#include "config.h"
#endif
+#include <stdlib.h>
+
+#include "ircons.h"
+
#include "bearch.h"
#include "belower.h"
#include "benode_t.h"
#include "besched.h"
#include "irgmod.h"
-#include "iredges.h"
+#include "iredges_t.h"
#include "irgwalk.h"
#undef is_Perm
#define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm)
+#undef is_Call
+#define is_Call(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_call)
+
/* lowering walker environment */
typedef struct _lower_env_t {
- be_chordal_env_t *chord_env;
- int do_copy;
+ be_chordal_env_t *chord_env;
+ int do_copy;
+ firm_dbg_module_t *dbg_module;
} lower_env_t;
/* holds a perm register pair */
PERM_COPY
} perm_type_t;
+/* structure to represent cycles or chains in a perm */
typedef struct _perm_cycle_t {
const arch_register_t **elems; /**< the registers in the cycle */
int n_elems; /**< number of elements in the cycle */
return -1;
}
+/* returns the number register pairs marked as checked */
static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
int i, n_checked = 0;
static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
int i;
- for (i = 0; i < n; i++) {
- if (in_out) {
+ if (in_out) {
+ for (i = 0; i < n; i++) {
/* out register matches */
if (pairs[i].out_reg->index == reg->index)
return pairs[i].out_node;
}
- else {
+ }
+ else {
+ for (i = 0; i < n; i++) {
/* in register matches */
if (pairs[i].in_reg->index == reg->index)
return pairs[i].in_node;
return NULL;
}
+/**
+ * Gets the index in the register pair array where the in/out register
+ * corresponds to reg_idx.
+ *
+ * @param pairs The array of register pairs
+ * @param n The number of pairs
+ * @param reg The register index to look for
+ * @param in_out 0 == look for IN register, 1 == look for OUT register
+ * @return The corresponding index in pairs or -1 if not found
+ */
+static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
+ int i;
+
+ if (in_out) {
+ for (i = 0; i < n; i++) {
+ /* out register matches */
+ if (pairs[i].out_reg->index == reg_idx)
+ return i;
+ }
+ }
+ else {
+ for (i = 0; i < n; i++) {
+ /* in register matches */
+ if (pairs[i].in_reg->index == reg_idx)
+ return i;
+ }
+ }
+
+ return -1;
+}
+
/**
* Gets an array of register pairs and tries to identify a cycle or chain starting
* at position start.
static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
int head = pairs[start].in_reg->index;
int cur_idx = pairs[start].out_reg->index;
+ int cur_pair_idx = start;
int n_pairs_done = get_n_checked_pairs(pairs, n) + 1;
int idx;
/* check for cycle or end of a chain */
while (cur_idx != head && n_pairs_done < n) {
/* goto next register in cycle or chain */
- cur_idx = pairs[cur_idx].out_reg->index;
+ cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
+
+ if (cur_pair_idx < 0)
+ break;
+
+ cur_idx = pairs[cur_pair_idx].out_reg->index;
/* it's not the first element: insert it */
if (cur_idx != head) {
- cycle->elems[idx++] = pairs[cur_idx].out_reg;
+ cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
cycle->n_elems++;
- pairs[cur_idx].checked = 1;
n_pairs_done++;
}
else {
/* we are there where we started -> CYCLE */
cycle->type = PERM_CYCLE;
}
+
+ /* mark the pair as checked */
+ pairs[cur_pair_idx].checked = 1;
}
return cycle;
/**
* Lowers a perm node. Resolves cycles and creates a bunch of
* copy and swap operations to permute registers.
+ * Note: The caller of this function has to make sure, that irn
+ * is a Perm node.
*
* @param irn The perm node
+ * @param block The block the perm node belongs to
* @param walk_env The environment
*/
-static void lower_perms_walker(ir_node *irn, void *walk_env) {
- const be_node_factory_t *fact;
+static void lower_perm_node(ir_node *irn, void *walk_env) {
const arch_register_class_t *reg_class;
const arch_env_t *arch_env;
+ firm_dbg_module_t *mod;
lower_env_t *env = walk_env;
reg_pair_t *pairs;
const ir_edge_t *edge;
perm_cycle_t *cycle;
int n, i, pn, do_copy;
- ir_node *block, *arg, *res, *sched_point, *in[2];
+ ir_node *sched_point, *block, *in[2];
+ ir_node *arg1, *arg2, *res1, *res2;
ir_node *cpyxchg = NULL;
- fact = env->chord_env->main_env->node_factory;
arch_env = env->chord_env->main_env->arch_env;
do_copy = env->do_copy;
-
- /* check if perm */
- if (! is_Perm(arch_env, irn))
- return;
-
- block = get_nodes_block(irn);
+ mod = env->dbg_module;
+ block = get_nodes_block(irn);
/*
Get the schedule predecessor node to the perm
should be ok.
*/
sched_point = sched_prev(irn);
+ DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
assert(sched_point && "Perm is not scheduled or has no predecessor");
n = get_irn_arity(irn);
/* Mark all equal pairs as checked, and exchange the OUT proj with
the IN node. */
- i = 0;
- while (pairs[i].in_reg->index == pairs[i].out_reg->index) {
- /* remove the proj from the schedule */
- sched_remove(pairs[i].out_node);
-
- /* remove the argument from schedule */
- sched_remove(pairs[i].in_node);
+ for (i = 0; i < n; i++) {
+ if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
+ DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
+ irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
+
+ /* We have to check for a special case:
+ The in-node could be a Proj from a Perm. In this case,
+ we need to correct the projnum */
+ if (is_Perm(arch_env, pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
+ set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
+ }
- /* exchange the proj with the argument */
- exchange(pairs[i].out_node, pairs[i].in_node);
+ /* remove the proj from the schedule */
+ sched_remove(pairs[i].out_node);
- /* add the argument after the magic scheduling point */
- sched_add_after(sched_point, pairs[i].in_node);
+ /* reroute the edges from the proj to the argument */
+ edges_reroute(pairs[i].out_node, pairs[i].in_node, env->chord_env->irg);
- pairs[i++].checked = 1;
+ pairs[i].checked = 1;
+ }
}
/* Set do_copy to 0 if it's on but we have no free register */
cycle = calloc(1, sizeof(*cycle));
cycle = get_perm_cycle(cycle, pairs, n, i);
-//todo: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
+ /* We don't need to do anything if we have a Perm with two
+ elements which represents a cycle, because those nodes
+ already represent exchange nodes */
+ if (n == 2 && cycle->type == PERM_CYCLE) {
+ free(cycle);
+ continue;
+ }
+
+//TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
// the copy cascade (this reduces the cycle into a chain)
/* build copy/swap nodes from back to front */
- for (i = cycle->n_elems - 2; i >= 0; i++) {
- arg = get_node_for_register(pairs, n, cycle->elems[i], 0);
- res = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
+ for (i = cycle->n_elems - 2; i >= 0; i--) {
+ arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
+ arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
+
+ res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
+ res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
/*
If we have a cycle and don't copy: we need to create exchange nodes
OUT_2 = out node with register i
*/
if (cycle->type == PERM_CYCLE && !do_copy) {
- in[0] = arg;
- in[1] = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
-
- cpyxchg = new_Perm(fact, reg_class, env->chord_env->irg, block, 2, in);
- set_Proj_pred(res, cpyxchg);
- set_Proj_proj(res, 0);
- set_Proj_pred(get_node_for_register(pairs, n, cycle->elems[i], 1), cpyxchg);
- set_Proj_proj(get_node_for_register(pairs, n, cycle->elems[i], 1), 1);
+ in[0] = arg1;
+ in[1] = arg2;
+
+ DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
+ irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
+ DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
+ irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
+
+ cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in);
+
+ sched_remove(res1);
+ sched_remove(res2);
+
+ set_Proj_pred(res2, cpyxchg);
+ set_Proj_proj(res2, 0);
+ set_Proj_pred(res1, cpyxchg);
+ set_Proj_proj(res1, 1);
+
+ sched_add_after(sched_point, res1);
+ sched_add_after(sched_point, res2);
+
+ arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
+ arch_set_irn_register(arch_env, res1, cycle->elems[i]);
}
else {
- cpyxchg = new_Copy(fact, reg_class, env->chord_env->irg, block, arg);
+ DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
+ irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
+
+ cpyxchg = be_new_Copy(reg_class, env->chord_env->irg, block, arg1);
arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
/* remove the proj from the schedule */
- sched_remove(res);
+ sched_remove(res2);
/* exchange copy node and proj */
- exchange(res, cpyxchg);
+ exchange(res2, cpyxchg);
}
/* insert the copy/exchange node in schedule after the magic schedule node (see above) */
sched_add_after(sched_point, cpyxchg);
+
+ DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
}
- free(cycle->elems);
+// free(cycle->elems);
free(cycle);
}
}
/**
- * Walks over all nodes in an irg and performs perm lowering.
+ * Adds Projs to keep nodes for each register class, which eats the
+ * caller saved registers.
+ * Note: The caller has to make sure, that call is a Call
+ *
+ * @param call The Call node
+ * @param walk_env The walker environment
+ */
+static void lower_call_node(ir_node *call, void *walk_env) {
+ lower_env_t *env = walk_env;
+ const arch_env_t *arch_env = env->chord_env->main_env->arch_env;
+ firm_dbg_module_t *mod = env->dbg_module;
+ const arch_register_class_t *reg_class;
+ int i, j, set_size = 0, pn, keep_arity;
+ arch_isa_t *isa = arch_env_get_isa(arch_env);
+ const ir_node *proj_T = NULL;
+ ir_node **in_keep, *block = get_nodes_block(call);
+ bitset_t *proj_set;
+ const ir_edge_t *edge;
+ const arch_register_t *reg;
+
+ /* Prepare the bitset where we store the projnums which are already in use*/
+ for (i = 0; i < arch_isa_get_n_reg_class(isa); i++) {
+ reg_class = arch_isa_get_reg_class(isa, i);
+ set_size += arch_register_class_n_regs(reg_class);
+ }
+
+ in_keep = malloc(set_size * sizeof(ir_node *));
+
+ proj_set = bitset_malloc(set_size);
+ bitset_clear_all(proj_set);
+
+ /* check if there is a ProjT node and which arguments are used */
+ foreach_out_edge(call, edge) {
+ if (get_irn_mode(get_edge_src_irn(edge)) == mode_T)
+ proj_T = get_edge_src_irn(edge);
+ }
+
+ /* set all used arguments */
+ if (proj_T) {
+ foreach_out_edge(proj_T, edge) {
+ bitset_set(proj_set, get_Proj_proj(get_edge_src_irn(edge)));
+ }
+ }
+
+ /* Create for each caller save register a proj (keep node arguement) */
+ /* if this proj is not already present */
+ for (i = 0; i < arch_isa_get_n_reg_class(isa); i++) {
+
+ /* reset the keep input, as we need one keep for each register class */
+ memset(in_keep, 0, set_size * sizeof(ir_node *));
+ keep_arity = 0;
+ reg_class = arch_isa_get_reg_class(isa, i);
+
+ for (j = 0; j < arch_register_class_n_regs(reg_class); j++) {
+ reg = arch_register_for_index(reg_class, j);
+
+ /* only check caller save registers */
+ if (arch_register_type_is(reg, caller_saved)) {
+ pn = isa->impl->get_projnum_for_register(isa, reg);
+ if (!bitset_is_set(proj_set, pn)) {
+ in_keep[keep_arity++] = new_r_Proj(current_ir_graph, block, (ir_node *)proj_T, mode_Is, pn);
+ }
+ }
+ }
+
+ /* ok, we found some caller save register which are not in use but must be saved */
+ if (keep_arity) {
+ be_new_Keep(reg_class, current_ir_graph, block, keep_arity, in_keep);
+ }
+ }
+
+ bitset_free(proj_set);
+ return;
+}
+
+/**
+ * Calls the corresponding lowering function for the node.
+ *
+ * @param irn The node to be checked for lowering
+ * @param walk_env The walker environment
+ */
+static void lower_nodes_walker(ir_node *irn, void *walk_env) {
+ lower_env_t *env = walk_env;
+ const arch_env_t *arch_env = env->chord_env->main_env->arch_env;
+
+ if (!is_Block(irn)) {
+ if (is_Perm(arch_env, irn) && ! is_Proj(irn)) {
+ lower_perm_node(irn, walk_env);
+ }
+ else if (is_Call(arch_env, irn)) {
+ lower_call_node(irn, walk_env);
+ }
+ }
+
+ return;
+}
+
+/**
+ * Walks over all blocks in an irg and performs some lowering.
*
* @param chord_env The chordal environment containing the irg
* @param do_copy 1 == resolve cycles with a free reg if available
*/
-void lower_perms(be_chordal_env_t *chord_env, int do_copy) {
+void lower_nodes(be_chordal_env_t *chord_env, int do_copy) {
lower_env_t env;
- env.chord_env = chord_env;
- env.do_copy = do_copy;
+ env.chord_env = chord_env;
+ env.do_copy = do_copy;
+ env.dbg_module = firm_dbg_register("ir.be.lower");
- irg_block_walk_graph(chord_env->irg, NULL, lower_perms_walker, &env);
+ irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_walker, &env);
}
+
+#undef is_Perm
+#undef is_Call