float [-0.0, -0.0] NAN ;; -0 or NAN. > >> > [frange] float [-5.0e+0, -1.0e+0] +NAN ;; [-5, -1] or +NAN > >> > [frange] float [-5.0e+0, -0.0] NAN ;; [-5, -0] or +-NAN > >> > [frange] float [-5.0e+0, -0.0] ;; [-5, -0] > >> > [frange] float [5.0e+0, 1.0e+1] ;; [5, 10] > >> > > >> > We could represent an unknown sign with +NAN -NAN if preferred. > >> > >> maybe -+NAN or +-NAN? I prefer to somehow show both signs for clarity > > > > Sure. > > > >> > >> > > >> > Notice the NAN signs are decoupled from the range, so we can represent > >> > a negative range with a positive NAN. For this range, > >> > frange::known_bit() would return false, as only when the signs of the > >> > NANs and range agree can we be certain. > >> > > >> > There is no longer any pessimization of ranges for intersects > >> > involving NANs. Also, union and intersect work with signed zeros: > >> > > >> > // [-0, x] U [+0, x] => [-0, x] > >> > // [ x, -0] U [ x, +0] => [ x, +0] > >> > // [-0, x] ^ [+0, x] => [+0, x] > >> > // [ x, -0] ^ [ x, +0] => [ x, -0] > >> > > >> > The special casing for signed zeros in the singleton code is gone in > >> > favor of just making sure the signs in the range agree, that is > >> > [-0, -0] for example. > >> > > >> > I have removed the idea that a known NAN is a "range", so a NAN is no > >> > longer in the endpoints itself. Requesting the bound of a known NAN > >> > is a hard fail. For that matter, we don't store the actual NAN in the > >> > range. The only information we have are the set of boolean flags. > >> > This way we make sure nothing seeps into the frange. This also means > >> > it's explicit that we don't track anything but the sign in NANs. We > >> > can revisit this if we desire to track signalling or whatever > >> > concoction y'all can imagine. > >> > > >> > All in all, I'm quite happy with this. It does look better, and we > >> > handle all the corner cases we couldn't before. Thanks for the > >> > suggestion. > >> > > >> > Regstrapped with mpfr tests on x86-64 and ppc64le Linux. Selftests > >> > were also run with -ffinite-math-only on x86-64. > >> > > >> > At Jakub's suggestion, I built lapack with associated tests. They > >> > pass on x86-64 and ppc64le Linux with no regressions from mainline. > >> > As a sanity check, I also ran them for -ffinite-math-only on x86 which > >> > (as expected) returned: > >> > > >> > NaN arithmetic did not perform per the ieee spec > >> > > >> > Otherwise, all tests pass for -ffinite-math-only. > >> > > >> > How does this look? > >> > >> Overall it looks good. > >> > >> Reading ::intersect and ::union I find it less clear to spread out the _nan > >> cases into separate functions. > > > > OK, will inline them. > > > >> > >> Can you add a comment to frange that its representation is > >> a single value-range specified by m_type, m_min, m_max > >> unioned with the set of { -NaN, +NaN }? Because somehow > >> the ::undefined_p vs. m_type == VR_UNDEFINED checks are > >> a bit confusing to the occasional reader can we instead use > >> ::nan_p to complement ::undefined_p? > > > > Wouldn't that just make nan_p the same as known_nan? Speaking of > > which, I'm not a big fan of known_nan. Perhaps we should rename all > > the known_foo variants to foo_p variants? Or...maybe even: > > > > // fpclassify like API > > bool isfinite () const; > > bool isinf () const; > > bool maybe_isinf () const; > > bool isnan () const; > > bool maybe_isnan () const; > > bool signbit_p (bool &signbit) const; > > > > That would make it clear how they map to the fpclassify API. And the > > signbit_p() follows what we do for singleton_p(tree *). > > > > isnan() would be your nan_p suggestion. > > FWIW, the reason I didn't do this with the poly_int stuff is that > it makes negative conditions harder to reason about. It's easy for > tired eyes to read: > > !isfinite() > > as meaning "is infinite", especially since there isn't a separate > isinfinite() query. But if isfinite() is equivalent to known_isfinite() > then !isfinite() instead means "might be infinite". !known_isfinite() > IMO makes that more explicit. Ughh, fair enough. I've settled on: bool known_isfinite () const; bool known_isnan () const; bool known_isinf () const; bool maybe_isnan () const; bool maybe_isinf () const; bool signbit_p (bool &signbit) const; Let me know what you think. In this next revision, I have addressed a few of the suggestions by Richi, though I have left the out-of-line handling of NANs for intersect/union because I still find them more readable (for now). This may get shuffled again if we implement frange_base and frange, so hang on. Also, I opted to implement ][ NAN with m_kind == VR_NAN instead of overloading VR_UNDEFINED. This makes the code less confusing. I have retested on x86-64 Linux. Let me know what y'all think. Thanks. Aldy