/* * Copyright (c) 2007, 2022, Oracle and/or its affiliates. All rights reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA * or visit www.oracle.com if you need additional information or have any * questions.
*/
// // S U P E R W O R D T R A N S F O R M //=============================================================================
//------------------------------SuperWord---------------------------
SuperWord::SuperWord(PhaseIdealLoop* phase) :
_phase(phase),
_arena(phase->C->comp_arena()),
_igvn(phase->_igvn),
_packset(arena(), 8, 0, NULL), // packs for the current block
_bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
_block(arena(), 8, 0, NULL), // nodes in current block
_post_block(arena(), 8, 0, NULL), // nodes common to current block which are marked as post loop vectorizable
_data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
_mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
_mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
_node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
_clone_map(phase->C->clone_map()), // map of nodes created in cloning
_cmovev_kit(_arena, this), // map to facilitate CMoveV creation
_align_to_ref(NULL), // memory reference to align vectors to
_disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
_dg(_arena), // dependence graph
_visited(arena()), // visited node set
_post_visited(arena()), // post visited node set
_n_idx_list(arena(), 8), // scratch list of (node,index) pairs
_nlist(arena(), 8, 0, NULL), // scratch list of nodes
_stk(arena(), 8, 0, NULL), // scratch stack of nodes
_lpt(NULL), // loop tree node
_lp(NULL), // CountedLoopNode
_pre_loop_end(NULL), // Pre loop CountedLoopEndNode
_bb(NULL), // basic block
_iv(NULL), // induction var
_race_possible(false), // cases where SDMU is true
_early_return(true), // analysis evaluations routine
_do_vector_loop(phase->C->do_vector_loop()), // whether to do vectorization/simd style
_do_reserve_copy(DoReserveCopyInSuperWord),
_num_work_vecs(0), // amount of vector work we have
_num_reductions(0), // amount of reduction work we have
_ii_first(-1), // first loop generation index - only if do_vector_loop()
_ii_last(-1), // last loop generation index - only if do_vector_loop()
_ii_order(arena(), 8, 0, 0)
{ #ifndef PRODUCT
_vector_loop_debug = 0; if (_phase->C->method() != NULL) {
_vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
}
#endif
}
staticconstbool _do_vector_loop_experimental = false; // Experimental vectorization which uses data from loop unrolling.
//------------------------------transform_loop--------------------------- bool SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
assert(UseSuperWord, "should be"); // SuperWord only works with power of two vector sizes. int vector_width = Matcher::vector_width_in_bytes(T_BYTE); if (vector_width < 2 || !is_power_of_2(vector_width)) { returnfalse;
}
if (!cl->is_valid_counted_loop(T_INT)) { returnfalse; // skip malformed counted loop
}
if (cl->is_rce_post_loop() && cl->is_reduction_loop()) { // Post loop vectorization doesn't support reductions returnfalse;
}
// skip any loop that has not been assigned max unroll by analysis if (do_optimization) { if (SuperWordLoopUnrollAnalysis && cl->slp_max_unroll() == 0) { returnfalse;
}
}
// Check for no control flow in body (other than exit)
Node *cl_exit = cl->loopexit(); if (cl->is_main_loop() && (cl_exit->in(0) != lpt->_head)) { #ifndef PRODUCT if (TraceSuperWord) {
tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
lpt->dump_head();
} #endif returnfalse;
}
// Make sure the are no extra control users of the loop backedge if (cl->back_control()->outcnt() != 1) { returnfalse;
}
// Skip any loops already optimized by slp if (cl->is_vectorized_loop()) { returnfalse;
}
if (cl->is_unroll_only()) { returnfalse;
}
if (cl->is_main_loop()) { // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
CountedLoopEndNode* pre_end = find_pre_loop_end(cl); if (pre_end == NULL) { returnfalse;
}
Node* pre_opaq1 = pre_end->limit(); if (pre_opaq1->Opcode() != Op_Opaque1) { returnfalse;
}
set_pre_loop_end(pre_end);
}
init(); // initialize data structures
set_lpt(lpt);
set_lp(cl);
// For now, define one block which is the entire loop body
set_bb(cl);
bool success = true; if (do_optimization) {
assert(_packset.length() == 0, "packset must be empty");
success = SLP_extract(); if (PostLoopMultiversioning) { if (cl->is_vectorized_loop() && cl->is_main_loop() && !cl->is_reduction_loop()) {
IdealLoopTree *lpt_next = cl->is_strip_mined() ? lpt->_parent->_next : lpt->_next;
CountedLoopNode *cl_next = lpt_next->_head->as_CountedLoop();
_phase->has_range_checks(lpt_next); // Main loop SLP works well for manually unrolled loops. But post loop // vectorization doesn't work for these. To bail out the optimization // earlier, we have range check and loop stride conditions below. if (cl_next->is_post_loop() && !cl_next->range_checks_present() &&
cl_next->stride_is_con() && abs(cl_next->stride_con()) == 1) { if (!cl_next->is_vectorized_loop()) { // Propagate some main loop attributes to its corresponding scalar // rce'd post loop for vectorization with vector masks
cl_next->set_slp_max_unroll(cl->slp_max_unroll());
cl_next->set_slp_pack_count(cl->slp_pack_count());
}
}
}
}
} return success;
}
//------------------------------max vector size------------------------------ int SuperWord::max_vector_size(BasicType bt) { int max_vector = Matcher::max_vector_size(bt); int sw_max_vector_limit = SuperWordMaxVectorSize / type2aelembytes(bt); if (max_vector > sw_max_vector_limit) {
max_vector = sw_max_vector_limit;
} return max_vector;
}
// First clear the entries for (uint i = 0; i < lpt()->_body.size(); i++) {
ignored_loop_nodes[i] = -1;
}
int max_vector = max_vector_size(T_BYTE);
// Process the loop, some/all of the stack entries will not be in order, ergo // need to preprocess the ignored initial state before we process the loop for (uint i = 0; i < lpt()->_body.size(); i++) {
Node* n = lpt()->_body.at(i); if (n == cl->incr() ||
n->is_reduction() ||
n->is_AddP() ||
n->is_Cmp() ||
n->is_Bool() ||
n->is_IfTrue() ||
n->is_CountedLoop() ||
(n == cl_exit)) {
ignored_loop_nodes[i] = n->_idx; continue;
}
if (n->is_If()) {
IfNode *iff = n->as_If(); if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) { if (lpt()->is_loop_exit(iff)) {
ignored_loop_nodes[i] = n->_idx; continue;
}
}
}
if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
Node* n_tail = n->in(LoopNode::LoopBackControl); if (n_tail != n->in(LoopNode::EntryControl)) { if (!n_tail->is_Mem()) {
is_slp = false; break;
}
}
}
// This must happen after check of phi/if if (n->is_Phi() || n->is_If()) {
ignored_loop_nodes[i] = n->_idx; continue;
}
if (n->is_Mem()) {
MemNode* current = n->as_Mem();
Node* adr = n->in(MemNode::Address);
Node* n_ctrl = _phase->get_ctrl(adr);
// save a queue of post process nodes if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) { // Process the memory expression int stack_idx = 0; bool have_side_effects = true; if (adr->is_AddP() == false) {
nstack.push(adr, stack_idx++);
} else { // Mark the components of the memory operation in nstack
SWPointer p1(current, this, &nstack, true);
have_side_effects = p1.node_stack()->is_nonempty();
}
// Process the pointer stack while (have_side_effects) {
Node* pointer_node = nstack.node(); for (uint j = 0; j < lpt()->_body.size(); j++) {
Node* cur_node = lpt()->_body.at(j); if (cur_node == pointer_node) {
ignored_loop_nodes[j] = cur_node->_idx; break;
}
}
nstack.pop();
have_side_effects = nstack.is_nonempty();
}
}
}
}
if (is_slp) { // In the main loop, SLP works well if parts of the operations in the loop body // are not vectorizable and those non-vectorizable parts will be unrolled only. // But in post loops with vector masks, we create singleton packs directly from // scalars so all operations should be vectorized together. This compares the // number of packs in the post loop with the main loop and bail out if the post // loop potentially has more packs. if (cl->is_rce_post_loop()) { for (uint i = 0; i < lpt()->_body.size(); i++) { if (ignored_loop_nodes[i] == -1) {
_post_block.at_put_grow(rpo_idx++, lpt()->_body.at(i));
}
} if (_post_block.length() > cl->slp_pack_count()) { // Clear local_loop_unroll_factor and bail out directly from here
local_loop_unroll_factor = 0;
cl->mark_was_slp();
cl->set_slp_max_unroll(0); return;
}
}
// Now we try to find the maximum supported consistent vector which the machine // description can use bool flag_small_bt = false; for (uint i = 0; i < lpt()->_body.size(); i++) { if (ignored_loop_nodes[i] != -1) continue;
BasicType bt;
Node* n = lpt()->_body.at(i); if (n->is_Mem()) {
bt = n->as_Mem()->memory_type();
} else {
bt = n->bottom_type()->basic_type();
}
if (is_java_primitive(bt) == false) continue;
int cur_max_vector = max_vector_size(bt);
// If a max vector exists which is not larger than _local_loop_unroll_factor // stop looking, we already have the max vector to map to. if (cur_max_vector < local_loop_unroll_factor) {
is_slp = false; if (TraceSuperWordLoopUnrollAnalysis) {
tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
} break;
}
// Map the maximal common vector except conversion nodes, because we can't get // the precise basic type for conversion nodes in the stage of early analysis. if (!VectorNode::is_convert_opcode(n->Opcode()) &&
VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) { if (cur_max_vector < max_vector && !flag_small_bt) {
max_vector = cur_max_vector;
} elseif (cur_max_vector > max_vector && UseSubwordForMaxVector) { // Analyse subword in the loop to set maximum vector size to take advantage of full vector width for subword types. // Here we analyze if narrowing is likely to happen and if it is we set vector size more aggressively. // We check for possibility of narrowing by looking through chain operations using subword types. if (is_subword_type(bt)) {
uint start, end;
VectorNode::vector_operands(n, &start, &end);
for (uint j = start; j < end; j++) {
Node* in = n->in(j); // Don't propagate through a memory if (!in->is_Mem() && in_bb(in) && in->bottom_type()->basic_type() == T_INT) { bool same_type = true; for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
Node *use = in->fast_out(k); if (!in_bb(use) && use->bottom_type()->basic_type() != bt) {
same_type = false; break;
}
} if (same_type) {
max_vector = cur_max_vector;
flag_small_bt = true;
cl->mark_subword_loop();
}
}
}
}
}
}
} if (is_slp) {
local_loop_unroll_factor = max_vector;
cl->mark_passed_slp();
}
cl->mark_was_slp(); if (cl->is_main_loop() || cl->is_rce_post_loop()) {
cl->set_slp_max_unroll(local_loop_unroll_factor);
}
}
}
//------------------------------SLP_extract--------------------------- // Extract the superword level parallelism // // 1) A reverse post-order of nodes in the block is constructed. By scanning // this list from first to last, all definitions are visited before their uses. // // 2) A point-to-point dependence graph is constructed between memory references. // This simplifies the upcoming "independence" checker. // // 3) The maximum depth in the node graph from the beginning of the block // to each node is computed. This is used to prune the graph search // in the independence checker. // // 4) For integer types, the necessary bit width is propagated backwards // from stores to allow packed operations on byte, char, and short // integers. This reverses the promotion to type "int" that javac // did for operations like: char c1,c2,c3; c1 = c2 + c3. // // 5) One of the memory references is picked to be an aligned vector reference. // The pre-loop trip count is adjusted to align this reference in the // unrolled body. // // 6) The initial set of pack pairs is seeded with memory references. // // 7) The set of pack pairs is extended by following use->def and def->use links. // // 8) The pairs are combined into vector sized packs. // // 9) Reorder the memory slices to co-locate members of the memory packs. // // 10) Generate ideal vector nodes for the final set of packs and where necessary, // inserting scalar promotion, vector creation from multiple scalars, and // extraction of scalar values from vectors. // bool SuperWord::SLP_extract() {
#ifndef PRODUCT if (_do_vector_loop && TraceSuperWord) {
tty->print("SuperWord::SLP_extract\n");
tty->print("input loop\n");
_lpt->dump_head();
_lpt->dump(); for (uint i = 0; i < _lpt->_body.size(); i++) {
_lpt->_body.at(i)->dump();
}
} #endif // Ready the block if (!construct_bb()) { returnfalse; // Exit if no interesting nodes or complex graph.
}
// build _dg, _disjoint_ptrs
dependence_graph();
// compute function depth(Node*)
compute_max_depth();
CountedLoopNode *cl = lpt()->_head->as_CountedLoop(); if (cl->is_main_loop()) { if (_do_vector_loop_experimental) { if (mark_generations() != -1) {
hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
if (!construct_bb()) { returnfalse; // Exit if no interesting nodes or complex graph.
}
dependence_graph();
compute_max_depth();
}
#ifndef PRODUCT if (TraceSuperWord) {
tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
_lpt->dump_head(); for (int j = 0; j < _block.length(); j++) {
Node* n = _block.at(j); int d = depth(n); for (int i = 0; i < d; i++) tty->print("%s", " ");
tty->print("%d :", d);
n->dump();
}
} #endif
}
compute_vector_element_type();
// Attempt vectorization
find_adjacent_refs();
if (align_to_ref() == NULL) { returnfalse; // Did not find memory reference to align vectors
}
extend_packlist();
if (_do_vector_loop_experimental) { if (_packset.length() == 0) { #ifndef PRODUCT if (TraceSuperWord) {
tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
} #endif
pack_parallel();
}
}
combine_packs();
construct_my_pack_map(); if (UseVectorCmov) {
merge_packs_to_cmove();
}
filter_packs();
schedule();
// Record eventual count of vector packs for checks in post loop vectorization if (PostLoopMultiversioning) {
cl->set_slp_pack_count(_packset.length());
}
} else {
assert(cl->is_rce_post_loop(), "Must be an rce'd post loop"); int saved_mapped_unroll_factor = cl->slp_max_unroll(); if (saved_mapped_unroll_factor) { int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
// now reset the slp_unroll_factor so that we can check the analysis mapped // what the vector loop was mapped to
cl->set_slp_max_unroll(0);
// do the analysis on the post loop
unrolling_analysis(vector_mapped_unroll_factor);
// if our analyzed loop is a canonical fit, start processing it if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) { // now add the vector nodes to packsets for (int i = 0; i < _post_block.length(); i++) {
Node* n = _post_block.at(i);
Node_List* singleton = new Node_List();
singleton->push(n);
_packset.append(singleton);
set_my_pack(n, singleton);
}
// map base types for vector usage
compute_vector_element_type();
} else { returnfalse;
}
} else { // for some reason we could not map the slp analysis state of the vectorized loop returnfalse;
}
}
return output();
}
//------------------------------find_adjacent_refs--------------------------- // Find the adjacent memory references and create pack pairs for them. // This is the initial set of packs that will then be extended by // following use->def and def->use links. The align positions are // assigned relative to the reference "align_to_ref" void SuperWord::find_adjacent_refs() { // Get list of memory operations
Node_List memops; for (int i = 0; i < _block.length(); i++) {
Node* n = _block.at(i); if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
is_java_primitive(n->as_Mem()->memory_type())) { int align = memory_alignment(n->as_Mem(), 0); if (align != bottom_align) {
memops.push(n);
}
}
} if (TraceSuperWord) {
tty->print_cr("\nfind_adjacent_refs found %d memops", memops.size());
}
Node_List align_to_refs; int max_idx; int best_iv_adjustment = 0;
MemNode* best_align_to_mem_ref = NULL;
while (memops.size() != 0) { // Find a memory reference to align to.
MemNode* mem_ref = find_align_to_ref(memops, max_idx); if (mem_ref == NULL) break;
align_to_refs.push(mem_ref); int iv_adjustment = get_iv_adjustment(mem_ref);
if (best_align_to_mem_ref == NULL) { // Set memory reference which is the best from all memory operations // to be used for alignment. The pre-loop trip count is modified to align // this reference to a vector-aligned address.
best_align_to_mem_ref = mem_ref;
best_iv_adjustment = iv_adjustment;
NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
}
SWPointer align_to_ref_p(mem_ref, this, NULL, false); // Set alignment relative to "align_to_ref" for all related memory operations. for (int i = memops.size() - 1; i >= 0; i--) {
MemNode* s = memops.at(i)->as_Mem(); if (isomorphic(s, mem_ref) &&
(!_do_vector_loop || same_origin_idx(s, mem_ref))) {
SWPointer p2(s, this, NULL, false); if (p2.comparable(align_to_ref_p)) { int align = memory_alignment(s, iv_adjustment);
set_alignment(s, align);
}
}
}
// Create initial pack pairs of memory operations for which // alignment is set and vectors will be aligned. bool create_pack = true; if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) { if (vectors_should_be_aligned()) { int vw = vector_width(mem_ref); int vw_best = vector_width(best_align_to_mem_ref); if (vw > vw_best) { // Do not vectorize a memory access with more elements per vector // if unaligned memory access is not allowed because number of // iterations in pre-loop will be not enough to align it.
create_pack = false;
} else {
SWPointer p2(best_align_to_mem_ref, this, NULL, false); if (!align_to_ref_p.invar_equals(p2)) { // Do not vectorize memory accesses with different invariants // if unaligned memory accesses are not allowed.
create_pack = false;
}
}
}
} else { if (same_velt_type(mem_ref, best_align_to_mem_ref)) { // Can't allow vectorization of unaligned memory accesses with the // same type since it could be overlapped accesses to the same array.
create_pack = false;
} else { // Allow independent (different type) unaligned memory operations // if HW supports them. if (vectors_should_be_aligned()) {
create_pack = false;
} else { // Check if packs of the same memory type but // with a different alignment were created before. for (uint i = 0; i < align_to_refs.size(); i++) {
MemNode* mr = align_to_refs.at(i)->as_Mem(); if (mr == mem_ref) { // Skip when we are looking at same memory operation. continue;
} if (same_velt_type(mr, mem_ref) &&
memory_alignment(mr, iv_adjustment) != 0)
create_pack = false;
}
}
}
} if (create_pack) { for (uint i = 0; i < memops.size(); i++) {
Node* s1 = memops.at(i); int align = alignment(s1); if (align == top_align) continue; for (uint j = 0; j < memops.size(); j++) {
Node* s2 = memops.at(j); if (alignment(s2) == top_align) continue; if (s1 != s2 && are_adjacent_refs(s1, s2)) { if (stmts_can_pack(s1, s2, align)) {
Node_List* pair = new Node_List();
pair->push(s1);
pair->push(s2); if (!_do_vector_loop || same_origin_idx(s1, s2)) {
_packset.append(pair);
}
}
}
}
}
} else { // Don't create unaligned pack // First, remove remaining memory ops of the same type from the list. for (int i = memops.size() - 1; i >= 0; i--) {
MemNode* s = memops.at(i)->as_Mem(); if (same_velt_type(s, mem_ref)) {
memops.remove(i);
}
}
// Second, remove already constructed packs of the same type. for (int i = _packset.length() - 1; i >= 0; i--) {
Node_List* p = _packset.at(i);
MemNode* s = p->at(0)->as_Mem(); if (same_velt_type(s, mem_ref)) {
remove_pack_at(i);
}
}
// If needed find the best memory reference for loop alignment again. if (same_velt_type(mem_ref, best_align_to_mem_ref)) { // Put memory ops from remaining packs back on memops list for // the best alignment search.
uint orig_msize = memops.size(); for (int i = 0; i < _packset.length(); i++) {
Node_List* p = _packset.at(i);
MemNode* s = p->at(0)->as_Mem();
assert(!same_velt_type(s, mem_ref), "sanity");
memops.push(s);
}
best_align_to_mem_ref = find_align_to_ref(memops, max_idx); if (best_align_to_mem_ref == NULL) { if (TraceSuperWord) {
tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
} // best_align_to_mem_ref will be used for adjusting the pre-loop limit in // SuperWord::align_initial_loop_index. Find one with the biggest vector size, // smallest data size and smallest iv offset from memory ops from remaining packs. if (_packset.length() > 0) { if (orig_msize == 0) {
best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
} else { for (uint i = 0; i < orig_msize; i++) {
memops.remove(0);
}
best_align_to_mem_ref = find_align_to_ref(memops, max_idx);
assert(best_align_to_mem_ref == NULL, "sanity");
best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
}
assert(best_align_to_mem_ref != NULL, "sanity");
} break;
}
best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);) // Restore list. while (memops.size() > orig_msize)
(void)memops.pop();
}
} // unaligned memory accesses
// Remove used mem nodes. for (int i = memops.size() - 1; i >= 0; i--) {
MemNode* m = memops.at(i)->as_Mem(); if (alignment(m) != top_align) {
memops.remove(i);
}
}
} // while (memops.size() != 0
set_align_to_ref(best_align_to_mem_ref);
if (TraceSuperWord) {
tty->print_cr("\nAfter find_adjacent_refs");
print_packset();
}
}
//------------------------------find_align_to_ref--------------------------- // Find a memory reference to align the loop induction variable to. // Looks first at stores then at loads, looking for a memory reference // with the largest number of references similar to it.
MemNode* SuperWord::find_align_to_ref(Node_List &memops, int &idx) {
GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
// Count number of comparable memory ops for (uint i = 0; i < memops.size(); i++) {
MemNode* s1 = memops.at(i)->as_Mem();
SWPointer p1(s1, this, NULL, false); // Only discard unalignable memory references if vector memory references // should be aligned on this platform. if (vectors_should_be_aligned() && !ref_is_alignable(p1)) {
*cmp_ct.adr_at(i) = 0; continue;
} for (uint j = i+1; j < memops.size(); j++) {
MemNode* s2 = memops.at(j)->as_Mem(); if (isomorphic(s1, s2)) {
SWPointer p2(s2, this, NULL, false); if (p1.comparable(p2)) {
(*cmp_ct.adr_at(i))++;
(*cmp_ct.adr_at(j))++;
}
}
}
}
// Find Store (or Load) with the greatest number of "comparable" references, // biggest vector size, smallest data size and smallest iv offset. int max_ct = 0; int max_vw = 0; int max_idx = -1; int min_size = max_jint; int min_iv_offset = max_jint; for (uint j = 0; j < memops.size(); j++) {
MemNode* s = memops.at(j)->as_Mem(); if (s->is_Store()) { int vw = vector_width_in_bytes(s);
assert(vw > 1, "sanity");
SWPointer p(s, this, NULL, false); if ( cmp_ct.at(j) > max_ct ||
(cmp_ct.at(j) == max_ct &&
( vw > max_vw ||
(vw == max_vw &&
( data_size(s) < min_size ||
(data_size(s) == min_size &&
p.offset_in_bytes() < min_iv_offset)))))) {
max_ct = cmp_ct.at(j);
max_vw = vw;
max_idx = j;
min_size = data_size(s);
min_iv_offset = p.offset_in_bytes();
}
}
} // If no stores, look at loads if (max_ct == 0) { for (uint j = 0; j < memops.size(); j++) {
MemNode* s = memops.at(j)->as_Mem(); if (s->is_Load()) { int vw = vector_width_in_bytes(s);
assert(vw > 1, "sanity");
SWPointer p(s, this, NULL, false); if ( cmp_ct.at(j) > max_ct ||
(cmp_ct.at(j) == max_ct &&
( vw > max_vw ||
(vw == max_vw &&
( data_size(s) < min_size ||
(data_size(s) == min_size &&
p.offset_in_bytes() < min_iv_offset)))))) {
max_ct = cmp_ct.at(j);
max_vw = vw;
max_idx = j;
min_size = data_size(s);
min_iv_offset = p.offset_in_bytes();
}
}
}
}
#ifdef ASSERT if (TraceSuperWord && Verbose) {
tty->print_cr("\nVector memops after find_align_to_ref"); for (uint i = 0; i < memops.size(); i++) {
MemNode* s = memops.at(i)->as_Mem();
s->dump();
}
} #endif
idx = max_idx; if (max_ct > 0) { #ifdef ASSERT if (TraceSuperWord) {
tty->print("\nVector align to node: ");
memops.at(max_idx)->as_Mem()->dump();
} #endif return memops.at(max_idx)->as_Mem();
} return NULL;
}
//------------------span_works_for_memory_size----------------------------- staticbool span_works_for_memory_size(MemNode* mem, int span, int mem_size, int offset) { bool span_matches_memory = false; if ((mem_size == type2aelembytes(T_BYTE) || mem_size == type2aelembytes(T_SHORT))
&& ABS(span) == type2aelembytes(T_INT)) { // There is a mismatch on span size compared to memory. for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
Node* use = mem->fast_out(j); if (!VectorNode::is_type_transition_to_int(use)) { returnfalse;
}
} // If all uses transition to integer, it means that we can successfully align even on mismatch. returntrue;
} else {
span_matches_memory = ABS(span) == mem_size;
} return span_matches_memory && (ABS(offset) % mem_size) == 0;
}
//------------------------------ref_is_alignable--------------------------- // Can the preloop align the reference to position zero in the vector? bool SuperWord::ref_is_alignable(SWPointer& p) { if (!p.has_iv()) { returntrue; // no induction variable
}
CountedLoopEndNode* pre_end = pre_loop_end();
assert(pre_end->stride_is_con(), "pre loop stride is constant"); int preloop_stride = pre_end->stride_con();
int span = preloop_stride * p.scale_in_bytes(); int mem_size = p.memory_size(); int offset = p.offset_in_bytes(); // Stride one accesses are alignable if offset is aligned to memory operation size. // Offset can be unaligned when UseUnalignedAccesses is used. if (span_works_for_memory_size(p.mem(), span, mem_size, offset)) { returntrue;
} // If the initial offset from start of the object is computable, // check if the pre-loop can align the final offset accordingly. // // In other words: Can we find an i such that the offset // after i pre-loop iterations is aligned to vw? // (init_offset + pre_loop) % vw == 0 (1) // where // pre_loop = i * span // is the number of bytes added to the offset by i pre-loop iterations. // // For this to hold we need pre_loop to increase init_offset by // pre_loop = vw - (init_offset % vw) // // This is only possible if pre_loop is divisible by span because each // pre-loop iteration increases the initial offset by 'span' bytes: // (vw - (init_offset % vw)) % span == 0 // int vw = vector_width_in_bytes(p.mem());
assert(vw > 1, "sanity");
Node* init_nd = pre_end->init_trip(); if (init_nd->is_Con() && p.invar() == NULL) { int init = init_nd->bottom_type()->is_int()->get_con(); int init_offset = init * p.scale_in_bytes() + offset; if (init_offset < 0) { // negative offset from object start? returnfalse; // may happen in dead loop
} if (vw % span == 0) { // If vm is a multiple of span, we use formula (1). if (span > 0) { return (vw - (init_offset % vw)) % span == 0;
} else {
assert(span < 0, "nonzero stride * scale"); return (init_offset % vw) % -span == 0;
}
} elseif (span % vw == 0) { // If span is a multiple of vw, we can simplify formula (1) to: // (init_offset + i * span) % vw == 0 // => // (init_offset % vw) + ((i * span) % vw) == 0 // => // init_offset % vw == 0 // // Because we add a multiple of vw to the initial offset, the final // offset is a multiple of vw if and only if init_offset is a multiple. // return (init_offset % vw) == 0;
}
} returnfalse;
} //---------------------------get_vw_bytes_special------------------------ int SuperWord::get_vw_bytes_special(MemNode* s) { // Get the vector width in bytes. int vw = vector_width_in_bytes(s);
// Check for special case where there is an MulAddS2I usage where short vectors are going to need combined.
BasicType btype = velt_basic_type(s); if (type2aelembytes(btype) == 2) { bool should_combine_adjacent = true; for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
Node* user = s->fast_out(i); if (!VectorNode::is_muladds2i(user)) {
should_combine_adjacent = false;
}
} if (should_combine_adjacent) {
vw = MIN2(max_vector_size(btype)*type2aelembytes(btype), vw * 2);
}
}
// Check for special case where there is a type conversion between different data size. int vectsize = max_vector_size_in_def_use_chain(s); if (vectsize < max_vector_size(btype)) {
vw = MIN2(vectsize * type2aelembytes(btype), vw);
}
return vw;
}
//---------------------------get_iv_adjustment--------------------------- // Calculate loop's iv adjustment for this memory ops. int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
SWPointer align_to_ref_p(mem_ref, this, NULL, false); int offset = align_to_ref_p.offset_in_bytes(); int scale = align_to_ref_p.scale_in_bytes(); int elt_size = align_to_ref_p.memory_size(); int vw = get_vw_bytes_special(mem_ref);
assert(vw > 1, "sanity"); int iv_adjustment; if (scale != 0) { int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1; // At least one iteration is executed in pre-loop by default. As result // several iterations are needed to align memory operations in main-loop even // if offset is 0. int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw)); // iv_adjustment_in_bytes must be a multiple of elt_size if vector memory // references should be aligned on this platform.
assert((ABS(iv_adjustment_in_bytes) % elt_size) == 0 || !vectors_should_be_aligned(), "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
iv_adjustment = iv_adjustment_in_bytes/elt_size;
} else { // This memory op is not dependent on iv (scale == 0)
iv_adjustment = 0;
}
//---------------------------dependence_graph--------------------------- // Construct dependency graph. // Add dependence edges to load/store nodes for memory dependence // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x) void SuperWord::dependence_graph() {
CountedLoopNode *cl = lpt()->_head->as_CountedLoop(); // First, assign a dependence node to each memory node for (int i = 0; i < _block.length(); i++ ) {
Node *n = _block.at(i); if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
_dg.make_node(n);
}
}
// For each memory slice, create the dependences for (int i = 0; i < _mem_slice_head.length(); i++) {
Node* n = _mem_slice_head.at(i);
Node* n_tail = _mem_slice_tail.at(i);
// Get slice in predecessor order (last is first) if (cl->is_main_loop()) {
mem_slice_preds(n_tail, n, _nlist);
}
#ifndef PRODUCT if(TraceSuperWord && Verbose) {
tty->print_cr("SuperWord::dependence_graph: built a new mem slice"); for (int j = _nlist.length() - 1; j >= 0 ; j--) {
_nlist.at(j)->dump();
}
} #endif // Make the slice dependent on the root
DepMem* slice = _dg.dep(n);
_dg.make_edge(_dg.root(), slice);
// Create a sink for the slice
DepMem* slice_sink = _dg.make_node(NULL);
_dg.make_edge(slice_sink, _dg.tail());
// Now visit each pair of memory ops, creating the edges for (int j = _nlist.length() - 1; j >= 0 ; j--) {
Node* s1 = _nlist.at(j);
// If no dependency yet, use slice if (_dg.dep(s1)->in_cnt() == 0) {
_dg.make_edge(slice, s1);
}
SWPointer p1(s1->as_Mem(), this, NULL, false); bool sink_dependent = true; for (int k = j - 1; k >= 0; k--) {
Node* s2 = _nlist.at(k); if (s1->is_Load() && s2->is_Load()) continue;
SWPointer p2(s2->as_Mem(), this, NULL, false);
int cmp = p1.cmp(p2); if (SuperWordRTDepCheck &&
p1.base() != p2.base() && p1.valid() && p2.valid()) { // Trace disjoint pointers
OrderedPair pp(p1.base(), p2.base());
_disjoint_ptrs.append_if_missing(pp);
} if (!SWPointer::not_equal(cmp)) { // Possibly same address
_dg.make_edge(s1, s2);
sink_dependent = false;
}
} if (sink_dependent) {
_dg.make_edge(s1, slice_sink);
}
}
if (TraceSuperWord) {
tty->print_cr("\nDependence graph for slice: %d", n->_idx); for (int q = 0; q < _nlist.length(); q++) {
_dg.print(_nlist.at(q));
}
tty->cr();
}
_nlist.clear();
}
if (TraceSuperWord) {
tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE"); for (int r = 0; r < _disjoint_ptrs.length(); r++) {
_disjoint_ptrs.at(r).print();
tty->cr();
}
tty->cr();
}
}
//---------------------------mem_slice_preds--------------------------- // Return a memory slice (node list) in predecessor order starting at "start" void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
assert(preds.length() == 0, "start empty");
Node* n = start;
Node* prev = NULL; while (true) {
NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
assert(in_bb(n), "must be in block"); for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* out = n->fast_out(i); if (out->is_Load()) { if (in_bb(out)) {
preds.push(out); if (TraceSuperWord && Verbose) {
tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
}
}
} else { // FIXME if (out->is_MergeMem() && !in_bb(out)) { // Either unrolling is causing a memory edge not to disappear, // or need to run igvn.optimize() again before SLP
} elseif (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) { // Ditto. Not sure what else to check further.
} elseif (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) { // StoreCM has an input edge used as a precedence edge. // Maybe an issue when oop stores are vectorized.
} else {
assert(out == prev || prev == NULL, "no branches off of store slice");
}
}//else
}//for if (n == stop) break;
preds.push(n); if (TraceSuperWord && Verbose) {
tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
}
prev = n;
assert(n->is_Mem(), "unexpected node %s", n->Name());
n = n->in(MemNode::Memory);
}
}
//------------------------------stmts_can_pack--------------------------- // Can s1 and s2 be in a pack with s1 immediately preceding s2 and // s1 aligned at "align" bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
// Do not use superword for non-primitives
BasicType bt1 = velt_basic_type(s1);
BasicType bt2 = velt_basic_type(s2); if(!is_java_primitive(bt1) || !is_java_primitive(bt2)) returnfalse;
BasicType longer_bt = longer_type_for_conversion(s1); if (max_vector_size(bt1) < 2 ||
(longer_bt != T_ILLEGAL && max_vector_size(longer_bt) < 2)) { returnfalse; // No vectors for this type
}
if (isomorphic(s1, s2)) { if ((independent(s1, s2) && have_similar_inputs(s1, s2)) || reduction(s1, s2)) { if (!exists_at(s1, 0) && !exists_at(s2, 1)) { if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) { int s1_align = alignment(s1); int s2_align = alignment(s2); if (s1_align == top_align || s1_align == align) { if (s2_align == top_align || s2_align == align + data_size(s1)) { returntrue;
}
}
}
}
}
} returnfalse;
}
//------------------------------exists_at--------------------------- // Does s exist in a pack at position pos? bool SuperWord::exists_at(Node* s, uint pos) { for (int i = 0; i < _packset.length(); i++) {
Node_List* p = _packset.at(i); if (p->at(pos) == s) { returntrue;
}
} returnfalse;
}
//------------------------------are_adjacent_refs--------------------------- // Is s1 immediately before s2 in memory? bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) { if (!s1->is_Mem() || !s2->is_Mem()) returnfalse; if (!in_bb(s1) || !in_bb(s2)) returnfalse;
// Do not use superword for non-primitives if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
!is_java_primitive(s2->as_Mem()->memory_type())) { returnfalse;
}
// FIXME - co_locate_pack fails on Stores in different mem-slices, so // only pack memops that are in the same alias set until that's fixed. if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
_phase->C->get_alias_index(s2->as_Mem()->adr_type())) returnfalse;
SWPointer p1(s1->as_Mem(), this, NULL, false);
SWPointer p2(s2->as_Mem(), this, NULL, false); if (p1.base() != p2.base() || !p1.comparable(p2)) returnfalse; int diff = p2.offset_in_bytes() - p1.offset_in_bytes(); return diff == data_size(s1);
}
//------------------------------isomorphic--------------------------- // Are s1 and s2 similar? bool SuperWord::isomorphic(Node* s1, Node* s2) { if (s1->Opcode() != s2->Opcode()) returnfalse; if (s1->req() != s2->req()) returnfalse; if (!same_velt_type(s1, s2)) returnfalse;
Node* s1_ctrl = s1->in(0);
Node* s2_ctrl = s2->in(0); // If the control nodes are equivalent, no further checks are required to test for isomorphism. if (s1_ctrl == s2_ctrl) { returntrue;
} else { bool s1_ctrl_inv = ((s1_ctrl == NULL) ? true : lpt()->is_invariant(s1_ctrl)); bool s2_ctrl_inv = ((s2_ctrl == NULL) ? true : lpt()->is_invariant(s2_ctrl)); // If the control nodes are not invariant for the loop, fail isomorphism test. if (!s1_ctrl_inv || !s2_ctrl_inv) { returnfalse;
} if(s1_ctrl != NULL && s2_ctrl != NULL) { if (s1_ctrl->is_Proj()) {
s1_ctrl = s1_ctrl->in(0);
assert(lpt()->is_invariant(s1_ctrl), "must be invariant");
} if (s2_ctrl->is_Proj()) {
s2_ctrl = s2_ctrl->in(0);
assert(lpt()->is_invariant(s2_ctrl), "must be invariant");
} if (!s1_ctrl->is_RangeCheck() || !s2_ctrl->is_RangeCheck()) { returnfalse;
}
} // Control nodes are invariant. However, we have no way of checking whether they resolve // in an equivalent manner. But, we know that invariant range checks are guaranteed to // throw before the loop (if they would have thrown). Thus, the loop would not have been reached. // Therefore, if the control nodes for both are range checks, we accept them to be isomorphic. for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
Node* t1 = s1->fast_out(i); for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
Node* t2 = s2->fast_out(j); if (VectorNode::is_muladds2i(t1) && VectorNode::is_muladds2i(t2)) { returntrue;
}
}
}
} returnfalse;
}
//------------------------------independent--------------------------- // Is there no data path from s1 to s2 or s2 to s1? bool SuperWord::independent(Node* s1, Node* s2) { // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first"); int d1 = depth(s1); int d2 = depth(s2); if (d1 == d2) return s1 != s2;
Node* deep = d1 > d2 ? s1 : s2;
Node* shallow = d1 > d2 ? s2 : s1;
visited_clear();
return independent_path(shallow, deep);
}
//--------------------------have_similar_inputs----------------------- // For a node pair (s1, s2) which is isomorphic and independent, // do s1 and s2 have similar input edges? bool SuperWord::have_similar_inputs(Node* s1, Node* s2) { // assert(isomorphic(s1, s2) == true, "check isomorphic"); // assert(independent(s1, s2) == true, "check independent"); if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) { for (uint i = 1; i < s1->req(); i++) {
Node* s1_in = s1->in(i);
Node* s2_in = s2->in(i); if (s1_in->is_Phi() && s2_in->is_Add() && s2_in->in(1) == s1_in) { // Special handling for expressions with loop iv, like "b[i] = a[i] * i". // In this case, one node has an input from the tripcount iv and another // node has an input from iv plus an offset. if (!s1_in->as_Phi()->is_tripcount(T_INT)) returnfalse;
} else { if (s1_in->Opcode() != s2_in->Opcode()) returnfalse;
}
}
} returntrue;
}
//------------------------------reduction--------------------------- // Is there a data path between s1 and s2 and the nodes reductions? bool SuperWord::reduction(Node* s1, Node* s2) { bool retValue = false; int d1 = depth(s1); int d2 = depth(s2); if (d2 > d1) { if (s1->is_reduction() && s2->is_reduction()) { // This is an ordered set, so s1 should define s2 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
Node* t1 = s1->fast_out(i); if (t1 == s2) { // both nodes are reductions and connected
retValue = true;
}
}
}
}
return retValue;
}
//------------------------------independent_path------------------------------ // Helper for independent bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) { if (dp >= 1000) returnfalse; // stop deep recursion
visited_set(deep); int shal_depth = depth(shallow);
assert(shal_depth <= depth(deep), "must be"); for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
Node* pred = preds.current(); if (in_bb(pred) && !visited_test(pred)) { if (shallow == pred) { returnfalse;
} if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) { returnfalse;
}
}
} returntrue;
}
//------------------------------data_size--------------------------- int SuperWord::data_size(Node* s) {
Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov if (UseVectorCmov) {
use = _cmovev_kit.is_Bool_candidate(s); if (use != NULL) { return data_size(use);
}
use = _cmovev_kit.is_Cmp_candidate(s); if (use != NULL) { return data_size(use);
}
}
//------------------------------extend_packlist--------------------------- // Extend packset by following use->def and def->use links from pack members. void SuperWord::extend_packlist() { bool changed; do {
packset_sort(_packset.length());
changed = false; for (int i = 0; i < _packset.length(); i++) {
Node_List* p = _packset.at(i);
changed |= follow_use_defs(p);
changed |= follow_def_uses(p);
}
} while (changed);
if (_race_possible) { for (int i = 0; i < _packset.length(); i++) {
Node_List* p = _packset.at(i);
order_def_uses(p);
}
}
if (TraceSuperWord) {
tty->print_cr("\nAfter extend_packlist");
print_packset();
}
}
//------------------------------adjust_alignment_for_type_conversion--------------------------------- // Adjust the target alignment if conversion between different data size exists in def-use nodes. int SuperWord::adjust_alignment_for_type_conversion(Node* s, Node* t, int align) { // Do not use superword for non-primitives
BasicType bt1 = velt_basic_type(s);
BasicType bt2 = velt_basic_type(t); if (!is_java_primitive(bt1) || !is_java_primitive(bt2)) { return align;
} if (longer_type_for_conversion(s) != T_ILLEGAL ||
longer_type_for_conversion(t) != T_ILLEGAL) {
align = align / data_size(s) * data_size(t);
} return align;
}
//------------------------------follow_use_defs--------------------------- // Extend the packset by visiting operand definitions of nodes in pack p bool SuperWord::follow_use_defs(Node_List* p) {
assert(p->size() == 2, "just checking");
Node* s1 = p->at(0);
Node* s2 = p->at(1);
assert(s1->req() == s2->req(), "just checking");
assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
//------------------------------follow_def_uses--------------------------- // Extend the packset by visiting uses of nodes in pack p bool SuperWord::follow_def_uses(Node_List* p) { bool changed = false;
Node* s1 = p->at(0);
Node* s2 = p->at(1);
assert(p->size() == 2, "just checking");
assert(s1->req() == s2->req(), "just checking");
assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
if (s1->is_Store()) returnfalse;
int align = alignment(s1);
NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);) int savings = -1; int num_s1_uses = 0;
Node* u1 = NULL;
Node* u2 = NULL; for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
Node* t1 = s1->fast_out(i);
num_s1_uses++; if (!in_bb(t1)) continue; for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
Node* t2 = s2->fast_out(j); if (!in_bb(t2)) continue; if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv if (!opnd_positions_match(s1, t1, s2, t2)) continue; int adjusted_align = alignment(s1);
adjusted_align = adjust_alignment_for_type_conversion(s1, t1, adjusted_align); if (stmts_can_pack(t1, t2, adjusted_align)) { int my_savings = est_savings(t1, t2); if (my_savings > savings) {
savings = my_savings;
u1 = t1;
u2 = t2;
align = adjusted_align;
}
}
}
} if (num_s1_uses > 1) {
_race_possible = true;
} if (savings >= 0) {
Node_List* pair = new Node_List();
pair->push(u1);
pair->push(u2);
_packset.append(pair);
NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
set_alignment(u1, u2, align);
changed = true;
} return changed;
}
//------------------------------order_def_uses--------------------------- // For extended packsets, ordinally arrange uses packset by major component void SuperWord::order_def_uses(Node_List* p) {
Node* s1 = p->at(0);
if (s1->is_Store()) return;
// reductions are always managed beforehand if (s1->is_reduction()) return;
for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
Node* t1 = s1->fast_out(i);
// Only allow operand swap on commuting operations if (!t1->is_Add() && !t1->is_Mul() && !VectorNode::is_muladds2i(t1)) { break;
}
// Now find t1's packset
Node_List* p2 = NULL; for (int j = 0; j < _packset.length(); j++) {
p2 = _packset.at(j);
Node* first = p2->at(0); if (t1 == first) { break;
}
p2 = NULL;
} // Arrange all sub components by the major component if (p2 != NULL) { for (uint j = 1; j < p->size(); j++) {
Node* d1 = p->at(j);
Node* u1 = p2->at(j);
opnd_positions_match(s1, t1, d1, u1);
}
}
}
}
//---------------------------opnd_positions_match------------------------- // Is the use of d1 in u1 at the same operand position as d2 in u2? bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) { // check reductions to see if they are marshalled to represent the reduction // operator in a specified opnd if (u1->is_reduction() && u2->is_reduction()) { // ensure reductions have phis and reduction definitions feeding the 1st operand
Node* first = u1->in(2); if (first->is_Phi() || first->is_reduction()) {
u1->swap_edges(1, 2);
} // ensure reductions have phis and reduction definitions feeding the 1st operand
first = u2->in(2); if (first->is_Phi() || first->is_reduction()) {
u2->swap_edges(1, 2);
} returntrue;
}
uint ct = u1->req(); if (ct != u2->req()) returnfalse;
uint i1 = 0;
uint i2 = 0; do { for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break; for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break; if (i1 != i2) { if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) { // Further analysis relies on operands position matching.
u2->swap_edges(i1, i2);
} elseif (VectorNode::is_muladds2i(u2) && u1 != u2) { if (i1 == 5 - i2) { // ((i1 == 3 && i2 == 2) || (i1 == 2 && i2 == 3) || (i1 == 1 && i2 == 4) || (i1 == 4 && i2 == 1))
u2->swap_edges(1, 2);
u2->swap_edges(3, 4);
} if (i1 == 3 - i2 || i1 == 7 - i2) { // ((i1 == 1 && i2 == 2) || (i1 == 2 && i2 == 1) || (i1 == 3 && i2 == 4) || (i1 == 4 && i2 == 3))
u2->swap_edges(2, 3);
u2->swap_edges(1, 4);
} returnfalse; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
} else { returnfalse;
}
} elseif (i1 == i2 && VectorNode::is_muladds2i(u2) && u1 != u2) {
u2->swap_edges(1, 3);
u2->swap_edges(2, 4); returnfalse; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
}
} while (i1 < ct); returntrue;
}
//------------------------------est_savings--------------------------- // Estimate the savings from executing s1 and s2 as a pack int SuperWord::est_savings(Node* s1, Node* s2) { int save_in = 2 - 1; // 2 operations per instruction in packed form
// inputs for (uint i = 1; i < s1->req(); i++) {
Node* x1 = s1->in(i);
Node* x2 = s2->in(i); if (x1 != x2) { if (are_adjacent_refs(x1, x2)) {
save_in += adjacent_profit(x1, x2);
} elseif (!in_packset(x1, x2)) {
save_in -= pack_cost(2);
} else {
save_in += unpack_cost(2);
}
}
}
// uses of result
uint ct = 0; int save_use = 0; for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
Node* s1_use = s1->fast_out(i); for (int j = 0; j < _packset.length(); j++) {
Node_List* p = _packset.at(j); if (p->at(0) == s1_use) { for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
Node* s2_use = s2->fast_out(k); if (p->at(p->size()-1) == s2_use) {
ct++; if (are_adjacent_refs(s1_use, s2_use)) {
save_use += adjacent_profit(s1_use, s2_use);
}
}
}
}
}
}
if (ct < s1->outcnt()) save_use += unpack_cost(1); if (ct < s2->outcnt()) save_use += unpack_cost(1);
return MAX2(save_in, save_use);
}
//------------------------------costs--------------------------- int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; } int SuperWord::pack_cost(int ct) { return ct; } int SuperWord::unpack_cost(int ct) { return ct; }
//------------------------------combine_packs--------------------------- // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last void SuperWord::combine_packs() { bool changed = true; // Combine packs regardless max vector size. while (changed) {
changed = false; for (int i = 0; i < _packset.length(); i++) {
Node_List* p1 = _packset.at(i); if (p1 == NULL) continue; // Because of sorting we can start at i + 1 for (int j = i + 1; j < _packset.length(); j++) {
Node_List* p2 = _packset.at(j); if (p2 == NULL) continue; if (i == j) continue; if (p1->at(p1->size()-1) == p2->at(0)) { for (uint k = 1; k < p2->size(); k++) {
p1->push(p2->at(k));
}
_packset.at_put(j, NULL);
changed = true;
}
}
}
}
// Split packs which have size greater then max vector size. for (int i = 0; i < _packset.length(); i++) {
Node_List* p1 = _packset.at(i); if (p1 != NULL) {
uint max_vlen = max_vector_size_in_def_use_chain(p1->at(0)); // Max elements in vector
assert(is_power_of_2(max_vlen), "sanity");
uint psize = p1->size(); if (!is_power_of_2(psize)) { // Skip pack which can't be vector. // case1: for(...) { a[i] = i; } elements values are different (i+x) // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
_packset.at_put(i, NULL); continue;
} if (psize > max_vlen) {
Node_List* pack = new Node_List(); for (uint j = 0; j < psize; j++) {
pack->push(p1->at(j)); if (pack->size() >= max_vlen) {
assert(is_power_of_2(pack->size()), "sanity");
_packset.append(pack); pack = new Node_List();
}
}
_packset.at_put(i, NULL);
}
}
}
// Compress list. for (int i = _packset.length() - 1; i >= 0; i--) {
Node_List* p1 = _packset.at(i); if (p1 == NULL) {
_packset.remove_at(i);
}
}
if (TraceSuperWord) {
tty->print_cr("\nAfter combine_packs");
print_packset();
}
}
//-----------------------------construct_my_pack_map-------------------------- // Construct the map from nodes to packs. Only valid after the // point where a node is only in one pack (after combine_packs). void SuperWord::construct_my_pack_map() {
Node_List* rslt = NULL; for (int i = 0; i < _packset.length(); i++) {
Node_List* p = _packset.at(i); for (uint j = 0; j < p->size(); j++) {
Node* s = p->at(j); #ifdef ASSERT if (my_pack(s) != NULL) {
s->dump(1);
tty->print_cr("packs[%d]:", i);
print_pack(p);
assert(false, "only in one pack");
} #endif
set_my_pack(s, p);
}
}
}
//------------------------------filter_packs--------------------------- // Remove packs that are not implemented or not profitable. void SuperWord::filter_packs() { // Remove packs that are not implemented for (int i = _packset.length() - 1; i >= 0; i--) {
Node_List* pk = _packset.at(i); bool impl = implemented(pk); if (!impl) { #ifndef PRODUCT if ((TraceSuperWord && Verbose) || _vector_loop_debug) {
tty->print_cr("Unimplemented");
pk->at(0)->dump();
} #endif
remove_pack_at(i);
}
Node *n = pk->at(0); if (n->is_reduction()) {
_num_reductions++;
} else {
_num_work_vecs++;
}
}
// Remove packs that are not profitable bool changed; do {
changed = false; for (int i = _packset.length() - 1; i >= 0; i--) {
Node_List* pk = _packset.at(i); bool prof = profitable(pk); if (!prof) { #ifndef PRODUCT if ((TraceSuperWord && Verbose) || _vector_loop_debug) {
tty->print_cr("Unprofitable");
pk->at(0)->dump();
} #endif
remove_pack_at(i);
changed = true;
}
}
} while (changed);
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.