On Tue, 16 Aug 2022, Jakub Jelinek wrote:
> On Mon, Aug 15, 2022 at 03:06:16PM +0200, Jakub Jelinek via Gcc-patches wrote:
> > On Mon, Aug 15, 2022 at 12:07:38PM +0000, Richard Biener wrote:
> > > Ah, I misread
> > >
> > > +static rtx
> > > +expand_builtin_issignaling (tree exp, rtx target)
> > > +{
> > > + if (!validate_arglist (exp, REAL_TYPE, VOID_TYPE))
> > > + return NULL_RTX;
> > > +
> > > + tree arg = CALL_EXPR_ARG (exp, 0);
> > > + scalar_float_mode fmode = SCALAR_FLOAT_TYPE_MODE (TREE_TYPE (arg));
> > > + const struct real_format *fmt = REAL_MODE_FORMAT (fmode);
> > > +
> > > + /* Expand the argument yielding a RTX expression. */
> > > + rtx temp = expand_normal (arg);
> > > +
> > > + /* If mode doesn't support NaN, always return 0. */
> > > + if (!HONOR_NANS (fmode))
> > > + {
> > > + emit_move_insn (target, const0_rtx);
> > > + return target;
> >
> > I think I can expand on the comment why HONOR_NANS instead of HONOR_SNANS
> > and also add comment to the folding case.
>
> So what about like in this incremental patch:
>
> --- gcc/builtins.cc.jj 2022-08-16 13:23:04.220103861 +0200
> +++ gcc/builtins.cc 2022-08-16 13:32:03.411257574 +0200
> @@ -2765,7 +2765,13 @@ expand_builtin_issignaling (tree exp, rt
> /* Expand the argument yielding a RTX expression. */
> rtx temp = expand_normal (arg);
>
> - /* If mode doesn't support NaN, always return 0. */
> + /* If mode doesn't support NaN, always return 0.
> + Don't use !HONOR_SNANS (fmode) here, so there is some possibility of
> + __builtin_issignaling working without -fsignaling-nans. Especially
> + when -fno-signaling-nans is the default.
> + On the other side, MODE_HAS_NANS (fmode) is unnecessary, with
> + -ffinite-math-only even __builtin_isnan or __builtin_fpclassify
> + fold to 0 or non-NaN/Inf classification. */
> if (!HONOR_NANS (fmode))
> {
> emit_move_insn (target, const0_rtx);
> @@ -9259,6 +9265,12 @@ fold_builtin_classify (location_t loc, t
> return fold_build2_loc (loc, UNORDERED_EXPR, type, arg, arg);
>
> case BUILT_IN_ISSIGNALING:
> + /* Folding to true for REAL_CST is done in fold_const_call_ss.
> + Don't use tree_expr_signaling_nan_p (arg) -> integer_one_node
> + and !tree_expr_maybe_signaling_nan_p (arg) -> integer_zero_node
> + here, so there is some possibility of __builtin_issignaling working
> + without -fsignaling-nans. Especially when -fno-signaling-nans is
> + the default. */
> if (!tree_expr_maybe_nan_p (arg))
> return omit_one_operand_loc (loc, type, integer_zero_node, arg);
> return NULL_TREE;
Can you also amend the extend.texi documentation? I think the
behavior will be special enough to worth mentioning it (I don't see
any of -ffinite-math-only effect on isnan/isinf mentioned though).
I'm OK with the rest of the patch if Joseph doesn't have comments
on the actual issignaling lowerings (which I didn't review for
correctness due to lack of knowledge).
> > > > That seems like a glibc bug/weird feature in the __MATH_TG macro
> > > > or _Generic.
> > > > When compiled with C++ it is rejected.
> > >
> > > So what about __builtin_issignaling then? Do we want to silently
> > > ignore errors there?
> >
> > I think we should just restrict it to the scalar floating point types.
> > After all, other typegeneric builtins that are or can be used similarly
> > do the same thing.
>
> Note, that is what the patch does currently (rejecting _Complex
> {float,double,long double} etc. arguments).
I see.
Richard.
