[Bug fortran/50514] New: gfortran should check ISHFT & ISHFTC aruments (r178939)

[Bug fortran/50514] New: gfortran should check ISHFT & ISHFTC aruments (r178939)

[zeccav at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50514

             Bug #: 50514
           Summary: gfortran should check ISHFT & ISHFTC aruments
                    (r178939)
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: minor
          Priority: P3
         Component: fortran
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zeccav@gmail.com


! gfortran should check ISHFT & ISHFTC aruments (r178939)
! gfortran should not accept SHIFT>BIT_SIZE(I)
      print *,ishft(I=m,SHIFT=640)
      print *,ishftc(I=m,SHIFT=640)
! abs(SHIFT) must be <= SIZE
      print *,ishftc(I=m,SHIFT=1,SIZE=0)
      end

[Bug fortran/50514] gfortran should check ISHFT & ISHFTC aruments (r178939)

[kargl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50514

kargl at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kargl at gcc dot gnu.org

--- Comment #1 from kargl at gcc dot gnu.org 2011-09-25 22:05:14 UTC ---
Technically, it the programmer that should check the arguments.
The prohibitions in the Fortran Standard are placed on the
programmer not the Fortran processor (ie., the compiler).

Adding the check into

do i = 1,1000000
   do j = 1, 1000000
      do k = 1, 1000000
         ! Some stuff here to compute k and n < bit_size(k)
         m = ishft(k,n)  ! for some k, n
         ! More stuff here to use m
      end do
   end do
end do

seems like useless overhead.

[Bug fortran/50514] gfortran should check ISHFT & ISHFTC aruments (r178939)

[zeccav at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50514

--- Comment #2 from Vittorio Zecca <zeccav at gmail dot com> 2011-09-28 09:20:40 UTC ---
I meant checking static expressions at compilation time, as in my example.
This has no cost at run time.
You proposed a run time check that still should be done if requested with a
kind of -fcheck option.
By the way, gfortran is already checking consistency of static arguments to
intrinsic functions, it is that just these one are left unchecked.

[Bug fortran/50514] gfortran should check ISHFT & ISHFTC aruments (r178939)

[sgk at troutmask dot apl.washington.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50514

--- Comment #3 from Steve Kargl <sgk at troutmask dot apl.washington.edu> 2011-09-28 19:45:48 UTC ---
On Wed, Sep 28, 2011 at 09:20:40AM +0000, zeccav at gmail dot com wrote:
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50514
> 
> --- Comment #2 from Vittorio Zecca <zeccav at gmail dot com> 2011-09-28 09:20:40 UTC ---
> I meant checking static expressions at compilation time, as in my example.
> This has no cost at run time.

I know want you meant.

> You proposed a run time check that still should be done if requested with a
> kind of -fcheck option.

I disagree.

> By the way, gfortran is already checking consistency of static arguments to
> intrinsic functions, it is that just these one are left unchecked.

Yes, I know.  I wrote some of the code.

Index: testsuite/gfortran.dg/ishft_3.f90
===================================================================
--- testsuite/gfortran.dg/ishft_3.f90    (revision 179208)
+++ testsuite/gfortran.dg/ishft_3.f90    (working copy)
@@ -1,11 +1,38 @@
 ! { dg-do compile }
+! PR fortran/50514
 program ishft_3
-  integer i, j
-  write(*,*) ishftc( 3, 2, 3 )
-  write(*,*) ishftc( 3, 2, i )
-  write(*,*) ishftc( 3, i, j )
-  write(*,*) ishftc( 3, 128 )     ! { dg-error "exceeds BIT_SIZE of first" }
-  write(*,*) ishftc( 3, 0, 128 )  ! { dg-error "exceeds BIT_SIZE of first" }
-  write(*,*) ishftc( 3, 0, 0 )    ! { dg-error "Invalid third argument" }
-  write(*,*) ishftc( 3, 3, 2 )    ! { dg-error "exceeds third argument" }
+
+   implicit none
+
+   integer j, m
+
+   m = 42
+   !
+   ! These should compile.
+   !
+   j = ishft(m, 16)
+   j = ishft(m, -16)
+   j = ishftc(m, 16)
+   j = ishftc(m, -16)
+   !
+   ! These should issue an error.
+   !
+   j = ishft(m, 640)    ! { dg-error "absolute value of SHIFT" }
+   j = ishftc(m, 640)   ! { dg-error "absolute value of SHIFT" }
+   j = ishft(m, -640)   ! { dg-error "absolute value of SHIFT" }
+   j = ishftc(m, -640)  ! { dg-error "absolute value of SHIFT" }
+
+   ! abs(SHIFT) must be <= SIZE
+
+   j = ishftc(m,  1, 2)
+   j = ishftc(m,  1, 2)
+   j = ishftc(m, -1, 2)
+   j = ishftc(m, -1, 2)
+
+   j = ishftc(m,  10, 2)! { dg-error "absolute value of SHIFT" }
+   j = ishftc(m,  10, 2)! { dg-error "absolute value of SHIFT" }
+   j = ishftc(m, -10, 2)! { dg-error "absolute value of SHIFT" }
+   j = ishftc(m, -10, 2)! { dg-error "absolute value of SHIFT" }
+
+   j = ishftc(m, 1, -2) ! { dg-error "must be positive" }
 end program

Index: fortran/check.c
===================================================================
--- fortran/check.c    (revision 179208)
+++ fortran/check.c    (working copy)
@@ -318,6 +318,22 @@ less_than_bitsize1 (const char *arg1, gf
     {
       gfc_extract_int (expr2, &i2);
       i3 = gfc_validate_kind (BT_INTEGER, expr1->ts.kind, false);
+ 
+      /* For ISHFT[C], |shift| <= bit_size(i).  */
+      if (strncmp (arg2, "ISHFT", 5) == 0)
+    {
+      if (i2 < 0)
+        i2 = -i2;
+
+      if (i2 > gfc_integer_kinds[i3].bit_size)
+        {
+          gfc_error ("The absolute value of SHIFT at %L must be less "
+             "than or equal to BIT_SIZE('%s')",
+             &expr2->where, arg1);
+          return FAILURE;
+        }
+    }
+
       if (or_equal)
     {
       if (i2 > gfc_integer_kinds[i3].bit_size)
@@ -1961,6 +1977,9 @@ gfc_check_ishft (gfc_expr *i, gfc_expr *
       || type_check (shift, 1, BT_INTEGER) == FAILURE)
     return FAILURE;

+  if (less_than_bitsize1 ("I", i, "ISHFT", shift, true) == FAILURE)
+    return FAILURE;
+
   return SUCCESS;
 }

@@ -1972,7 +1991,35 @@ gfc_check_ishftc (gfc_expr *i, gfc_expr 
       || type_check (shift, 1, BT_INTEGER) == FAILURE)
     return FAILURE;

-  if (size != NULL && type_check (size, 2, BT_INTEGER) == FAILURE)
+  if (size != NULL) 
+    {
+      int i2, i3;
+
+      if (type_check (size, 2, BT_INTEGER) == FAILURE)
+    return FAILURE;
+
+      if (less_than_bitsize1 ("I", i, "SIZE", size, true) == FAILURE)
+    return FAILURE;
+
+      gfc_extract_int (size, &i3);
+      if (i3 <= 0)
+    {
+      gfc_error ("SIZE at %L must be positive", &size->where);
+      return FAILURE;
+    }
+
+      gfc_extract_int (shift, &i2);
+      if (i2 < 0)
+    i2 = -i2;
+
+      if (i2 > i3)
+    {
+      gfc_error ("The absolute value of SHIFT at %L must be less than "
+             "or equal to SIZE at %L", &shift->where, &size->where);
+      return FAILURE;
+    }
+    }
+  else if (less_than_bitsize1 ("I", i, "ISHFTC", shift, true) == FAILURE)
     return FAILURE;

   return SUCCESS;

[Bug fortran/50514] gfortran should check ISHFT & ISHFTC aruments (r178939)

[zeccav at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50514

--- Comment #4 from Vittorio Zecca <zeccav at gmail dot com> 2011-09-29 06:58:24 UTC ---
About run time checking: I believe the bit size of k is known at compile time,
and the overhead to check n against it is negligible as compared to computing
ishft itself and maybe n.
Of course when I am asking -fcheck I am prepared to slower execution, but it
may well pay off, if I find a bug. I believe programmer (debugging) time is now
costlier than hardware time.

[Bug fortran/50514] gfortran should check ISHFT & ISHFTC aruments (r178939)

[kargl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50514

kargl at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |FIXED

--- Comment #5 from kargl at gcc dot gnu.org 2011-10-20 18:11:33 UTC ---
Fixed on trunk for static checking.  Runtime checking is not
going to happen due to the overhead.