#include "iredges_t.h"
#include "irgwalk.h"
+#ifdef _WIN32
+#include <malloc.h>
+#else
+#include <alloca.h>
+#endif
+
#undef is_Perm
#define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm)
int cur_idx = pairs[start].out_reg->index;
int cur_pair_idx = start;
int n_pairs_done = get_n_checked_pairs(pairs, n);
- int idx, done = 0;
+ int idx;
perm_type_t cycle_tp = PERM_CYCLE;
/* We could be right in the middle of a chain, so we need to find the start */
- while (head != cur_idx && !done) {
+ while (head != cur_idx) {
/* goto previous register in cycle or chain */
cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
if (cur_pair_idx < 0) {
cycle_tp = PERM_CHAIN;
- done = 1;
+ break;
}
else {
head = pairs[cur_pair_idx].in_reg->index;
}
/* assume worst case: all remaining pairs build a cycle or chain */
- cycle->elems = calloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
+ cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
cycle->n_elems = 2; /* initial number of elements is 2 */
cycle->elems[0] = pairs[start].in_reg;
cycle->elems[1] = pairs[start].out_reg;
cycle->type = cycle_tp;
- n_pairs_done++;
+ cur_idx = pairs[start].out_reg->index;
idx = 2;
/* check for cycle or end of a chain */
- while (cur_idx != head && n_pairs_done < n) {
+ while (cur_idx != head) {
/* goto next register in cycle or chain */
cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
if (cur_idx != head) {
cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
cycle->n_elems++;
-
- n_pairs_done++;
}
else {
/* we are there where we started -> CYCLE */
assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
- pairs = calloc(n, sizeof(pairs[0]));
+ pairs = alloca(n * sizeof(pairs[0]));
/* build the list of register pairs (in, out) */
i = 0;
/* go to the first not-checked pair */
while (pairs[i].checked) i++;
- cycle = calloc(1, sizeof(*cycle));
+ cycle = xcalloc(1, sizeof(*cycle));
cycle = get_perm_cycle(cycle, pairs, n, i);
DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
//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 */
- for (i = 0; i < cycle->n_elems - 1; i++) {
+ /* build copy/swap nodes from back to front */
+ 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);
in[0] = arg1;
in[1] = arg2;
+ /* At this point we have to handle the following problem: */
+ /* */
+ /* If we have a cycle with more than two elements, then */
+ /* this could correspond to the following Perm node: */
+ /* */
+ /* +----+ +----+ +----+ */
+ /* | r1 | | r2 | | r3 | */
+ /* +-+--+ +-+--+ +--+-+ */
+ /* | | | */
+ /* | | | */
+ /* +-+--------+---------+-+ */
+ /* | Perm | */
+ /* +-+--------+---------+-+ */
+ /* | | | */
+ /* | | | */
+ /* +-+--+ +-+--+ +--+-+ */
+ /* |Proj| |Proj| |Proj| */
+ /* | r2 | | r3 | | r1 | */
+ /* +----+ +----+ +----+ */
+ /* */
+ /* This node is about to be split up into two 2x Perm's */
+ /* for which we need 4 Proj's and the one additional Proj */
+ /* of the first Perm has to be one IN of the second. So in */
+ /* general we need to create one additional Proj for each */
+ /* "middle" Perm and set this to one in node of the successor */
+ /* Perm. */
+
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",
cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in);
+ if (i > 0) {
+ /* cycle is not done yet */
+ int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
+
+ /* create intermediate proj */
+ res2 = new_r_Proj(get_irn_irg(irn), block, cpyxchg, get_irn_mode(res1), 0);
+
+ /* set as in for next Perm */
+ pairs[pidx].in_node = res2;
+ }
+ else {
+ sched_remove(res2);
+ }
+
sched_remove(res1);
- sched_remove(res2);
set_Proj_pred(res2, cpyxchg);
set_Proj_proj(res2, 0);
/* insert the copy/exchange node in schedule after the magic schedule node (see above) */
sched_add_after(sched_point, cpyxchg);
+ /* set the new scheduling point */
+ sched_point = cpyxchg;
DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
}
* @param walk_env The walker environment
*/
static void lower_call_node(ir_node *call, const void *walk_env) {
- const arch_env_t *arch_env = walk_env;
+ const arch_env_t *arch_env = walk_env;
+ int bitset_idx = 0;
+ int set_size = 0;
+ arch_isa_t *isa = arch_env_get_isa(arch_env);
+ const ir_node *proj_T = NULL;
+ ir_node *block = get_nodes_block(call);
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);
+ int i, j, pn, keep_arity;
+ ir_node **in_keep;
bitset_t *proj_set;
const ir_edge_t *edge;
const arch_register_t *reg;
set_size += arch_register_class_n_regs(reg_class);
}
- in_keep = malloc(set_size * sizeof(ir_node *));
+ in_keep = xmalloc(set_size * sizeof(ir_node *));
proj_set = bitset_malloc(set_size);
bitset_clear_all(proj_set);
ir_node *proj = get_edge_src_irn(edge);
assert(is_Proj(proj));
- bitset_set(proj_set, get_Proj_proj(proj));
+ pn = isa->impl->handle_call_proj(isa, proj, 0);
+ bitset_set(proj_set, pn);
}
}
else {
/* 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)) {
- ir_node *proj = new_r_Proj(current_ir_graph, block, (ir_node *)proj_T, mode_Is, pn);
+ /* Only create new proj, iff not already present */
+ if (!bitset_is_set(proj_set, bitset_idx)) {
+ ir_node *proj = new_r_Proj(current_ir_graph, block, (ir_node *)proj_T, mode_Is, bitset_idx);
+
+ pn = isa->impl->handle_call_proj(isa, proj, 1);
in_keep[keep_arity++] = proj;
}
+
+ bitset_idx++;
}
}
* @param irn The node to be checked for lowering
* @param walk_env The walker environment
*/
-static void lower_nodes_before_sched_walker(ir_node *irn, const void *walk_env) {
+static void lower_nodes_before_sched_walker(ir_node *irn, void *walk_env) {
const arch_env_t *arch_env = walk_env;
if (!is_Block(irn) && !is_Proj(irn)) {
* @param do_copy 1 == resolve cycles with a free reg if available
*/
void lower_nodes_before_sched(ir_graph *irg, const void *env) {
- irg_walk_blkwise_graph(irg, NULL, lower_nodes_before_sched_walker, env);
+ irg_walk_blkwise_graph(irg, NULL, lower_nodes_before_sched_walker, (void *)env);
}
env.chord_env = chord_env;
env.do_copy = do_copy;
- env.dbg_module = firm_dbg_register("ir.be.lower");
+ env.dbg_module = firm_dbg_register("firm.be.lower");
irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env);
}
projects / libfirm / blobdiff