On Wed, May 19, 2021 at 10:15:53AM +0200, Christophe Lyon via Gcc-patches wrote:
> After this update, the test fails on arm and aarch64: according to the
> logs, the optimization is still performed 14 times.
Seems this is because
if (change
&& !flag_syntax_only
&& (load_extend_op (TYPE_MODE (TREE_TYPE (and0)))
== ZERO_EXTEND))
{
tree uns = unsigned_type_for (TREE_TYPE (and0));
and0 = fold_convert_loc (loc, uns, and0);
and1 = fold_convert_loc (loc, uns, and1);
}
in fold-const.c adds on these targets extra casts that prevent the
optimizations, instead of
_1 = __v._M_value;
_2 = (int) _1;
_3 = _2 & 1;
_4 = __v._M_value;
_5 = (int) _4;
D.8503 = _3 == _5;
on x86_64 it is:
_1 = __v._M_value;
_2 = (unsigned char) _1;
_3 = (int) _2;
_4 = _3 & 1;
_5 = __v._M_value;
_6 = (int) _5;
D.10471 = _4 == _6;
Before fre1 it is
__v$_M_value_9 = __v._M_value;
_1 = __v$_M_value_9;
_2 = (int) _1;
_8 = _1 & 1;
_3 = (int) _8;
_4 = __v$_M_value_9;
_5 = (int) _4;
_7 = _4 == _8;
which fre1 using match.pd optimizes into
__v$_M_value_9 = __v._M_value;
_10 = __v$_M_value_9 & -2;
_7 = _10 == 0;
but on aarch64 before fre1 we have:
__v$_M_value_10 = __v._M_value;
_1 = __v$_M_value_10;
_2 = (unsigned char) _1;
_3 = (int) _2;
_9 = _2 & 1;
_4 = (int) _9;
_5 = __v$_M_value_10;
_6 = (int) _5;
_8 = _4 == _6;
which is not optimized.
Jakub
