1.1 --- a/src/share/vm/opto/loopTransform.cpp Fri Sep 28 14:24:22 2018 +0200
1.2 +++ b/src/share/vm/opto/loopTransform.cpp Fri Dec 14 11:22:26 2018 +0100
1.3 @@ -1537,13 +1537,20 @@
1.4
1.5 //------------------------------adjust_limit-----------------------------------
1.6 // Helper function for add_constraint().
1.7 -Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1.8 +Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up) {
1.9 // Compute "I :: (limit-offset)/scale"
1.10 Node *con = new (C) SubINode(rc_limit, offset);
1.11 register_new_node(con, pre_ctrl);
1.12 Node *X = new (C) DivINode(0, con, scale);
1.13 register_new_node(X, pre_ctrl);
1.14
1.15 + // When the absolute value of scale is greater than one, the integer
1.16 + // division may round limit down so add one to the limit.
1.17 + if (round_up) {
1.18 + X = new (C) AddINode(X, _igvn.intcon(1));
1.19 + register_new_node(X, pre_ctrl);
1.20 + }
1.21 +
1.22 // Adjust loop limit
1.23 loop_limit = (stride_con > 0)
1.24 ? (Node*)(new (C) MinINode(loop_limit, X))
1.25 @@ -1584,7 +1591,7 @@
1.26 // (upper_limit-offset) may overflow or underflow.
1.27 // But it is fine since main loop will either have
1.28 // less iterations or will be skipped in such case.
1.29 - *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1.30 + *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl, false);
1.31
1.32 // The underflow limit: low_limit <= scale*I+offset.
1.33 // For pre-loop compute
1.34 @@ -1620,7 +1627,8 @@
1.35 // max(pre_limit, original_limit) is used in do_range_check().
1.36 }
1.37 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1.38 - *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1.39 + *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl,
1.40 + scale_con > 1 && stride_con > 0);
1.41
1.42 } else { // stride_con*scale_con < 0
1.43 // For negative stride*scale pre-loop checks for overflow and
1.44 @@ -1646,7 +1654,8 @@
1.45 Node *plus_one = new (C) AddINode(offset, one);
1.46 register_new_node( plus_one, pre_ctrl );
1.47 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1.48 - *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1.49 + *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl,
1.50 + scale_con < -1 && stride_con > 0);
1.51
1.52 if (low_limit->get_int() == -max_jint) {
1.53 if (!RangeLimitCheck) return;
1.54 @@ -1681,7 +1690,8 @@
1.55 // I > (low_limit-(offset+1))/scale
1.56 // )
1.57
1.58 - *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1.59 + *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl,
1.60 + false);
1.61 }
1.62 }
1.63
