[PATCH] D77379: [FPEnv] Use single enum to represent rounding mode

[Serge Pavlov via Phabricator via llvm-commits]

sepavloff marked an inline comment as done.
sepavloff added inline comments.


================
Comment at: llvm/include/llvm/ADT/FloatingPointMode.h:26
+/// assigned to the rounding modes must agree with the values used by FLT_ROUNDS
+/// (C11, 5.2.4.2.2p8).
+enum class RoundingMode : int8_t {
----------------
rjmccall wrote:
> sepavloff wrote:
> > rjmccall wrote:
> > > I agree that we should use one enum across LLVM and Clang.  I'm not sure that using the `FLT_ROUNDS` values is worthwhile, especially since (1) `FLT_ROUNDS` doesn't specify a value for some of these (like `NearestTiesToAway`) and (2) some of the values it does use (e.g. for "indeterminable") make this actively more awkward to store.  And the most useful thing we could do — matching the values of `FE_TONEAREST` and so on — isn't possible because those values are unportable.  I'd rather we just pick arbitrary, non-ABI-stable values, like we normally would, and then make the places that rely on matching some external schema translate.
> > >  (1) FLT_ROUNDS doesn't specify a value for some of these (like NearestTiesToAway)
> > 
> > In the recent C standard draft (http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2454.pdf), there is support of all 5 rounding modes including values returned by `FLT_ROUNDS` (5.2.4.2.2p11), values used by `fegetround` and `fesetround` (FE_TONEARESTFROMZERO in 7.6p13)
> > 
> > >  (2) some of the values it does use (e.g. for "indeterminable") make this actively more awkward to store.
> > 
> > This is not a rounding mode value, it is just error indicator returned by intrinsic functions, it does not need to be stored. Added comment about that.
> > 
> > > And the most useful thing we could do — matching the values of FE_TONEAREST and so on — isn't possible because those values are unportable.
> > 
> > I am working on patch that implements `fesetround` as intrinsic function. It introduces two intrinsic functions, one is `llvm.set_rounding` (D74729 [FPEnv] Intrinsic for setting rounding mode) and the other is `llvm.convert_rounding` (unpublished yet). The latter translates platform-dependent values like FE_DOWNWARD into platform independent representation, which is the same as used by FLT_ROUNDS.
> > 
> > Actually the motivation for this patch was just the need to have platform-independent representation of rounding mode that could be used in IR, which is more or less platform-independent. The representation used by `FLT_ROUNDS` fits these purposes because:
> > * it is platform-neutral,
> > * it is defined by standard, and
> > * it encodes all IEEE rounding modes. 
> Okay.  I'm just worried that trying to match `FLT_ROUNDS` in our internal representation is ultimately going to cause unnecessary problems.  If C has standardized a value for NearesetTiesToAway, that certainly helps avoid that.  `FLT_ROUNDS` allows targets to add implementation-defined values; are there any targets that support other rounding modes that aren't currently described?
There are 10 possible "arithmetic" rounding modes (https://upload.wikimedia.org/wikipedia/commons/8/8a/Comparison_rounding_graphs_SMIL.svg).  Current implementation of glibc does not define anything beyond the standard. I know there are cores that use non-IEEE modes "stochastic rounding to nearest" and "away from zero", don't know if they expose these modes through standard interface like FLT_ROUNDS. It looks like there is no precedent of extending the value set returned by FLT_ROUNDS.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D77379/new/

https://reviews.llvm.org/D77379