https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116142
Bug ID: 116142
Summary: vec_widen_smult_{odd,even}_M ineffective for a simple
widening dot product (vect_used_by_reduction is not
set?)
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: xry111 at gcc dot gnu.org
Target Milestone: ---
On LoongArch, for this test case:
short x[8], y[8];
int dot() {
int ret = 0;
for (int i = 0; i < 8; i++)
ret += x[i] * y[i];
return ret;
}
The compiler produces:
vld $vr1,$r12,0
vld $vr0,$r12,32
vslti.h $vr2,$vr1,0
vslti.h $vr3,$vr0,0
vilvh.h $vr4,$vr2,$vr1
vilvh.h $vr5,$vr3,$vr0
vilvl.h $vr1,$vr2,$vr1
vilvl.h $vr0,$vr3,$vr0
vmul.w $vr0,$vr0,$vr1
vmadd.w $vr0,$vr5,$vr4
vhaddw.d.w $vr0,$vr0,$vr0
vhaddw.q.d $vr0,$vr0,$vr0
vpickve2gr.w $r4,$vr0,0
slli.w $r4,$r4,0
jr $r1
This is stupid and we just want:
vld $vr1,$r12,0
vld $vr0,$r12,32
vmulwev.w.h $vr2, $vr1, $vr0
vmulwod.w.h $vr3, $vr1, $vr0
vadd.w $vr2, $vr2, $vr3
vhaddw.d.w $vr2,$vr2,$vr2
vhaddw.q.d $vr2,$vr2,$vr2
vpickve2gr.w $r4,$vr2,0
jr $r1
After reading GCC internal I found we missed vec_widen_smult_even_v8hi and
vec_widen_smult_odd_v8hi. So I added them:
+(define_expand "vec_widen_smult_even_v8hi"
+ [(match_operand:V4SI 0 "register_operand" "=f")
+ (match_operand:V8HI 1 "register_operand" " f")
+ (match_operand:V8HI 2 "register_operand" " f")]
+ "ISA_HAS_LSX"
+{
+ emit_insn (gen_lsx_vmulwev_w_h (operands[0], operands[1], operands[2]));
+ DONE;
+})
+
+(define_expand "vec_widen_smult_odd_v8hi"
+ [(match_operand:V4SI 0 "register_operand" "=f")
+ (match_operand:V8HI 1 "register_operand" " f")
+ (match_operand:V8HI 2 "register_operand" " f")]
+ "ISA_HAS_LSX"
+{
+ emit_insn (gen_lsx_vmulwod_w_h (operands[0], operands[1], operands[2]));
+ DONE;
+})
+
But they are not used at all, despite some comment in tree-vect-stmts.cc
suggests this approach should work:
However, in the special case that the result of the widening
operation is used in a reduction computation only, the order doesn't
matter (because when vectorizing a reduction we change the order of
the computation). Some targets can take advantage of this and
generate more efficient code. For example, targets like Altivec,
that support widen_mult using a sequence of {mult_even,mult_odd}
generate the following vectors:
vect1: [res1,res3,res5,res7],
vect2: [res2,res4,res6,res8].
... ...
if (vect_loop
&& STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction
&& !nested_in_vect_loop_p (vect_loop, stmt_info)
&& supportable_widening_operation (vinfo, VEC_WIDEN_MULT_EVEN_EXPR,
stmt_info, vectype_out,
vectype_in, code1,
code2, multi_step_cvt,
interm_types))
{
/* Elements in a vector with vect_used_by_reduction property cannot
be reordered if the use chain with this property does not have the
same operation. One such an example is s += a * b, where elements
in a and b cannot be reordered. Here we check if the vector
defined
by STMT is only directly used in the reduction statement. */
tree lhs = gimple_assign_lhs (stmt_info->stmt);
stmt_vec_info use_stmt_info = loop_info->lookup_single_use (lhs);
if (use_stmt_info
&& STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def)
return true;
}
c1 = VEC_WIDEN_MULT_LO_EXPR;
c2 = VEC_WIDEN_MULT_HI_EXPR;
break;
When I use gdb to debug cc1 I found STMT_VINFO_RELEVANT (stmt_info) is never
vect_used_by_reduction for my test case, but (obviously) the result of the
multiplication is only used by a reduction. So IMO there's a bug in tree
optimization. Or am I missing something?