[Bug 1250904] [NEW] variable is optimized out by mistake

[Shkarnikov Sergey]

Public bug reported:

OS: Ubuntu 12.04.3 LTS
gcc version: 4.6.3-1ubuntu5

It's a  modificated test case from gcc-4.6.3 testsuite (reassoc-11.c):

int main(int a, int b, int c, int d){
 int e = (a ^ b) ^ (c ^ d);
 int f = (c ^ a) ^ (b ^ d);
 return ( * ((int *)(((long unsigned ) & c) + (((long unsigned ) & e) - ((long unsigned ) & c)) *
         (f <= (~(f ^ (~((b << 2 | f << 2) - (( b | f) << 2))))))))) ^ f;
}

Those terrifying expression in braces should be evaluated to "e"  ( "(f
<= (~(f ^ (~((b << 2 | f << 2) - (( b | f) << 2))))))" is a predicate
and always equal to 1) and so test should return 0. But if the
optimization is applied (O1 and greater) test returns some other value.
Decompilation shows that "e" and "f" are optimized out. "f" value is
evaluated in place, but evaluation of "e" is also optimized out. Then
expression in braces is  evaluated to some garbage value (interesting
that compiler handles somehow dereferencing of "e" even after it was
optimized out).

Example is quite exotic, but, in my opinion, reveals incorrect
optimizer's behavior.

** Affects: gcc-4.6 (Ubuntu)
     Importance: Undecided
         Status: New


** Tags: optimizer

** Description changed:

  OS: Ubuntu 12.04.3 LTS
  gcc version: 4.6.3-1ubuntu5
  
  It's a  modificated test case from gcc-4.6.3 testsuite (reassoc-11.c):
  
  int main(int a, int b, int c, int d){
- 	int e = (a ^ b) ^ (c ^ d);
- 	int f = (c ^ a) ^ (b ^ d);
- 	return ( * ((int *)(((long unsigned ) & c) + (((long unsigned ) & e) - ((long unsigned ) & c)) *
-                                                                                                                             (f <= (~(f ^ (~((b << 2 | f << 2) - (( b | f) << 2))))))))) ^ f;
+  int e = (a ^ b) ^ (c ^ d);
+  int f = (c ^ a) ^ (b ^ d);
+  return ( * ((int *)(((long unsigned ) & c) + (((long unsigned ) & e) - ((long unsigned ) & c)) *
+          (f <= (~(f ^ (~((b << 2 | f << 2) - (( b | f) << 2))))))))) ^ f;
  }
  
  Those terrifying expression in braces should be evaluated to "e"  ( "(f
  <= (~(f ^ (~((b << 2 | f << 2) - (( b | f) << 2))))))" is a predicate
  and always equal to 1) and so test should return 0. But if the
  optimization is applied (O1 and greater) test returns some other value.
  Decompilation shows that "e" and "f" are optimized out. "f" value is
  evaluated in place, but evaluation of "e" is also optimized out. Then
  expression in braces is  evaluated to some garbage value (interesting
  that compiler handles somehow dereferencing of "e" even after it was
  optimized out).
  
  Example is quite exotic, but, in my opinion, reveals incorrect
  optimizer's behavior.

-- 
You received this bug notification because you are a member of Ubuntu
Foundations Bugs, which is subscribed to gcc-4.6 in Ubuntu.
https://bugs.launchpad.net/bugs/1250904

Title:
  variable is optimized out by mistake

Status in “gcc-4.6” package in Ubuntu:
  New

Bug description:
  OS: Ubuntu 12.04.3 LTS
  gcc version: 4.6.3-1ubuntu5

  It's a  modificated test case from gcc-4.6.3 testsuite (reassoc-11.c):

  int main(int a, int b, int c, int d){
   int e = (a ^ b) ^ (c ^ d);
   int f = (c ^ a) ^ (b ^ d);
   return ( * ((int *)(((long unsigned ) & c) + (((long unsigned ) & e) - ((long unsigned ) & c)) *
           (f <= (~(f ^ (~((b << 2 | f << 2) - (( b | f) << 2))))))))) ^ f;
  }

  Those terrifying expression in braces should be evaluated to "e"  (
  "(f <= (~(f ^ (~((b << 2 | f << 2) - (( b | f) << 2))))))" is a
  predicate and always equal to 1) and so test should return 0. But if
  the optimization is applied (O1 and greater) test returns some other
  value. Decompilation shows that "e" and "f" are optimized out. "f"
  value is evaluated in place, but evaluation of "e" is also optimized
  out. Then expression in braces is  evaluated to some garbage value
  (interesting that compiler handles somehow dereferencing of "e" even
  after it was optimized out).

  Example is quite exotic, but, in my opinion, reveals incorrect
  optimizer's behavior.

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/gcc-4.6/+bug/1250904/+subscriptions