Implement __builtin_issignaling
Commit Message
Hi!
The following patch implements a new builtin, __builtin_issignaling,
which can be used to implement the ISO/IEC TS 18661-1 issignaling
macro.
It is implemented as type-generic function, so there is just one
builtin, not many with various suffixes.
This patch doesn't address PR56831 nor PR58416, but I think compared to
using glibc issignaling macro could make some cases better (as
the builtin is expanded always inline and for SFmode/DFmode just
reinterprets a memory or pseudo register as SImode/DImode, so could
avoid some raising of exception + turning sNaN into qNaN before the
builtin can analyze it).
For floading point modes that do not have NaNs it will return 0,
otherwise I've tried to implement this for all the other supported
real formats.
It handles both the MIPS/PA floats where a sNaN has the mantissa
MSB set and the rest where a sNaN has it cleared, with the exception
of format which are known never to be in the MIPS/PA form.
The MIPS/PA floats are handled using a test like
(x & mask) == mask,
the other usually as
((x ^ bit) & mask) > val
where bit, mask and val are some constants.
IBM double double is done by doing DFmode test on the most significant
half, and Intel/Motorola extended (12 or 16 bytes) and IEEE quad are
handled by extracting 32-bit/16-bit words or 64-bit parts from the
value and testing those.
On x86, XFmode is handled by a special optab so that even pseudo numbers
are considered signaling, like in glibc and like the i386 specific testcase
tests.
Bootstrapped/regtested on x86_64-linux, i686-linux, powerpc64le-linux and
powerpc64-linux (the last tested with -m32/-m64), ok for trunk?
2022-08-15 Jakub Jelinek <jakub@redhat.com>
gcc/
* builtins.def (BUILT_IN_ISSIGNALING): New built-in.
* builtins.cc (expand_builtin_issignaling): New function.
(expand_builtin_signbit): Don't overwrite target.
(expand_builtin): Handle BUILT_IN_ISSIGNALING.
(fold_builtin_classify): Likewise.
(fold_builtin_1): Likewise.
* optabs.def (issignaling_optab): New.
* fold-const-call.cc (fold_const_call_ss): Handle
BUILT_IN_ISSIGNALING.
* config/i386/i386.md (issignalingxf2): New expander.
* doc/extend.texi (__builtin_issignaling): Document.
* doc/md.texi (issignaling<mode>2): Likewise.
gcc/c-family/
* c-common.cc (check_builtin_function_arguments): Handle
BUILT_IN_ISSIGNALING.
gcc/c/
* c-typeck.cc (convert_arguments): Handle BUILT_IN_ISSIGNALING.
gcc/fortran/
* f95-lang.cc (gfc_init_builtin_functions): Initialize
BUILT_IN_ISSIGNALING.
gcc/testsuite/
* gcc.dg/torture/builtin-issignaling-1.c: New test.
* gcc.dg/torture/builtin-issignaling-2.c: New test.
* gcc.target/i386/builtin-issignaling-1.c: New test.
Jakub
Comments
On Mon, 15 Aug 2022, Jakub Jelinek wrote:
> Hi!
>
> The following patch implements a new builtin, __builtin_issignaling,
> which can be used to implement the ISO/IEC TS 18661-1 issignaling
> macro.
>
> It is implemented as type-generic function, so there is just one
> builtin, not many with various suffixes.
> This patch doesn't address PR56831 nor PR58416, but I think compared to
> using glibc issignaling macro could make some cases better (as
> the builtin is expanded always inline and for SFmode/DFmode just
> reinterprets a memory or pseudo register as SImode/DImode, so could
> avoid some raising of exception + turning sNaN into qNaN before the
> builtin can analyze it).
>
> For floading point modes that do not have NaNs it will return 0,
> otherwise I've tried to implement this for all the other supported
> real formats.
> It handles both the MIPS/PA floats where a sNaN has the mantissa
> MSB set and the rest where a sNaN has it cleared, with the exception
> of format which are known never to be in the MIPS/PA form.
> The MIPS/PA floats are handled using a test like
> (x & mask) == mask,
> the other usually as
> ((x ^ bit) & mask) > val
> where bit, mask and val are some constants.
> IBM double double is done by doing DFmode test on the most significant
> half, and Intel/Motorola extended (12 or 16 bytes) and IEEE quad are
> handled by extracting 32-bit/16-bit words or 64-bit parts from the
> value and testing those.
> On x86, XFmode is handled by a special optab so that even pseudo numbers
> are considered signaling, like in glibc and like the i386 specific testcase
> tests.
>
> Bootstrapped/regtested on x86_64-linux, i686-linux, powerpc64le-linux and
> powerpc64-linux (the last tested with -m32/-m64), ok for trunk?
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).
Generally it looks OK - what does it do to size optimized code?
glibc 2.31 seems to silently accept
#include <tgmath.h>
int foo(_Complex double x)
{
return issignaling (x);
}
for vector double we get
t.c:5:23: error: incompatible type for argument 1 of ‘__issignalingf’
5 | return issignaling (x);
| ^
| |
| v2df {aka __vector(2) double}
/usr/include/bits/mathcalls-helper-functions.h:42:1: note: expected
‘float’ but argument is of type ‘v2df’ {aka ‘__vector(2) double’}
42 | __MATHDECL_1 (int, __issignaling,, (_Mdouble_ __value))
| ^
as far as I can see your __builtin silently accepts all arguments
without diagnostics and eventually dispatches to 'issignaling'
which isn't in glibc which instead seems to have __issignaling?
Richard.
> 2022-08-15 Jakub Jelinek <jakub@redhat.com>
>
> gcc/
> * builtins.def (BUILT_IN_ISSIGNALING): New built-in.
> * builtins.cc (expand_builtin_issignaling): New function.
> (expand_builtin_signbit): Don't overwrite target.
> (expand_builtin): Handle BUILT_IN_ISSIGNALING.
> (fold_builtin_classify): Likewise.
> (fold_builtin_1): Likewise.
> * optabs.def (issignaling_optab): New.
> * fold-const-call.cc (fold_const_call_ss): Handle
> BUILT_IN_ISSIGNALING.
> * config/i386/i386.md (issignalingxf2): New expander.
> * doc/extend.texi (__builtin_issignaling): Document.
> * doc/md.texi (issignaling<mode>2): Likewise.
> gcc/c-family/
> * c-common.cc (check_builtin_function_arguments): Handle
> BUILT_IN_ISSIGNALING.
> gcc/c/
> * c-typeck.cc (convert_arguments): Handle BUILT_IN_ISSIGNALING.
> gcc/fortran/
> * f95-lang.cc (gfc_init_builtin_functions): Initialize
> BUILT_IN_ISSIGNALING.
> gcc/testsuite/
> * gcc.dg/torture/builtin-issignaling-1.c: New test.
> * gcc.dg/torture/builtin-issignaling-2.c: New test.
> * gcc.target/i386/builtin-issignaling-1.c: New test.
>
> --- gcc/builtins.def.jj 2022-01-11 23:11:21.548301986 +0100
> +++ gcc/builtins.def 2022-08-11 12:15:14.200908656 +0200
> @@ -908,6 +908,7 @@ DEF_GCC_BUILTIN (BUILT_IN_ISLESS,
> DEF_GCC_BUILTIN (BUILT_IN_ISLESSEQUAL, "islessequal", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> DEF_GCC_BUILTIN (BUILT_IN_ISLESSGREATER, "islessgreater", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> DEF_GCC_BUILTIN (BUILT_IN_ISUNORDERED, "isunordered", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> +DEF_GCC_BUILTIN (BUILT_IN_ISSIGNALING, "issignaling", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> DEF_LIB_BUILTIN (BUILT_IN_LABS, "labs", BT_FN_LONG_LONG, ATTR_CONST_NOTHROW_LEAF_LIST)
> DEF_C99_BUILTIN (BUILT_IN_LLABS, "llabs", BT_FN_LONGLONG_LONGLONG, ATTR_CONST_NOTHROW_LEAF_LIST)
> DEF_GCC_BUILTIN (BUILT_IN_LONGJMP, "longjmp", BT_FN_VOID_PTR_INT, ATTR_NORETURN_NOTHROW_LIST)
> --- gcc/builtins.cc.jj 2022-07-26 10:32:23.250277352 +0200
> +++ gcc/builtins.cc 2022-08-12 17:13:06.158423558 +0200
> @@ -123,6 +123,7 @@ static rtx expand_builtin_fegetround (tr
> static rtx expand_builtin_feclear_feraise_except (tree, rtx, machine_mode,
> optab);
> static rtx expand_builtin_cexpi (tree, rtx);
> +static rtx expand_builtin_issignaling (tree, rtx);
> static rtx expand_builtin_int_roundingfn (tree, rtx);
> static rtx expand_builtin_int_roundingfn_2 (tree, rtx);
> static rtx expand_builtin_next_arg (void);
> @@ -2747,6 +2748,294 @@ build_call_nofold_loc (location_t loc, t
> return fn;
> }
>
> +/* Expand the __builtin_issignaling builtin. This needs to handle
> + all floating point formats that do support NaNs (for those that
> + don't it just sets target to 0). */
> +
> +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;
> + }
> +
> + /* Check if the back end provides an insn that handles issignaling for the
> + argument's mode. */
> + enum insn_code icode = optab_handler (issignaling_optab, fmode);
> + if (icode != CODE_FOR_nothing)
> + {
> + rtx_insn *last = get_last_insn ();
> + rtx this_target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
> + if (maybe_emit_unop_insn (icode, this_target, temp, UNKNOWN))
> + return this_target;
> + delete_insns_since (last);
> + }
> +
> + if (DECIMAL_FLOAT_MODE_P (fmode))
> + {
> + scalar_int_mode imode;
> + rtx hi;
> + switch (fmt->ieee_bits)
> + {
> + case 32:
> + case 64:
> + imode = int_mode_for_mode (fmode).require ();
> + temp = gen_lowpart (imode, temp);
> + break;
> + case 128:
> + imode = int_mode_for_size (64, 1).require ();
> + hi = NULL_RTX;
> + /* For decimal128, TImode support isn't always there and even when
> + it is, working on the DImode high part is usually better. */
> + if (!MEM_P (temp))
> + {
> + if (rtx t = simplify_gen_subreg (imode, temp, fmode,
> + subreg_highpart_offset (imode,
> + fmode)))
> + hi = t;
> + else
> + {
> + scalar_int_mode imode2;
> + if (int_mode_for_mode (fmode).exists (&imode2))
> + {
> + rtx temp2 = gen_lowpart (imode2, temp);
> + poly_uint64 off = subreg_highpart_offset (imode, imode2);
> + if (rtx t = simplify_gen_subreg (imode, temp2,
> + imode2, off))
> + hi = t;
> + }
> + }
> + if (!hi)
> + {
> + rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + }
> + if (!hi)
> + {
> + poly_int64 offset
> + = subreg_highpart_offset (imode, GET_MODE (temp));
> + hi = adjust_address (temp, imode, offset);
> + }
> + temp = hi;
> + break;
> + default:
> + gcc_unreachable ();
> + }
> + /* In all of decimal{32,64,128}, there is MSB sign bit and sNaN
> + have 6 bits below it all set. */
> + rtx val
> + = GEN_INT (HOST_WIDE_INT_C (0x3f) << (GET_MODE_BITSIZE (imode) - 7));
> + temp = expand_binop (imode, and_optab, temp, val,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, EQ, temp, val, imode, 1, 1);
> + return temp;
> + }
> +
> + /* Only PDP11 has these defined differently but doesn't support NaNs. */
> + gcc_assert (FLOAT_WORDS_BIG_ENDIAN == WORDS_BIG_ENDIAN);
> + gcc_assert (fmt->signbit_ro > 0 && fmt->b == 2);
> + gcc_assert (MODE_COMPOSITE_P (fmode)
> + || (fmt->pnan == fmt->p
> + && fmt->signbit_ro == fmt->signbit_rw));
> +
> + switch (fmt->p)
> + {
> + case 106: /* IBM double double */
> + /* For IBM double double, recurse on the most significant double. */
> + gcc_assert (MODE_COMPOSITE_P (fmode));
> + temp = convert_modes (DFmode, fmode, temp, 0);
> + fmode = DFmode;
> + fmt = REAL_MODE_FORMAT (DFmode);
> + /* FALLTHRU */
> + case 8: /* bfloat */
> + case 11: /* IEEE half */
> + case 24: /* IEEE single */
> + case 53: /* IEEE double or Intel extended with rounding to double */
> + if (fmt->p == 53 && fmt->signbit_ro == 79)
> + goto extended;
> + {
> + scalar_int_mode imode = int_mode_for_mode (fmode).require ();
> + temp = gen_lowpart (imode, temp);
> + rtx val = GEN_INT ((HOST_WIDE_INT_M1U << (fmt->p - 2))
> + & ~(HOST_WIDE_INT_M1U << fmt->signbit_ro));
> + if (fmt->qnan_msb_set)
> + {
> + rtx mask = GEN_INT (~(HOST_WIDE_INT_M1U << fmt->signbit_ro));
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << (fmt->p - 2));
> + /* For non-MIPS/PA IEEE single/double/half or bfloat, expand to:
> + ((temp ^ bit) & mask) > val. */
> + temp = expand_binop (imode, xor_optab, temp, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, and_optab, temp, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, GTU, temp, val, imode,
> + 1, 1);
> + }
> + else
> + {
> + /* For MIPS/PA IEEE single/double, expand to:
> + (temp & val) == val. */
> + temp = expand_binop (imode, and_optab, temp, val,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, EQ, temp, val, imode,
> + 1, 1);
> + }
> + }
> + break;
> + case 113: /* IEEE quad */
> + {
> + rtx hi = NULL_RTX, lo = NULL_RTX;
> + scalar_int_mode imode = int_mode_for_size (64, 1).require ();
> + /* For IEEE quad, TImode support isn't always there and even when
> + it is, working on DImode parts is usually better. */
> + if (!MEM_P (temp))
> + {
> + hi = simplify_gen_subreg (imode, temp, fmode,
> + subreg_highpart_offset (imode, fmode));
> + lo = simplify_gen_subreg (imode, temp, fmode,
> + subreg_lowpart_offset (imode, fmode));
> + if (!hi || !lo)
> + {
> + scalar_int_mode imode2;
> + if (int_mode_for_mode (fmode).exists (&imode2))
> + {
> + rtx temp2 = gen_lowpart (imode2, temp);
> + hi = simplify_gen_subreg (imode, temp2, imode2,
> + subreg_highpart_offset (imode,
> + imode2));
> + lo = simplify_gen_subreg (imode, temp2, imode2,
> + subreg_lowpart_offset (imode,
> + imode2));
> + }
> + }
> + if (!hi || !lo)
> + {
> + rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + }
> + if (!hi || !lo)
> + {
> + poly_int64 offset
> + = subreg_highpart_offset (imode, GET_MODE (temp));
> + hi = adjust_address (temp, imode, offset);
> + offset = subreg_lowpart_offset (imode, GET_MODE (temp));
> + lo = adjust_address (temp, imode, offset);
> + }
> + rtx val = GEN_INT ((HOST_WIDE_INT_M1U << (fmt->p - 2 - 64))
> + & ~(HOST_WIDE_INT_M1U << (fmt->signbit_ro - 64)));
> + if (fmt->qnan_msb_set)
> + {
> + rtx mask = GEN_INT (~(HOST_WIDE_INT_M1U << (fmt->signbit_ro
> + - 64)));
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << (fmt->p - 2 - 64));
> + /* For non-MIPS/PA IEEE quad, expand to:
> + (((hi ^ bit) | ((lo | -lo) >> 63)) & mask) > val. */
> + rtx nlo = expand_unop (imode, neg_optab, lo, NULL_RTX, 0);
> + lo = expand_binop (imode, ior_optab, lo, nlo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + lo = expand_shift (RSHIFT_EXPR, imode, lo, 63, NULL_RTX, 1);
> + temp = expand_binop (imode, xor_optab, hi, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, ior_optab, temp, lo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, and_optab, temp, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, GTU, temp, val, imode,
> + 1, 1);
> + }
> + else
> + {
> + /* For MIPS/PA IEEE quad, expand to:
> + (hi & val) == val. */
> + temp = expand_binop (imode, and_optab, hi, val,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, EQ, temp, val, imode,
> + 1, 1);
> + }
> + }
> + break;
> + case 64: /* Intel or Motorola extended */
> + extended:
> + {
> + rtx ex, hi, lo;
> + scalar_int_mode imode = int_mode_for_size (32, 1).require ();
> + scalar_int_mode iemode = int_mode_for_size (16, 1).require ();
> + if (!MEM_P (temp))
> + {
> + rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + if (fmt->signbit_ro == 95)
> + {
> + /* Motorola, always big endian, with 16-bit gap in between
> + 16-bit sign+exponent and 64-bit mantissa. */
> + ex = adjust_address (temp, iemode, 0);
> + hi = adjust_address (temp, imode, 4);
> + lo = adjust_address (temp, imode, 8);
> + }
> + else if (!WORDS_BIG_ENDIAN)
> + {
> + /* Intel little endian, 64-bit mantissa followed by 16-bit
> + sign+exponent and then either 16 or 48 bits of gap. */
> + ex = adjust_address (temp, iemode, 8);
> + hi = adjust_address (temp, imode, 4);
> + lo = adjust_address (temp, imode, 0);
> + }
> + else
> + {
> + /* Big endian Itanium. */
> + ex = adjust_address (temp, iemode, 0);
> + hi = adjust_address (temp, imode, 2);
> + lo = adjust_address (temp, imode, 6);
> + }
> + rtx val = GEN_INT (HOST_WIDE_INT_M1U << 30);
> + gcc_assert (fmt->qnan_msb_set);
> + rtx mask = GEN_INT (0x7fff);
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << 30);
> + /* For Intel/Motorola extended format, expand to:
> + (ex & mask) == mask && ((hi ^ bit) | ((lo | -lo) >> 31)) > val. */
> + rtx nlo = expand_unop (imode, neg_optab, lo, NULL_RTX, 0);
> + lo = expand_binop (imode, ior_optab, lo, nlo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + lo = expand_shift (RSHIFT_EXPR, imode, lo, 31, NULL_RTX, 1);
> + temp = expand_binop (imode, xor_optab, hi, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, ior_optab, temp, lo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, GTU, temp, val, imode, 1, 1);
> + ex = expand_binop (iemode, and_optab, ex, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + ex = emit_store_flag_force (gen_reg_rtx (GET_MODE (temp)), EQ,
> + ex, mask, iemode, 1, 1);
> + temp = expand_binop (GET_MODE (temp), and_optab, temp, ex,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + }
> + break;
> + default:
> + gcc_unreachable ();
> + }
> +
> + return temp;
> +}
> +
> /* Expand a call to one of the builtin rounding functions gcc defines
> as an extension (lfloor and lceil). As these are gcc extensions we
> do not need to worry about setting errno to EDOM.
> @@ -5508,9 +5797,9 @@ expand_builtin_signbit (tree exp, rtx ta
> if (icode != CODE_FOR_nothing)
> {
> rtx_insn *last = get_last_insn ();
> - target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
> - if (maybe_emit_unop_insn (icode, target, temp, UNKNOWN))
> - return target;
> + rtx this_target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
> + if (maybe_emit_unop_insn (icode, this_target, temp, UNKNOWN))
> + return this_target;
> delete_insns_since (last);
> }
>
> @@ -7120,6 +7409,12 @@ expand_builtin (tree exp, rtx target, rt
> return target;
> break;
>
> + case BUILT_IN_ISSIGNALING:
> + target = expand_builtin_issignaling (exp, target);
> + if (target)
> + return target;
> + break;
> +
> CASE_FLT_FN (BUILT_IN_ICEIL):
> CASE_FLT_FN (BUILT_IN_LCEIL):
> CASE_FLT_FN (BUILT_IN_LLCEIL):
> @@ -8963,6 +9258,11 @@ fold_builtin_classify (location_t loc, t
> arg = builtin_save_expr (arg);
> return fold_build2_loc (loc, UNORDERED_EXPR, type, arg, arg);
>
> + case BUILT_IN_ISSIGNALING:
> + if (!tree_expr_maybe_nan_p (arg))
> + return omit_one_operand_loc (loc, type, integer_zero_node, arg);
> + return NULL_TREE;
> +
> default:
> gcc_unreachable ();
> }
> @@ -9399,6 +9699,9 @@ fold_builtin_1 (location_t loc, tree exp
> case BUILT_IN_ISNAND128:
> return fold_builtin_classify (loc, fndecl, arg0, BUILT_IN_ISNAN);
>
> + case BUILT_IN_ISSIGNALING:
> + return fold_builtin_classify (loc, fndecl, arg0, BUILT_IN_ISSIGNALING);
> +
> case BUILT_IN_FREE:
> if (integer_zerop (arg0))
> return build_empty_stmt (loc);
> --- gcc/optabs.def.jj 2022-02-04 14:36:55.424599447 +0100
> +++ gcc/optabs.def 2022-08-11 13:06:09.888416939 +0200
> @@ -313,6 +313,7 @@ OPTAB_D (fmod_optab, "fmod$a3")
> OPTAB_D (hypot_optab, "hypot$a3")
> OPTAB_D (ilogb_optab, "ilogb$a2")
> OPTAB_D (isinf_optab, "isinf$a2")
> +OPTAB_D (issignaling_optab, "issignaling$a2")
> OPTAB_D (ldexp_optab, "ldexp$a3")
> OPTAB_D (log10_optab, "log10$a2")
> OPTAB_D (log1p_optab, "log1p$a2")
> --- gcc/fold-const-call.cc.jj 2022-01-18 11:58:59.510983085 +0100
> +++ gcc/fold-const-call.cc 2022-08-11 12:31:07.294918860 +0200
> @@ -952,6 +952,10 @@ fold_const_call_ss (wide_int *result, co
> *result = wi::shwi (real_isfinite (arg) ? 1 : 0, precision);
> return true;
>
> + case CFN_BUILT_IN_ISSIGNALING:
> + *result = wi::shwi (real_issignaling_nan (arg) ? 1 : 0, precision);
> + return true;
> +
> CASE_CFN_ISINF:
> case CFN_BUILT_IN_ISINFD32:
> case CFN_BUILT_IN_ISINFD64:
> --- gcc/config/i386/i386.md.jj 2022-08-10 09:06:51.463232943 +0200
> +++ gcc/config/i386/i386.md 2022-08-12 11:56:14.763951760 +0200
> @@ -24720,6 +24720,58 @@ (define_expand "spaceshipxf3"
> DONE;
> })
>
> +;; Defined because the generic expand_builtin_issignaling for XFmode
> +;; only tests for sNaNs, but i387 treats also pseudo numbers as always
> +;; signaling.
> +(define_expand "issignalingxf2"
> + [(match_operand:SI 0 "register_operand")
> + (match_operand:XF 1 "general_operand")]
> + ""
> +{
> + rtx temp = operands[1];
> + if (!MEM_P (temp))
> + {
> + rtx mem = assign_stack_temp (XFmode, GET_MODE_SIZE (XFmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + rtx ex = adjust_address (temp, HImode, 8);
> + rtx hi = adjust_address (temp, SImode, 4);
> + rtx lo = adjust_address (temp, SImode, 0);
> + rtx val = GEN_INT (HOST_WIDE_INT_M1U << 30);
> + rtx mask = GEN_INT (0x7fff);
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << 30);
> + /* Expand to:
> + ((ex & mask) && (int) hi >= 0)
> + || ((ex & mask) == mask && ((hi ^ bit) | ((lo | -lo) >> 31)) > val). */
> + rtx nlo = expand_unop (SImode, neg_optab, lo, NULL_RTX, 0);
> + lo = expand_binop (SImode, ior_optab, lo, nlo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + lo = expand_shift (RSHIFT_EXPR, SImode, lo, 31, NULL_RTX, 1);
> + temp = expand_binop (SImode, xor_optab, hi, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (SImode, ior_optab, temp, lo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (gen_reg_rtx (SImode), GTU, temp, val,
> + SImode, 1, 1);
> + ex = expand_binop (HImode, and_optab, ex, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + rtx temp2 = emit_store_flag_force (gen_reg_rtx (SImode), NE,
> + ex, const0_rtx, SImode, 1, 1);
> + ex = emit_store_flag_force (gen_reg_rtx (SImode), EQ,
> + ex, mask, HImode, 1, 1);
> + temp = expand_binop (SImode, and_optab, temp, ex,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + rtx temp3 = emit_store_flag_force (gen_reg_rtx (SImode), GE,
> + hi, const0_rtx, SImode, 0, 1);
> + temp2 = expand_binop (SImode, and_optab, temp2, temp3,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (SImode, ior_optab, temp, temp2,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + emit_move_insn (operands[0], temp);
> + DONE;
> +})
> +
> (include "mmx.md")
> (include "sse.md")
> (include "sync.md")
> --- gcc/doc/extend.texi.jj 2022-07-26 10:32:23.642272293 +0200
> +++ gcc/doc/extend.texi 2022-08-11 21:57:06.727147454 +0200
> @@ -13001,6 +13001,7 @@ is called and the @var{flag} argument pa
> @findex __builtin_isless
> @findex __builtin_islessequal
> @findex __builtin_islessgreater
> +@findex __builtin_issignaling
> @findex __builtin_isunordered
> @findex __builtin_object_size
> @findex __builtin_powi
> @@ -14489,6 +14490,14 @@ Similar to @code{__builtin_nans}, except
> @code{_Float@var{n}x}.
> @end deftypefn
>
> +@deftypefn {Built-in Function} int __builtin_issignaling (...)
> +Return non-zero if the argument is a signaling NaN and zero otherwise.
> +Note while the parameter list is an
> +ellipsis, this function only accepts exactly one floating-point
> +argument. GCC treats this parameter as type-generic, which means it
> +does not do default promotion from float to double.
> +@end deftypefn
> +
> @deftypefn {Built-in Function} int __builtin_ffs (int x)
> Returns one plus the index of the least significant 1-bit of @var{x}, or
> if @var{x} is zero, returns zero.
> --- gcc/doc/md.texi.jj 2022-06-27 11:18:02.610059335 +0200
> +++ gcc/doc/md.texi 2022-08-11 22:00:11.470708501 +0200
> @@ -6184,6 +6184,10 @@ floating-point mode.
>
> This pattern is not allowed to @code{FAIL}.
>
> +@cindex @code{issignaling@var{m}2} instruction pattern
> +@item @samp{issignaling@var{m}2}
> +Set operand 0 to 1 if operand 1 is a signaling NaN and to 0 otherwise.
> +
> @cindex @code{cadd90@var{m}3} instruction pattern
> @item @samp{cadd90@var{m}3}
> Perform vector add and subtract on even/odd number pairs. The operation being
> --- gcc/c-family/c-common.cc.jj 2022-08-10 09:06:51.214236184 +0200
> +++ gcc/c-family/c-common.cc 2022-08-11 12:19:06.471714333 +0200
> @@ -6294,6 +6294,7 @@ check_builtin_function_arguments (locati
> case BUILT_IN_ISINF_SIGN:
> case BUILT_IN_ISNAN:
> case BUILT_IN_ISNORMAL:
> + case BUILT_IN_ISSIGNALING:
> case BUILT_IN_SIGNBIT:
> if (builtin_function_validate_nargs (loc, fndecl, nargs, 1))
> {
> --- gcc/c/c-typeck.cc.jj 2022-08-10 09:06:51.331234661 +0200
> +++ gcc/c/c-typeck.cc 2022-08-11 12:16:20.586995677 +0200
> @@ -3546,6 +3546,7 @@ convert_arguments (location_t loc, vec<l
> case BUILT_IN_ISINF_SIGN:
> case BUILT_IN_ISNAN:
> case BUILT_IN_ISNORMAL:
> + case BUILT_IN_ISSIGNALING:
> case BUILT_IN_FPCLASSIFY:
> type_generic_remove_excess_precision = true;
> break;
> --- gcc/fortran/f95-lang.cc.jj 2022-06-03 11:20:13.207071074 +0200
> +++ gcc/fortran/f95-lang.cc 2022-08-11 12:20:57.243190944 +0200
> @@ -1013,6 +1013,8 @@ gfc_init_builtin_functions (void)
> "__builtin_isnan", ATTR_CONST_NOTHROW_LEAF_LIST);
> gfc_define_builtin ("__builtin_isnormal", ftype, BUILT_IN_ISNORMAL,
> "__builtin_isnormal", ATTR_CONST_NOTHROW_LEAF_LIST);
> + gfc_define_builtin ("__builtin_issignaling", ftype, BUILT_IN_ISSIGNALING,
> + "__builtin_issignaling", ATTR_CONST_NOTHROW_LEAF_LIST);
> gfc_define_builtin ("__builtin_signbit", ftype, BUILT_IN_SIGNBIT,
> "__builtin_signbit", ATTR_CONST_NOTHROW_LEAF_LIST);
>
> --- gcc/testsuite/gcc.dg/torture/builtin-issignaling-1.c.jj 2022-08-11 21:20:48.790870018 +0200
> +++ gcc/testsuite/gcc.dg/torture/builtin-issignaling-1.c 2022-08-12 12:11:34.800702002 +0200
> @@ -0,0 +1,95 @@
> +/* { dg-do run } */
> +/* { dg-add-options ieee } */
> +/* { dg-additional-options "-fsignaling-nans" } */
> +/* Workaround for PR57484 on ia32: */
> +/* { dg-additional-options "-msse2 -mfpmath=sse" { target { ia32 && sse2_runtime } } } */
> +
> +int
> +f1 (void)
> +{
> + return __builtin_issignaling (__builtin_nansf (""));
> +}
> +
> +int
> +f2 (void)
> +{
> + return __builtin_issignaling (__builtin_nan (""));
> +}
> +
> +int
> +f3 (void)
> +{
> + return __builtin_issignaling (0.0L);
> +}
> +
> +int
> +f4 (float x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f5 (double x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f6 (long double x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +#ifdef __SIZEOF_FLOAT128__
> +int
> +f7 (_Float128 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +#endif
> +
> +float x;
> +double y;
> +long double z;
> +#ifdef __SIZEOF_FLOAT128__
> +_Float128 w;
> +#endif
> +
> +int
> +main ()
> +{
> + if (!f1 () || f2 () || f3 ())
> + __builtin_abort ();
> + asm volatile ("" : : : "memory");
> + if (f4 (x) || !f4 (__builtin_nansf ("0x123")) || f4 (42.0f) || f4 (__builtin_nanf ("0x234"))
> + || f4 (__builtin_inff ()) || f4 (-__builtin_inff ()) || f4 (-42.0f) || f4 (-0.0f) || f4 (0.0f))
> + __builtin_abort ();
> + x = __builtin_nansf ("");
> + asm volatile ("" : : : "memory");
> + if (!f4 (x))
> + __builtin_abort ();
> + if (f5 (y) || !f5 (__builtin_nans ("0x123")) || f5 (42.0) || f5 (__builtin_nan ("0x234"))
> + || f5 (__builtin_inf ()) || f5 (-__builtin_inf ()) || f5 (-42.0) || f5 (-0.0) || f5 (0.0))
> + __builtin_abort ();
> + y = __builtin_nans ("");
> + asm volatile ("" : : : "memory");
> + if (!f5 (y))
> + __builtin_abort ();
> + if (f6 (z) || !f6 (__builtin_nansl ("0x123")) || f6 (42.0L) || f6 (__builtin_nanl ("0x234"))
> + || f6 (__builtin_infl ()) || f6 (-__builtin_infl ()) || f6 (-42.0L) || f6 (-0.0L) || f6 (0.0L))
> + __builtin_abort ();
> + z = __builtin_nansl ("");
> + asm volatile ("" : : : "memory");
> + if (!f6 (z))
> + __builtin_abort ();
> +#ifdef __SIZEOF_FLOAT128__
> + if (f7 (w) || !f7 (__builtin_nansf128 ("0x123")) || f7 (42.0Q) || f7 (__builtin_nanf128 ("0x234"))
> + || f7 (__builtin_inff128 ()) || f7 (-__builtin_inff128 ()) || f7 (-42.0Q) || f7 (-0.0Q) || f7 (0.0Q))
> + __builtin_abort ();
> + w = __builtin_nansf128 ("");
> + asm volatile ("" : : : "memory");
> + if (!f7 (w))
> + __builtin_abort ();
> +#endif
> + return 0;
> +}
> --- gcc/testsuite/gcc.dg/torture/builtin-issignaling-2.c.jj 2022-08-11 21:21:34.673265283 +0200
> +++ gcc/testsuite/gcc.dg/torture/builtin-issignaling-2.c 2022-08-12 12:13:10.349427951 +0200
> @@ -0,0 +1,73 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target dfp } */
> +/* { dg-additional-options "-fsignaling-nans" } */
> +
> +int
> +f1 (void)
> +{
> + return __builtin_issignaling (__builtin_nansd32 (""));
> +}
> +
> +int
> +f2 (void)
> +{
> + return __builtin_issignaling (__builtin_nand64 (""));
> +}
> +
> +int
> +f3 (void)
> +{
> + return __builtin_issignaling (0.0DD);
> +}
> +
> +int
> +f4 (_Decimal32 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f5 (_Decimal64 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f6 (_Decimal128 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +_Decimal32 x;
> +_Decimal64 y;
> +_Decimal128 z;
> +
> +int
> +main ()
> +{
> + if (!f1 () || f2 () || f3 ())
> + __builtin_abort ();
> + asm volatile ("" : : : "memory");
> + if (f4 (x) || !f4 (__builtin_nansd32 ("0x123")) || f4 (42.0DF) || f4 (__builtin_nand32 ("0x234"))
> + || f4 (__builtin_infd32 ()) || f4 (-__builtin_infd32 ()) || f4 (-42.0DF) || f4 (-0.0DF) || f4 (0.0DF))
> + __builtin_abort ();
> + x = __builtin_nansd32 ("");
> + asm volatile ("" : : : "memory");
> + if (!f4 (x))
> + __builtin_abort ();
> + if (f5 (y) || !f5 (__builtin_nansd64 ("0x123")) || f5 (42.0DD) || f5 (__builtin_nand64 ("0x234"))
> + || f5 (__builtin_infd64 ()) || f5 (-__builtin_infd64 ()) || f5 (-42.0DD) || f5 (-0.0DD) || f5 (0.0DD))
> + __builtin_abort ();
> + y = __builtin_nansd64 ("");
> + asm volatile ("" : : : "memory");
> + if (!f5 (y))
> + __builtin_abort ();
> + if (f6 (z) || !f6 (__builtin_nansd128 ("0x123")) || f6 (42.0DL) || f6 (__builtin_nand128 ("0x234"))
> + || f6 (__builtin_infd128 ()) || f6 (-__builtin_infd128 ()) || f6 (-42.0DL) || f6 (-0.0DL) || f6 (0.0DL))
> + __builtin_abort ();
> + z = __builtin_nansd128 ("");
> + asm volatile ("" : : : "memory");
> + if (!f6 (z))
> + __builtin_abort ();
> + return 0;
> +}
> --- gcc/testsuite/gcc.target/i386/builtin-issignaling-1.c.jj 2022-08-12 12:19:19.723502685 +0200
> +++ gcc/testsuite/gcc.target/i386/builtin-issignaling-1.c 2022-08-12 12:40:51.499218604 +0200
> @@ -0,0 +1,80 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -fsignaling-nans" } */
> +
> +#if __LDBL_MANT_DIG__ == 64
> +union U { struct { unsigned long long m; unsigned short e; } p; long double l; };
> +union U zero = { { 0, 0 } };
> +union U mzero = { { 0, 0x8000 } };
> +union U denorm = { { 42, 0 } };
> +union U mdenorm = { { 42, 0x8000 } };
> +union U pseudodenorm = { { 0x8000000000000000ULL, 0 } };
> +union U mpseudodenorm = { { 0x8000000000000000ULL, 0x8000 } };
> +union U pseudodenorm1 = { { 0x8000000000000042ULL, 0 } };
> +union U mpseudodenorm1 = { { 0x8000000000000042ULL, 0x8000 } };
> +union U pseudoinf = { { 0, 0x7fff } };
> +union U mpseudoinf = { { 0, 0xffff } };
> +union U pseudonan = { { 42, 0x7fff } };
> +union U mpseudonan = { { 42, 0xffff } };
> +union U pseudonan1 = { { 0x4000000000000000ULL, 0x7fff } };
> +union U mpseudonan1 = { { 0x4000000000000000ULL, 0xffff } };
> +union U pseudonan2 = { { 0x4000000000000042ULL, 0x7fff } };
> +union U mpseudonan2 = { { 0x4000000000000042ULL, 0xffff } };
> +union U inf = { { 0x8000000000000000ULL, 0x7fff } };
> +union U minf = { { 0x8000000000000000ULL, 0xffff } };
> +union U snan = { { 0x8000000000000042ULL, 0x7fff } };
> +union U msnan = { { 0x8000000000000042ULL, 0xffff } };
> +union U indefinite = { { 0xc000000000000000ULL, 0x7fff } };
> +union U mindefinite = { { 0xc000000000000000ULL, 0xffff } };
> +union U qnan = { { 0xc000000000000042ULL, 0x7fff } };
> +union U mqnan = { { 0xc000000000000042ULL, 0xffff } };
> +union U unnormal = { { 0, 0x42 } };
> +union U munnormal = { { 0, 0x8042 } };
> +union U unnormal1 = { { 42, 0x42 } };
> +union U munnormal1 = { { 42, 0x8042 } };
> +union U normal = { { 0x8000000000000000ULL, 0x42 } };
> +union U mnormal = { { 0x8000000000000000ULL, 0x8042 } };
> +union U normal1 = { { 0x8000000000000042ULL, 0x42 } };
> +union U mnormal1 = { { 0x8000000000000042ULL, 0x8042 } };
> +#endif
> +
> +int
> +main ()
> +{
> +#if __LDBL_MANT_DIG__ == 64
> + asm volatile ("" : : : "memory");
> + if (__builtin_issignaling (zero.l)
> + || __builtin_issignaling (mzero.l)
> + || __builtin_issignaling (denorm.l)
> + || __builtin_issignaling (mdenorm.l)
> + || __builtin_issignaling (pseudodenorm.l)
> + || __builtin_issignaling (mpseudodenorm.l)
> + || __builtin_issignaling (pseudodenorm1.l)
> + || __builtin_issignaling (mpseudodenorm1.l)
> + || !__builtin_issignaling (pseudoinf.l)
> + || !__builtin_issignaling (mpseudoinf.l)
> + || !__builtin_issignaling (pseudonan.l)
> + || !__builtin_issignaling (mpseudonan.l)
> + || !__builtin_issignaling (pseudonan1.l)
> + || !__builtin_issignaling (mpseudonan1.l)
> + || !__builtin_issignaling (pseudonan2.l)
> + || !__builtin_issignaling (mpseudonan2.l)
> + || __builtin_issignaling (inf.l)
> + || __builtin_issignaling (minf.l)
> + || !__builtin_issignaling (snan.l)
> + || !__builtin_issignaling (msnan.l)
> + || __builtin_issignaling (indefinite.l)
> + || __builtin_issignaling (mindefinite.l)
> + || __builtin_issignaling (qnan.l)
> + || __builtin_issignaling (mqnan.l)
> + || !__builtin_issignaling (unnormal.l)
> + || !__builtin_issignaling (munnormal.l)
> + || !__builtin_issignaling (unnormal1.l)
> + || !__builtin_issignaling (munnormal1.l)
> + || __builtin_issignaling (normal.l)
> + || __builtin_issignaling (mnormal.l)
> + || __builtin_issignaling (normal1.l)
> + || __builtin_issignaling (mnormal1.l))
> + __builtin_abort ();
> +#endif
> + return 0;
> +}
>
> Jakub
>
>
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.
> 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.
Jakub
On Aug 15 2022, Jakub Jelinek via Gcc-patches wrote:
> That seems like a glibc bug/weird feature in the __MATH_TG macro
> or _Generic.
__MATH_TG is only defined for real floating types, since all of the type
generic macros in <math.h> only accept real floating types.
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.
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.
> which doesn't use HONOR_SNANS but still HONOR_NANS and thus
> -ffinite-math-only. You possibly want MODE_HAS_NANS instead here?
But I'm not sure we want this. With -ffast-math/-ffinite-math-only etc.,
__builtin_isnan or __builtin_fpclassify for NaNs/Infs will just return 0,
so it would be strange if __builtin_issignaling didn't.
People usually only call __builtin_issignaling when they know they
have a NaN, so they want it guarded with
__builtin_isnan/__builtin_fpclassify or say <=> unordered.
> > 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.
Jakub
Hi Jakub,
Thank you for taking this on (this patch and the Fortran ones), it is really appreciated and will improve gfortran's IEEE conformance.
- "Implement __builtin_issignaling”: is approved for the tiny Fortran part
- "libgfortran: Use __builtin_issignaling in libgfortran”: was approved by Thomas
- I will have to look at the next two patches (ieee_value and ieee_class) in the next few days. I think we can make adjustments to the runtime library as well.
Related: there are several tests in our testsuite that have { dg-require-effective-target issignaling }
This is checked by check_effective_target_issignaling in gcc/testsuite/lib/target-supports.exp, and probes the support for the issignaling macro in <math.h>. I think it would make sense to adjust those tests to run on a wider range of targets, using the new built-in.
FX
On Mon, Aug 15, 2022 at 06:14:22PM +0200, FX wrote:
> Thank you for taking this on (this patch and the Fortran ones), it is really appreciated and will improve gfortran's IEEE conformance.
>
> - "Implement __builtin_issignaling”: is approved for the tiny Fortran part
> - "libgfortran: Use __builtin_issignaling in libgfortran”: was approved by Thomas
Thanks.
> - I will have to look at the next two patches (ieee_value and ieee_class) in the next few days. I think we can make adjustments to the runtime library as well.
I think we can, but we need to make it in an ABI compatible way until
libgfortran bumps soname.
That is why I haven't used _gfortrani_whatever as name etc.
>
> Related: there are several tests in our testsuite that have { dg-require-effective-target issignaling }
> This is checked by check_effective_target_issignaling in gcc/testsuite/lib/target-supports.exp, and probes the support for the issignaling macro in <math.h>. I think it would make sense to adjust those tests to run on a wider range of targets, using the new built-in.
We could replace it with some other effective target that would test if the
target supports sNaNs (or perhaps just NaNs might be enough).
Jakub
On Mon, Aug 15, 2022 at 12:12 PM Jakub Jelinek via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Hi!
>
> The following patch implements a new builtin, __builtin_issignaling,
> which can be used to implement the ISO/IEC TS 18661-1 issignaling
> macro.
>
> It is implemented as type-generic function, so there is just one
> builtin, not many with various suffixes.
> This patch doesn't address PR56831 nor PR58416, but I think compared to
> using glibc issignaling macro could make some cases better (as
> the builtin is expanded always inline and for SFmode/DFmode just
> reinterprets a memory or pseudo register as SImode/DImode, so could
> avoid some raising of exception + turning sNaN into qNaN before the
> builtin can analyze it).
>
> For floading point modes that do not have NaNs it will return 0,
> otherwise I've tried to implement this for all the other supported
> real formats.
> It handles both the MIPS/PA floats where a sNaN has the mantissa
> MSB set and the rest where a sNaN has it cleared, with the exception
> of format which are known never to be in the MIPS/PA form.
> The MIPS/PA floats are handled using a test like
> (x & mask) == mask,
> the other usually as
> ((x ^ bit) & mask) > val
> where bit, mask and val are some constants.
> IBM double double is done by doing DFmode test on the most significant
> half, and Intel/Motorola extended (12 or 16 bytes) and IEEE quad are
> handled by extracting 32-bit/16-bit words or 64-bit parts from the
> value and testing those.
> On x86, XFmode is handled by a special optab so that even pseudo numbers
> are considered signaling, like in glibc and like the i386 specific testcase
> tests.
>
> Bootstrapped/regtested on x86_64-linux, i686-linux, powerpc64le-linux and
> powerpc64-linux (the last tested with -m32/-m64), ok for trunk?
>
> 2022-08-15 Jakub Jelinek <jakub@redhat.com>
>
> gcc/
> * builtins.def (BUILT_IN_ISSIGNALING): New built-in.
> * builtins.cc (expand_builtin_issignaling): New function.
> (expand_builtin_signbit): Don't overwrite target.
> (expand_builtin): Handle BUILT_IN_ISSIGNALING.
> (fold_builtin_classify): Likewise.
> (fold_builtin_1): Likewise.
> * optabs.def (issignaling_optab): New.
> * fold-const-call.cc (fold_const_call_ss): Handle
> BUILT_IN_ISSIGNALING.
> * config/i386/i386.md (issignalingxf2): New expander.
> * doc/extend.texi (__builtin_issignaling): Document.
> * doc/md.texi (issignaling<mode>2): Likewise.
> gcc/c-family/
> * c-common.cc (check_builtin_function_arguments): Handle
> BUILT_IN_ISSIGNALING.
> gcc/c/
> * c-typeck.cc (convert_arguments): Handle BUILT_IN_ISSIGNALING.
> gcc/fortran/
> * f95-lang.cc (gfc_init_builtin_functions): Initialize
> BUILT_IN_ISSIGNALING.
> gcc/testsuite/
> * gcc.dg/torture/builtin-issignaling-1.c: New test.
> * gcc.dg/torture/builtin-issignaling-2.c: New test.
> * gcc.target/i386/builtin-issignaling-1.c: New test.
OK for x86 part.
Thanks,
Uros.
>
> --- gcc/builtins.def.jj 2022-01-11 23:11:21.548301986 +0100
> +++ gcc/builtins.def 2022-08-11 12:15:14.200908656 +0200
> @@ -908,6 +908,7 @@ DEF_GCC_BUILTIN (BUILT_IN_ISLESS,
> DEF_GCC_BUILTIN (BUILT_IN_ISLESSEQUAL, "islessequal", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> DEF_GCC_BUILTIN (BUILT_IN_ISLESSGREATER, "islessgreater", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> DEF_GCC_BUILTIN (BUILT_IN_ISUNORDERED, "isunordered", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> +DEF_GCC_BUILTIN (BUILT_IN_ISSIGNALING, "issignaling", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
> DEF_LIB_BUILTIN (BUILT_IN_LABS, "labs", BT_FN_LONG_LONG, ATTR_CONST_NOTHROW_LEAF_LIST)
> DEF_C99_BUILTIN (BUILT_IN_LLABS, "llabs", BT_FN_LONGLONG_LONGLONG, ATTR_CONST_NOTHROW_LEAF_LIST)
> DEF_GCC_BUILTIN (BUILT_IN_LONGJMP, "longjmp", BT_FN_VOID_PTR_INT, ATTR_NORETURN_NOTHROW_LIST)
> --- gcc/builtins.cc.jj 2022-07-26 10:32:23.250277352 +0200
> +++ gcc/builtins.cc 2022-08-12 17:13:06.158423558 +0200
> @@ -123,6 +123,7 @@ static rtx expand_builtin_fegetround (tr
> static rtx expand_builtin_feclear_feraise_except (tree, rtx, machine_mode,
> optab);
> static rtx expand_builtin_cexpi (tree, rtx);
> +static rtx expand_builtin_issignaling (tree, rtx);
> static rtx expand_builtin_int_roundingfn (tree, rtx);
> static rtx expand_builtin_int_roundingfn_2 (tree, rtx);
> static rtx expand_builtin_next_arg (void);
> @@ -2747,6 +2748,294 @@ build_call_nofold_loc (location_t loc, t
> return fn;
> }
>
> +/* Expand the __builtin_issignaling builtin. This needs to handle
> + all floating point formats that do support NaNs (for those that
> + don't it just sets target to 0). */
> +
> +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;
> + }
> +
> + /* Check if the back end provides an insn that handles issignaling for the
> + argument's mode. */
> + enum insn_code icode = optab_handler (issignaling_optab, fmode);
> + if (icode != CODE_FOR_nothing)
> + {
> + rtx_insn *last = get_last_insn ();
> + rtx this_target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
> + if (maybe_emit_unop_insn (icode, this_target, temp, UNKNOWN))
> + return this_target;
> + delete_insns_since (last);
> + }
> +
> + if (DECIMAL_FLOAT_MODE_P (fmode))
> + {
> + scalar_int_mode imode;
> + rtx hi;
> + switch (fmt->ieee_bits)
> + {
> + case 32:
> + case 64:
> + imode = int_mode_for_mode (fmode).require ();
> + temp = gen_lowpart (imode, temp);
> + break;
> + case 128:
> + imode = int_mode_for_size (64, 1).require ();
> + hi = NULL_RTX;
> + /* For decimal128, TImode support isn't always there and even when
> + it is, working on the DImode high part is usually better. */
> + if (!MEM_P (temp))
> + {
> + if (rtx t = simplify_gen_subreg (imode, temp, fmode,
> + subreg_highpart_offset (imode,
> + fmode)))
> + hi = t;
> + else
> + {
> + scalar_int_mode imode2;
> + if (int_mode_for_mode (fmode).exists (&imode2))
> + {
> + rtx temp2 = gen_lowpart (imode2, temp);
> + poly_uint64 off = subreg_highpart_offset (imode, imode2);
> + if (rtx t = simplify_gen_subreg (imode, temp2,
> + imode2, off))
> + hi = t;
> + }
> + }
> + if (!hi)
> + {
> + rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + }
> + if (!hi)
> + {
> + poly_int64 offset
> + = subreg_highpart_offset (imode, GET_MODE (temp));
> + hi = adjust_address (temp, imode, offset);
> + }
> + temp = hi;
> + break;
> + default:
> + gcc_unreachable ();
> + }
> + /* In all of decimal{32,64,128}, there is MSB sign bit and sNaN
> + have 6 bits below it all set. */
> + rtx val
> + = GEN_INT (HOST_WIDE_INT_C (0x3f) << (GET_MODE_BITSIZE (imode) - 7));
> + temp = expand_binop (imode, and_optab, temp, val,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, EQ, temp, val, imode, 1, 1);
> + return temp;
> + }
> +
> + /* Only PDP11 has these defined differently but doesn't support NaNs. */
> + gcc_assert (FLOAT_WORDS_BIG_ENDIAN == WORDS_BIG_ENDIAN);
> + gcc_assert (fmt->signbit_ro > 0 && fmt->b == 2);
> + gcc_assert (MODE_COMPOSITE_P (fmode)
> + || (fmt->pnan == fmt->p
> + && fmt->signbit_ro == fmt->signbit_rw));
> +
> + switch (fmt->p)
> + {
> + case 106: /* IBM double double */
> + /* For IBM double double, recurse on the most significant double. */
> + gcc_assert (MODE_COMPOSITE_P (fmode));
> + temp = convert_modes (DFmode, fmode, temp, 0);
> + fmode = DFmode;
> + fmt = REAL_MODE_FORMAT (DFmode);
> + /* FALLTHRU */
> + case 8: /* bfloat */
> + case 11: /* IEEE half */
> + case 24: /* IEEE single */
> + case 53: /* IEEE double or Intel extended with rounding to double */
> + if (fmt->p == 53 && fmt->signbit_ro == 79)
> + goto extended;
> + {
> + scalar_int_mode imode = int_mode_for_mode (fmode).require ();
> + temp = gen_lowpart (imode, temp);
> + rtx val = GEN_INT ((HOST_WIDE_INT_M1U << (fmt->p - 2))
> + & ~(HOST_WIDE_INT_M1U << fmt->signbit_ro));
> + if (fmt->qnan_msb_set)
> + {
> + rtx mask = GEN_INT (~(HOST_WIDE_INT_M1U << fmt->signbit_ro));
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << (fmt->p - 2));
> + /* For non-MIPS/PA IEEE single/double/half or bfloat, expand to:
> + ((temp ^ bit) & mask) > val. */
> + temp = expand_binop (imode, xor_optab, temp, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, and_optab, temp, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, GTU, temp, val, imode,
> + 1, 1);
> + }
> + else
> + {
> + /* For MIPS/PA IEEE single/double, expand to:
> + (temp & val) == val. */
> + temp = expand_binop (imode, and_optab, temp, val,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, EQ, temp, val, imode,
> + 1, 1);
> + }
> + }
> + break;
> + case 113: /* IEEE quad */
> + {
> + rtx hi = NULL_RTX, lo = NULL_RTX;
> + scalar_int_mode imode = int_mode_for_size (64, 1).require ();
> + /* For IEEE quad, TImode support isn't always there and even when
> + it is, working on DImode parts is usually better. */
> + if (!MEM_P (temp))
> + {
> + hi = simplify_gen_subreg (imode, temp, fmode,
> + subreg_highpart_offset (imode, fmode));
> + lo = simplify_gen_subreg (imode, temp, fmode,
> + subreg_lowpart_offset (imode, fmode));
> + if (!hi || !lo)
> + {
> + scalar_int_mode imode2;
> + if (int_mode_for_mode (fmode).exists (&imode2))
> + {
> + rtx temp2 = gen_lowpart (imode2, temp);
> + hi = simplify_gen_subreg (imode, temp2, imode2,
> + subreg_highpart_offset (imode,
> + imode2));
> + lo = simplify_gen_subreg (imode, temp2, imode2,
> + subreg_lowpart_offset (imode,
> + imode2));
> + }
> + }
> + if (!hi || !lo)
> + {
> + rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + }
> + if (!hi || !lo)
> + {
> + poly_int64 offset
> + = subreg_highpart_offset (imode, GET_MODE (temp));
> + hi = adjust_address (temp, imode, offset);
> + offset = subreg_lowpart_offset (imode, GET_MODE (temp));
> + lo = adjust_address (temp, imode, offset);
> + }
> + rtx val = GEN_INT ((HOST_WIDE_INT_M1U << (fmt->p - 2 - 64))
> + & ~(HOST_WIDE_INT_M1U << (fmt->signbit_ro - 64)));
> + if (fmt->qnan_msb_set)
> + {
> + rtx mask = GEN_INT (~(HOST_WIDE_INT_M1U << (fmt->signbit_ro
> + - 64)));
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << (fmt->p - 2 - 64));
> + /* For non-MIPS/PA IEEE quad, expand to:
> + (((hi ^ bit) | ((lo | -lo) >> 63)) & mask) > val. */
> + rtx nlo = expand_unop (imode, neg_optab, lo, NULL_RTX, 0);
> + lo = expand_binop (imode, ior_optab, lo, nlo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + lo = expand_shift (RSHIFT_EXPR, imode, lo, 63, NULL_RTX, 1);
> + temp = expand_binop (imode, xor_optab, hi, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, ior_optab, temp, lo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, and_optab, temp, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, GTU, temp, val, imode,
> + 1, 1);
> + }
> + else
> + {
> + /* For MIPS/PA IEEE quad, expand to:
> + (hi & val) == val. */
> + temp = expand_binop (imode, and_optab, hi, val,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, EQ, temp, val, imode,
> + 1, 1);
> + }
> + }
> + break;
> + case 64: /* Intel or Motorola extended */
> + extended:
> + {
> + rtx ex, hi, lo;
> + scalar_int_mode imode = int_mode_for_size (32, 1).require ();
> + scalar_int_mode iemode = int_mode_for_size (16, 1).require ();
> + if (!MEM_P (temp))
> + {
> + rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + if (fmt->signbit_ro == 95)
> + {
> + /* Motorola, always big endian, with 16-bit gap in between
> + 16-bit sign+exponent and 64-bit mantissa. */
> + ex = adjust_address (temp, iemode, 0);
> + hi = adjust_address (temp, imode, 4);
> + lo = adjust_address (temp, imode, 8);
> + }
> + else if (!WORDS_BIG_ENDIAN)
> + {
> + /* Intel little endian, 64-bit mantissa followed by 16-bit
> + sign+exponent and then either 16 or 48 bits of gap. */
> + ex = adjust_address (temp, iemode, 8);
> + hi = adjust_address (temp, imode, 4);
> + lo = adjust_address (temp, imode, 0);
> + }
> + else
> + {
> + /* Big endian Itanium. */
> + ex = adjust_address (temp, iemode, 0);
> + hi = adjust_address (temp, imode, 2);
> + lo = adjust_address (temp, imode, 6);
> + }
> + rtx val = GEN_INT (HOST_WIDE_INT_M1U << 30);
> + gcc_assert (fmt->qnan_msb_set);
> + rtx mask = GEN_INT (0x7fff);
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << 30);
> + /* For Intel/Motorola extended format, expand to:
> + (ex & mask) == mask && ((hi ^ bit) | ((lo | -lo) >> 31)) > val. */
> + rtx nlo = expand_unop (imode, neg_optab, lo, NULL_RTX, 0);
> + lo = expand_binop (imode, ior_optab, lo, nlo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + lo = expand_shift (RSHIFT_EXPR, imode, lo, 31, NULL_RTX, 1);
> + temp = expand_binop (imode, xor_optab, hi, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (imode, ior_optab, temp, lo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (target, GTU, temp, val, imode, 1, 1);
> + ex = expand_binop (iemode, and_optab, ex, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + ex = emit_store_flag_force (gen_reg_rtx (GET_MODE (temp)), EQ,
> + ex, mask, iemode, 1, 1);
> + temp = expand_binop (GET_MODE (temp), and_optab, temp, ex,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + }
> + break;
> + default:
> + gcc_unreachable ();
> + }
> +
> + return temp;
> +}
> +
> /* Expand a call to one of the builtin rounding functions gcc defines
> as an extension (lfloor and lceil). As these are gcc extensions we
> do not need to worry about setting errno to EDOM.
> @@ -5508,9 +5797,9 @@ expand_builtin_signbit (tree exp, rtx ta
> if (icode != CODE_FOR_nothing)
> {
> rtx_insn *last = get_last_insn ();
> - target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
> - if (maybe_emit_unop_insn (icode, target, temp, UNKNOWN))
> - return target;
> + rtx this_target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
> + if (maybe_emit_unop_insn (icode, this_target, temp, UNKNOWN))
> + return this_target;
> delete_insns_since (last);
> }
>
> @@ -7120,6 +7409,12 @@ expand_builtin (tree exp, rtx target, rt
> return target;
> break;
>
> + case BUILT_IN_ISSIGNALING:
> + target = expand_builtin_issignaling (exp, target);
> + if (target)
> + return target;
> + break;
> +
> CASE_FLT_FN (BUILT_IN_ICEIL):
> CASE_FLT_FN (BUILT_IN_LCEIL):
> CASE_FLT_FN (BUILT_IN_LLCEIL):
> @@ -8963,6 +9258,11 @@ fold_builtin_classify (location_t loc, t
> arg = builtin_save_expr (arg);
> return fold_build2_loc (loc, UNORDERED_EXPR, type, arg, arg);
>
> + case BUILT_IN_ISSIGNALING:
> + if (!tree_expr_maybe_nan_p (arg))
> + return omit_one_operand_loc (loc, type, integer_zero_node, arg);
> + return NULL_TREE;
> +
> default:
> gcc_unreachable ();
> }
> @@ -9399,6 +9699,9 @@ fold_builtin_1 (location_t loc, tree exp
> case BUILT_IN_ISNAND128:
> return fold_builtin_classify (loc, fndecl, arg0, BUILT_IN_ISNAN);
>
> + case BUILT_IN_ISSIGNALING:
> + return fold_builtin_classify (loc, fndecl, arg0, BUILT_IN_ISSIGNALING);
> +
> case BUILT_IN_FREE:
> if (integer_zerop (arg0))
> return build_empty_stmt (loc);
> --- gcc/optabs.def.jj 2022-02-04 14:36:55.424599447 +0100
> +++ gcc/optabs.def 2022-08-11 13:06:09.888416939 +0200
> @@ -313,6 +313,7 @@ OPTAB_D (fmod_optab, "fmod$a3")
> OPTAB_D (hypot_optab, "hypot$a3")
> OPTAB_D (ilogb_optab, "ilogb$a2")
> OPTAB_D (isinf_optab, "isinf$a2")
> +OPTAB_D (issignaling_optab, "issignaling$a2")
> OPTAB_D (ldexp_optab, "ldexp$a3")
> OPTAB_D (log10_optab, "log10$a2")
> OPTAB_D (log1p_optab, "log1p$a2")
> --- gcc/fold-const-call.cc.jj 2022-01-18 11:58:59.510983085 +0100
> +++ gcc/fold-const-call.cc 2022-08-11 12:31:07.294918860 +0200
> @@ -952,6 +952,10 @@ fold_const_call_ss (wide_int *result, co
> *result = wi::shwi (real_isfinite (arg) ? 1 : 0, precision);
> return true;
>
> + case CFN_BUILT_IN_ISSIGNALING:
> + *result = wi::shwi (real_issignaling_nan (arg) ? 1 : 0, precision);
> + return true;
> +
> CASE_CFN_ISINF:
> case CFN_BUILT_IN_ISINFD32:
> case CFN_BUILT_IN_ISINFD64:
> --- gcc/config/i386/i386.md.jj 2022-08-10 09:06:51.463232943 +0200
> +++ gcc/config/i386/i386.md 2022-08-12 11:56:14.763951760 +0200
> @@ -24720,6 +24720,58 @@ (define_expand "spaceshipxf3"
> DONE;
> })
>
> +;; Defined because the generic expand_builtin_issignaling for XFmode
> +;; only tests for sNaNs, but i387 treats also pseudo numbers as always
> +;; signaling.
> +(define_expand "issignalingxf2"
> + [(match_operand:SI 0 "register_operand")
> + (match_operand:XF 1 "general_operand")]
> + ""
> +{
> + rtx temp = operands[1];
> + if (!MEM_P (temp))
> + {
> + rtx mem = assign_stack_temp (XFmode, GET_MODE_SIZE (XFmode));
> + emit_move_insn (mem, temp);
> + temp = mem;
> + }
> + rtx ex = adjust_address (temp, HImode, 8);
> + rtx hi = adjust_address (temp, SImode, 4);
> + rtx lo = adjust_address (temp, SImode, 0);
> + rtx val = GEN_INT (HOST_WIDE_INT_M1U << 30);
> + rtx mask = GEN_INT (0x7fff);
> + rtx bit = GEN_INT (HOST_WIDE_INT_1U << 30);
> + /* Expand to:
> + ((ex & mask) && (int) hi >= 0)
> + || ((ex & mask) == mask && ((hi ^ bit) | ((lo | -lo) >> 31)) > val). */
> + rtx nlo = expand_unop (SImode, neg_optab, lo, NULL_RTX, 0);
> + lo = expand_binop (SImode, ior_optab, lo, nlo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + lo = expand_shift (RSHIFT_EXPR, SImode, lo, 31, NULL_RTX, 1);
> + temp = expand_binop (SImode, xor_optab, hi, bit,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (SImode, ior_optab, temp, lo,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = emit_store_flag_force (gen_reg_rtx (SImode), GTU, temp, val,
> + SImode, 1, 1);
> + ex = expand_binop (HImode, and_optab, ex, mask,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + rtx temp2 = emit_store_flag_force (gen_reg_rtx (SImode), NE,
> + ex, const0_rtx, SImode, 1, 1);
> + ex = emit_store_flag_force (gen_reg_rtx (SImode), EQ,
> + ex, mask, HImode, 1, 1);
> + temp = expand_binop (SImode, and_optab, temp, ex,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + rtx temp3 = emit_store_flag_force (gen_reg_rtx (SImode), GE,
> + hi, const0_rtx, SImode, 0, 1);
> + temp2 = expand_binop (SImode, and_optab, temp2, temp3,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + temp = expand_binop (SImode, ior_optab, temp, temp2,
> + NULL_RTX, 1, OPTAB_LIB_WIDEN);
> + emit_move_insn (operands[0], temp);
> + DONE;
> +})
> +
> (include "mmx.md")
> (include "sse.md")
> (include "sync.md")
> --- gcc/doc/extend.texi.jj 2022-07-26 10:32:23.642272293 +0200
> +++ gcc/doc/extend.texi 2022-08-11 21:57:06.727147454 +0200
> @@ -13001,6 +13001,7 @@ is called and the @var{flag} argument pa
> @findex __builtin_isless
> @findex __builtin_islessequal
> @findex __builtin_islessgreater
> +@findex __builtin_issignaling
> @findex __builtin_isunordered
> @findex __builtin_object_size
> @findex __builtin_powi
> @@ -14489,6 +14490,14 @@ Similar to @code{__builtin_nans}, except
> @code{_Float@var{n}x}.
> @end deftypefn
>
> +@deftypefn {Built-in Function} int __builtin_issignaling (...)
> +Return non-zero if the argument is a signaling NaN and zero otherwise.
> +Note while the parameter list is an
> +ellipsis, this function only accepts exactly one floating-point
> +argument. GCC treats this parameter as type-generic, which means it
> +does not do default promotion from float to double.
> +@end deftypefn
> +
> @deftypefn {Built-in Function} int __builtin_ffs (int x)
> Returns one plus the index of the least significant 1-bit of @var{x}, or
> if @var{x} is zero, returns zero.
> --- gcc/doc/md.texi.jj 2022-06-27 11:18:02.610059335 +0200
> +++ gcc/doc/md.texi 2022-08-11 22:00:11.470708501 +0200
> @@ -6184,6 +6184,10 @@ floating-point mode.
>
> This pattern is not allowed to @code{FAIL}.
>
> +@cindex @code{issignaling@var{m}2} instruction pattern
> +@item @samp{issignaling@var{m}2}
> +Set operand 0 to 1 if operand 1 is a signaling NaN and to 0 otherwise.
> +
> @cindex @code{cadd90@var{m}3} instruction pattern
> @item @samp{cadd90@var{m}3}
> Perform vector add and subtract on even/odd number pairs. The operation being
> --- gcc/c-family/c-common.cc.jj 2022-08-10 09:06:51.214236184 +0200
> +++ gcc/c-family/c-common.cc 2022-08-11 12:19:06.471714333 +0200
> @@ -6294,6 +6294,7 @@ check_builtin_function_arguments (locati
> case BUILT_IN_ISINF_SIGN:
> case BUILT_IN_ISNAN:
> case BUILT_IN_ISNORMAL:
> + case BUILT_IN_ISSIGNALING:
> case BUILT_IN_SIGNBIT:
> if (builtin_function_validate_nargs (loc, fndecl, nargs, 1))
> {
> --- gcc/c/c-typeck.cc.jj 2022-08-10 09:06:51.331234661 +0200
> +++ gcc/c/c-typeck.cc 2022-08-11 12:16:20.586995677 +0200
> @@ -3546,6 +3546,7 @@ convert_arguments (location_t loc, vec<l
> case BUILT_IN_ISINF_SIGN:
> case BUILT_IN_ISNAN:
> case BUILT_IN_ISNORMAL:
> + case BUILT_IN_ISSIGNALING:
> case BUILT_IN_FPCLASSIFY:
> type_generic_remove_excess_precision = true;
> break;
> --- gcc/fortran/f95-lang.cc.jj 2022-06-03 11:20:13.207071074 +0200
> +++ gcc/fortran/f95-lang.cc 2022-08-11 12:20:57.243190944 +0200
> @@ -1013,6 +1013,8 @@ gfc_init_builtin_functions (void)
> "__builtin_isnan", ATTR_CONST_NOTHROW_LEAF_LIST);
> gfc_define_builtin ("__builtin_isnormal", ftype, BUILT_IN_ISNORMAL,
> "__builtin_isnormal", ATTR_CONST_NOTHROW_LEAF_LIST);
> + gfc_define_builtin ("__builtin_issignaling", ftype, BUILT_IN_ISSIGNALING,
> + "__builtin_issignaling", ATTR_CONST_NOTHROW_LEAF_LIST);
> gfc_define_builtin ("__builtin_signbit", ftype, BUILT_IN_SIGNBIT,
> "__builtin_signbit", ATTR_CONST_NOTHROW_LEAF_LIST);
>
> --- gcc/testsuite/gcc.dg/torture/builtin-issignaling-1.c.jj 2022-08-11 21:20:48.790870018 +0200
> +++ gcc/testsuite/gcc.dg/torture/builtin-issignaling-1.c 2022-08-12 12:11:34.800702002 +0200
> @@ -0,0 +1,95 @@
> +/* { dg-do run } */
> +/* { dg-add-options ieee } */
> +/* { dg-additional-options "-fsignaling-nans" } */
> +/* Workaround for PR57484 on ia32: */
> +/* { dg-additional-options "-msse2 -mfpmath=sse" { target { ia32 && sse2_runtime } } } */
> +
> +int
> +f1 (void)
> +{
> + return __builtin_issignaling (__builtin_nansf (""));
> +}
> +
> +int
> +f2 (void)
> +{
> + return __builtin_issignaling (__builtin_nan (""));
> +}
> +
> +int
> +f3 (void)
> +{
> + return __builtin_issignaling (0.0L);
> +}
> +
> +int
> +f4 (float x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f5 (double x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f6 (long double x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +#ifdef __SIZEOF_FLOAT128__
> +int
> +f7 (_Float128 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +#endif
> +
> +float x;
> +double y;
> +long double z;
> +#ifdef __SIZEOF_FLOAT128__
> +_Float128 w;
> +#endif
> +
> +int
> +main ()
> +{
> + if (!f1 () || f2 () || f3 ())
> + __builtin_abort ();
> + asm volatile ("" : : : "memory");
> + if (f4 (x) || !f4 (__builtin_nansf ("0x123")) || f4 (42.0f) || f4 (__builtin_nanf ("0x234"))
> + || f4 (__builtin_inff ()) || f4 (-__builtin_inff ()) || f4 (-42.0f) || f4 (-0.0f) || f4 (0.0f))
> + __builtin_abort ();
> + x = __builtin_nansf ("");
> + asm volatile ("" : : : "memory");
> + if (!f4 (x))
> + __builtin_abort ();
> + if (f5 (y) || !f5 (__builtin_nans ("0x123")) || f5 (42.0) || f5 (__builtin_nan ("0x234"))
> + || f5 (__builtin_inf ()) || f5 (-__builtin_inf ()) || f5 (-42.0) || f5 (-0.0) || f5 (0.0))
> + __builtin_abort ();
> + y = __builtin_nans ("");
> + asm volatile ("" : : : "memory");
> + if (!f5 (y))
> + __builtin_abort ();
> + if (f6 (z) || !f6 (__builtin_nansl ("0x123")) || f6 (42.0L) || f6 (__builtin_nanl ("0x234"))
> + || f6 (__builtin_infl ()) || f6 (-__builtin_infl ()) || f6 (-42.0L) || f6 (-0.0L) || f6 (0.0L))
> + __builtin_abort ();
> + z = __builtin_nansl ("");
> + asm volatile ("" : : : "memory");
> + if (!f6 (z))
> + __builtin_abort ();
> +#ifdef __SIZEOF_FLOAT128__
> + if (f7 (w) || !f7 (__builtin_nansf128 ("0x123")) || f7 (42.0Q) || f7 (__builtin_nanf128 ("0x234"))
> + || f7 (__builtin_inff128 ()) || f7 (-__builtin_inff128 ()) || f7 (-42.0Q) || f7 (-0.0Q) || f7 (0.0Q))
> + __builtin_abort ();
> + w = __builtin_nansf128 ("");
> + asm volatile ("" : : : "memory");
> + if (!f7 (w))
> + __builtin_abort ();
> +#endif
> + return 0;
> +}
> --- gcc/testsuite/gcc.dg/torture/builtin-issignaling-2.c.jj 2022-08-11 21:21:34.673265283 +0200
> +++ gcc/testsuite/gcc.dg/torture/builtin-issignaling-2.c 2022-08-12 12:13:10.349427951 +0200
> @@ -0,0 +1,73 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target dfp } */
> +/* { dg-additional-options "-fsignaling-nans" } */
> +
> +int
> +f1 (void)
> +{
> + return __builtin_issignaling (__builtin_nansd32 (""));
> +}
> +
> +int
> +f2 (void)
> +{
> + return __builtin_issignaling (__builtin_nand64 (""));
> +}
> +
> +int
> +f3 (void)
> +{
> + return __builtin_issignaling (0.0DD);
> +}
> +
> +int
> +f4 (_Decimal32 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f5 (_Decimal64 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +int
> +f6 (_Decimal128 x)
> +{
> + return __builtin_issignaling (x);
> +}
> +
> +_Decimal32 x;
> +_Decimal64 y;
> +_Decimal128 z;
> +
> +int
> +main ()
> +{
> + if (!f1 () || f2 () || f3 ())
> + __builtin_abort ();
> + asm volatile ("" : : : "memory");
> + if (f4 (x) || !f4 (__builtin_nansd32 ("0x123")) || f4 (42.0DF) || f4 (__builtin_nand32 ("0x234"))
> + || f4 (__builtin_infd32 ()) || f4 (-__builtin_infd32 ()) || f4 (-42.0DF) || f4 (-0.0DF) || f4 (0.0DF))
> + __builtin_abort ();
> + x = __builtin_nansd32 ("");
> + asm volatile ("" : : : "memory");
> + if (!f4 (x))
> + __builtin_abort ();
> + if (f5 (y) || !f5 (__builtin_nansd64 ("0x123")) || f5 (42.0DD) || f5 (__builtin_nand64 ("0x234"))
> + || f5 (__builtin_infd64 ()) || f5 (-__builtin_infd64 ()) || f5 (-42.0DD) || f5 (-0.0DD) || f5 (0.0DD))
> + __builtin_abort ();
> + y = __builtin_nansd64 ("");
> + asm volatile ("" : : : "memory");
> + if (!f5 (y))
> + __builtin_abort ();
> + if (f6 (z) || !f6 (__builtin_nansd128 ("0x123")) || f6 (42.0DL) || f6 (__builtin_nand128 ("0x234"))
> + || f6 (__builtin_infd128 ()) || f6 (-__builtin_infd128 ()) || f6 (-42.0DL) || f6 (-0.0DL) || f6 (0.0DL))
> + __builtin_abort ();
> + z = __builtin_nansd128 ("");
> + asm volatile ("" : : : "memory");
> + if (!f6 (z))
> + __builtin_abort ();
> + return 0;
> +}
> --- gcc/testsuite/gcc.target/i386/builtin-issignaling-1.c.jj 2022-08-12 12:19:19.723502685 +0200
> +++ gcc/testsuite/gcc.target/i386/builtin-issignaling-1.c 2022-08-12 12:40:51.499218604 +0200
> @@ -0,0 +1,80 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -fsignaling-nans" } */
> +
> +#if __LDBL_MANT_DIG__ == 64
> +union U { struct { unsigned long long m; unsigned short e; } p; long double l; };
> +union U zero = { { 0, 0 } };
> +union U mzero = { { 0, 0x8000 } };
> +union U denorm = { { 42, 0 } };
> +union U mdenorm = { { 42, 0x8000 } };
> +union U pseudodenorm = { { 0x8000000000000000ULL, 0 } };
> +union U mpseudodenorm = { { 0x8000000000000000ULL, 0x8000 } };
> +union U pseudodenorm1 = { { 0x8000000000000042ULL, 0 } };
> +union U mpseudodenorm1 = { { 0x8000000000000042ULL, 0x8000 } };
> +union U pseudoinf = { { 0, 0x7fff } };
> +union U mpseudoinf = { { 0, 0xffff } };
> +union U pseudonan = { { 42, 0x7fff } };
> +union U mpseudonan = { { 42, 0xffff } };
> +union U pseudonan1 = { { 0x4000000000000000ULL, 0x7fff } };
> +union U mpseudonan1 = { { 0x4000000000000000ULL, 0xffff } };
> +union U pseudonan2 = { { 0x4000000000000042ULL, 0x7fff } };
> +union U mpseudonan2 = { { 0x4000000000000042ULL, 0xffff } };
> +union U inf = { { 0x8000000000000000ULL, 0x7fff } };
> +union U minf = { { 0x8000000000000000ULL, 0xffff } };
> +union U snan = { { 0x8000000000000042ULL, 0x7fff } };
> +union U msnan = { { 0x8000000000000042ULL, 0xffff } };
> +union U indefinite = { { 0xc000000000000000ULL, 0x7fff } };
> +union U mindefinite = { { 0xc000000000000000ULL, 0xffff } };
> +union U qnan = { { 0xc000000000000042ULL, 0x7fff } };
> +union U mqnan = { { 0xc000000000000042ULL, 0xffff } };
> +union U unnormal = { { 0, 0x42 } };
> +union U munnormal = { { 0, 0x8042 } };
> +union U unnormal1 = { { 42, 0x42 } };
> +union U munnormal1 = { { 42, 0x8042 } };
> +union U normal = { { 0x8000000000000000ULL, 0x42 } };
> +union U mnormal = { { 0x8000000000000000ULL, 0x8042 } };
> +union U normal1 = { { 0x8000000000000042ULL, 0x42 } };
> +union U mnormal1 = { { 0x8000000000000042ULL, 0x8042 } };
> +#endif
> +
> +int
> +main ()
> +{
> +#if __LDBL_MANT_DIG__ == 64
> + asm volatile ("" : : : "memory");
> + if (__builtin_issignaling (zero.l)
> + || __builtin_issignaling (mzero.l)
> + || __builtin_issignaling (denorm.l)
> + || __builtin_issignaling (mdenorm.l)
> + || __builtin_issignaling (pseudodenorm.l)
> + || __builtin_issignaling (mpseudodenorm.l)
> + || __builtin_issignaling (pseudodenorm1.l)
> + || __builtin_issignaling (mpseudodenorm1.l)
> + || !__builtin_issignaling (pseudoinf.l)
> + || !__builtin_issignaling (mpseudoinf.l)
> + || !__builtin_issignaling (pseudonan.l)
> + || !__builtin_issignaling (mpseudonan.l)
> + || !__builtin_issignaling (pseudonan1.l)
> + || !__builtin_issignaling (mpseudonan1.l)
> + || !__builtin_issignaling (pseudonan2.l)
> + || !__builtin_issignaling (mpseudonan2.l)
> + || __builtin_issignaling (inf.l)
> + || __builtin_issignaling (minf.l)
> + || !__builtin_issignaling (snan.l)
> + || !__builtin_issignaling (msnan.l)
> + || __builtin_issignaling (indefinite.l)
> + || __builtin_issignaling (mindefinite.l)
> + || __builtin_issignaling (qnan.l)
> + || __builtin_issignaling (mqnan.l)
> + || !__builtin_issignaling (unnormal.l)
> + || !__builtin_issignaling (munnormal.l)
> + || !__builtin_issignaling (unnormal1.l)
> + || !__builtin_issignaling (munnormal1.l)
> + || __builtin_issignaling (normal.l)
> + || __builtin_issignaling (mnormal.l)
> + || __builtin_issignaling (normal1.l)
> + || __builtin_issignaling (mnormal1.l))
> + __builtin_abort ();
> +#endif
> + return 0;
> +}
>
> Jakub
>
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;
> > > 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).
Jakub
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.
On Tue, Aug 16, 2022 at 11:41:06AM +0000, Richard Biener wrote:
> 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).
Like this?
--- gcc/doc/extend.texi.jj 2022-08-16 13:23:04.227103773 +0200
+++ gcc/doc/extend.texi 2022-08-16 13:56:01.250769807 +0200
@@ -13557,6 +13557,8 @@ In the same fashion, GCC provides @code{
@code{isinf_sign}, @code{isnormal} and @code{signbit} built-ins used with
@code{__builtin_} prefixed. The @code{isinf} and @code{isnan}
built-in functions appear both with and without the @code{__builtin_} prefix.
+With @code{-ffinite-math-only} option the @code{isinf} and @code{isnan}
+built-in functions will always return 0.
GCC provides built-in versions of the ISO C99 floating-point rounding and
exceptions handling functions @code{fegetround}, @code{feclearexcept} and
@@ -14496,6 +14498,12 @@ Note while the parameter list is an
ellipsis, this function only accepts exactly one floating-point
argument. GCC treats this parameter as type-generic, which means it
does not do default promotion from float to double.
+This built-in function can work even without the non-default
+@code{-fsignaling-nans} option, although if a signaling NaN is computed,
+stored or passed as argument to some function other than this built-in
+in the current translation unit, it is safer to use @code{-fsignaling-nans}.
+With @code{-ffinite-math-only} option this built-in function will always
+return 0.
@end deftypefn
@deftypefn {Built-in Function} int __builtin_ffs (int x)
Jakub
On Tue, 16 Aug 2022, Jakub Jelinek wrote:
> On Tue, Aug 16, 2022 at 11:41:06AM +0000, Richard Biener wrote:
> > 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).
>
> Like this?
Yes.
Thanks,
Richard.
> --- gcc/doc/extend.texi.jj 2022-08-16 13:23:04.227103773 +0200
> +++ gcc/doc/extend.texi 2022-08-16 13:56:01.250769807 +0200
> @@ -13557,6 +13557,8 @@ In the same fashion, GCC provides @code{
> @code{isinf_sign}, @code{isnormal} and @code{signbit} built-ins used with
> @code{__builtin_} prefixed. The @code{isinf} and @code{isnan}
> built-in functions appear both with and without the @code{__builtin_} prefix.
> +With @code{-ffinite-math-only} option the @code{isinf} and @code{isnan}
> +built-in functions will always return 0.
>
> GCC provides built-in versions of the ISO C99 floating-point rounding and
> exceptions handling functions @code{fegetround}, @code{feclearexcept} and
> @@ -14496,6 +14498,12 @@ Note while the parameter list is an
> ellipsis, this function only accepts exactly one floating-point
> argument. GCC treats this parameter as type-generic, which means it
> does not do default promotion from float to double.
> +This built-in function can work even without the non-default
> +@code{-fsignaling-nans} option, although if a signaling NaN is computed,
> +stored or passed as argument to some function other than this built-in
> +in the current translation unit, it is safer to use @code{-fsignaling-nans}.
> +With @code{-ffinite-math-only} option this built-in function will always
> +return 0.
> @end deftypefn
>
> @deftypefn {Built-in Function} int __builtin_ffs (int x)
>
> Jakub
>
>
On Tue, Aug 16, 2022 at 11:41:06AM +0000, Richard Biener wrote:
> 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).
I'd like to ping this patch.
Patch: https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599697.html
with
https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599794.html
https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599800.html
incremental additions.
gcc/testsuite/gcc.dg/torture/builtin-issignaling-1.c's testcase f4-f7
functions (i.e. functions which just return __builtin_issignaling of
the argument which is float, double, long double and _Float128 on
x86_64-linux) with -O2 are in assembly:
.p2align 4
.globl f4
.type f4, @function
f4:
.LFB0:
.cfi_startproc
movd %xmm0, %eax
xorl $4194304, %eax
andl $2147483647, %eax
cmpl $2143289344, %eax
seta %al
movzbl %al, %eax
ret
.cfi_endproc
.LFE0:
.size f4, .-f4
.p2align 4
.globl f5
.type f5, @function
f5:
.LFB1:
.cfi_startproc
movabsq $9221120237041090560, %rdx
movq %xmm0, %rax
btcq $51, %rax
btrq $63, %rax
cmpq %rax, %rdx
setb %al
movzbl %al, %eax
ret
.cfi_endproc
.LFE1:
.size f5, .-f5
.p2align 4
.globl f6
.type f6, @function
f6:
.LFB2:
.cfi_startproc
fldt 8(%rsp)
fstpt -24(%rsp)
movl -20(%rsp), %edx
movl -24(%rsp), %eax
movl %edx, %ecx
negl %eax
orl -24(%rsp), %eax
notl %edx
xorl $1073741824, %ecx
shrl $31, %eax
orl %ecx, %eax
movzwl -16(%rsp), %ecx
cmpl $-1073741824, %eax
movl %ecx, %esi
seta %al
notl %ecx
andl $32767, %esi
movzbl %al, %eax
negl %esi
testw $32767, %cx
sete %cl
andl %esi, %edx
movzbl %cl, %ecx
shrl $31, %edx
andl %ecx, %eax
orl %edx, %eax
ret
.cfi_endproc
.LFE2:
.size f6, .-f6
.p2align 4
.globl f7
.type f7, @function
f7:
.LFB3:
.cfi_startproc
movaps %xmm0, -24(%rsp)
movq -24(%rsp), %rsi
movq -16(%rsp), %rdi
movq %rsi, %rdx
movq %rdi, %rax
negq %rdx
btcq $47, %rax
orq %rsi, %rdx
shrq $63, %rdx
orq %rdx, %rax
movabsq $9223231299366420480, %rdx
btrq $63, %rax
cmpq %rax, %rdx
setb %al
movzbl %al, %eax
ret
.cfi_endproc
.LFE3:
.size f7, .-f7
Thanks
Jakub
On Tue, 23 Aug 2022, Jakub Jelinek via Gcc-patches wrote:
> On Tue, Aug 16, 2022 at 11:41:06AM +0000, Richard Biener wrote:
> > 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).
>
> I'd like to ping this patch.
> Patch: https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599697.html
> with
> https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599794.html
> https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599800.html
> incremental additions.
One testsuite comment:
> +#ifdef __SIZEOF_FLOAT128__
> + if (f7 (w) || !f7 (__builtin_nansf128 ("0x123")) || f7 (42.0Q) || f7 (__builtin_nanf128 ("0x234"))
> + || f7 (__builtin_inff128 ()) || f7 (-__builtin_inff128 ()) || f7 (-42.0Q) || f7 (-0.0Q) || f7 (0.0Q))
> + __builtin_abort ();
__SIZEOF_FLOAT128__ is a target-specific macro for two targets. I think
it's better to model things on the existing _FloatN and _FloatNx tests, so
have such a test for each such type, preferably with most of the test
implementation type-generic (see e.g. floatn-builtin.h) and then with the
individual tests including e.g.
/* { dg-add-options float128 } */
/* { dg-require-effective-target float128_runtime } */
to enable and test for the relevant support. That would mean _Float128
gets tested wherever supported - and also that the support is properly
tested for _Float16, which doesn't look like it's covered by the tests in
this patch at all at present.
On Mon, Aug 15, 2022 at 12:12:02PM +0200, Jakub Jelinek via Gcc-patches wrote:
> Hi!
>
> The following patch implements a new builtin, __builtin_issignaling,
> which can be used to implement the ISO/IEC TS 18661-1 issignaling
> macro.
I haven't looked in detail at the patch, but from the description I think it
needs a way for machine dependent parts to optimize this for a given mode when
various switches are used.
For example on the PowerPC (and at least older AMD processors) it can often
times be expensive to move data from a floating point register to a GPR
register to do the masking operation. I implemented special purpose
recognizers a few years ago to speed up the SFmode math library to recognize
when the glibc tests were being done.
If the machine has a faster way to implement it, it should be allowed to do
this.
On some of the PowerPC's, I might want to generate code like (note, I might
have some of the tests backwards, but the idea is test if the value is a NaN,
and if it is a NaN, look at the signalling bit. Otherwise return 0:
if (__builtin_isnan (x)) {
long long mantissa = <extract mantissa>
return ~((mantissa >> 52) & 1);
} else {
return 0;
}
That way it doesn't have to do a direct move to a GPR register to do the
extraction in general.
For SFmode/DFmode, you should be able to optimize it to something like to avoid
any explicit jumps on a power10 (note, I may have some of the tests/shifts
wrong, but that is the general method you could use):
static int inline
is_signalling_nan (double value)
{
int exponent = __builtin_vec_scalar_extract_exp (value);
long long mantissa = __builtin_vec_scalar_extract_sig (value);
return (mantissa >> 52) & (exponent == 0);
}
On Thu, Aug 25, 2022 at 03:23:12PM -0400, Michael Meissner wrote:
> On Mon, Aug 15, 2022 at 12:12:02PM +0200, Jakub Jelinek via Gcc-patches wrote:
> > Hi!
> >
> > The following patch implements a new builtin, __builtin_issignaling,
> > which can be used to implement the ISO/IEC TS 18661-1 issignaling
> > macro.
>
> I haven't looked in detail at the patch, but from the description I think it
> needs a way for machine dependent parts to optimize this for a given mode when
> various switches are used.
There is an optab which takes priority over the generic code.
So if you have something better than the generic expansion, feel free to add
the named pattern(s).
Currently it is used just on x86 for xf mode.
Jakub
On Thu, Aug 25, 2022 at 09:56:18PM +0200, Jakub Jelinek wrote:
> On Thu, Aug 25, 2022 at 03:23:12PM -0400, Michael Meissner wrote:
> > On Mon, Aug 15, 2022 at 12:12:02PM +0200, Jakub Jelinek via Gcc-patches wrote:
> > > Hi!
> > >
> > > The following patch implements a new builtin, __builtin_issignaling,
> > > which can be used to implement the ISO/IEC TS 18661-1 issignaling
> > > macro.
> >
> > I haven't looked in detail at the patch, but from the description I think it
> > needs a way for machine dependent parts to optimize this for a given mode when
> > various switches are used.
>
> There is an optab which takes priority over the generic code.
> So if you have something better than the generic expansion, feel free to add
> the named pattern(s).
> Currently it is used just on x86 for xf mode.
That is good to know. When I looked into before, I didn't notice there was an
optab for the case.
@@ -908,6 +908,7 @@ DEF_GCC_BUILTIN (BUILT_IN_ISLESS,
DEF_GCC_BUILTIN (BUILT_IN_ISLESSEQUAL, "islessequal", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
DEF_GCC_BUILTIN (BUILT_IN_ISLESSGREATER, "islessgreater", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
DEF_GCC_BUILTIN (BUILT_IN_ISUNORDERED, "isunordered", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
+DEF_GCC_BUILTIN (BUILT_IN_ISSIGNALING, "issignaling", BT_FN_INT_VAR, ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF)
DEF_LIB_BUILTIN (BUILT_IN_LABS, "labs", BT_FN_LONG_LONG, ATTR_CONST_NOTHROW_LEAF_LIST)
DEF_C99_BUILTIN (BUILT_IN_LLABS, "llabs", BT_FN_LONGLONG_LONGLONG, ATTR_CONST_NOTHROW_LEAF_LIST)
DEF_GCC_BUILTIN (BUILT_IN_LONGJMP, "longjmp", BT_FN_VOID_PTR_INT, ATTR_NORETURN_NOTHROW_LIST)
@@ -123,6 +123,7 @@ static rtx expand_builtin_fegetround (tr
static rtx expand_builtin_feclear_feraise_except (tree, rtx, machine_mode,
optab);
static rtx expand_builtin_cexpi (tree, rtx);
+static rtx expand_builtin_issignaling (tree, rtx);
static rtx expand_builtin_int_roundingfn (tree, rtx);
static rtx expand_builtin_int_roundingfn_2 (tree, rtx);
static rtx expand_builtin_next_arg (void);
@@ -2747,6 +2748,294 @@ build_call_nofold_loc (location_t loc, t
return fn;
}
+/* Expand the __builtin_issignaling builtin. This needs to handle
+ all floating point formats that do support NaNs (for those that
+ don't it just sets target to 0). */
+
+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;
+ }
+
+ /* Check if the back end provides an insn that handles issignaling for the
+ argument's mode. */
+ enum insn_code icode = optab_handler (issignaling_optab, fmode);
+ if (icode != CODE_FOR_nothing)
+ {
+ rtx_insn *last = get_last_insn ();
+ rtx this_target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
+ if (maybe_emit_unop_insn (icode, this_target, temp, UNKNOWN))
+ return this_target;
+ delete_insns_since (last);
+ }
+
+ if (DECIMAL_FLOAT_MODE_P (fmode))
+ {
+ scalar_int_mode imode;
+ rtx hi;
+ switch (fmt->ieee_bits)
+ {
+ case 32:
+ case 64:
+ imode = int_mode_for_mode (fmode).require ();
+ temp = gen_lowpart (imode, temp);
+ break;
+ case 128:
+ imode = int_mode_for_size (64, 1).require ();
+ hi = NULL_RTX;
+ /* For decimal128, TImode support isn't always there and even when
+ it is, working on the DImode high part is usually better. */
+ if (!MEM_P (temp))
+ {
+ if (rtx t = simplify_gen_subreg (imode, temp, fmode,
+ subreg_highpart_offset (imode,
+ fmode)))
+ hi = t;
+ else
+ {
+ scalar_int_mode imode2;
+ if (int_mode_for_mode (fmode).exists (&imode2))
+ {
+ rtx temp2 = gen_lowpart (imode2, temp);
+ poly_uint64 off = subreg_highpart_offset (imode, imode2);
+ if (rtx t = simplify_gen_subreg (imode, temp2,
+ imode2, off))
+ hi = t;
+ }
+ }
+ if (!hi)
+ {
+ rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
+ emit_move_insn (mem, temp);
+ temp = mem;
+ }
+ }
+ if (!hi)
+ {
+ poly_int64 offset
+ = subreg_highpart_offset (imode, GET_MODE (temp));
+ hi = adjust_address (temp, imode, offset);
+ }
+ temp = hi;
+ break;
+ default:
+ gcc_unreachable ();
+ }
+ /* In all of decimal{32,64,128}, there is MSB sign bit and sNaN
+ have 6 bits below it all set. */
+ rtx val
+ = GEN_INT (HOST_WIDE_INT_C (0x3f) << (GET_MODE_BITSIZE (imode) - 7));
+ temp = expand_binop (imode, and_optab, temp, val,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = emit_store_flag_force (target, EQ, temp, val, imode, 1, 1);
+ return temp;
+ }
+
+ /* Only PDP11 has these defined differently but doesn't support NaNs. */
+ gcc_assert (FLOAT_WORDS_BIG_ENDIAN == WORDS_BIG_ENDIAN);
+ gcc_assert (fmt->signbit_ro > 0 && fmt->b == 2);
+ gcc_assert (MODE_COMPOSITE_P (fmode)
+ || (fmt->pnan == fmt->p
+ && fmt->signbit_ro == fmt->signbit_rw));
+
+ switch (fmt->p)
+ {
+ case 106: /* IBM double double */
+ /* For IBM double double, recurse on the most significant double. */
+ gcc_assert (MODE_COMPOSITE_P (fmode));
+ temp = convert_modes (DFmode, fmode, temp, 0);
+ fmode = DFmode;
+ fmt = REAL_MODE_FORMAT (DFmode);
+ /* FALLTHRU */
+ case 8: /* bfloat */
+ case 11: /* IEEE half */
+ case 24: /* IEEE single */
+ case 53: /* IEEE double or Intel extended with rounding to double */
+ if (fmt->p == 53 && fmt->signbit_ro == 79)
+ goto extended;
+ {
+ scalar_int_mode imode = int_mode_for_mode (fmode).require ();
+ temp = gen_lowpart (imode, temp);
+ rtx val = GEN_INT ((HOST_WIDE_INT_M1U << (fmt->p - 2))
+ & ~(HOST_WIDE_INT_M1U << fmt->signbit_ro));
+ if (fmt->qnan_msb_set)
+ {
+ rtx mask = GEN_INT (~(HOST_WIDE_INT_M1U << fmt->signbit_ro));
+ rtx bit = GEN_INT (HOST_WIDE_INT_1U << (fmt->p - 2));
+ /* For non-MIPS/PA IEEE single/double/half or bfloat, expand to:
+ ((temp ^ bit) & mask) > val. */
+ temp = expand_binop (imode, xor_optab, temp, bit,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = expand_binop (imode, and_optab, temp, mask,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = emit_store_flag_force (target, GTU, temp, val, imode,
+ 1, 1);
+ }
+ else
+ {
+ /* For MIPS/PA IEEE single/double, expand to:
+ (temp & val) == val. */
+ temp = expand_binop (imode, and_optab, temp, val,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = emit_store_flag_force (target, EQ, temp, val, imode,
+ 1, 1);
+ }
+ }
+ break;
+ case 113: /* IEEE quad */
+ {
+ rtx hi = NULL_RTX, lo = NULL_RTX;
+ scalar_int_mode imode = int_mode_for_size (64, 1).require ();
+ /* For IEEE quad, TImode support isn't always there and even when
+ it is, working on DImode parts is usually better. */
+ if (!MEM_P (temp))
+ {
+ hi = simplify_gen_subreg (imode, temp, fmode,
+ subreg_highpart_offset (imode, fmode));
+ lo = simplify_gen_subreg (imode, temp, fmode,
+ subreg_lowpart_offset (imode, fmode));
+ if (!hi || !lo)
+ {
+ scalar_int_mode imode2;
+ if (int_mode_for_mode (fmode).exists (&imode2))
+ {
+ rtx temp2 = gen_lowpart (imode2, temp);
+ hi = simplify_gen_subreg (imode, temp2, imode2,
+ subreg_highpart_offset (imode,
+ imode2));
+ lo = simplify_gen_subreg (imode, temp2, imode2,
+ subreg_lowpart_offset (imode,
+ imode2));
+ }
+ }
+ if (!hi || !lo)
+ {
+ rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
+ emit_move_insn (mem, temp);
+ temp = mem;
+ }
+ }
+ if (!hi || !lo)
+ {
+ poly_int64 offset
+ = subreg_highpart_offset (imode, GET_MODE (temp));
+ hi = adjust_address (temp, imode, offset);
+ offset = subreg_lowpart_offset (imode, GET_MODE (temp));
+ lo = adjust_address (temp, imode, offset);
+ }
+ rtx val = GEN_INT ((HOST_WIDE_INT_M1U << (fmt->p - 2 - 64))
+ & ~(HOST_WIDE_INT_M1U << (fmt->signbit_ro - 64)));
+ if (fmt->qnan_msb_set)
+ {
+ rtx mask = GEN_INT (~(HOST_WIDE_INT_M1U << (fmt->signbit_ro
+ - 64)));
+ rtx bit = GEN_INT (HOST_WIDE_INT_1U << (fmt->p - 2 - 64));
+ /* For non-MIPS/PA IEEE quad, expand to:
+ (((hi ^ bit) | ((lo | -lo) >> 63)) & mask) > val. */
+ rtx nlo = expand_unop (imode, neg_optab, lo, NULL_RTX, 0);
+ lo = expand_binop (imode, ior_optab, lo, nlo,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ lo = expand_shift (RSHIFT_EXPR, imode, lo, 63, NULL_RTX, 1);
+ temp = expand_binop (imode, xor_optab, hi, bit,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = expand_binop (imode, ior_optab, temp, lo,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = expand_binop (imode, and_optab, temp, mask,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = emit_store_flag_force (target, GTU, temp, val, imode,
+ 1, 1);
+ }
+ else
+ {
+ /* For MIPS/PA IEEE quad, expand to:
+ (hi & val) == val. */
+ temp = expand_binop (imode, and_optab, hi, val,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = emit_store_flag_force (target, EQ, temp, val, imode,
+ 1, 1);
+ }
+ }
+ break;
+ case 64: /* Intel or Motorola extended */
+ extended:
+ {
+ rtx ex, hi, lo;
+ scalar_int_mode imode = int_mode_for_size (32, 1).require ();
+ scalar_int_mode iemode = int_mode_for_size (16, 1).require ();
+ if (!MEM_P (temp))
+ {
+ rtx mem = assign_stack_temp (fmode, GET_MODE_SIZE (fmode));
+ emit_move_insn (mem, temp);
+ temp = mem;
+ }
+ if (fmt->signbit_ro == 95)
+ {
+ /* Motorola, always big endian, with 16-bit gap in between
+ 16-bit sign+exponent and 64-bit mantissa. */
+ ex = adjust_address (temp, iemode, 0);
+ hi = adjust_address (temp, imode, 4);
+ lo = adjust_address (temp, imode, 8);
+ }
+ else if (!WORDS_BIG_ENDIAN)
+ {
+ /* Intel little endian, 64-bit mantissa followed by 16-bit
+ sign+exponent and then either 16 or 48 bits of gap. */
+ ex = adjust_address (temp, iemode, 8);
+ hi = adjust_address (temp, imode, 4);
+ lo = adjust_address (temp, imode, 0);
+ }
+ else
+ {
+ /* Big endian Itanium. */
+ ex = adjust_address (temp, iemode, 0);
+ hi = adjust_address (temp, imode, 2);
+ lo = adjust_address (temp, imode, 6);
+ }
+ rtx val = GEN_INT (HOST_WIDE_INT_M1U << 30);
+ gcc_assert (fmt->qnan_msb_set);
+ rtx mask = GEN_INT (0x7fff);
+ rtx bit = GEN_INT (HOST_WIDE_INT_1U << 30);
+ /* For Intel/Motorola extended format, expand to:
+ (ex & mask) == mask && ((hi ^ bit) | ((lo | -lo) >> 31)) > val. */
+ rtx nlo = expand_unop (imode, neg_optab, lo, NULL_RTX, 0);
+ lo = expand_binop (imode, ior_optab, lo, nlo,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ lo = expand_shift (RSHIFT_EXPR, imode, lo, 31, NULL_RTX, 1);
+ temp = expand_binop (imode, xor_optab, hi, bit,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = expand_binop (imode, ior_optab, temp, lo,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = emit_store_flag_force (target, GTU, temp, val, imode, 1, 1);
+ ex = expand_binop (iemode, and_optab, ex, mask,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ ex = emit_store_flag_force (gen_reg_rtx (GET_MODE (temp)), EQ,
+ ex, mask, iemode, 1, 1);
+ temp = expand_binop (GET_MODE (temp), and_optab, temp, ex,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ }
+ break;
+ default:
+ gcc_unreachable ();
+ }
+
+ return temp;
+}
+
/* Expand a call to one of the builtin rounding functions gcc defines
as an extension (lfloor and lceil). As these are gcc extensions we
do not need to worry about setting errno to EDOM.
@@ -5508,9 +5797,9 @@ expand_builtin_signbit (tree exp, rtx ta
if (icode != CODE_FOR_nothing)
{
rtx_insn *last = get_last_insn ();
- target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
- if (maybe_emit_unop_insn (icode, target, temp, UNKNOWN))
- return target;
+ rtx this_target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
+ if (maybe_emit_unop_insn (icode, this_target, temp, UNKNOWN))
+ return this_target;
delete_insns_since (last);
}
@@ -7120,6 +7409,12 @@ expand_builtin (tree exp, rtx target, rt
return target;
break;
+ case BUILT_IN_ISSIGNALING:
+ target = expand_builtin_issignaling (exp, target);
+ if (target)
+ return target;
+ break;
+
CASE_FLT_FN (BUILT_IN_ICEIL):
CASE_FLT_FN (BUILT_IN_LCEIL):
CASE_FLT_FN (BUILT_IN_LLCEIL):
@@ -8963,6 +9258,11 @@ fold_builtin_classify (location_t loc, t
arg = builtin_save_expr (arg);
return fold_build2_loc (loc, UNORDERED_EXPR, type, arg, arg);
+ case BUILT_IN_ISSIGNALING:
+ if (!tree_expr_maybe_nan_p (arg))
+ return omit_one_operand_loc (loc, type, integer_zero_node, arg);
+ return NULL_TREE;
+
default:
gcc_unreachable ();
}
@@ -9399,6 +9699,9 @@ fold_builtin_1 (location_t loc, tree exp
case BUILT_IN_ISNAND128:
return fold_builtin_classify (loc, fndecl, arg0, BUILT_IN_ISNAN);
+ case BUILT_IN_ISSIGNALING:
+ return fold_builtin_classify (loc, fndecl, arg0, BUILT_IN_ISSIGNALING);
+
case BUILT_IN_FREE:
if (integer_zerop (arg0))
return build_empty_stmt (loc);
@@ -313,6 +313,7 @@ OPTAB_D (fmod_optab, "fmod$a3")
OPTAB_D (hypot_optab, "hypot$a3")
OPTAB_D (ilogb_optab, "ilogb$a2")
OPTAB_D (isinf_optab, "isinf$a2")
+OPTAB_D (issignaling_optab, "issignaling$a2")
OPTAB_D (ldexp_optab, "ldexp$a3")
OPTAB_D (log10_optab, "log10$a2")
OPTAB_D (log1p_optab, "log1p$a2")
@@ -952,6 +952,10 @@ fold_const_call_ss (wide_int *result, co
*result = wi::shwi (real_isfinite (arg) ? 1 : 0, precision);
return true;
+ case CFN_BUILT_IN_ISSIGNALING:
+ *result = wi::shwi (real_issignaling_nan (arg) ? 1 : 0, precision);
+ return true;
+
CASE_CFN_ISINF:
case CFN_BUILT_IN_ISINFD32:
case CFN_BUILT_IN_ISINFD64:
@@ -24720,6 +24720,58 @@ (define_expand "spaceshipxf3"
DONE;
})
+;; Defined because the generic expand_builtin_issignaling for XFmode
+;; only tests for sNaNs, but i387 treats also pseudo numbers as always
+;; signaling.
+(define_expand "issignalingxf2"
+ [(match_operand:SI 0 "register_operand")
+ (match_operand:XF 1 "general_operand")]
+ ""
+{
+ rtx temp = operands[1];
+ if (!MEM_P (temp))
+ {
+ rtx mem = assign_stack_temp (XFmode, GET_MODE_SIZE (XFmode));
+ emit_move_insn (mem, temp);
+ temp = mem;
+ }
+ rtx ex = adjust_address (temp, HImode, 8);
+ rtx hi = adjust_address (temp, SImode, 4);
+ rtx lo = adjust_address (temp, SImode, 0);
+ rtx val = GEN_INT (HOST_WIDE_INT_M1U << 30);
+ rtx mask = GEN_INT (0x7fff);
+ rtx bit = GEN_INT (HOST_WIDE_INT_1U << 30);
+ /* Expand to:
+ ((ex & mask) && (int) hi >= 0)
+ || ((ex & mask) == mask && ((hi ^ bit) | ((lo | -lo) >> 31)) > val). */
+ rtx nlo = expand_unop (SImode, neg_optab, lo, NULL_RTX, 0);
+ lo = expand_binop (SImode, ior_optab, lo, nlo,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ lo = expand_shift (RSHIFT_EXPR, SImode, lo, 31, NULL_RTX, 1);
+ temp = expand_binop (SImode, xor_optab, hi, bit,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = expand_binop (SImode, ior_optab, temp, lo,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = emit_store_flag_force (gen_reg_rtx (SImode), GTU, temp, val,
+ SImode, 1, 1);
+ ex = expand_binop (HImode, and_optab, ex, mask,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ rtx temp2 = emit_store_flag_force (gen_reg_rtx (SImode), NE,
+ ex, const0_rtx, SImode, 1, 1);
+ ex = emit_store_flag_force (gen_reg_rtx (SImode), EQ,
+ ex, mask, HImode, 1, 1);
+ temp = expand_binop (SImode, and_optab, temp, ex,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ rtx temp3 = emit_store_flag_force (gen_reg_rtx (SImode), GE,
+ hi, const0_rtx, SImode, 0, 1);
+ temp2 = expand_binop (SImode, and_optab, temp2, temp3,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = expand_binop (SImode, ior_optab, temp, temp2,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ emit_move_insn (operands[0], temp);
+ DONE;
+})
+
(include "mmx.md")
(include "sse.md")
(include "sync.md")
@@ -13001,6 +13001,7 @@ is called and the @var{flag} argument pa
@findex __builtin_isless
@findex __builtin_islessequal
@findex __builtin_islessgreater
+@findex __builtin_issignaling
@findex __builtin_isunordered
@findex __builtin_object_size
@findex __builtin_powi
@@ -14489,6 +14490,14 @@ Similar to @code{__builtin_nans}, except
@code{_Float@var{n}x}.
@end deftypefn
+@deftypefn {Built-in Function} int __builtin_issignaling (...)
+Return non-zero if the argument is a signaling NaN and zero otherwise.
+Note while the parameter list is an
+ellipsis, this function only accepts exactly one floating-point
+argument. GCC treats this parameter as type-generic, which means it
+does not do default promotion from float to double.
+@end deftypefn
+
@deftypefn {Built-in Function} int __builtin_ffs (int x)
Returns one plus the index of the least significant 1-bit of @var{x}, or
if @var{x} is zero, returns zero.
@@ -6184,6 +6184,10 @@ floating-point mode.
This pattern is not allowed to @code{FAIL}.
+@cindex @code{issignaling@var{m}2} instruction pattern
+@item @samp{issignaling@var{m}2}
+Set operand 0 to 1 if operand 1 is a signaling NaN and to 0 otherwise.
+
@cindex @code{cadd90@var{m}3} instruction pattern
@item @samp{cadd90@var{m}3}
Perform vector add and subtract on even/odd number pairs. The operation being
@@ -6294,6 +6294,7 @@ check_builtin_function_arguments (locati
case BUILT_IN_ISINF_SIGN:
case BUILT_IN_ISNAN:
case BUILT_IN_ISNORMAL:
+ case BUILT_IN_ISSIGNALING:
case BUILT_IN_SIGNBIT:
if (builtin_function_validate_nargs (loc, fndecl, nargs, 1))
{
@@ -3546,6 +3546,7 @@ convert_arguments (location_t loc, vec<l
case BUILT_IN_ISINF_SIGN:
case BUILT_IN_ISNAN:
case BUILT_IN_ISNORMAL:
+ case BUILT_IN_ISSIGNALING:
case BUILT_IN_FPCLASSIFY:
type_generic_remove_excess_precision = true;
break;
@@ -1013,6 +1013,8 @@ gfc_init_builtin_functions (void)
"__builtin_isnan", ATTR_CONST_NOTHROW_LEAF_LIST);
gfc_define_builtin ("__builtin_isnormal", ftype, BUILT_IN_ISNORMAL,
"__builtin_isnormal", ATTR_CONST_NOTHROW_LEAF_LIST);
+ gfc_define_builtin ("__builtin_issignaling", ftype, BUILT_IN_ISSIGNALING,
+ "__builtin_issignaling", ATTR_CONST_NOTHROW_LEAF_LIST);
gfc_define_builtin ("__builtin_signbit", ftype, BUILT_IN_SIGNBIT,
"__builtin_signbit", ATTR_CONST_NOTHROW_LEAF_LIST);
@@ -0,0 +1,95 @@
+/* { dg-do run } */
+/* { dg-add-options ieee } */
+/* { dg-additional-options "-fsignaling-nans" } */
+/* Workaround for PR57484 on ia32: */
+/* { dg-additional-options "-msse2 -mfpmath=sse" { target { ia32 && sse2_runtime } } } */
+
+int
+f1 (void)
+{
+ return __builtin_issignaling (__builtin_nansf (""));
+}
+
+int
+f2 (void)
+{
+ return __builtin_issignaling (__builtin_nan (""));
+}
+
+int
+f3 (void)
+{
+ return __builtin_issignaling (0.0L);
+}
+
+int
+f4 (float x)
+{
+ return __builtin_issignaling (x);
+}
+
+int
+f5 (double x)
+{
+ return __builtin_issignaling (x);
+}
+
+int
+f6 (long double x)
+{
+ return __builtin_issignaling (x);
+}
+
+#ifdef __SIZEOF_FLOAT128__
+int
+f7 (_Float128 x)
+{
+ return __builtin_issignaling (x);
+}
+#endif
+
+float x;
+double y;
+long double z;
+#ifdef __SIZEOF_FLOAT128__
+_Float128 w;
+#endif
+
+int
+main ()
+{
+ if (!f1 () || f2 () || f3 ())
+ __builtin_abort ();
+ asm volatile ("" : : : "memory");
+ if (f4 (x) || !f4 (__builtin_nansf ("0x123")) || f4 (42.0f) || f4 (__builtin_nanf ("0x234"))
+ || f4 (__builtin_inff ()) || f4 (-__builtin_inff ()) || f4 (-42.0f) || f4 (-0.0f) || f4 (0.0f))
+ __builtin_abort ();
+ x = __builtin_nansf ("");
+ asm volatile ("" : : : "memory");
+ if (!f4 (x))
+ __builtin_abort ();
+ if (f5 (y) || !f5 (__builtin_nans ("0x123")) || f5 (42.0) || f5 (__builtin_nan ("0x234"))
+ || f5 (__builtin_inf ()) || f5 (-__builtin_inf ()) || f5 (-42.0) || f5 (-0.0) || f5 (0.0))
+ __builtin_abort ();
+ y = __builtin_nans ("");
+ asm volatile ("" : : : "memory");
+ if (!f5 (y))
+ __builtin_abort ();
+ if (f6 (z) || !f6 (__builtin_nansl ("0x123")) || f6 (42.0L) || f6 (__builtin_nanl ("0x234"))
+ || f6 (__builtin_infl ()) || f6 (-__builtin_infl ()) || f6 (-42.0L) || f6 (-0.0L) || f6 (0.0L))
+ __builtin_abort ();
+ z = __builtin_nansl ("");
+ asm volatile ("" : : : "memory");
+ if (!f6 (z))
+ __builtin_abort ();
+#ifdef __SIZEOF_FLOAT128__
+ if (f7 (w) || !f7 (__builtin_nansf128 ("0x123")) || f7 (42.0Q) || f7 (__builtin_nanf128 ("0x234"))
+ || f7 (__builtin_inff128 ()) || f7 (-__builtin_inff128 ()) || f7 (-42.0Q) || f7 (-0.0Q) || f7 (0.0Q))
+ __builtin_abort ();
+ w = __builtin_nansf128 ("");
+ asm volatile ("" : : : "memory");
+ if (!f7 (w))
+ __builtin_abort ();
+#endif
+ return 0;
+}
@@ -0,0 +1,73 @@
+/* { dg-do run } */
+/* { dg-require-effective-target dfp } */
+/* { dg-additional-options "-fsignaling-nans" } */
+
+int
+f1 (void)
+{
+ return __builtin_issignaling (__builtin_nansd32 (""));
+}
+
+int
+f2 (void)
+{
+ return __builtin_issignaling (__builtin_nand64 (""));
+}
+
+int
+f3 (void)
+{
+ return __builtin_issignaling (0.0DD);
+}
+
+int
+f4 (_Decimal32 x)
+{
+ return __builtin_issignaling (x);
+}
+
+int
+f5 (_Decimal64 x)
+{
+ return __builtin_issignaling (x);
+}
+
+int
+f6 (_Decimal128 x)
+{
+ return __builtin_issignaling (x);
+}
+
+_Decimal32 x;
+_Decimal64 y;
+_Decimal128 z;
+
+int
+main ()
+{
+ if (!f1 () || f2 () || f3 ())
+ __builtin_abort ();
+ asm volatile ("" : : : "memory");
+ if (f4 (x) || !f4 (__builtin_nansd32 ("0x123")) || f4 (42.0DF) || f4 (__builtin_nand32 ("0x234"))
+ || f4 (__builtin_infd32 ()) || f4 (-__builtin_infd32 ()) || f4 (-42.0DF) || f4 (-0.0DF) || f4 (0.0DF))
+ __builtin_abort ();
+ x = __builtin_nansd32 ("");
+ asm volatile ("" : : : "memory");
+ if (!f4 (x))
+ __builtin_abort ();
+ if (f5 (y) || !f5 (__builtin_nansd64 ("0x123")) || f5 (42.0DD) || f5 (__builtin_nand64 ("0x234"))
+ || f5 (__builtin_infd64 ()) || f5 (-__builtin_infd64 ()) || f5 (-42.0DD) || f5 (-0.0DD) || f5 (0.0DD))
+ __builtin_abort ();
+ y = __builtin_nansd64 ("");
+ asm volatile ("" : : : "memory");
+ if (!f5 (y))
+ __builtin_abort ();
+ if (f6 (z) || !f6 (__builtin_nansd128 ("0x123")) || f6 (42.0DL) || f6 (__builtin_nand128 ("0x234"))
+ || f6 (__builtin_infd128 ()) || f6 (-__builtin_infd128 ()) || f6 (-42.0DL) || f6 (-0.0DL) || f6 (0.0DL))
+ __builtin_abort ();
+ z = __builtin_nansd128 ("");
+ asm volatile ("" : : : "memory");
+ if (!f6 (z))
+ __builtin_abort ();
+ return 0;
+}
@@ -0,0 +1,80 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -fsignaling-nans" } */
+
+#if __LDBL_MANT_DIG__ == 64
+union U { struct { unsigned long long m; unsigned short e; } p; long double l; };
+union U zero = { { 0, 0 } };
+union U mzero = { { 0, 0x8000 } };
+union U denorm = { { 42, 0 } };
+union U mdenorm = { { 42, 0x8000 } };
+union U pseudodenorm = { { 0x8000000000000000ULL, 0 } };
+union U mpseudodenorm = { { 0x8000000000000000ULL, 0x8000 } };
+union U pseudodenorm1 = { { 0x8000000000000042ULL, 0 } };
+union U mpseudodenorm1 = { { 0x8000000000000042ULL, 0x8000 } };
+union U pseudoinf = { { 0, 0x7fff } };
+union U mpseudoinf = { { 0, 0xffff } };
+union U pseudonan = { { 42, 0x7fff } };
+union U mpseudonan = { { 42, 0xffff } };
+union U pseudonan1 = { { 0x4000000000000000ULL, 0x7fff } };
+union U mpseudonan1 = { { 0x4000000000000000ULL, 0xffff } };
+union U pseudonan2 = { { 0x4000000000000042ULL, 0x7fff } };
+union U mpseudonan2 = { { 0x4000000000000042ULL, 0xffff } };
+union U inf = { { 0x8000000000000000ULL, 0x7fff } };
+union U minf = { { 0x8000000000000000ULL, 0xffff } };
+union U snan = { { 0x8000000000000042ULL, 0x7fff } };
+union U msnan = { { 0x8000000000000042ULL, 0xffff } };
+union U indefinite = { { 0xc000000000000000ULL, 0x7fff } };
+union U mindefinite = { { 0xc000000000000000ULL, 0xffff } };
+union U qnan = { { 0xc000000000000042ULL, 0x7fff } };
+union U mqnan = { { 0xc000000000000042ULL, 0xffff } };
+union U unnormal = { { 0, 0x42 } };
+union U munnormal = { { 0, 0x8042 } };
+union U unnormal1 = { { 42, 0x42 } };
+union U munnormal1 = { { 42, 0x8042 } };
+union U normal = { { 0x8000000000000000ULL, 0x42 } };
+union U mnormal = { { 0x8000000000000000ULL, 0x8042 } };
+union U normal1 = { { 0x8000000000000042ULL, 0x42 } };
+union U mnormal1 = { { 0x8000000000000042ULL, 0x8042 } };
+#endif
+
+int
+main ()
+{
+#if __LDBL_MANT_DIG__ == 64
+ asm volatile ("" : : : "memory");
+ if (__builtin_issignaling (zero.l)
+ || __builtin_issignaling (mzero.l)
+ || __builtin_issignaling (denorm.l)
+ || __builtin_issignaling (mdenorm.l)
+ || __builtin_issignaling (pseudodenorm.l)
+ || __builtin_issignaling (mpseudodenorm.l)
+ || __builtin_issignaling (pseudodenorm1.l)
+ || __builtin_issignaling (mpseudodenorm1.l)
+ || !__builtin_issignaling (pseudoinf.l)
+ || !__builtin_issignaling (mpseudoinf.l)
+ || !__builtin_issignaling (pseudonan.l)
+ || !__builtin_issignaling (mpseudonan.l)
+ || !__builtin_issignaling (pseudonan1.l)
+ || !__builtin_issignaling (mpseudonan1.l)
+ || !__builtin_issignaling (pseudonan2.l)
+ || !__builtin_issignaling (mpseudonan2.l)
+ || __builtin_issignaling (inf.l)
+ || __builtin_issignaling (minf.l)
+ || !__builtin_issignaling (snan.l)
+ || !__builtin_issignaling (msnan.l)
+ || __builtin_issignaling (indefinite.l)
+ || __builtin_issignaling (mindefinite.l)
+ || __builtin_issignaling (qnan.l)
+ || __builtin_issignaling (mqnan.l)
+ || !__builtin_issignaling (unnormal.l)
+ || !__builtin_issignaling (munnormal.l)
+ || !__builtin_issignaling (unnormal1.l)
+ || !__builtin_issignaling (munnormal1.l)
+ || __builtin_issignaling (normal.l)
+ || __builtin_issignaling (mnormal.l)
+ || __builtin_issignaling (normal1.l)
+ || __builtin_issignaling (mnormal1.l))
+ __builtin_abort ();
+#endif
+ return 0;
+}