Message ID | 8b49a9906a0d1019bd877bf526f71ab5321550fe.camel@mengyan1223.wang |
---|---|
State | New |
Headers | show |
Series | fold-const: do not fold 'inf/inf' with -ftrapping-math [PR95115] | expand |
On Sun, Jan 30, 2022 at 7:39 PM Xi Ruoyao via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > 'inf/inf' should raise an invalid operation exception at runtime. So it > should not be folded during compilation unless -fno-trapping-math is > used. I wonder if it would make sense to handle it similar to inexact = real_arithmetic (&value, code, &d1, &d2); real_convert (&result, mode, &value); /* Don't constant fold this floating point operation if the result has overflowed and flag_trapping_math. */ if (flag_trapping_math && MODE_HAS_INFINITIES (mode) && REAL_VALUE_ISINF (result) && !REAL_VALUE_ISINF (d1) && !REAL_VALUE_ISINF (d2)) return NULL_TREE; thus whenever we fold to NaN but flag_trapping_math is set, do not fold? We already exclude NaN operands earlier (but unconditionally return NaN even for NaN + 1. though that should raise invalid as well, no?) I suppose Joseph should know best what the correct semantics are here. Thanks, Richard. > gcc/ > PR middle-end/95115 > * fold-const.cc (const_binop): Do not fold "inf/inf". > > gcc/testsuite > * gcc.dg/pr95115.c: New test. > --- > gcc/fold-const.cc | 8 ++++++++ > gcc/testsuite/gcc.dg/pr95115.c | 25 +++++++++++++++++++++++++ > 2 files changed, 33 insertions(+) > create mode 100644 gcc/testsuite/gcc.dg/pr95115.c > > diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc > index fd9c6352d4f..0e99d5a2e3d 100644 > --- a/gcc/fold-const.cc > +++ b/gcc/fold-const.cc > @@ -1283,6 +1283,14 @@ const_binop (enum tree_code code, tree arg1, tree arg2) > && (flag_trapping_math || ! MODE_HAS_INFINITIES (mode))) > return NULL_TREE; > > + /* Don't perform "inf/inf", because it would raise an invalid > + operation exception (IEEE 754 section 7.2 clause d). */ > + if (code == RDIV_EXPR > + && flag_trapping_math > + && REAL_VALUE_ISINF (d1) > + && REAL_VALUE_ISINF (d2)) > + return NULL_TREE; > + > /* If either operand is a NaN, just return it. Otherwise, set up > for floating-point trap; we return an overflow. */ > if (REAL_VALUE_ISNAN (d1)) > diff --git a/gcc/testsuite/gcc.dg/pr95115.c b/gcc/testsuite/gcc.dg/pr95115.c > new file mode 100644 > index 00000000000..46a95dfb698 > --- /dev/null > +++ b/gcc/testsuite/gcc.dg/pr95115.c > @@ -0,0 +1,25 @@ > +/* { dg-do run } */ > +/* { dg-options "-O2 -ftrapping-math" } */ > +/* { dg-add-options ieee } */ > +/* { dg-require-effective-target fenv_exceptions } */ > + > +#include <fenv.h> > +#include <stdlib.h> > + > +double > +x (void) > +{ > + double d = __builtin_inf (); > + return d / d; > +} > + > +int > +main (void) > +{ > + double r = x (); > + if (!__builtin_isnan (r)) > + abort (); > + if (!fetestexcept (FE_INVALID)) > + abort (); > + exit (0); > +} > -- > 2.35.1 > >
On Mon, 2022-01-31 at 10:35 +0100, Richard Biener wrote: > On Sun, Jan 30, 2022 at 7:39 PM Xi Ruoyao via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: > > > > 'inf/inf' should raise an invalid operation exception at runtime. > > So it > > should not be folded during compilation unless -fno-trapping-math is > > used. > > I wonder if it would make sense to handle it similar to > > inexact = real_arithmetic (&value, code, &d1, &d2); > real_convert (&result, mode, &value); > > /* Don't constant fold this floating point operation if > the result has overflowed and flag_trapping_math. */ > if (flag_trapping_math > && MODE_HAS_INFINITIES (mode) > && REAL_VALUE_ISINF (result) > && !REAL_VALUE_ISINF (d1) > && !REAL_VALUE_ISINF (d2)) > return NULL_TREE; > thus whenever we fold to NaN but flag_trapping_math is set, > do not fold? Like this? inexact = real_arithmetic (&value, code, &d1, &d2); real_convert (&result, mode, &value); + /* Don't constant fold this floating point operation if + both operands are not NaN but the result is NaN, and + flag_trapping_math. Such operations should raise an + invalid operation exception. */ + if (flag_trapping_math + && MODE_HAS_NANS (mode) + && REAL_VALUE_ISNAN (result) + && !REAL_VALUE_ISNAN (d1) + && !REAL_VALUE_ISNAN (d2)) + return NULL_TREE; + /* Don't constant fold this floating point operation if the result has overflowed and flag_trapping_math. */ if (flag_trapping_math > We already exclude NaN operands earlier > (but unconditionally return NaN even for NaN + 1. though > that should raise invalid as well, no?) qNaN + 1 (or even qNaN + qNaN) should produce a qNaN but should not raise any exception. OTOH sNaN + 1 should produce a qNaN and raise an invalid operation exception. For sNaN operands we already ruled out the const folding earlier: > /* Don't perform operation if we honor signaling NaNs and > either operand is a signaling NaN. */ > if (HONOR_SNANS (mode) > && (REAL_VALUE_ISSIGNALING_NAN (d1) > || REAL_VALUE_ISSIGNALING_NAN (d2))) > return NULL_TREE; > I suppose Joseph should know best what the correct semantics > are here. > > Thanks, > Richard. > > > gcc/ > > PR middle-end/95115 > > * fold-const.cc (const_binop): Do not fold "inf/inf". > > > > gcc/testsuite > > * gcc.dg/pr95115.c: New test. > > --- > > gcc/fold-const.cc | 8 ++++++++ > > gcc/testsuite/gcc.dg/pr95115.c | 25 +++++++++++++++++++++++++ > > 2 files changed, 33 insertions(+) > > create mode 100644 gcc/testsuite/gcc.dg/pr95115.c > > > > diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc > > index fd9c6352d4f..0e99d5a2e3d 100644 > > --- a/gcc/fold-const.cc > > +++ b/gcc/fold-const.cc > > @@ -1283,6 +1283,14 @@ const_binop (enum tree_code code, tree arg1, > > tree arg2) > > && (flag_trapping_math || ! MODE_HAS_INFINITIES (mode))) > > return NULL_TREE; > > > > + /* Don't perform "inf/inf", because it would raise an invalid > > + operation exception (IEEE 754 section 7.2 clause d). */ > > + if (code == RDIV_EXPR > > + && flag_trapping_math > > + && REAL_VALUE_ISINF (d1) > > + && REAL_VALUE_ISINF (d2)) > > + return NULL_TREE; > > + > > /* If either operand is a NaN, just return it. Otherwise, > > set up > > for floating-point trap; we return an overflow. */ > > if (REAL_VALUE_ISNAN (d1)) > > diff --git a/gcc/testsuite/gcc.dg/pr95115.c > > b/gcc/testsuite/gcc.dg/pr95115.c > > new file mode 100644 > > index 00000000000..46a95dfb698 > > --- /dev/null > > +++ b/gcc/testsuite/gcc.dg/pr95115.c > > @@ -0,0 +1,25 @@ > > +/* { dg-do run } */ > > +/* { dg-options "-O2 -ftrapping-math" } */ > > +/* { dg-add-options ieee } */ > > +/* { dg-require-effective-target fenv_exceptions } */ > > + > > +#include <fenv.h> > > +#include <stdlib.h> > > + > > +double > > +x (void) > > +{ > > + double d = __builtin_inf (); > > + return d / d; > > +} > > + > > +int > > +main (void) > > +{ > > + double r = x (); > > + if (!__builtin_isnan (r)) > > + abort (); > > + if (!fetestexcept (FE_INVALID)) > > + abort (); > > + exit (0); > > +} > > -- > > 2.35.1 > > > >
On Mon, Jan 31, 2022 at 11:07 AM Xi Ruoyao <xry111@mengyan1223.wang> wrote: > > On Mon, 2022-01-31 at 10:35 +0100, Richard Biener wrote: > > On Sun, Jan 30, 2022 at 7:39 PM Xi Ruoyao via Gcc-patches > > <gcc-patches@gcc.gnu.org> wrote: > > > > > > 'inf/inf' should raise an invalid operation exception at runtime. > > > So it > > > should not be folded during compilation unless -fno-trapping-math is > > > used. > > > > I wonder if it would make sense to handle it similar to > > > > inexact = real_arithmetic (&value, code, &d1, &d2); > > real_convert (&result, mode, &value); > > > > /* Don't constant fold this floating point operation if > > the result has overflowed and flag_trapping_math. */ > > if (flag_trapping_math > > && MODE_HAS_INFINITIES (mode) > > && REAL_VALUE_ISINF (result) > > && !REAL_VALUE_ISINF (d1) > > && !REAL_VALUE_ISINF (d2)) > > return NULL_TREE; > > > thus whenever we fold to NaN but flag_trapping_math is set, > > do not fold? > > Like this? > > inexact = real_arithmetic (&value, code, &d1, &d2); > real_convert (&result, mode, &value); > > + /* Don't constant fold this floating point operation if > + both operands are not NaN but the result is NaN, and > + flag_trapping_math. Such operations should raise an > + invalid operation exception. */ > + if (flag_trapping_math > + && MODE_HAS_NANS (mode) > + && REAL_VALUE_ISNAN (result) > + && !REAL_VALUE_ISNAN (d1) > + && !REAL_VALUE_ISNAN (d2)) > + return NULL_TREE; > + > /* Don't constant fold this floating point operation if > the result has overflowed and flag_trapping_math. */ > if (flag_trapping_math > > > We already exclude NaN operands earlier > > (but unconditionally return NaN even for NaN + 1. though > > that should raise invalid as well, no?) > > qNaN + 1 (or even qNaN + qNaN) should produce a qNaN but should not > raise any exception. OTOH sNaN + 1 should produce a qNaN and raise an > invalid operation exception. Ah, I see. So yes, like that. Joseph, do you agree? Thanks, Richard. > For sNaN operands we already ruled out the > const folding earlier: > > > /* Don't perform operation if we honor signaling NaNs and > > either operand is a signaling NaN. */ > > if (HONOR_SNANS (mode) > > && (REAL_VALUE_ISSIGNALING_NAN (d1) > > || REAL_VALUE_ISSIGNALING_NAN (d2))) > > return NULL_TREE; > > > I suppose Joseph should know best what the correct semantics > > are here. > > > > Thanks, > > Richard. > > > > > gcc/ > > > PR middle-end/95115 > > > * fold-const.cc (const_binop): Do not fold "inf/inf". > > > > > > gcc/testsuite > > > * gcc.dg/pr95115.c: New test. > > > --- > > > gcc/fold-const.cc | 8 ++++++++ > > > gcc/testsuite/gcc.dg/pr95115.c | 25 +++++++++++++++++++++++++ > > > 2 files changed, 33 insertions(+) > > > create mode 100644 gcc/testsuite/gcc.dg/pr95115.c > > > > > > diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc > > > index fd9c6352d4f..0e99d5a2e3d 100644 > > > --- a/gcc/fold-const.cc > > > +++ b/gcc/fold-const.cc > > > @@ -1283,6 +1283,14 @@ const_binop (enum tree_code code, tree arg1, > > > tree arg2) > > > && (flag_trapping_math || ! MODE_HAS_INFINITIES (mode))) > > > return NULL_TREE; > > > > > > + /* Don't perform "inf/inf", because it would raise an invalid > > > + operation exception (IEEE 754 section 7.2 clause d). */ > > > + if (code == RDIV_EXPR > > > + && flag_trapping_math > > > + && REAL_VALUE_ISINF (d1) > > > + && REAL_VALUE_ISINF (d2)) > > > + return NULL_TREE; > > > + > > > /* If either operand is a NaN, just return it. Otherwise, > > > set up > > > for floating-point trap; we return an overflow. */ > > > if (REAL_VALUE_ISNAN (d1)) > > > diff --git a/gcc/testsuite/gcc.dg/pr95115.c > > > b/gcc/testsuite/gcc.dg/pr95115.c > > > new file mode 100644 > > > index 00000000000..46a95dfb698 > > > --- /dev/null > > > +++ b/gcc/testsuite/gcc.dg/pr95115.c > > > @@ -0,0 +1,25 @@ > > > +/* { dg-do run } */ > > > +/* { dg-options "-O2 -ftrapping-math" } */ > > > +/* { dg-add-options ieee } */ > > > +/* { dg-require-effective-target fenv_exceptions } */ > > > + > > > +#include <fenv.h> > > > +#include <stdlib.h> > > > + > > > +double > > > +x (void) > > > +{ > > > + double d = __builtin_inf (); > > > + return d / d; > > > +} > > > + > > > +int > > > +main (void) > > > +{ > > > + double r = x (); > > > + if (!__builtin_isnan (r)) > > > + abort (); > > > + if (!fetestexcept (FE_INVALID)) > > > + abort (); > > > + exit (0); > > > +} > > > -- > > > 2.35.1 > > > > > > > > -- > Xi Ruoyao <xry111@mengyan1223.wang> > School of Aerospace Science and Technology, Xidian University
On Mon, 31 Jan 2022, Richard Biener via Gcc-patches wrote: > I wonder if it would make sense to handle it similar to > > inexact = real_arithmetic (&value, code, &d1, &d2); > real_convert (&result, mode, &value); > > /* Don't constant fold this floating point operation if > the result has overflowed and flag_trapping_math. */ > if (flag_trapping_math > && MODE_HAS_INFINITIES (mode) > && REAL_VALUE_ISINF (result) > && !REAL_VALUE_ISINF (d1) > && !REAL_VALUE_ISINF (d2)) > return NULL_TREE; > > thus whenever we fold to NaN but flag_trapping_math is set, > do not fold? We already exclude NaN operands earlier Yes, it's true in general that a NaN result for non-NaN operands means "invalid" is raised. While an infinite result for finite operands means either "overflow" (if inexact) or "divide by zero" (if exact). (Underflow is more complicated.)
diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc index fd9c6352d4f..0e99d5a2e3d 100644 --- a/gcc/fold-const.cc +++ b/gcc/fold-const.cc @@ -1283,6 +1283,14 @@ const_binop (enum tree_code code, tree arg1, tree arg2) && (flag_trapping_math || ! MODE_HAS_INFINITIES (mode))) return NULL_TREE; + /* Don't perform "inf/inf", because it would raise an invalid + operation exception (IEEE 754 section 7.2 clause d). */ + if (code == RDIV_EXPR + && flag_trapping_math + && REAL_VALUE_ISINF (d1) + && REAL_VALUE_ISINF (d2)) + return NULL_TREE; + /* If either operand is a NaN, just return it. Otherwise, set up for floating-point trap; we return an overflow. */ if (REAL_VALUE_ISNAN (d1)) diff --git a/gcc/testsuite/gcc.dg/pr95115.c b/gcc/testsuite/gcc.dg/pr95115.c new file mode 100644 index 00000000000..46a95dfb698 --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr95115.c @@ -0,0 +1,25 @@ +/* { dg-do run } */ +/* { dg-options "-O2 -ftrapping-math" } */ +/* { dg-add-options ieee } */ +/* { dg-require-effective-target fenv_exceptions } */ + +#include <fenv.h> +#include <stdlib.h> + +double +x (void) +{ + double d = __builtin_inf (); + return d / d; +} + +int +main (void) +{ + double r = x (); + if (!__builtin_isnan (r)) + abort (); + if (!fetestexcept (FE_INVALID)) + abort (); + exit (0); +}