+2005-03-21 Mostafa Hagog <mustafa@il.ibm.com>
+
+ PR middle-end/20177
+ * ddg.c (create_ddg_dependence): Ignore reg-anti dependency.
+ * modulo-sched.c (const_iteration_count): Return on NULL
+ pre-header.
+ (print_node_sched_params): Return on NULL dump_file.
+ (generate_reg_moves): Handle reg-anti dependencies and disregard
+ closing branch when generating register moves.
+ (sms_schedule): Mark the SMSed block dirty.
+ * passes.c (rest_of_handle_sms): Call update_life_info for all
+ basic-blocks.
+ * testsuite/gcc.dg/20050321-1.c: New test.
+
2005-03-21 Bob Wilson <bob.wilson@acm.org>
* config/xtensa/lib1funcs.asm (__mulsi3): Use symbolic name for ACCLO.
{
rtx insn;
rtx head, tail;
+
+ if (! pre_header)
+ return NULL_RTX;
+
get_block_head_tail (pre_header->index, &head, &tail);
for (insn = tail; insn != PREV_INSN (head); insn = PREV_INSN (insn))
{
int i;
+ if (! dump_file)
+ return;
for (i = 0; i < num_nodes; i++)
{
node_sched_params_ptr nsp = &node_sched_params[i];
return maxii;
}
-
-/* Given the partial schedule, generate register moves when the length
- of the register live range is more than ii; the number of moves is
- determined according to the following equation:
- SCHED_TIME (use) - SCHED_TIME (def) { 1 broken loop-carried
- nreg_moves = ----------------------------------- - { dependence.
- ii { 0 if not.
- This handles the modulo-variable-expansions (mve's) needed for the ps. */
+/*
+ Breaking intra-loop register anti-dependences:
+ Each intra-loop register anti-dependence implies a cross-iteration true
+ dependence of distance 1. Therefore, we can remove such false dependencies
+ and figure out if the partial schedule broke them by checking if (for a
+ true-dependence of distance 1): SCHED_TIME (def) < SCHED_TIME (use) and
+ if so generate a register move. The number of such moves is equal to:
+ SCHED_TIME (use) - SCHED_TIME (def) { 0 broken
+ nreg_moves = ----------------------------------- + 1 - { dependecnce.
+ ii { 1 if not.
+*/
static void
generate_reg_moves (partial_schedule_ptr ps)
{
{
int nreg_moves4e = (SCHED_TIME (e->dest) - SCHED_TIME (e->src)) / ii;
+ if (e->distance == 1)
+ nreg_moves4e = (SCHED_TIME (e->dest) - SCHED_TIME (e->src) + ii) / ii;
+
/* If dest precedes src in the schedule of the kernel, then dest
will read before src writes and we can save one reg_copy. */
if (SCHED_ROW (e->dest) == SCHED_ROW (e->src)
{
int dest_copy = (SCHED_TIME (e->dest) - SCHED_TIME (e->src)) / ii;
+ if (e->distance == 1)
+ dest_copy = (SCHED_TIME (e->dest) - SCHED_TIME (e->src) + ii) / ii;
+
if (SCHED_ROW (e->dest) == SCHED_ROW (e->src)
&& SCHED_COLUMN (e->dest) < SCHED_COLUMN (e->src))
dest_copy--;
int amount = PS_MIN_CYCLE (ps);
int ii = ps->ii;
- for (i = 0; i < g->num_nodes; i++)
+ /* Don't include the closing branch assuming that it is the last node. */
+ for (i = 0; i < g->num_nodes - 1; i++)
{
ddg_node_ptr u = &g->nodes[i];
int normalized_time = SCHED_TIME (u) - amount;
/* SCHED_STAGE (u_node) >= from_stage == 0. Generate increasing
number of reg_moves starting with the second occurrence of
u_node, which is generated if its SCHED_STAGE <= to_stage. */
- i_reg_moves = to_stage - SCHED_STAGE (u_node);
+ i_reg_moves = to_stage - SCHED_STAGE (u_node) + 1;
i_reg_moves = MAX (i_reg_moves, 0);
i_reg_moves = MIN (i_reg_moves, SCHED_NREG_MOVES (u_node));
scheduling passes doesn't touch it. */
if (! flag_resched_modulo_sched)
g->bb->flags |= BB_DISABLE_SCHEDULE;
+ /* The life-info is not valid any more. */
+ g->bb->flags |= BB_DIRTY;
generate_reg_moves (ps);
if (dump_file)
static void
rest_of_handle_sms (void)
{
+ sbitmap blocks;
+
timevar_push (TV_SMS);
open_dump_file (DFI_sms, current_function_decl);
/* Update the life information, because we add pseudos. */
max_regno = max_reg_num ();
allocate_reg_info (max_regno, FALSE, FALSE);
- update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
- (PROP_DEATH_NOTES
- | PROP_KILL_DEAD_CODE
- | PROP_SCAN_DEAD_CODE));
+ blocks = sbitmap_alloc (last_basic_block);
+ sbitmap_ones (blocks);
+ update_life_info (blocks, UPDATE_LIFE_GLOBAL_RM_NOTES,
+ (PROP_DEATH_NOTES
+ | PROP_REG_INFO
+ | PROP_KILL_DEAD_CODE
+ | PROP_SCAN_DEAD_CODE));
+
no_new_pseudos = 1;
ggc_collect ();