[Bug c/37569] New: Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

The loop in the example is incorrectly replaced with endless loop when
compiling with -Os, -O2 or -O3. It is compiled correctly with -O0 or -O1.

uint16 sum=0;
uint16 *i;

/*sum all words in inspected area*/
i = (uint16 *)0x0FFE;
do
{
--i;
sum = sum + *(i);
} while ( i != 0 );

(see the full source in the attachment)

Target: i686-pc-cygwin
Configured with: ../gcc-4.3.2/configure
Thread model: single
gcc version 4.3.2 (GCC) 

(the same problem occurs with earlier versions of GCC (e.g. 3.4.3) and G++ )


-- 
   Summary: Loop incorrectly optimized to endless loop with -Os/-
O2/-O3
   Product: gcc
   Version: 4.3.2
Status: UNCONFIRMED
  Severity: major
  Priority: P3
 Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: ppr at melexis dot com
  GCC host triplet: i686-pc-cygwin


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #1 from ppr at melexis dot com  2008-09-18 07:52 ---
Created an attachment (id=16352)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16352&action=view)
preprocessed source of the test case


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #2 from ppr at melexis dot com  2008-09-18 07:53 ---
Created an attachment (id=16353)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16353&action=view)
The produced assembler source


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #3 from ppr at melexis dot com  2008-09-18 07:53 ---
Created an attachment (id=16354)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16354&action=view)
The compiler output


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #4 from ppr at melexis dot com  2008-09-18 07:58 ---
The problem seems to be related with the i pointer reaching 0. If the loop
condition is e.g. "while ( i != 2 )" everything is ok.
The same if the "sum = sum + *(i)" is replaced e.g. with "sum = sum + *(i+1)"


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #6 from ppr at melexis dot com  2008-09-18 08:22 ---
(In reply to comment #5)
> another vrp bug? -fno-tree-vrp helps.
> 
Yes, -fno-tree-vrp does help


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #9 from ppr at melexis dot com  2008-09-18 10:20 ---
(In reply to comment #8)
> I think this code is invalid.  I seem to recall the argument that it's
> undefined behavior to increment or decrement a pointer to 0.

Well, the pointer is not NULL initially, it's just decremented until it reaches
 zero. I agree this code doesn't make much sense in i686-pc environment, but
I've discovered the issue while working with GCC port for 16-bit
microcontroller. The routine is used to calculate ROM checksum and address 0 is
a perfectly valid address at this target.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #11 from ppr at melexis dot com  2008-09-18 10:57 ---
(In reply to comment #10)

This workaround doesn't work with GCC 3.4.3 (the base for our 16-bit port). It
does work with the latest GCC4.3 though.

Anyway, thanks for your support, and for explaining the reasons behind the
compiler's behaviour.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569

[Bug c/37569] Loop incorrectly optimized to endless loop with -Os/-O2/-O3

[ppr at melexis dot com]

--- Comment #13 from ppr at melexis dot com  2008-09-18 12:15 ---
(In reply to comment #12)
> Note you can do the same as AVR does - force flag_delete_null_pointer_checks 
> to
> zero:
> 

Thanks for the hint, I'll use it!

Plamen


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37569