#include <assert.h>
#include "tailrec.h"
#include "array.h"
-#include "irprog.h"
+#include "irprog_t.h"
#include "irgwalk.h"
#include "irgmod.h"
#include "irop.h"
#include "irgraph_t.h"
#include "ircons.h"
#include "irflag.h"
-#include "firmstat.h"
-
-static int n_opt_applications = 0;
+#include "trouts.h"
+#include "return.h"
+#include "scalar_replace.h"
+#include "irouts.h"
+#include "irhooks.h"
/**
- * the environment for colelcting data
+ * the environment for collecting data
*/
typedef struct _collect_t {
- ir_node *proj_X; /**< initial exec proj */
- ir_node *block; /**< old first block */
- int blk_idx; /**< cfgpred index of the initial exec in block */
- ir_node *proj_m; /**< linked list of memory from start proj's */
- ir_node *proj_data; /**< linked list of all parameter access proj's */
+ ir_node *proj_X; /**< initial exec proj */
+ ir_node *block; /**< old first block */
+ int blk_idx; /**< cfgpred index of the initial exec in block */
+ ir_node *proj_m; /**< linked list of memory from start proj's */
+ ir_node *proj_data; /**< linked list of all parameter access proj's */
} collect_t;
/**
- * walker for collecting data
+ * walker for collecting data, fills a collect_t environment
*/
static void collect_data(ir_node *node, void *env)
{
ir_node *start = get_Proj_pred(pred);
if (get_irn_op(start) == op_Start) {
- if (get_Proj_proj(pred) == pn_Start_T_args) {
- /* found Proj(ProjT(Start)) */
- set_irn_link(node, data->proj_data);
- data->proj_data = node;
- }
+ if (get_Proj_proj(pred) == pn_Start_T_args) {
+ /* found Proj(ProjT(Start)) */
+ set_irn_link(node, data->proj_data);
+ data->proj_data = node;
+ }
}
}
else if (op == op_Start) {
switch (get_Proj_proj(node)) {
- case pn_Start_M:
- /* found ProjM(Start) */
- set_irn_link(node, data->proj_m);
- data->proj_m = node;
- break;
- case pn_Start_X_initial_exec:
- /* found ProjX(Start) */
- data->proj_X = node;
- break;
- default:
- break;
+ case pn_Start_M:
+ /* found ProjM(Start) */
+ set_irn_link(node, data->proj_m);
+ data->proj_m = node;
+ break;
+ case pn_Start_X_initial_exec:
+ /* found ProjX(Start) */
+ data->proj_X = node;
+ break;
+ default:
+ break;
}
}
break;
/*
* the first block has the initial exec as cfg predecessor
*/
- if (node != current_ir_graph->start_block) {
+ if (node != get_irg_start_block(current_ir_graph)) {
for (i = 0; i < n_pred; ++i) {
- if (get_Block_cfgpred(node, i) == data->proj_X) {
- data->block = node;
- data->blk_idx = i;
- break;
- }
+ if (get_Block_cfgpred(node, i) == data->proj_X) {
+ data->block = node;
+ data->blk_idx = i;
+ break;
+ }
}
}
break;
*/
static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls)
{
- ir_graph *rem_irg = current_ir_graph;
- ir_node *end_block = irg->end_block;
+ ir_node *end_block = get_irg_end_block(irg);
ir_node *block, *jmp, *call, *calls;
ir_node **in;
ir_node **phis;
ir_node ***call_params;
- ir_node *p;
+ ir_node *p, *n;
int i, j, n_params;
collect_t data;
- int rem = get_optimize();
+ int rem = get_optimize();
+ entity *ent = get_irg_entity(irg);
+ ir_type *method_tp = get_entity_type(ent);
assert(n_tail_calls);
- /* we add nwe nodes, so the outs are inconsistant */
+ /* we add new nodes, so the outs are inconsistent */
set_irg_outs_inconsistent(irg);
/* we add new blocks and change the control flow */
- set_irg_dom_inconsistent(irg);
+ set_irg_doms_inconsistent(irg);
+ set_irg_extblk_inconsistent(irg);
/* we add a new loop */
set_irg_loopinfo_inconsistent(irg);
+ /* calls are removed */
+ set_trouts_inconsistent();
+
+ /* we must build some new nodes WITHOUT CSE */
set_optimize(0);
/* collect needed data */
assert(data.proj_m && "Could not find ProjM(Start)");
assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
- current_ir_graph = irg;
-
/* allocate in's for phi and block construction */
NEW_ARR_A(ir_node *, in, n_tail_calls + 1);
in[0] = data.proj_X;
/* turn Return's into Jmp's */
- for (i = 1, p = rets; p; p = get_irn_link(p)) {
- ir_node *jmp;
+ for (i = 1, p = rets; p; p = n) {
+ ir_node *block = get_nodes_block(p);
- set_cur_block(get_nodes_block(p));
- jmp = new_Jmp();
+ n = get_irn_link(p);
+ in[i++] = new_r_Jmp(irg, block);
- exchange(p, new_Bad());
- in[i++] = jmp;
+ exchange(p, new_r_Bad(irg));
- add_End_keepalive(get_irg_end(irg), jmp);
+ /* we might generate an endless loop, so add
+ * the block to the keep-alive list */
+ add_End_keepalive(get_irg_end(irg), block);
}
/* create a new block at start */
- block = new_Block(n_tail_calls + 1, in);
- jmp = new_Jmp();
+ block = new_r_Block(irg, n_tail_calls + 1, in);
+ jmp = new_r_Jmp(irg, block);
/* the old first block is now the second one */
set_Block_cfgpred(data.block, data.blk_idx, jmp);
/* build the memory phi */
i = 0;
- in[i] = new_rd_Proj(NULL, irg, irg->start_block, irg->start, mode_M, pn_Start_M);
+ in[i] = new_r_Proj(irg, get_irg_start_block(irg), get_irg_start(irg), mode_M, pn_Start_M);
++i;
for (calls = call; calls; calls = get_irn_link(calls)) {
}
assert(i == n_tail_calls + 1);
- phis[0] = new_rd_Phi(NULL, irg, block, n_tail_calls + 1, in, mode_M);
+ phis[0] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode_M);
- /* build the data phi's */
+ /* build the data Phi's */
if (n_params > 0) {
ir_node *calls;
+ ir_node *args;
+ ir_node *args_bl;
- NEW_ARR_A(ir_node **, call_params, n_params);
+ NEW_ARR_A(ir_node **, call_params, n_tail_calls);
/* collect all parameters */
for (i = 0, calls = call; calls; calls = get_irn_link(calls)) {
++i;
}
- /* build new projs and Phi's */
+ /* build new Proj's and Phi's */
+ args = get_irg_args(irg);
+ args_bl = get_nodes_block(args);
for (i = 0; i < n_params; ++i) {
- ir_mode *mode = get_irn_mode(call_params[0][i]);
+ ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
- in[0] = new_rd_Proj(NULL, irg, block, irg->args, mode, i);
+ in[0] = new_r_Proj(irg, args_bl, args, mode, i);
for (j = 0; j < n_tail_calls; ++j)
- in[j + 1] = call_params[j][i];
+ in[j + 1] = call_params[j][i];
- phis[i + 1] = new_rd_Phi(NULL, irg, block, n_tail_calls + 1, in, mode);
+ phis[i + 1] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode);
}
}
* ok, we are here, so we have build and collected all needed Phi's
* now exchange all Projs into links to Phi
*/
- for (p = data.proj_m; p; p = get_irn_link(p)) {
+ for (p = data.proj_m; p; p = n) {
+ n = get_irn_link(p);
exchange(p, phis[0]);
}
- for (p = data.proj_data; p; p = get_irn_link(p)) {
+ for (p = data.proj_data; p; p = n) {
long proj = get_Proj_proj(p);
assert(0 <= proj && proj < n_params);
+ n = get_irn_link(p);
exchange(p, phis[proj + 1]);
}
- current_ir_graph = rem_irg;
+ /* tail recursion was done, all info is invalid */
+ set_irg_doms_inconsistent(irg);
+ set_irg_outs_inconsistent(irg);
+ set_irg_extblk_inconsistent(irg);
+ set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
+ set_trouts_inconsistent();
+ set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
+
set_optimize(rem);
}
+/**
+ * Check the lifetime of locals in the given graph.
+ * Tail recursion can only be done, if we can prove that
+ * the lifetime of locals end with the recursive call.
+ * We do this by checking that no address of a local variable is
+ * stored or transmitted as an argument to a call.
+ *
+ * @return non-zero if it's ok to do tail recursion
+ */
+static int check_lifetime_of_locals(ir_graph *irg)
+{
+ ir_node *irg_frame = get_irg_frame(irg);
+ int i;
+
+ if (get_irg_outs_state(irg) != outs_consistent)
+ compute_irg_outs(irg);
+
+ for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
+ ir_node *succ = get_irn_out(irg_frame, i);
+
+ if (is_Sel(succ) && is_address_taken(succ))
+ return 0;
+ }
+ return 1;
+}
+
/*
* convert simple tail-calls into loops
*/
-void opt_tail_rec_irg(ir_graph *irg)
+int opt_tail_rec_irg(ir_graph *irg)
{
- ir_node *end_block = irg->end_block;
- int n_preds;
+ ir_node *end_block;
int i, n_tail_calls = 0;
ir_node *rets = NULL;
+ ir_type *mtd_type, *call_type;
if (! get_opt_tail_recursion() || ! get_opt_optimize())
- return;
+ return 0;
+
+ if (! check_lifetime_of_locals(irg))
+ return 0;
+
+ /*
+ * This tail recursion optimization works best
+ * if the Returns are normalized.
+ */
+ normalize_n_returns(irg);
+ end_block = get_irg_end_block(irg);
set_irn_link(end_block, NULL);
- n_preds = get_Block_n_cfgpreds(end_block);
- for (i = 0; i < n_preds; ++i) {
+ for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
ir_node *ret = get_Block_cfgpred(end_block, i);
ir_node *call, *call_ptr;
entity *ent;
- int j, n_ress;
+ int j;
ir_node **ress;
- /* search all returns of a block */
- if (get_irn_op(ret) != op_Return)
+ /* search all Returns of a block */
+ if (! is_Return(ret))
continue;
/* check, if it's a Return self() */
call = skip_Proj(get_Return_mem(ret));
- if (get_irn_op(call) != op_Call)
+ if (! is_Call(call))
continue;
/* check if it's a recursive call */
continue;
/* ok, mem is routed to a recursive call, check return args */
- n_ress = get_Return_n_ress(ret);
ress = get_Return_res_arr(ret);
-
- for (j = 0; j < n_ress; ++j) {
+ for (j = get_Return_n_ress(ret) - 1; j >= 0; --j) {
ir_node *irn = skip_Proj(skip_Proj(ress[j]));
if (irn != call) {
- /* not routed to a call */
- break;
+ /* not routed to a call */
+ break;
}
}
- if (j < n_ress)
+ if (j >= 0)
continue;
+ /*
+ * Check, that the types match. At least in C
+ * this might fail.
+ */
+ mtd_type = get_entity_type(ent);
+ call_type = get_Call_type(call);
+
+ if (mtd_type != call_type) {
+ /*
+ * Hmm, the types did not match, bad.
+ * This can happen in C when no prototype is given
+ * or K&R style is used.
+ */
+#if 0
+ printf("Warning: Tail recursion fails because of different method and call types:\n");
+ dump_type(mtd_type);
+ dump_type(call_type);
+#endif
+ return 0;
+ }
+
/* here, we have found a call */
set_irn_link(call, get_irn_link(end_block));
set_irn_link(end_block, call);
/* now, end_block->link contains the list of all tail calls */
if (! n_tail_calls)
- return;
+ return 0;
if (get_opt_tail_recursion_verbose() && get_firm_verbosity() > 1)
printf(" Performing tail recursion for graph %s and %d Calls\n",
- get_entity_ld_name(get_irg_entity(irg)), n_tail_calls);
- n_opt_applications++;
+ get_entity_ld_name(get_irg_entity(irg)), n_tail_calls);
+ hook_tail_rec(irg, n_tail_calls);
do_opt_tail_rec(irg, rets, n_tail_calls);
+
+ return n_tail_calls;
}
/*
void opt_tail_recursion(void)
{
int i;
+ int n_opt_applications = 0;
+ ir_graph *irg;
if (! get_opt_tail_recursion() || ! get_opt_optimize())
return;
- n_opt_applications = 0;
-
- for (i = 0; i < get_irp_n_irgs(); i++) {
- current_ir_graph = get_irp_irg(i);
+ for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
+ irg = get_irp_irg(i);
- opt_tail_rec_irg(current_ir_graph);
+ if (opt_tail_rec_irg(irg))
+ ++n_opt_applications;
}
if (get_opt_tail_recursion_verbose())