[Bug tree-optimization/80738] New: dead first stmt in a=0;a=b;b=0 whatever the aliasing

[glisse at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80738

            Bug ID: 80738
           Summary: dead first stmt in a=0;a=b;b=0 whatever the aliasing
           Product: gcc
           Version: 8.0
            Status: UNCONFIRMED
          Keywords: alias, missed-optimization
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org
  Target Milestone: ---

In C++, with move/swap, we regularly see the following appear

void f(int&a,int&b){
  a=0;
  a=b;
  b=0;
}

If a and b do not alias, the first statement is clearly killed by the second.
If a and b are the same, the last statement kills all the rest. So in all
cases, the first statement is dead. (we could imagine that removing it means
that in some cases, we will copy uninitialized memory to itself, but that
doesn't seem so bad)

This is yet another case where the same optimization is valid whether 2
pointers alias or not, like PR 66261 or PR 80617, but the others were more
about propagation and this one is about dead code...

It seems hard to detect (involves 4 gimple statements, and plenty of
intermediate statements could interfere), the gain may not be worth it. If we
split the paths just right: if(&a!=&b){a=0;a=b;}else{a=0;a=b;}b=0; becomes
if(&a!=&b)a=b; b=0; and the if could be done unconditionally... Seems unlikely.