[Bug tree-optimization/69845] New: Expression getting incorrectly optimized after being rewritten by compiler

[mdaniels at qnx dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69845

            Bug ID: 69845
           Summary: Expression getting incorrectly optimized after being
                    rewritten by compiler
           Product: gcc
           Version: 5.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: mdaniels at qnx dot com
  Target Milestone: ---

Created attachment 37709
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37709&action=edit
Test case

In the given example, when evaluating the expression for 'C', there should be
no overflow. I am however, seeing the compiler rewrite the expressions into a
form that will overflow.

I have only tested this on x86 and x86_64, and according to 5.1.2.3p14, since
these archs do not trap on overflow this is legal behaviour. Because of the
phrasing that the expression can only be rewritten if the same results occur, I
have to assume this means the newly added overflow does not cause undefined
behaviour. The compiler added it knowing that it will not trap and that
positive and negative overflow cancel after all.

With that in mind, it looks like one of the optimization passes is taking
advantage of this and causing incorrect results (optimizes everything to a call
to puts("failed\n")). When I compile the test case with -O0 it will pass, but
-O1 and above will cause the test case to fail.

Given the problem description, it should come as no surprise that the problem
is not seen when -fwrapv or -ftrapv are used.

I have tested on Ubuntu 14.04 x86_64 with GCC 4.8.3, 4.9.3, and 5.3.0. 4.8.3
does not reproduce, 4.9.3 and 5.3.0 do.