Mercurial > hg > graal-compiler
diff src/share/vm/opto/loopTransform.cpp @ 4137:04b9a2566eec
Merge with hsx23/hotspot.
author | Thomas Wuerthinger <thomas.wuerthinger@oracle.com> |
---|---|
date | Sat, 17 Dec 2011 21:40:27 +0100 |
parents | da6a29fb0da5 |
children | 5e990493719e |
line wrap: on
line diff
--- a/src/share/vm/opto/loopTransform.cpp Sat Dec 17 20:50:09 2011 +0100 +++ b/src/share/vm/opto/loopTransform.cpp Sat Dec 17 21:40:27 2011 +0100 @@ -509,14 +509,13 @@ // backedges) and then map to the new peeled iteration. This leaves // the pre-loop with only 1 user (the new peeled iteration), but the // peeled-loop backedge has 2 users. - Node* new_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx]; - new_exit_value = move_loop_predicates(entry, new_exit_value, !counted_loop); + Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; _igvn.hash_delete(head); - head->set_req(LoopNode::EntryControl, new_exit_value); + head->set_req(LoopNode::EntryControl, new_entry); for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { Node* old = head->fast_out(j); if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { - new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; + Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; if (!new_exit_value ) // Backedge value is ALSO loop invariant? // Then loop body backedge value remains the same. new_exit_value = old->in(LoopNode::LoopBackControl); @@ -710,10 +709,13 @@ // Adjust body_size to determine if we unroll or not uint body_size = _body.size(); + // Key test to unroll loop in CRC32 java code + int xors_in_loop = 0; // Also count ModL, DivL and MulL which expand mightly for (uint k = 0; k < _body.size(); k++) { Node* n = _body.at(k); switch (n->Opcode()) { + case Op_XorI: xors_in_loop++; break; // CRC32 java code case Op_ModL: body_size += 30; break; case Op_DivL: body_size += 30; break; case Op_MulL: body_size += 10; break; @@ -730,7 +732,8 @@ // Check for being too big if (body_size > (uint)LoopUnrollLimit) { - // Normal case: loop too big + if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; + // Normal case: loop too big return false; } @@ -824,13 +827,23 @@ //------------------------------clone_up_backedge_goo-------------------------- // If Node n lives in the back_ctrl block and cannot float, we clone a private // version of n in preheader_ctrl block and return that, otherwise return n. -Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) { +Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { if( get_ctrl(n) != back_ctrl ) return n; + // Only visit once + if (visited.test_set(n->_idx)) { + Node *x = clones.find(n->_idx); + if (x != NULL) + return x; + return n; + } + Node *x = NULL; // If required, a clone of 'n' // Check for 'n' being pinned in the backedge. if( n->in(0) && n->in(0) == back_ctrl ) { + assert(clones.find(n->_idx) == NULL, "dead loop"); x = n->clone(); // Clone a copy of 'n' to preheader + clones.push(x, n->_idx); x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader } @@ -838,10 +851,13 @@ // If there are no changes we can just return 'n', otherwise // we need to clone a private copy and change it. for( uint i = 1; i < n->req(); i++ ) { - Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) ); + Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); if( g != n->in(i) ) { - if( !x ) + if( !x ) { + assert(clones.find(n->_idx) == NULL, "dead loop"); x = n->clone(); + clones.push(x, n->_idx); + } x->set_req(i, g); } } @@ -960,6 +976,9 @@ post_head->set_req(LoopNode::EntryControl, zer_taken); set_idom(post_head, zer_taken, dd_main_exit); + Arena *a = Thread::current()->resource_area(); + VectorSet visited(a); + Node_Stack clones(a, main_head->back_control()->outcnt()); // Step A3: Make the fall-in values to the post-loop come from the // fall-out values of the main-loop. for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { @@ -968,7 +987,8 @@ Node *post_phi = old_new[main_phi->_idx]; Node *fallmain = clone_up_backedge_goo(main_head->back_control(), post_head->init_control(), - main_phi->in(LoopNode::LoopBackControl)); + main_phi->in(LoopNode::LoopBackControl), + visited, clones); _igvn.hash_delete(post_phi); post_phi->set_req( LoopNode::EntryControl, fallmain ); } @@ -1032,6 +1052,8 @@ main_head->set_req(LoopNode::EntryControl, min_taken); set_idom(main_head, min_taken, dd_main_head); + visited.Clear(); + clones.clear(); // Step B3: Make the fall-in values to the main-loop come from the // fall-out values of the pre-loop. for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { @@ -1040,7 +1062,8 @@ Node *pre_phi = old_new[main_phi->_idx]; Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), main_head->init_control(), - pre_phi->in(LoopNode::LoopBackControl)); + pre_phi->in(LoopNode::LoopBackControl), + visited, clones); _igvn.hash_delete(main_phi); main_phi->set_req( LoopNode::EntryControl, fallpre ); } @@ -2080,7 +2103,7 @@ if (!_head->is_CountedLoop()) return false; // Dead loop CountedLoopNode *cl = _head->as_CountedLoop(); - if (!cl->loopexit()) + if (!cl->is_valid_counted_loop()) return false; // Malformed loop if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) return false; // Infinite loop @@ -2236,7 +2259,7 @@ } CountedLoopNode *cl = _head->as_CountedLoop(); - if (!cl->loopexit()) return true; // Ignore various kinds of broken loops + if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops // Do nothing special to pre- and post- loops if (cl->is_pre_loop() || cl->is_post_loop()) return true; @@ -2617,7 +2640,7 @@ // Must have constant stride CountedLoopNode* head = lpt->_head->as_CountedLoop(); - if (!head->stride_is_con() || !head->is_normal_loop()) { + if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { return false; }