Re: [PATCH] Implement __builtin_issignaling

[Richard Biener <rguenther@suse.de>]

On Mon, 15 Aug 2022, Jakub Jelinek wrote:

> On Mon, Aug 15, 2022 at 11:24:14AM +0000, Richard Biener wrote:
> > Unlike the issignalling macro from glibc the builtin will return
> > false for sNaN arguments when -fno-signalling-nans is used (similar
> > to isinf, isnan, etc.).  I think this deserves mentioning in the
> > documentation (and I have my reservations about this long-time
> > behavior of FP classification builtins we have).
> 
> I have actually tried to make the builtin working even with
> -fno-signaling-nans (i.e. the default).
> That is why the folding is done only if the argument is REAL_CST
> or if !tree_expr_maybe_nan_p (arg).
> At one point I was doing the folding when
> tree_expr_signaling_nan_p (arg) (to true) or
> !tree_expr_maybe_signaling_nan_p (arg) (to false) and in that
> case indeed -fsignaling-nans was a requirement.
> -fsignaling-nans is used in the tests nevertheless because the
> tests really do care about sNaNs, so I've turned on the option
> that says they should be honored.

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;

which doesn't use HONOR_SNANS but still HONOR_NANS and thus
-ffinite-math-only.  You possibly want MODE_HAS_NANS instead here?

> > Generally it looks OK - what does it do to size optimized code?
> 
> The problem is that except for the glibc __issignaling{f,,l,f128}
> entrypoints, other C libraries don't implement it, so there is nothing to
> fallback to (unless we want to also implement it in libgcc.a).
> 
> For float/double, it is relatively short:
>         movd    %xmm0, %eax
>         xorl    $4194304, %eax
>         andl    $2147483647, %eax
>         cmpl    $2143289344, %eax
>         seta    %al
>         movzbl  %al, %eax
> which e.g. for if (__builtin_issignaling (arg)) could be even simplified
> further by just doing ja or jna, resp.
>         movabsq $9221120237041090560, %rdx
>         movq    %xmm0, %rax
>         btcq    $51, %rax
>         btrq    $63, %rax
>         cmpq    %rax, %rdx
>         setb    %al
>         movzbl  %al, %eax
> For long double (especially Intel) / _Float128 it is larger (26 insns for XFmode,
> 15 for _Float128), sure.
> 
> > glibc 2.31 seems to silently accept
> > 
> > #include <tgmath.h>
> > 
> > int foo(_Complex double x)
> > {
> >   return issignaling (x);
> > }
> 
> 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?

Richard.