diff mbox series

Fortran: error recovery while simplifying expressions [PR103707,PR106987]

Message ID trinity-168a3a2a-979c-498c-acae-4e12dc114be7-1709673831988@3c-app-gmx-bs32
State New
Headers show
Series Fortran: error recovery while simplifying expressions [PR103707,PR106987] | expand

Commit Message

Harald Anlauf March 5, 2024, 9:23 p.m. UTC
Dear all,

error recovery on arithmetic errors during simplification has bugged
me for a long time, especially since the occurence of ICEs depended
on whether -frange-check is specified or not, whether array ctors
were involved, etc.

I've now come up with the attached patch that classifies the arithmetic
result codes into "hard" and "soft" errors.

A "soft" error means that it is an overflow or other exception (e.g. NaN)
that is ignored with -fno-range-check.  After the patch, a soft error
will not stop simplification (a hard one will), and error status will be
passed along.

I took this opportunity to change the emitted error for division by zero
for real and complex division dependent on whether the numerator is
regular or not.  This makes e.g. (0.)/0 a NaN and now says so, in
accordance with some other brands.

Regtested on x86_64-pc-linux-gnu.  OK for mainline?

Other comments?

Thanks,
Harald

Comments

Paul Richard Thomas March 6, 2024, 10:51 a.m. UTC | #1
Hi Harald,

This all looks good to me. OK for mainline and, according to intestinal
fortitude on your part, earlier branches.

Thanks

Paul


On Tue, 5 Mar 2024 at 21:24, Harald Anlauf <anlauf@gmx.de> wrote:

> Dear all,
>
> error recovery on arithmetic errors during simplification has bugged
> me for a long time, especially since the occurence of ICEs depended
> on whether -frange-check is specified or not, whether array ctors
> were involved, etc.
>
> I've now come up with the attached patch that classifies the arithmetic
> result codes into "hard" and "soft" errors.
>
> A "soft" error means that it is an overflow or other exception (e.g. NaN)
> that is ignored with -fno-range-check.  After the patch, a soft error
> will not stop simplification (a hard one will), and error status will be
> passed along.
>
> I took this opportunity to change the emitted error for division by zero
> for real and complex division dependent on whether the numerator is
> regular or not.  This makes e.g. (0.)/0 a NaN and now says so, in
> accordance with some other brands.
>
> Regtested on x86_64-pc-linux-gnu.  OK for mainline?
>
> Other comments?
>
> Thanks,
> Harald
>
>
Harald Anlauf March 6, 2024, 5:09 p.m. UTC | #2
Hi Paul,

thanks for reviewing the patch, and your trust in me :-)

Backporting to 13-branch seems easily feasible (needs another small
queued backport on which this patch depends), but going further is
definitely out of the question...  Will wait a couple of weeks though.

Harald

On 3/6/24 11:51, Paul Richard Thomas wrote:
> Hi Harald,
>
> This all looks good to me. OK for mainline and, according to intestinal
> fortitude on your part, earlier branches.
>
> Thanks
>
> Paul
>
>
> On Tue, 5 Mar 2024 at 21:24, Harald Anlauf <anlauf@gmx.de> wrote:
>
>> Dear all,
>>
>> error recovery on arithmetic errors during simplification has bugged
>> me for a long time, especially since the occurence of ICEs depended
>> on whether -frange-check is specified or not, whether array ctors
>> were involved, etc.
>>
>> I've now come up with the attached patch that classifies the arithmetic
>> result codes into "hard" and "soft" errors.
>>
>> A "soft" error means that it is an overflow or other exception (e.g. NaN)
>> that is ignored with -fno-range-check.  After the patch, a soft error
>> will not stop simplification (a hard one will), and error status will be
>> passed along.
>>
>> I took this opportunity to change the emitted error for division by zero
>> for real and complex division dependent on whether the numerator is
>> regular or not.  This makes e.g. (0.)/0 a NaN and now says so, in
>> accordance with some other brands.
>>
>> Regtested on x86_64-pc-linux-gnu.  OK for mainline?
>>
>> Other comments?
>>
>> Thanks,
>> Harald
>>
>>
>
diff mbox series

Patch

From d9b87bea6af77fbc794e1f21cfecb0468c68cb72 Mon Sep 17 00:00:00 2001
From: Harald Anlauf <anlauf@gmx.de>
Date: Tue, 5 Mar 2024 21:54:26 +0100
Subject: [PATCH] Fortran: error recovery while simplifying expressions
 [PR103707,PR106987]

When an exception is encountered during simplification of arithmetic
expressions, the result may depend on whether range-checking is active
(-frange-check) or not.  However, the code path in the front-end should
stay the same for "soft" errors for which the exception is triggered by the
check, while "hard" errors should always terminate the simplification, so
that error recovery is independent of the flag.  Separation of arithmetic
error codes into "hard" and "soft" errors shall be done consistently via
is_hard_arith_error().

	PR fortran/103707
	PR fortran/106987

gcc/fortran/ChangeLog:

	* arith.cc (is_hard_arith_error): New helper function to determine
	whether an arithmetic error is "hard" or not.
	(check_result): Use it.
	(gfc_arith_divide): Set "Division by zero" only for regular
	numerators of real and complex divisions.
	(reduce_unary): Use is_hard_arith_error to determine whether a hard
	or (recoverable) soft error was encountered.  Terminate immediately
	on hard error, otherwise remember code of first soft error.
	(reduce_binary_ac): Likewise.
	(reduce_binary_ca): Likewise.
	(reduce_binary_aa): Likewise.

gcc/testsuite/ChangeLog:

	* gfortran.dg/pr99350.f90:
	* gfortran.dg/arithmetic_overflow_3.f90: New test.
---
 gcc/fortran/arith.cc                          | 134 ++++++++++++------
 .../gfortran.dg/arithmetic_overflow_3.f90     |  48 +++++++
 gcc/testsuite/gfortran.dg/pr99350.f90         |   2 +-
 3 files changed, 143 insertions(+), 41 deletions(-)
 create mode 100644 gcc/testsuite/gfortran.dg/arithmetic_overflow_3.f90

diff --git a/gcc/fortran/arith.cc b/gcc/fortran/arith.cc
index d17d1aaa1d9..b373c25e5e1 100644
--- a/gcc/fortran/arith.cc
+++ b/gcc/fortran/arith.cc
@@ -130,6 +130,30 @@  gfc_arith_error (arith code)
 }


+/* Check if a certain arithmetic error code is severe enough to prevent
+   further simplification, as opposed to errors thrown by the range check
+   (e.g. overflow) or arithmetic exceptions that are tolerated with
+   -fno-range-check.  */
+
+static bool
+is_hard_arith_error (arith code)
+{
+  switch (code)
+    {
+    case ARITH_OK:
+    case ARITH_OVERFLOW:
+    case ARITH_UNDERFLOW:
+    case ARITH_NAN:
+    case ARITH_DIV0:
+    case ARITH_ASYMMETRIC:
+      return false;
+
+    default:
+      return true;
+    }
+}
+
+
 /* Get things ready to do math.  */

 void
@@ -579,10 +603,10 @@  check_result (arith rc, gfc_expr *x, gfc_expr *r, gfc_expr **rp)
       val = ARITH_OK;
     }

-  if (val == ARITH_OK || val == ARITH_OVERFLOW)
-    *rp = r;
-  else
+  if (is_hard_arith_error (val))
     gfc_free_expr (r);
+  else
+    *rp = r;

   return val;
 }
@@ -792,23 +816,26 @@  gfc_arith_divide (gfc_expr *op1, gfc_expr *op2, gfc_expr **resultp)
       break;

     case BT_REAL:
-      if (mpfr_sgn (op2->value.real) == 0 && flag_range_check == 1)
-	{
-	  rc = ARITH_DIV0;
-	  break;
-	}
+      /* Set "Division by zero" only for regular numerator.  */
+      if (flag_range_check == 1
+	  && mpfr_zero_p (op2->value.real)
+	  && mpfr_regular_p (op1->value.real))
+	rc = ARITH_DIV0;

       mpfr_div (result->value.real, op1->value.real, op2->value.real,
 	       GFC_RND_MODE);
       break;

     case BT_COMPLEX:
-      if (mpc_cmp_si_si (op2->value.complex, 0, 0) == 0
-	  && flag_range_check == 1)
-	{
-	  rc = ARITH_DIV0;
-	  break;
-	}
+      /* Set "Division by zero" only for regular numerator.  */
+      if (flag_range_check == 1
+	  && mpfr_zero_p (mpc_realref (op2->value.complex))
+	  && mpfr_zero_p (mpc_imagref (op2->value.complex))
+	  && ((mpfr_regular_p (mpc_realref (op1->value.complex))
+	       && mpfr_number_p (mpc_imagref (op1->value.complex)))
+	      || (mpfr_regular_p (mpc_imagref (op1->value.complex))
+		  && mpfr_number_p (mpc_realref (op1->value.complex)))))
+	rc = ARITH_DIV0;

       gfc_set_model (mpc_realref (op1->value.complex));
       if (mpc_cmp_si_si (op2->value.complex, 0, 0) == 0)
@@ -1323,7 +1350,6 @@  reduce_unary (arith (*eval) (gfc_expr *, gfc_expr **), gfc_expr *op,
   gfc_constructor *c;
   gfc_expr *r;
   arith rc;
-  bool ov = false;

   if (op->expr_type == EXPR_CONSTANT)
     return eval (op, result);
@@ -1335,19 +1361,22 @@  reduce_unary (arith (*eval) (gfc_expr *, gfc_expr **), gfc_expr *op,
   head = gfc_constructor_copy (op->value.constructor);
   for (c = gfc_constructor_first (head); c; c = gfc_constructor_next (c))
     {
-      rc = reduce_unary (eval, c->expr, &r);
+      arith rc_tmp = reduce_unary (eval, c->expr, &r);

-      /* Remember any overflow encountered during reduction and continue,
-	 but terminate on serious errors.  */
-      if (rc == ARITH_OVERFLOW)
-	ov = true;
-      else if (rc != ARITH_OK)
-	break;
+      /* Remember first recoverable ("soft") error encountered during
+	 reduction and continue, but terminate on serious errors.  */
+      if (is_hard_arith_error (rc_tmp))
+	{
+	  rc = rc_tmp;
+	  break;
+	}
+      else if (rc_tmp != ARITH_OK && rc == ARITH_OK)
+	rc = rc_tmp;

       gfc_replace_expr (c->expr, r);
     }

-  if (rc != ARITH_OK && rc != ARITH_OVERFLOW)
+  if (is_hard_arith_error (rc))
     gfc_constructor_free (head);
   else
     {
@@ -1368,7 +1397,7 @@  reduce_unary (arith (*eval) (gfc_expr *, gfc_expr **), gfc_expr *op,
       *result = r;
     }

-  return ov ? ARITH_OVERFLOW : rc;
+  return rc;
 }


@@ -1384,22 +1413,31 @@  reduce_binary_ac (arith (*eval) (gfc_expr *, gfc_expr *, gfc_expr **),
   head = gfc_constructor_copy (op1->value.constructor);
   for (c = gfc_constructor_first (head); c; c = gfc_constructor_next (c))
     {
+      arith rc_tmp;
+
       gfc_simplify_expr (c->expr, 0);

       if (c->expr->expr_type == EXPR_CONSTANT)
-        rc = eval (c->expr, op2, &r);
+	rc_tmp = eval (c->expr, op2, &r);
       else if (c->expr->expr_type != EXPR_ARRAY)
-	rc = ARITH_NOT_REDUCED;
+	rc_tmp = ARITH_NOT_REDUCED;
       else
-	rc = reduce_binary_ac (eval, c->expr, op2, &r);
+	rc_tmp = reduce_binary_ac (eval, c->expr, op2, &r);

-      if (rc != ARITH_OK)
-	break;
+      /* Remember first recoverable ("soft") error encountered during
+	 reduction and continue, but terminate on serious errors.  */
+      if (is_hard_arith_error (rc_tmp))
+	{
+	  rc = rc_tmp;
+	  break;
+	}
+      else if (rc_tmp != ARITH_OK && rc == ARITH_OK)
+	rc = rc_tmp;

       gfc_replace_expr (c->expr, r);
     }

-  if (rc != ARITH_OK)
+  if (is_hard_arith_error (rc))
     gfc_constructor_free (head);
   else
     {
@@ -1438,22 +1476,31 @@  reduce_binary_ca (arith (*eval) (gfc_expr *, gfc_expr *, gfc_expr **),
   head = gfc_constructor_copy (op2->value.constructor);
   for (c = gfc_constructor_first (head); c; c = gfc_constructor_next (c))
     {
+      arith rc_tmp;
+
       gfc_simplify_expr (c->expr, 0);

       if (c->expr->expr_type == EXPR_CONSTANT)
-	rc = eval (op1, c->expr, &r);
+	rc_tmp = eval (op1, c->expr, &r);
       else if (c->expr->expr_type != EXPR_ARRAY)
-	rc = ARITH_NOT_REDUCED;
+	rc_tmp = ARITH_NOT_REDUCED;
       else
-	rc = reduce_binary_ca (eval, op1, c->expr, &r);
+	rc_tmp = reduce_binary_ca (eval, op1, c->expr, &r);

-      if (rc != ARITH_OK)
-	break;
+      /* Remember first recoverable ("soft") error encountered during
+	 reduction and continue, but terminate on serious errors.  */
+      if (is_hard_arith_error (rc_tmp))
+	{
+	  rc = rc_tmp;
+	  break;
+	}
+      else if (rc_tmp != ARITH_OK && rc == ARITH_OK)
+	rc = rc_tmp;

       gfc_replace_expr (c->expr, r);
     }

-  if (rc != ARITH_OK)
+  if (is_hard_arith_error (rc))
     gfc_constructor_free (head);
   else
     {
@@ -1503,10 +1550,17 @@  reduce_binary_aa (arith (*eval) (gfc_expr *, gfc_expr *, gfc_expr **),
        c && d;
        c = gfc_constructor_next (c), d = gfc_constructor_next (d))
     {
-      rc = reduce_binary (eval, c->expr, d->expr, &r);
+      arith rc_tmp = reduce_binary (eval, c->expr, d->expr, &r);

-      if (rc != ARITH_OK)
-	break;
+      /* Remember first recoverable ("soft") error encountered during
+	 reduction and continue, but terminate on serious errors.  */
+      if (is_hard_arith_error (rc_tmp))
+	{
+	  rc = rc_tmp;
+	  break;
+	}
+      else if (rc_tmp != ARITH_OK && rc == ARITH_OK)
+	rc = rc_tmp;

       gfc_replace_expr (c->expr, r);
     }
@@ -1514,7 +1568,7 @@  reduce_binary_aa (arith (*eval) (gfc_expr *, gfc_expr *, gfc_expr **),
   if (rc == ARITH_OK && (c || d))
     rc = ARITH_INCOMMENSURATE;

-  if (rc != ARITH_OK)
+  if (is_hard_arith_error (rc))
     gfc_constructor_free (head);
   else
     {
diff --git a/gcc/testsuite/gfortran.dg/arithmetic_overflow_3.f90 b/gcc/testsuite/gfortran.dg/arithmetic_overflow_3.f90
new file mode 100644
index 00000000000..4dc552742a3
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/arithmetic_overflow_3.f90
@@ -0,0 +1,48 @@ 
+! { dg-do compile }
+! { dg-additional-options "-frange-check" }
+!
+! PR fortran/103707
+! PR fortran/106987
+!
+! Check error recovery on arithmetic exceptions
+
+program p
+  implicit none
+  integer, parameter :: a(3) = [30,31,32]
+  integer, parameter :: e(1) = 2
+  print *, 2 ** a       ! { dg-error "Arithmetic overflow" }
+  print *, e ** 31      ! { dg-error "Arithmetic overflow" }
+end
+
+! { dg-prune-output "Result of exponentiation" }
+
+subroutine s
+  implicit none
+  real, parameter :: inf = real (z'7F800000')
+  real, parameter :: nan = real (z'7FC00000')
+
+  ! Unary operators
+  print *, -[inf,nan]           ! { dg-error "Arithmetic overflow" }
+  print *, -[nan,inf]           ! { dg-error "Arithmetic NaN" }
+
+  ! Binary operators
+  print *, [1.]/[0.]            ! { dg-error "Division by zero" }
+  print *, [0.]/[0.]            ! { dg-error "Arithmetic NaN" }
+  print *, 0. / [(0.,0.)]       ! { dg-error "Arithmetic NaN" }
+  print *, [1.,0.]/[0.,0.]      ! { dg-error "Division by zero" }
+  print *, [(1.,1.)]/[0.]       ! { dg-error "Division by zero" }
+  print *, [(1.,0.)]/[0.]       ! { dg-error "Division by zero" }
+  print *, [(0.,0.)]/[0.]       ! { dg-error "Arithmetic NaN" }
+  print *, - [1./0.]/[0.]       ! { dg-error "Division by zero" }
+  print *, - [ 1/0 ] * 1        ! { dg-error "Division by zero" }
+
+  ! Binary operators, exceptional input
+  print *, 1. / nan             ! { dg-error "Arithmetic NaN" }
+  print *, [inf] / inf          ! { dg-error "Arithmetic NaN" }
+  print *, inf + [nan]          ! { dg-error "Arithmetic NaN" }
+  print *, [(1.,0.)]/[(nan,0.)] ! { dg-error "Arithmetic NaN" }
+  print *, [(1.,0.)]/[(0.,nan)] ! { dg-error "Arithmetic NaN" }
+  print *, [(1.,0.)]/[(inf,0.)] ! OK
+  print *, [nan,inf] / (0.)     ! { dg-error "Arithmetic NaN" }
+  print *, [inf,nan] / (0.)     ! { dg-error "Arithmetic overflow" }
+end
diff --git a/gcc/testsuite/gfortran.dg/pr99350.f90 b/gcc/testsuite/gfortran.dg/pr99350.f90
index 7f751b9fdcc..ec198810f1c 100644
--- a/gcc/testsuite/gfortran.dg/pr99350.f90
+++ b/gcc/testsuite/gfortran.dg/pr99350.f90
@@ -7,7 +7,7 @@  program p
       character(:), pointer :: a
    end type
    type(t) :: z
-   character((0.)/0), target :: c = 'abc' ! { dg-error "Division by zero" }
+   character((0.)/0), target :: c = 'abc' ! { dg-error "Arithmetic NaN" }
    z%a => c
 ! The associate statement was not needed to trigger the ICE.
    associate (y => z%a)
--
2.35.3