https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111751
--- Comment #7 from JuzheZhong <juzhe.zhong at rivai dot ai> ---
(In reply to Andrew Pinski from comment #6)
> (In reply to JuzheZhong from comment #5)
> > (In reply to Andrew Pinski from comment #4)
> > > The issue for aarch64 with SVE is that MASK_LOAD is not optimized:
> > >
> > > ic = "\x00\x03\x06\t\f\x0f\x12\x15\x18\x1b\x1e!$\'*-";
> > > ib = "\x00\x03\x06\t\f\x0f\x12\x15\x18\x1b\x1e!$\'*-";
> > > vect__1.7_9 = .MASK_LOAD (&ib, 8B, { -1, -1, -1, -1, -1, -1, -1, -1, -1,
> > > -1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
> > > 0,
> > > ... });
> > > vect__2.10_35 = .MASK_LOAD (&ic, 8B, { -1, -1, -1, -1, -1, -1, -1, -1,
> > > -1,
> > > -1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
> > > 0,
> > > ... });
> >
> > I don't ARM SVE has issues ...
>
> It does as I mentioned if you use -fno-vect-cost-model, you get the above
> issue which should be optimized really to a constant vector ...
After investigation:
I found it failed to recognize its CONST_VECTOR value in FRE
/* Visit a load from a reference operator RHS, part of STMT, value number it,
and return true if the value number of the LHS has changed as a result. */
static bool
visit_reference_op_load (tree lhs, tree op, gimple *stmt)
{
bool changed = false;
tree result;
vn_reference_t res;
tree vuse = gimple_vuse (stmt);
tree last_vuse = vuse;
result = vn_reference_lookup (op, vuse, default_vn_walk_kind, &res, true,
&last_vuse);
/* We handle type-punning through unions by value-numbering based
on offset and size of the access. Be prepared to handle a
type-mismatch here via creating a VIEW_CONVERT_EXPR. */
if (result
&& !useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (op)))
{
/* Avoid the type punning in case the result mode has padding where
the op we lookup has not. */
if (maybe_lt (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (result))),
GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))
result = NULL_TREE;
....
The result is BLKmode, op is V16QImode
Then reach
/* Avoid the type punning in case the result mode has padding where
the op we lookup has not. */
if (maybe_lt (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (result))),
GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))
result = NULL_TREE;
If I delete this code, RVV can optimize it.
Do you have any suggestion ?
This is my observation:
Breakpoint 6, visit_reference_op_load (lhs=0x7ffff68364c8, op=0x7ffff6874410,
stmt=0x7ffff6872640) at ../../../../gcc/gcc/tree-ssa-sccvn.cc:5740
5740 result = vn_reference_lookup (op, vuse, default_vn_walk_kind, &res,
true, &last_vuse);
(gdb) c
Continuing.
Breakpoint 6, visit_reference_op_load (lhs=0x7ffff68364c8, op=0x7ffff6874410,
stmt=0x7ffff6872640) at ../../../../gcc/gcc/tree-ssa-sccvn.cc:5740
5740 result = vn_reference_lookup (op, vuse, default_vn_walk_kind, &res,
true, &last_vuse);
(gdb) n
5746 && !useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE
(op)))
(gdb) p debug (result)
"\x00\x03\x06\t\f\x0f\x12\x15\x18\x1b\x1e!$\'*-"
$9 = void
(gdb) p op->typed.type->type_common.mode
$10 = E_V16QImode
