1422 // qword_count - 64-bits element count
1423 // to - scratch
1424 // L_copy_bytes - entry label
1425 // L_copy_8_bytes - exit label
1426 //
1427 void copy_bytes_backward(Register from, Register dest,
1428 Register qword_count, Register to,
1429 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1430 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1431 Label L_loop;
1432 __ align(OptoLoopAlignment);
1433 if (UseUnalignedLoadStores) {
1434 Label L_end;
1435 if (UseAVX > 2) {
1436 __ movl(to, 0xffff);
1437 __ kmovql(k1, to);
1438 }
1439 // Copy 64-bytes per iteration
1440 __ BIND(L_loop);
1441 if (UseAVX > 2) {
1442 __ evmovdqul(xmm0, Address(from, qword_count, Address::times_8, 32), Assembler::AVX_512bit);
1443 __ evmovdqul(Address(dest, qword_count, Address::times_8, 32), xmm0, Assembler::AVX_512bit);
1444 } else if (UseAVX == 2) {
1445 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1446 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1447 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1448 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1449 } else {
1450 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
1451 __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
1452 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
1453 __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
1454 __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
1455 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
1456 __ movdqu(xmm3, Address(from, qword_count, Address::times_8, 0));
1457 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm3);
1458 }
1459 __ BIND(L_copy_bytes);
1460 __ subptr(qword_count, 8);
1461 __ jcc(Assembler::greaterEqual, L_loop);
1462
1463 __ addptr(qword_count, 4); // add(8) and sub(4)
|
1422 // qword_count - 64-bits element count
1423 // to - scratch
1424 // L_copy_bytes - entry label
1425 // L_copy_8_bytes - exit label
1426 //
1427 void copy_bytes_backward(Register from, Register dest,
1428 Register qword_count, Register to,
1429 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1430 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1431 Label L_loop;
1432 __ align(OptoLoopAlignment);
1433 if (UseUnalignedLoadStores) {
1434 Label L_end;
1435 if (UseAVX > 2) {
1436 __ movl(to, 0xffff);
1437 __ kmovql(k1, to);
1438 }
1439 // Copy 64-bytes per iteration
1440 __ BIND(L_loop);
1441 if (UseAVX > 2) {
1442 __ evmovdqul(xmm0, Address(from, qword_count, Address::times_8, 0), Assembler::AVX_512bit);
1443 __ evmovdqul(Address(dest, qword_count, Address::times_8, 0), xmm0, Assembler::AVX_512bit);
1444 } else if (UseAVX == 2) {
1445 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1446 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1447 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1448 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1449 } else {
1450 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
1451 __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
1452 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
1453 __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
1454 __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
1455 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
1456 __ movdqu(xmm3, Address(from, qword_count, Address::times_8, 0));
1457 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm3);
1458 }
1459 __ BIND(L_copy_bytes);
1460 __ subptr(qword_count, 8);
1461 __ jcc(Assembler::greaterEqual, L_loop);
1462
1463 __ addptr(qword_count, 4); // add(8) and sub(4)
|