https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101403
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Another question is whether we shouldn't extend the short_shift stuff in
build_binary_op also to COMPOUND_EXPRs with INTEGER_CST on the ultimate rhs.
unsigned int f1 (unsigned short x) { return x >> 12; }
unsigned int f2 (unsigned short x) { static int u; return x >> (u, 12); }
The C++ FE already handles these the same, as it does:
tree const_op1 = fold_for_warn (op1);
if (TREE_CODE (const_op1) != INTEGER_CST)
const_op1 = op1;
But it won't handle
int f ();
unsigned int f3 (unsigned short x) { return x >> (f (), 12); }
either. The C FE doesn't try to simplify op1 in any way.
I guess in those cases we don't really care about any side-effects in the
expression, what we are looking for is just whether the value is known
constant.
For diagnostics, it would be nice if we handled x >> (f (), 37) like we handle
x >> 37 if it is an out of bounds shift.
For the optimization, maybe the short_shift stuff could be repeated in
match.pd.
