[Bug tree-optimization/78947] New: sub-optimal code for (bool)(int ? int : int)

["peter at cordes dot ca" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78947

            Bug ID: 78947
           Summary: sub-optimal code for (bool)(int ? int : int)
           Product: gcc
           Version: 6.3.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: peter at cordes dot ca
  Target Milestone: ---
            Target: x86_64-*-*, i?86-*-*

Full version:
http://stackoverflow.com/questions/41323911/why-the-difference-in-code-generation-for-bool-bool-int-int


bool condSet(int cond, int a, int b) {
    return cond ? a : b;
}

clang and icc use a cmov to select a or b, and then booleanize that.

g++6.3 -O3 decides to booleanize b, then check the condition and if necessary
booleanize a into %al.  This is pretty obviously sub-optimal, even if a branch
is better than a cmov for a predictable condition.  (And it differs from what
gcc does if the return type is int)

condSet(int, int, int):
        testl   %edx, %edx    # b
        setne   %al           #, <retval>
        testl   %edi, %edi    # cond
        jne     .L6           #,
        rep ret
.L6:
        testl   %esi, %esi    # a
        setne   %al           #, <retval>
        ret



Writing the function as 

  int foo = cond ? a : b;
  return foo;

gets g++ to select an operand with cmov and then booleanize it, like clang does
in the first place (looks optimal to me):

        testl   %edi, %edi
        cmovel  %edx, %esi
        testl   %esi, %esi
        setne   %al
        retq