Re: c/2454

["Joseph S. Myers" <jsm28@cam.ac.uk>]

The following reply was made to PR c/2454; it has been noted by GNATS.

From: "Joseph S. Myers" <jsm28@cam.ac.uk>
To: Neil Booth <neil@daikokuya.demon.co.uk>
Cc: <gcc-gnats@gcc.gnu.org>,  <gcc-patches@gcc.gnu.org>
Subject: Re: c/2454
Date: Wed, 6 Jun 2001 21:32:12 +0100 (BST)

 On 6 Jun 2001, Neil Booth wrote:
 
 >  jsm28@gcc.gnu.org wrote:-
 >  
 >  >      Confirmed as a bug, and a regression from 2.95.  A small
 >  >      testcase is
 >  >      
 >  >      static int i;
 >  >      
 >  >      int
 >  >      main (void)
 >  >      {
 >  >        i = -1;
 >  >        switch((signed char) i) {
 >  >        case 255:
 >  >          abort ();
 >  >        default:
 >  >          exit (0);
 >  >        }
 >  >      }
 >  
 >  Ugh.  These signedness-of-less-than-int issues seem to be endless.
 
 This bug is more complicated than I'd hoped.  I think the patch below 
 addresses a C front end bug that is part of the problem - we shouldn't 
 strip type conversions from the switch case expression.  (2.95 did strip 
 them, but then gave a warning that the case value was out of range.  3.0 
 converts the case value, as required by the C standard - but the 
 conversion should be to the promoted type, int, not signed char.)  
 However, after this patch the testcase still fails with -O2, -Os, -O3.  
 Could someone familiar with the optimizers take a look at this?  This is a 
 regression, but not currently marked as "high".
 
 2001-06-06  Joseph S. Myers  <jsm28@cam.ac.uk>
 
 	* c-typeck.c (c_start_case): Don't strip conversions from the
 	controlling expression.  Partially fixes PR c/2454.
 
 2001-06-06  Joseph S. Myers  <jsm28@cam.ac.uk>
 
 	* gcc.c-torture/execute/20010606-1.c: New test.
 
 --- c-typeck.c.orig	Tue Jun  5 07:05:47 2001
 +++ c-typeck.c	Wed Jun  6 16:54:11 2001
 @@ -7044,7 +7044,6 @@
  	}
        else
  	{
 -	  tree index;
  	  type = TYPE_MAIN_VARIANT (TREE_TYPE (exp));
  
  	  if (warn_traditional && !in_system_header
 @@ -7054,14 +7053,6 @@
  
  	  exp = default_conversion (exp);
  	  type = TREE_TYPE (exp);
 -	  index = get_unwidened (exp, NULL_TREE);
 -	  /* We can't strip a conversion from a signed type to an
 -	     unsigned, because if we did, int_fits_type_p would do the
 -	     wrong thing when checking case values for being in range,
 -	     and it's too hard to do the right thing.  */
 -	  if (TREE_UNSIGNED (TREE_TYPE (exp))
 -	      == TREE_UNSIGNED (TREE_TYPE (index)))
 -	    exp = index;
  	}
      }
  
 --- testsuite/gcc.c-torture/execute/20010606-1.c	Mon Mar 26 23:57:02 2001
 +++ testsuite/gcc.c-torture/execute/20010606-1.c	Wed Jun  6 16:53:01 2001
 @@ -0,0 +1,22 @@
 +/* Origin: Joseph Myers <jsm28@cam.ac.uk>.  */
 +/* Case labels in a switch statement are converted to the promoted
 +   type of the controlling expression, not an unpromoted version.
 +   Reported as PR c/2454 by
 +   Andreas Krakowczyk <Andreas.Krakowczyk@fujitsu-siemens.com>.  */
 +
 +extern void exit (int);
 +extern void abort (void);
 +
 +static int i;
 +
 +int
 +main (void)
 +{
 +  i = -1;
 +  switch((signed char) i) {
 +  case 255:
 +    abort ();
 +  default:
 +    exit (0);
 +  }
 +}
 
 -- 
 Joseph S. Myers
 jsm28@cam.ac.uk