https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111591
--- Comment #16 from Kewen Lin <linkw at gcc dot gnu.org> ---
Tracing down it with template specialization, the aborting happens on
auto vn_b = Load(dn, in_b.get());
HWY_ASSERT_VEC_EQ(
dw, vw_signed_max,
SatWidenMulPairwiseAdd(
dw, InterleaveLower(dn_u, BitCast(dn_u, vn_b), vn_unsigned_max),
InterleaveLower(dn, vn_b, vn_signed_max)));
with "void operator()(int8_t, CappedTag<int8_t, 8> dn)"
by isolating, it doesn't get the expected result on "b0" for function
template <class DI16, class VU8, class VI8>
HWY_API Vec<DI16> SatWidenMulPairwiseAdd(DI16 di16, VU8 a, VI8 b) {
RebindToUnsigned<decltype(di16)> du16;
auto a0 = And(BitCast(di16, a), Set(di16, 255));
auto b0 = ShiftRight<8>(ShiftLeft<8>(BitCast(di16, b)));
auto a1 = BitCast(di16, ShiftRight<8>(BitCast(du16, a)));
auto b1 = ShiftRight<8>(BitCast(di16, b));
return SaturatedAdd(Mul(a0, b0), Mul(a1, b1));
}
specialized with
template <> HWY_API Vec128<int16_t, 4> SatWidenMulPairwiseAdd(Simd<int16_t, 4,
0> di16, Vec128<uint8_t, 8> a, Vec128<int8_t, 8> b)
further found that the unexpected values are from ShiftLeft<8>, the tree
optimized code looks expected but the final insn sequence look in wrong order.
Either -fdisable-rtl-sched2 or -fdisable-rtl-sched1 can make it pass. With
counter, I see an unexpected insn movement in sched2 on insn 395.
...
1436: %10:DI=0x70
REG_EQUIV 0x70
1438: %9:DI=0xc0
REG_EQUIV 0xc0
1437: %8:DI=0x1e0
REG_EQUIV 0x1e0
1441: %7:DI=0xd0
REG_EQUIV 0xd0
389: %0:V2DI=[%1:DI+%9:DI]
REG_DEAD %9:DI
REG_EQUAL [sfp:DI+0xc0]
1445: %5:DI=0xb0
REG_EQUIV 0xb0
1714: %9:DI=0xff0000
REG_EQUIV 0xff0000
373: [%1:DI+0x70]=%4:DI
REG_DEAD %4:DI
375: [%1:DI+0x78]=%6:DI
REG_DEAD %6:DI
1715: %9:DI=%9:DI|0xff
1785: %25:DI=high(unspec[`*.LC8',%2:DI] 47)
1716: %9:DI=%9:DI&0xffffffff|%9:DI<<0x20
REG_EQUIV 0xff00ff00ff00ff
410: %28:DI=%1:DI+0xae
REG_EQUAL sfp:DI+0xae
6: %31:SI=0
REG_EQUAL 0
1786: %25:DI=%25:DI+low(unspec[`*.LC8',%2:DI] 47)
REG_DEAD %2:DI
REG_EQUAL `*.LC8'
392: [%1:DI+%7:DI]=%0:V2DI
REG_DEAD %7:DI
// unexpected version having insn 395
moved here.
1738: %12:V2DI=[%1:DI+%10:DI]
376: [%1:DI+%8:DI]=%12:V2DI
REG_DEAD %12:V2DI
REG_DEAD %8:DI
REG_EQUIV [sfp:DI+%8:DI]
REG_EQUAL [sfp:DI+0x70]
390: [%1:DI+%10:DI]=%0:V2DI // since this store updates
[%1:DI+0x70] in 16 bytes, so the read
// can't pass this
REG_DEAD %0:V2DI
395: %4:DI=zero_extend([%1:DI+0x70]) // <------ this is expected
398: %6:DI=zero_extend([%1:DI+0x72])
401: %7:DI=zero_extend([%1:DI+0x74])
404: %8:DI=zero_extend([%1:DI+0x76])
396: %4:SI=%4:SI<<0x8
399: %6:SI=%6:SI<<0x8
402: %7:SI=%7:SI<<0x8
405: %8:SI=%8:SI<<0x8
....
the tree optimized IR for this part looks expected?
<bb 51> [local count: 119292722]:
v = a;
MEM <unsigned char[16]> [(char * {ref-all})&D.38735] = MEM <unsigned
char[16]> [(char * {ref-all})&v];
v ={v} {CLOBBER(eol)};
vect_a_raw_0_1121.562_722 = MEM <vector(4) short int> [(short int
*)&D.38735];
_215 = VIEW_CONVERT_EXPR<long unsigned int>(vect_a_raw_0_1121.562_722);
_830 = _215 & 71777214294589695;
_1549 = BIT_FIELD_REF <_830, 16, 32>;
_1537 = BIT_FIELD_REF <_830, 16, 16>;
_323 = BIT_FIELD_REF <_830, 16, 0>;
v = b;
MEM <unsigned char[16]> [(char * {ref-all})&b00] = MEM <unsigned char[16]>
[(char * {ref-all})&v];
==> ref-all here, so should be executed before any
reads below?
v ={v} {CLOBBER(eol)};
v = b00;
raw_u_1323 = v.raw[0];
_1324 = raw_u_1323 << 8;
v.raw[0] = _1324;
raw_u_1403 = v.raw[1];
_1404 = raw_u_1403 << 8;
v.raw[1] = _1404;
raw_u_1447 = v.raw[2];
_1448 = raw_u_1447 << 8;
v.raw[2] = _1448;
raw_u_128 = v.raw[3];
_129 = raw_u_128 << 8;
v.raw[3] = _129;
b01 = v;
v ={v} {CLOBBER(eol)};
ivtmp.577_734 = (unsigned long) &MEM <struct Vec128> [(void *)&b01 + -2B];
...
I guess there is some way to keep this kind of aliasing information after
expanding, need more investigations why sched considers it's safe to move.