#include "irpass.h"
#include "valueset.h"
#include "irloop.h"
+#include "firmstat_t.h"
#include "irgraph_t.h"
#include "irnode_t.h"
#define MAX_ANTIC_ITER 10
#define MAX_INSERT_ITER 3
-/* Infinite loops will be unrolled during antic iteration and
- will iterate until otherwise stopped.
- This also leaves every possible values of iteration variables in antic_in.
-*/
+/* Stops antic iteration from processing endless loops. */
+#define IGNORE_INF_LOOPS 1
+/* Stops antic information to flow over infinite loop backedge */
#define NO_INF_LOOPS 0
/* Attempt to reduce register pressure and reduce code size
for hoisted nodes. */
-#define HOIST_HIGH 0
+#define HOIST_HIGH 1
#define COMMON_DOM 1
/* Seamless implementation of handling loads and generally memory
dependent nodes with GVN-PRE. */
-#define LOADS 0
+#define LOADS 1
#define DIVMODS 0
/* Experimental */
-#define MIN_CUT 0
-#define OLD_DIVMODS 0
-#define NO_INF_LOOPS2 0
+/* Only optimize node directly after phi node if node is only successor. */
+#define MIN_CUT 0
-/* NIY Choose to be optimized nodes in a more sophisticated way
- to reduce number of newly introduced phi nodes. */
-#define BETTER_GREED 0
+/* GVN is very limited. This enables optimize_node during value recognition.
+ GVN ist still very limited, since a+1+b+1 doesn't equal a+b+2.
+ TODO Broken for yet unknown reasons. */
+#define OPTIMIZE_NODES 0
/** Additional info we need for every block. */
char changes; /* flag for fixed point iterations - non-zero if changes occurred */
char first_iter; /* non-zero for first fixed point iteration */
int iteration; /* iteration counter */
+#if OPTIMIZE_NODES
+ pset *value_table; /* standard value table*/
+ pset *gvnpre_values; /* gvnpre value table */
+#endif
} pre_env;
static pre_env *environ;
int infinite_loops;
} gvnpre_statistics;
-gvnpre_statistics *gvnpre_stats = NULL;
+static gvnpre_statistics *gvnpre_stats = NULL;
-static void init_stats()
+static void init_stats(void)
{
gvnpre_stats = XMALLOCZ(gvnpre_statistics);
}
-static void free_stats()
+static void free_stats(void)
{
free(gvnpre_stats);
gvnpre_stats = NULL;
}
-static void print_stats()
+static void print_stats(void)
{
gvnpre_statistics *stats = gvnpre_stats;
DB((dbg, LEVEL_1, "replaced : %d\n", stats->replaced));
/* memops are not the same, even if we want to optimize them
we have to take the order in account */
- if (is_memop(a) || is_memop(b)) {
+ if (is_memop(a) || is_memop(b) ||
+ get_irn_mode(a) == mode_T ||
+ get_irn_mode(b) == mode_T) {
/* Loads with the same predecessors are the same value;
this should only happen after phi translation. */
- if (! is_Load(a) || ! is_Load(b))
+ if ((! is_Load(a) || ! is_Load(b)) && (! is_Store(a) || ! is_Store(b)))
return 1;
}
in[i] = pred_value;
}
- if (changed) {
+ if (changed && !is_memop(irn) && get_irn_mode(irn) != mode_X) {
/* create representative for */
ir_node *nn = new_ir_node(
get_irn_dbg_info(irn),
in);
copy_node_attr(environ->graph, irn, nn);
+#if OPTIMIZE_NODES
+ /* TODO optimize_node() uses the valuetable and thus the custom
+ identify_cmp() and will fail trying. */
+ environ->graph->value_table = environ->value_table;
+ nn = optimize_node(nn);
+ environ->graph->value_table = environ->gvnpre_values;
+#endif
+
/* now the value can be determined because the
predecessors are the leaders */
value = identify_remember(nn);
ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK);
}
-#if NO_INF_LOOPS || NO_INF_LOOPS2
+#if IGNORE_INF_LOOPS || NO_INF_LOOPS
/**
* Returns non-zero if block is part of an infinite loop.
*/
if (is_Phi(n))
return 1;
-#if LOADS || OLD_DIVMODS || DIVMODS
+#if LOADS || DIVMODS
if (is_Proj(n) && mode != mode_X && mode != mode_T)
return 1;
#else
#if LOADS
if (is_Load(n))
return get_Load_volatility(n) == volatility_non_volatile;
+ if (is_Store(n))
+ return get_Store_volatility(n) == volatility_non_volatile;
#endif
if (get_irn_pinned(n) == op_pin_state_pinned)
/* filter loads with no phi predecessor from antic_in */
if (is_Load(n) && ! is_Phi(get_Load_mem(n)))
return 0;
+ if (is_Store(n) && ! is_Phi(get_Store_mem(n)))
+ return 0;
#endif
#if DIVMODS
if (is_Block(irn))
return;
+#if OPTIMIZE_NODES
+ environ->graph->value_table = environ->value_table;
+ identify_remember(irn);
+ environ->graph->value_table = environ->gvnpre_values;
+#endif
+
/* GVN step: remember the value. */
value = remember(irn);
- /* values that are not in antic_in also dont't need to be in any other set */
- if (! is_nice_value(irn))
- return;
-
if (is_irn_constlike(irn))
return;
block = get_nodes_block(irn);
info = get_block_info(block);
- ir_valueset_insert(info->avail_out, value, irn);
+ if (get_irn_mode(irn) != mode_X)
+ ir_valueset_insert(info->avail_out, value, irn);
+
+ /* values that are not in antic_in also dont't need to be in any other set */
+
+ if (! is_nice_value(irn))
+ return;
if (is_clean_in_block(irn, block, info->exp_gen)) {
DB((dbg, LEVEL_3, "%+F clean in block %+F\n", irn, block));
ir_nodehashmap_insert(map, node, trans);
}
-#if OLD_DIVMODS
-/* Helper function to compare the values of pred and avail_pred. */
-static unsigned match_pred(ir_node *pred, ir_node *avail_pred, ir_node *block, int pos)
-{
- ir_node *avail_value = identify(avail_pred);
- ir_node *pred_block = get_Block_cfgpred_block(block, pos);
- ir_node *trans_pred = get_translated(pred_block, pred);
- ir_node *value;
-
- if (trans_pred == NULL)
- trans_pred = pred;
- value = identify(trans_pred);
-
- DB((dbg, LEVEL_3, "manual compare %+F %+F\n", pred, avail_pred));
-
- return (value == avail_value);
-}
-
-/**
- * Does phi translation for redundant Div/Mod nodes only.
- * Returns NULL for non-redundant node, which needs to be phi translated.
- */
-static ir_node *phi_translate_divmod(ir_node *divmod, ir_node *block, int pos)
-{
- ir_node *mem = get_memop_mem(divmod);
- ir_node *trans = get_translated_pred(block, pos, mem);
-
- if (trans == NULL)
- trans = mem;
-
- /* no partial redundancy if this is a mode_M phi */
- if (is_Proj(trans)) {
- /* The last memory operation in predecessor block */
- ir_node *avail_op = get_Proj_pred(trans);
-
- if (get_irn_op(divmod) == get_irn_op(avail_op)) {
- unsigned left, right;
-
- if (is_Div(avail_op)) {
- if (get_Div_resmode(divmod) == get_Div_resmode(avail_op) &&
- get_Div_no_remainder(divmod) == get_Div_no_remainder(avail_op)) {
-
- left = match_pred(get_Div_left(divmod), get_Div_left(avail_op), block, pos);
- right = match_pred(get_Div_right(divmod), get_Div_right(avail_op), block, pos);
-
- if (left && right)
- return avail_op;
- }
- } else if (is_Mod(avail_op)) {
- if (get_Mod_resmode(divmod) == get_Mod_resmode(avail_op)) {
-
- left = match_pred(get_Mod_left(divmod), get_Mod_left(avail_op), block, pos);
- right = match_pred(get_Mod_right(divmod), get_Mod_right(avail_op), block, pos);
-
- if (left && right)
- return avail_op;
- }
- }
- }
- }
- return NULL;
-}
-#endif
-
/**
* Translates an expression above a Phi.
*
}
arity = get_irn_arity(node);
-#if OLD_DIVMODS
- if (is_Div(node) || is_Mod(node)) {
- ir_node *avail_op = phi_translate_divmod(node, block, pos);
- if (avail_op)
- return avail_op;
- }
-#endif
-
needed = 0;
in = ALLOCANZ(ir_node *, arity);
skip them and use the loads memory. */
if (is_Proj(pred_trans) && get_irn_mode(pred_trans) == mode_M) {
#if LOADS || DIVMODS
- ir_node *load = get_Proj_pred(pred_trans);
+ ir_node *loadstore = get_Proj_pred(pred_trans);
/* If we do not translate this node, we will get its value wrong. */
needed |= 1;
- if (is_Load(load)) {
+ if (is_Load(loadstore)) {
/* Put new load under the adjacent loads memory edge
such that GVN may compare them. */
- new_pred = get_Load_mem(load);
+ new_pred = get_Load_mem(loadstore);
+ } else if (is_Store(loadstore)) {
+ new_pred = get_Store_mem(loadstore);
}
#endif
} else {
/* add exp_gen */
if (env->first_iter) {
-#if NO_INF_LOOPS
+#if IGNORE_INF_LOOPS
/* keep antic_in of infinite loops empty */
if (! is_in_infinite_loop(block)) {
foreach_valueset(info->exp_gen, value, expr, iter) {
if (trans == NULL)
trans = phi_translate(expr, succ, pos, get_block_info(succ)->antic_in);
+
/* create new value if necessary */
trans_value = identify_or_remember(trans);
represent = expr;
if (is_clean_in_block(expr, block, info->antic_in)) {
-#if NO_INF_LOOPS2
- /* no flow over the backedge of endless loops */
- if (env->iteration <= 2 || ! (! is_in_infinite_loop(succ) || ! is_backedge(succ, pos))) {
+#if NO_INF_LOOPS
+ /* Prevent information flow over the backedge of endless loops. */
+ if (env->iteration <= 2 || (is_backedge(succ, pos) && !is_in_infinite_loop(succ))) {
ir_valueset_replace(info->antic_in, trans_value, represent);
}
#else
}
}
-#if BETTER_GREED
- /* value is redundant from last iteration,
- but has not been removed from antic_in (is not optimized) */
- if (! environ->first_iter && is_redundant(block, expr))
- return mode;
-#endif
-
/* If it is not the same value already existing along every predecessor
and it is defined by some predecessor then it is partially redundant. */
if (! partially_redundant || fully_redundant)
#endif
} /* update_new_set */
-#if BETTER_GREED
-/*
- * Returns redundant flag of node irn in block block.
- */
-static unsigned is_redundant(ir_node *block, ir_node *irn)
-{
- (void) block;
- (void) irn;
-
- /* TODO Needs to use a flag, because antic_done should only be used
- if node is finally processed by insert_nodes. */
- return 0;
-}
-#endif
-
/**
* Checks if hoisting irn is greedy.
* Greedy hoisting means that there are non partially redundant nodes
avail = ir_valueset_lookup(pred_info->avail_out, trans_val);
+ DB((dbg, LEVEL_3, "avail %+F\n", avail));
if (! avail)
return 1;
#if MIN_CUT
ir_node *idom;
int pos;
ir_valueset_iterator_t iter;
-#if BETTER_GREED
- plist_t *stack;
-#endif
/* only blocks */
if (! is_Block(block))
return;
}
-#if BETTER_GREED
- stack = plist_new();
- foreach_valueset(info->antic_in, value, expr, iter) {
- /* inverse topologic */
- plist_insert_front(stack, expr);
- }
-#endif
-
/* This is the main reason antic_in is preverred over antic_out;
we may iterate over every anticipated value first and not
over the predecessor blocks. */
continue;
}
-#if !BETTER_GREED
if (is_hoisting_greedy(expr, block)) {
DB((dbg, LEVEL_2, "greedy\n"));
continue;
}
-#endif
mode = is_partially_redundant(block, expr, value);
if (mode == NULL)
continue;
-#if BETTER_GREED
- if (is_hoisting_greedy(expr, block)) {
- DB((dbg, LEVEL_2, "Better greed: greedy\n"));
- continue;
- }
-#endif
-
-#if LOADS || OLD_DIVMODS || DIVMODS
+#if LOADS || DIVMODS
/* save old mode_M phis to remove keepalive edges later */
if (is_memop(expr)) {
ir_node *mem = get_memop_mem(expr);
inc_stats(gvnpre_stats->first_iter_found);
inc_stats(gvnpre_stats->partially);
}
- if (is_Load(expr))
+ if (is_Load(expr) || is_Store(expr))
inc_stats(gvnpre_stats->loads);
else if (is_Div(expr) || is_Mod(expr))
inc_stats(gvnpre_stats->divmods);
ir_valueset_insert(info->antic_done, value, expr);
env->changes |= 1;
}
-
-#if BETTER_GREED
- /* TODO Unfinished
- Better greed first determines which values are redundant
- and decides then which to take.
- insert_nodes needs to be split up for that. The cycle could be
- for each block: flag redundant nodes,
- use heuristic to adjust these flags (also consider antic_done),
- do insert nodes.
- This way we could decide if we should hoist a non redundant node,
- if all its successors are redundant.
- Or we might try to minimize the cut along hoisted nodes and their
- non redundant successors.
- */
- if (env->changes) {
- plist_element_t *it;
- /* iterate in inverse topological order */
- foreach_plist(stack, it) {
- ir_node *irn = (ir_node *)plist_element_get_value(it);
- ir_node *block = get_nodes_block(irn);
- int j;
- char redundant = 1;
-
- /* does irn only have redundant successors? */
-
- foreach_out_edge(irn, edge) {
- ir_node *succ = get_edge_src_irn(edge);
-
- /* if succ and irn are in the same block */
- if (get_nodes_block(succ) == block && is_redundant(block, succ)) {
- continue;
- } else {
- redundant = 0;
- break;
- }
- }
-
- if (redundant)
- flag_redundant(irn, 1);
- }
- }
- plist_free(stack);
-#endif
-
}
#if HOIST_HIGH
be available from this point on. Currently we do not push
the availability information through during the walk. */
ir_valueset_insert(target_info->new_set, value, nn);
+ ir_valueset_insert(target_info->avail_out, value, nn);
}
}
}
/* Detect and set links of infinite loops to non-zero. */
analyse_loops(irg);
+#if OPTIMIZE_NODES
+ new_identities(irg);
+ env.value_table = irg->value_table;
+ irg->value_table = NULL;
+#endif
+
/* Switch on GCSE. We need it to correctly compute
the value of a node, which is independent from
its block. */
set_opt_global_cse(1);
- /* new_identities */
+ /* new_identities() */
if (irg->value_table != NULL)
del_pset(irg->value_table);
/* initially assumed nodes in pset are 512 */
irg->value_table = new_pset(compare_gvn_identities, 512);
+#if OPTIMIZE_NODES
+ env.gvnpre_values = irg->value_table;
+#endif
/* do GVN-PRE pass */
gvn_pre(irg, &env);
restore_optimization_state(&state);
confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
+#if OPTIMIZE_NODES
+ irg->value_table = env.value_table;
+ del_pset(irg->value_table);
+ irg->value_table = env.gvnpre_values;
+#endif
+
/* TODO There seem to be optimizations that try to use the existing
value_table. */
new_identities(irg);