*
* Implementation of the RA-Interface for an external, (non-SSA) register allocator.
*
- * The external register allocator is a program taking 2 arguments:
- * 1) An input file in which the cfg is defined
- * 2) An output file containing the essential actions performed during allocation
- */
+ * The external register allocator is a program:
+ * PROG -i INPUTFILE -o OUTPUTFILE
+ *
+ * 1) Input file defines the interference graph
+ * 2) Output file contains the instructions to perform
+ *
-#if 0
-===============================================================================
-===============================================================================
The input file format
----------------------
-inputfile ::= regs cfg .
+inputfile ::= regs nodes interf affinities .
regs ::= 'regs' regcount . // Anzahl der register (0..regcount-1), die zur Verfuegung stehen
-cfg ::= 'cfg' ident '{' block* edge* '}' . // Steuerflussgraph der Prozedur
-
-block ::= 'block' block-nr '{' insn* '}' . // Grundblock im cfg versehen mit einer nummer
+nodes ::= 'nodes' '{' node* '}' . // All nodes in the graph
-edge ::= 'cf-edge' block-nr block-nr . // Steuerflusskante src-->tgt
+node ::= node-info
+ | node-info '<' reg-nr '>' . // Reg-nr is present in case of constraints
-insn ::= gen-insn // Befehl in einem block
- | copy-insn
+node-info ::= node-nr spill-costs .
-gen-insn ::= 'insn' insn-nr '{' uses defs '}' .
-copy-insn ::= 'copy' insn-nr '{' uses defs '}' .
+interf ::= 'interferences' '{' i-edge* '}' . // Interference edges of the graph
-defs ::= 'def' var-list . // Liste der definierten/verwendeten Variablen
-uses ::= 'use' var-list .
+i-edge ::= '(' node-nr ',' node-nr ')' .
-var-list ::= var-ref
- | var-ref var-list
+affinities ::= 'affinities' '{' a-edge* '}' . // Affinity edges of the graph
-var-ref ::= var-nr
- | var-nr '<' reg-nr '>' . // reg-nr gibt register constraint an.
+a-edge ::= '(' node-nr ',' node-nr ',' weight ')' .
-ident ::= non-whitespace-char*
-regcount, block-nr, insn-nr, reg-nr, var-nr ::= integer
-
-===============================================================================
-===============================================================================
+weight, regcount, node-nr ::= int32 .
+spill-costs ::= int32 . // negative spill costs indicate unspillable
The output file format
-----------------------
-outputfile ::= 'actions' '{' action-list '}'
-TODO
+outputfile ::= spills | allocs .
+
+spills ::= 'spills' node-nr+ .
+
+allocs ::= 'allocs' alloc* .
+
+alloc ::= node-nr reg-nr .
-===============================================================================
-===============================================================================
-#endif /* documentation of file formats */
+
+******** End of file format docu ********/
#ifdef HAVE_CONFIG_H
#include "config.h"
#include <stdio.h>
#include <stdlib.h>
+#include <limits.h>
+#ifdef WITH_LIBCORE
+#include <libcore/lc_opts.h>
+#include <libcore/lc_opts_enum.h>
+#endif
-#include "pmap.h"
+#include "set.h"
#include "pset.h"
+#include "pmap.h"
#include "bitset.h"
#include "irprintf_t.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irgwalk.h"
+#include "iredges_t.h"
+#include "irdom_t.h"
#include "phiclass.h"
#include "beraextern.h"
#include "bearch.h"
#include "benode_t.h"
+#include "beirgmod.h"
#include "besched.h"
#include "beutil.h"
+#include "belive_t.h"
+
+#define DBG_LEVEL 2
+
+typedef struct _var_info_t var_info_t;
+/**
+ * Environment with all the needed stuff
+ */
typedef struct _be_raext_env_t {
+ firm_dbg_module_t *dbg;
arch_env_t *aenv;
const arch_register_class_t *cls;
ir_graph *irg;
+ dom_front_info_t *dom_info;
- FILE *f; /**< file handle used for out- and input file */
- pmap *vars; /**< maps variable numbers (int) to the corresponding SSA-values (pset of irns) */
- pmap *blocks; /**< maps block numbers (int) to the block (ir_node*) having that node_nr */
+ FILE *f; /**< file handle used for out- and input file */
+ set *vars; /**< contains all var_info_t */
+ int n_cls_vars; /**< length of the array cls_vars */
+ var_info_t **cls_vars; /**< only the var_infos for current cls. needed for double iterating */
} be_raext_env_t;
-/**
- * Some little helpers
- */
-#define pmap_insert_sth(pmap, key, val) pmap_insert(pmap, (void *)key, (void *)val)
-#define pmap_get_sth(pmap, key) pmap_get(pmap, (void *)key)
-#define set_var_nr(irn, nr) set_irn_link(irn, INT_TO_PTR(nr))
-#define get_var_nr(irn) (PTR_TO_INT(get_irn_link(irn)))
+/******************************************************************************
+ _ _ _
+ | | | | | |
+ | |__| | ___| |_ __ ___ _ __ ___
+ | __ |/ _ \ | '_ \ / _ \ '__/ __|
+ | | | | __/ | |_) | __/ | \__ \
+ |_| |_|\___|_| .__/ \___|_| |___/
+ | |
+ |_|
+ *****************************************************************************/
+
+
+#define pset_foreach(pset, irn) for(irn=pset_first(pset); irn; irn=pset_next(pset))
+#define set_foreach(set, e) for(e=set_first(set); e; e=set_next(set))
/**
* Checks if _the_ result of the irn belongs to the
*/
#define is_res_in_reg_class(irn) arch_irn_has_reg_class(raenv->aenv, irn, -1, raenv->cls)
+static INLINE ir_node *get_first_non_phi(pset *s) {
+ ir_node *irn;
+
+ pset_foreach(s, irn)
+ if (!is_Phi(irn)) {
+ pset_break(s);
+ return irn;
+ }
+
+ assert(0 && "There must be a non-phi-irn in this");
+ return NULL;
+}
+
+static INLINE ir_node *get_first_phi(pset *s) {
+ ir_node *irn;
+
+ pset_foreach(s, irn)
+ if (is_Phi(irn)) {
+ pset_break(s);
+ return irn;
+ }
+
+ assert(0 && "There must be a phi in this");
+ return NULL;
+}
+
+static int get_loop_weight(ir_node *irn) {
+ int cost = 0;
+ ir_loop *loop = get_irn_loop(get_nodes_block(irn));
+
+ if (loop) {
+ int d = get_loop_depth(loop);
+ cost = d*d;
+ }
+ return cost+1;
+}
+
+#define get_const_weight(irn) (1)
+
+#define get_spill_weight(irn) get_loop_weight(irn)
+#define get_reload_weight(irn) get_loop_weight(irn)
+#define get_affinity_weight(irn) get_loop_weight(irn)
+
+/******************************************************************************
+ _____ _ _____ _
+ / ____| | | / ____| (_)
+ | | ___ _ __ ___| |_ _ __ | | ___ _ __ _ ___ ___
+ | | / _ \| '_ \/ __| __| '__| | | / _ \| '_ \| |/ _ \/ __|
+ | |___| (_) | | | \__ \ |_| | | |___| (_) | |_) | | __/\__ \
+ \_____\___/|_| |_|___/\__|_| \_____\___/| .__/|_|\___||___/
+ | |
+ |_|
+ *****************************************************************************/
+
+static void handle_constraints_walker(ir_node *irn, void *env) {
+ be_raext_env_t *raenv = env;
+ arch_register_req_t req;
+ int pos, max;
+
+ /* handle output constraints
+ * user -> irn becomes user -> cpy -> irn
+ */
+ arch_get_register_req(raenv->aenv, &req, irn, -1);
+ if (arch_register_req_is(&req, limited)) {
+ ir_node *cpy = be_new_Copy(req.cls, raenv->irg, get_nodes_block(irn), irn);
+ const ir_edge_t *edge;
+
+ /* all users of the irn use the copy instead */
+ sched_add_after(irn, cpy);
+ foreach_out_edge(irn, edge)
+ set_irn_n(edge->src, edge->pos, cpy);
+ }
+
+
+ /* handle input constraints by converting them into output constraints
+ * of copies of the former argument
+ * irn -> arg becomes irn -> copy -> arg
+ */
+ for (pos = 0, max = get_irn_arity(irn); pos<max; ++pos) {
+ arch_get_register_req(raenv->aenv, &req, irn, pos);
+ if (arch_register_req_is(&req, limited)) {
+ ir_node *arg = get_irn_n(irn, pos);
+ ir_node *cpy = be_new_Copy(req.cls, raenv->irg, get_nodes_block(irn), arg);
+
+ /* use the copy instead */
+ sched_add_before(irn, cpy);
+ set_irn_n(irn, pos, cpy);
+
+ /* set an out constraint for the copy */
+ be_set_constr_limited(cpy, -1, &req);
+ }
+ }
+}
+
+static void handle_constraints(be_raext_env_t *raenv) {
+ irg_block_walk_graph(raenv->irg, NULL, handle_constraints_walker, raenv);
+}
+
+
+/******************************************************************************
+ _____ _____ _____ _
+ / ____/ ____| /\ | __ \ | |
+ | (___| (___ / \ ______| | | | ___ ___| |_ _ __
+ \___ \\___ \ / /\ \______| | | |/ _ \/ __| __| '__|
+ ____) |___) / ____ \ | |__| | __/\__ \ |_| |
+ |_____/_____/_/ \_\ |_____/ \___||___/\__|_|
+
+ *****************************************************************************/
+
+#define mark_as_done(irn, pos) set_irn_link(irn, INT_TO_PTR(pos+1))
+#define has_been_done(irn, pos) (PTR_TO_INT(get_irn_link(irn)) > pos)
/**
- * Checks if the irn uses or defines values of the
- * current register class (raenv->cls)
+ * Insert a copy for the argument of @p start_phi found at position @p pos.
+ * Also searches a phi-loop of arbitrary length to detect and resolve
+ * the class of phi-swap-problems. To search for a loop recursion is used.
+ *
+ * 1) Simplest case (phi with a non-phi arg):
+ * A single copy is inserted.
+ *
+ * 2) Phi chain (phi (with phi-arg)* with non=phi arg):
+ * Several copies are placed, each after returning from recursion.
+ *
+ * 3) Phi-loop:
+ * On detection a loop breaker is inserted, which is a copy of the start_phi.
+ * This copy then pretends beeing the argumnent of the last phi.
+ * Now case 2) can be used.
+ *
+ * The values of @p start_phi and @p pos never change during recursion.
+ *
+ * @p raenv Environment with all the stuff needed
+ * @p start_phi Phi node to process
+ * @p pos Argument position to insert copy/copies for
+ * @p curr_phi Phi node currently processed during recursion. Equals start_phi on initial call
+ *
+ * @return NULL If no copy is necessary
+ * NULL If the phi has already been processed at this pos
+ * Link field is used to keep track of processed positions
+ * In all other cases the ir_node *copy which was placed is returned.
*/
-static INLINE int is_sth_in_reg_class(be_raext_env_t *raenv, const ir_node *irn) {
- int max, i;
-
- /* check arguments */
- for (i=0, max=get_irn_arity(irn); i<max; ++i)
- if (arch_irn_has_reg_class(raenv->aenv, get_irn_n(irn, i), -1, raenv->cls))
- return 1;
-
- /* check result(s) */
- if (get_irn_mode(irn) == mode_T) {
- ir_node *proj;
- for (proj = sched_next(irn); is_Proj(proj); proj = sched_next(proj))
- if (arch_irn_has_reg_class(raenv->aenv, proj, -1, raenv->cls))
- return 1;
- return 0;
- } else {
- return arch_irn_has_reg_class(raenv->aenv, irn, -1, raenv->cls);
+static ir_node *insert_copies(be_raext_env_t *raenv, ir_node *start_phi, int pos, ir_node *curr_phi) {
+ ir_node *arg = get_irn_n(curr_phi, pos);
+ ir_node *arg_blk = get_nodes_block(arg);
+ ir_node *pred_blk = get_Block_cfgpred_block(get_nodes_block(curr_phi), pos);
+ ir_node *curr_cpy, *last_cpy;
+
+ assert(is_Phi(start_phi) && is_Phi(curr_phi));
+
+ if (has_been_done(start_phi, pos))
+ return NULL;
+
+ /* In case this is a 'normal' phi we insert into
+ * the schedule before the pred_blk irn */
+ last_cpy = pred_blk;
+
+ /* If we detect a loop stop recursion. */
+ if (arg == start_phi) {
+ ir_node *loop_breaker;
+ if (start_phi == curr_phi) {
+ /* Phi directly uses itself. No copy necessary */
+ return NULL;
+ }
+
+ /* At least 2 phis are involved */
+ /* Insert a loop breaking copy (an additional variable T) */
+ loop_breaker = be_new_Copy(raenv->cls, raenv->irg, pred_blk, start_phi);
+ sched_add_before(pred_blk, loop_breaker);
+
+ arg = loop_breaker;
}
- assert(0 && "Where did you come from???");
+ /* If arg is a phi in the same block we have to continue search */
+ if (is_Phi(arg) && arg_blk == get_nodes_block(start_phi))
+ last_cpy = insert_copies(raenv, start_phi, pos, arg);
+
+ /* Insert copy of argument (may be the loop-breaker) */
+ curr_cpy = be_new_Copy(raenv->cls, raenv->irg, pred_blk, arg);
+ set_irn_n(curr_phi, pos, curr_cpy);
+ mark_as_done(curr_phi, pos);
+ sched_add_before(last_cpy, curr_cpy);
+ return curr_cpy;
}
/**
- * Perform simple SSA-destruction with copies
- * TODO: Phi-Swap-Problem
+ * Perform simple SSA-destruction with copies.
+ * The order of processing _must_ be
+ * for all positions {
+ * for all phis {
+ * doit
+ * }
+ * }
+ * else the magic to keep track of processed phi-positions will fail in
+ * function 'insert_copies'
*/
-static void ssa_destr_simple(ir_node *blk, void *env) {
+static void ssa_destr_simple_walker(ir_node *blk, void *env) {
be_raext_env_t *raenv = env;
+ int pos, max;
ir_node *phi;
- /* for all phi nodes (which are scheduled at first) */
- sched_foreach(blk, phi) {
- int i, max;
- const arch_register_class_t *cls;
+ /* for all argument positions of the phis */
+ for (pos=0, max=get_irn_arity(blk); pos<max; ++pos) {
- if (!is_Phi(phi))
- break;
+ /* for all phi nodes (which are scheduled first) */
+ sched_foreach(blk, phi) {
+ if (!is_Phi(phi))
+ break;
- cls = arch_get_irn_reg_class(raenv->aenv, phi, -1);
-
- /* for all args of these phis */
- for (i=0, max=get_irn_arity(phi); i<max; ++i) {
- ir_node *arg = get_irn_n(phi, i);
- ir_node *pred_blk = get_Block_cfgpred_block(blk, i);
- ir_node *cpy = be_new_Copy(cls, raenv->irg, pred_blk, arg);
- set_irn_n(phi, i, cpy);
- sched_add_before(pred_blk, cpy);
+ raenv->cls = arch_get_irn_reg_class(raenv->aenv, phi, -1);
+ insert_copies(raenv, phi, pos, phi);
}
}
}
+static void ssa_destr_simple(be_raext_env_t *raenv) {
+ be_clear_links(raenv->irg);
+ irg_block_walk_graph(raenv->irg, ssa_destr_simple_walker, NULL, raenv);
+}
+
+
+static void ssa_destr_rastello(be_raext_env_t *raenv) {
+ assert(0 && "NYI");
+ exit(0xDeadBeef);
+ /*
+ phi_class_compute(raenv->irg);
+ irg_block_walk_graph(irg, ssa_destr_rastello, NULL, &raenv);
+ */
+}
+
+/******************************************************************************
+ __ __ _ ___ __ __
+ \ \ / / | | |__ \ \ \ / /
+ \ \ / /_ _| |___ ) | \ \ / /_ _ _ __ ___
+ \ \/ / _` | / __| / / \ \/ / _` | '__/ __|
+ \ / (_| | \__ \ / /_ \ / (_| | | \__ \
+ \/ \__,_|_|___/ |____| \/ \__,_|_| |___/
+ *****************************************************************************/
+
+/**
+ * This struct maps a variable (nr) to the values belonging to this variable
+ */
+struct _var_info_t {
+ int var_nr; /* the key */
+ pset *values; /* the ssa-values belonging to this variable */
+};
+
+#define SET_REMOVED -1
+
+/**
+ * The link field of an irn points to the var_info struct
+ * representing the corresponding variable.
+ */
+#define set_var_info(irn, vi) set_irn_link(irn, vi)
+#define get_var_info(irn) ((var_info_t *)get_irn_link(irn))
+
+#define HASH_VAR_NR(var_nr) var_nr
+
+static int compare_var_infos(const void *e1, const void *e2, size_t size) {
+ const var_info_t *v1 = e1;
+ const var_info_t *v2 = e2;
+
+ if (v1->var_nr == SET_REMOVED || v2->var_nr == SET_REMOVED)
+ return 1;
+
+ return v1->var_nr != v2->var_nr;
+}
+
+static INLINE var_info_t *var_find(set *vars, int var_nr) {
+ var_info_t vi;
+ vi.var_nr = var_nr;
+
+ return set_find(vars, &vi, sizeof(vi), HASH_VAR_NR(var_nr));
+}
+
+static INLINE var_info_t *var_find_or_insert(set *vars, int var_nr) {
+ var_info_t vi, *found;
+ memset(&vi, 0, sizeof(vi));
+ vi.var_nr = var_nr;
+
+ found = set_insert(vars, &vi, sizeof(vi), HASH_VAR_NR(var_nr));
+
+ if (!found->values)
+ found->values = pset_new_ptr(1);
+
+ return found;
+}
+
+/**
+ * Adds a value to a variable. Sets all pointers accordingly.
+ */
+static INLINE var_info_t *var_add_value(be_raext_env_t *raenv, int var_nr, ir_node *irn) {
+ var_info_t *vi = var_find_or_insert(raenv->vars, var_nr);
+
+ /* var 2 value mapping */
+ pset_insert_ptr(vi->values, irn);
+
+ /* value 2 var mapping */
+ set_var_info(irn, vi);
+
+ return vi;
+}
+
+static INLINE pset *get_var_values(be_raext_env_t *raenv, int var_nr) {
+ var_info_t *vi = var_find(raenv->vars, var_nr);
+ assert(vi && "Variable does not exist");
+ return vi->values;
+}
+
/**
* Define variables (numbers) for all SSA-values.
* All values in a phi class get assigned the same variable name.
*/
static void values_to_vars(ir_node *irn, void *env) {
be_raext_env_t *raenv = env;
- ir_node *n;
int nr;
pset *vals;
+ if(arch_get_irn_reg_class(raenv->aenv, irn, -1) == NULL)
+ return;
+
vals = get_phi_class(irn);
- if (!vals) {
+ if (vals) {
+ nr = get_irn_node_nr(get_first_phi(vals));
+ } else {
/* not a phi class member, value == var */
+ nr = get_irn_node_nr(irn);
vals = pset_new_ptr(1);
pset_insert_ptr(vals, irn);
}
- /* value to var mapping */
- n = pset_first(vals);
- nr = get_irn_node_nr(n);
- for (; n; n=pset_next(vals))
- set_var_nr(irn, nr);
+ /* values <--> var mapping */
+ pset_foreach(vals, irn)
+ var_add_value(raenv, nr, irn);
+}
+
+
+/******************************************************************************
+ _____
+ | __ \
+ | | | |_ _ _ __ ___ _ __ ___ _ __
+ | | | | | | | '_ ` _ \| '_ \ / _ \ '__|
+ | |__| | |_| | | | | | | |_) | __/ |
+ |_____/ \__,_|_| |_| |_| .__/ \___|_|
+ | |
+ |_|
+ *****************************************************************************/
+
+
+static void extract_vars_of_cls(be_raext_env_t *raenv) {
+ int count = 0;
+ var_info_t *vi;
+
+ raenv->cls_vars = malloc(set_count(raenv->vars) * sizeof(*raenv->cls_vars));
+ assert(raenv->cls_vars);
- /* var to values mapping */
- pmap_insert_sth(raenv->vars, nr, vals);
+ set_foreach(raenv->vars, vi)
+ if (is_res_in_reg_class(get_first_non_phi(vi->values)))
+ raenv->cls_vars[count++] = vi;
+
+ raenv->cls_vars = realloc(raenv->cls_vars, count * sizeof(*raenv->cls_vars));
+ assert(raenv->cls_vars);
+
+ raenv->n_cls_vars = count;
}
+
/**
* Check if node irn has a limited-constraint at position pos.
* If yes, dump it to FILE raenv->f
arch_get_register_req(raenv->aenv, &req, irn, pos);
if (arch_register_req_is(&req, limited)) {
int reg_nr;
- req.limited(irn, pos, bs);
+ req.limited(req.limited_env, bs);
reg_nr = bitset_next_set(bs, 0);
- fprintf(raenv->f, " <%d>", reg_nr);
+ fprintf(raenv->f, "<%d>", reg_nr);
assert(-1 == bitset_next_set(bs, reg_nr+1) && "Constraints with more than 1 possible register are not supported");
}
}
-
-/**
- * Dump all blocks and instructions in that block
- */
-static void dump_blocks(ir_node *blk, void *env) {
- be_raext_env_t *raenv = env;
+static INLINE unsigned int get_spill_costs(be_raext_env_t *raenv, var_info_t *vi) {
ir_node *irn;
+ int c_spills=0, c_reloads=0;
+
+ pset_foreach(vi->values, irn) {
+ if (arch_irn_is_ignore(raenv->aenv, irn)) {
+ pset_break(vi->values);
+ return -1;
+ }
+
+ if (is_Phi(irn)) {
+ /* number of reloads is the number of non-phi uses of all values of this var */
+ const ir_edge_t *edge;
+ foreach_out_edge(irn, edge)
+ if (!is_Phi(edge->src))
+ c_reloads += get_reload_weight(edge->src);
+ } else {
+ /* number of spills is the number of non-phi values for this var */
+ c_spills += get_spill_weight(irn);
+ }
+ }
+
+ return c_spills + c_reloads;
+}
+
+static void dump_nodes(be_raext_env_t *raenv) {
FILE *f = raenv->f;
- int nr = get_irn_node_nr(blk);
+ int i;
- pmap_insert_sth(raenv->blocks, nr, blk);
+ fprintf(f, "\nnodes {\n");
- /* begin block scope */
- fprintf(f, "\n");
- fprintf(f, " block %d {\n", nr);
+ for (i=0; i<raenv->n_cls_vars; ++i) {
+ var_info_t *vi = raenv->cls_vars[i];
- /* for each instruction */
- for(irn=sched_first(blk); !sched_is_end(irn); irn=sched_next(irn)) {
- int max, i;
- if (is_Phi(irn) || !is_sth_in_reg_class(raenv, irn))
+ if (vi->var_nr == SET_REMOVED)
continue;
- fprintf(f, " insn %ld {\n", get_irn_node_nr(irn));
-
- /*
- * print all uses
- */
- fprintf(f, " use");
- for (i=0, max=get_irn_arity(irn); i<max; ++i) {
- ir_node *arg = get_irn_n(irn, i);
- if (arch_irn_has_reg_class(raenv->aenv, arg, -1, raenv->cls)) {
- fprintf(f, " %d", get_var_nr(arg));
- dump_constraint(raenv, irn, i);
- }
- }
- fprintf(f,"\n");
-
- /*
- * print all defs
- */
- fprintf(f, " def");
- /* special handling of projs */
- if (get_irn_mode(irn) == mode_T) {
- for (irn = sched_next(irn); is_Proj(irn); irn = sched_next(irn))
- if (arch_irn_has_reg_class(raenv->aenv, irn, -1, raenv->cls)) {
- fprintf(f, " %d", get_var_nr(irn));
- dump_constraint(raenv, irn, -1);
- }
- irn = sched_prev(irn); /* for outer loop */
- } else {
- if (arch_irn_has_reg_class(raenv->aenv, irn, -1, raenv->cls)) {
- fprintf(f, " %d", get_var_nr(irn));
- dump_constraint(raenv, irn, -1);
- }
- }
- fprintf(f,"\n");
-
- fprintf(f, " }\n");
+ fprintf(f, "%d %u", vi->var_nr, get_spill_costs(raenv, vi));
+ dump_constraint(raenv, get_first_non_phi(vi->values), -1);
+ fprintf(f, "\n");
}
- /* end the block scope */
- fprintf(f, " }\n");
+ fprintf(f, "}\n");
}
-/**
- * Dump all control flow edges of this irg
- */
-static void dump_edges(ir_node *blk, void *env) {
+static void dump_interferences(be_raext_env_t *raenv) {
+ int i,o;
+ var_info_t *vi1, *vi2;
+ ir_node *irn1, *irn2;
+ FILE *f = raenv->f;
+
+ fprintf(f, "\ninterferences {\n");
+
+ for (i=0; i<raenv->n_cls_vars; ++i) {
+ vi1 = raenv->cls_vars[i];
+
+ if (vi1->var_nr == SET_REMOVED)
+ continue;
+
+ for (o=i+1; o<raenv->n_cls_vars; ++o) {
+ vi2 = raenv->cls_vars[o];
+
+ if (vi2->var_nr == SET_REMOVED)
+ continue;
+
+ pset_foreach(vi1->values, irn1)
+ pset_foreach(vi2->values, irn2)
+ if (values_interfere(irn1, irn2)) {
+ pset_break(vi1->values);
+ pset_break(vi2->values);
+ fprintf(f, "(%d, %d)\n", vi1->var_nr, vi2->var_nr);
+ }
+ }
+ }
+ fprintf(f, "}\n");
+}
+
+static void dump_affinities_walker(ir_node *irn, void *env) {
be_raext_env_t *raenv = env;
- int i, max;
+ arch_register_req_t req;
+ int pos, max;
+ var_info_t *vi1, *vi2;
- if (get_irg_start_block(get_irn_irg(blk)) == blk)
+ if (arch_get_irn_reg_class(raenv->aenv, irn, -1) == NULL || arch_irn_is_ignore(raenv->aenv, irn))
return;
- /* dump cf edges in the flow-order "pred succ" */
- for (i=0, max=get_irn_arity(blk); i<max; ++i) {
- ir_node *pred = get_Block_cfgpred_block(blk, i);
- fprintf(raenv->f, " cf_edge %ld %ld\n", get_irn_node_nr(pred), get_irn_node_nr(blk));
+ vi1 = get_var_info(irn);
+
+ /* copies have affinities */
+ if (arch_irn_classify(raenv->aenv, irn) == arch_irn_class_copy) {
+ ir_node *other = get_irn_n(irn, 0);
+
+ if (! arch_irn_is_ignore(raenv->aenv, other)) {
+ vi2 = get_var_info(other);
+
+ fprintf(raenv->f, "(%d, %d, %d)\n", vi1->var_nr, vi2->var_nr, get_affinity_weight(irn));
+ }
+ }
+
+
+ /* should_be_equal constraints are affinites */
+ for (pos = 0, max = get_irn_arity(irn); pos<max; ++pos) {
+ arch_get_register_req(raenv->aenv, &req, irn, pos);
+
+ if (arch_register_req_is(&req, should_be_same) && arch_irn_is_ignore(raenv->aenv, req.other_same)) {
+ vi2 = get_var_info(req.other_same);
+
+ fprintf(raenv->f, "(%d, %d, %d)\n", vi1->var_nr, vi2->var_nr, get_affinity_weight(irn));
+ }
}
}
+static void dump_affinities(be_raext_env_t *raenv) {
+ fprintf(raenv->f, "\naffinities {\n");
+ irg_walk_graph(raenv->irg, NULL, dump_affinities_walker, raenv);
+ fprintf(raenv->f, "}\n");
+}
+
/**
* Dump all information needed by the external
* register allocator to a single file.
if (!(f = fopen(filename, "wt"))) {
fprintf(stderr, "Could not open file %s for writing\n", filename);
- exit(1);
+ assert(0);
+ exit(0xdeadbeef);
}
raenv->f = f;
+ /* dump register info */
fprintf(f, "regs %d\n", arch_register_class_n_regs(raenv->cls));
- fprintf(f, "cfg %s {\n", filename);
- irg_block_walk_graph(raenv->irg, NULL, dump_blocks, raenv);
- irg_block_walk_graph(raenv->irg, NULL, dump_edges, raenv);
-
- fprintf(f, "}\n");
+ /* dump the interference graph */
+ dump_nodes(raenv);
+ dump_interferences(raenv);
+ dump_affinities(raenv);
fclose(f);
}
+/******************************************************************************
+ ______ _
+ | ____| | |
+ | |__ __ _____ ___ _ _| |_ ___
+ | __| \ \/ / _ \/ __| | | | __/ _ \
+ | |____ > < __/ (__| |_| | || __/
+ |______/_/\_\___|\___|\__,_|\__\___|
+ *****************************************************************************/
/**
* Execute the external register allocator specified in the
- * firm-option TODO
+ * firm-option firm.be.ra.ext.callee
*/
-static void execute(char *out_file, char *result_file) {
+static void execute(char *prog_to_call, char *out_file, char *result_file) {
char cmd_line[1024];
- char *app_name = "echo"; //TODO get from firm-options file
int ret_status;
- snprintf(cmd_line, sizeof(cmd_line), "%s %s %s", app_name, out_file, result_file);
+ snprintf(cmd_line, sizeof(cmd_line), "%s -i %s -o %s", prog_to_call, out_file, result_file);
ret_status = system(cmd_line);
assert(ret_status != -1 && "Invokation of external register allocator failed");
}
+/******************************************************************************
+ _ _____ _ _
+ /\ | | | __ \ | | |
+ / \ _ __ _ __ | |_ _ | |__) |___ ___ _ _| | |_
+ / /\ \ | '_ \| '_ \| | | | | | _ // _ \/ __| | | | | __|
+ / ____ \| |_) | |_) | | |_| | | | \ \ __/\__ \ |_| | | |_
+ /_/ \_\ .__/| .__/|_|\__, | |_| \_\___||___/\__,_|_|\__|
+ | | | | __/ |
+ |_| |_| |___/
+ *****************************************************************************/
+
+/**
+ * Spill a variable and add reloads before all uses.
+ */
+static INLINE void var_add_spills_and_reloads(be_raext_env_t *raenv, int var_nr) {
+ var_info_t *vi = var_find(raenv->vars, var_nr);
+ ir_node *spill=NULL, *ctx, *irn;
+ const ir_edge_t *edge, *ne;
+ pset *spills = pset_new_ptr(4); /* the spills of this variable */
+ pset *reloads = pset_new_ptr(4); /* the reloads of this variable */
+ int new_size, n_spills, n_reloads;
+
+ assert(vi && "Variable nr does not exist!");
+ assert(pset_count(vi->values) && "There are no values associated to this variable");
+
+ /* the spill context is set to an arbitrary node of the phi-class,
+ * or the node itself if it is not member of a phi class
+ */
+ if (pset_count(vi->values) == 1)
+ ctx = get_first_non_phi(vi->values);
+ else
+ ctx = get_first_phi(vi->values);
+
+ DBG((raenv->dbg, LEVEL_2, "Spill context: %+F\n", ctx));
+
+ /* for each value of this variable insert the spills */
+ pset_foreach(vi->values, irn) {
+ if (is_Phi(irn)) {
+ sched_remove(irn);
+ continue;
+ }
+
+ /* all ordinary nodes must be spilled */
+ DBG((raenv->dbg, LEVEL_2, " spilling %+F\n", irn));
+ spill = be_spill(raenv->aenv, irn, ctx);
+
+ /* remember the spill */
+ pset_insert_ptr(spills, spill);
+ }
+
+ assert(spill && "There must be at least one non-phi-node");
+
+ /* insert reloads and wire them arbitrary*/
+ pset_foreach(vi->values, irn)
+ foreach_out_edge_safe(irn, edge, ne) {
+ ir_node *reload, *src = edge->src;
+ if (is_Phi(src) || be_is_Spill(src))
+ continue;
+
+ /* all real uses must be reloaded */
+ DBG((raenv->dbg, LEVEL_2, " reloading before %+F\n", src));
+ reload = be_reload(raenv->aenv, raenv->cls, edge->src, edge->pos, get_irn_mode(get_irn_n(spill, 0)), spill);
+
+ /* remember the reload */
+ pset_insert_ptr(reloads, reload);
+ }
+
+ /* correct the reload->spill pointers... */
+ be_ssa_constr_set(raenv->dom_info, spills);
+
+
+ /****** correct the variable <--> values mapping: ******
+ *
+ * - if we had a phi class it gets split into several new variables
+ * - all reloads are new variables
+ */
+ n_spills = pset_count(spills);
+ n_reloads = pset_count(reloads);
+
+ /* first make room for new pointers in the cls_var array */
+ new_size = raenv->n_cls_vars + n_reloads + ((n_spills>1) ? n_spills : 0);
+ raenv->cls_vars = realloc(raenv->cls_vars, (new_size) * sizeof(*raenv->cls_vars));
+ assert(raenv->cls_vars && "Out of mem!?");
+
+ /* if we had a real phi-class, we must... */
+ if (pset_count(spills) > 1) {
+ /* ...remove the old variable corresponding to the phi class */
+ vi->var_nr = SET_REMOVED;
+
+ /* ...add new vars for each non-phi-member */
+ pset_foreach(spills, irn) {
+ ir_node *spilled = get_irn_n(irn, 0);
+ raenv->cls_vars[raenv->n_cls_vars++] = var_add_value(raenv, get_irn_node_nr(spilled), spilled);
+ }
+ }
+
+ /* add new variables for all reloads */
+ pset_foreach(reloads, irn)
+ raenv->cls_vars[raenv->n_cls_vars++] = var_add_value(raenv, get_irn_node_nr(irn), irn);
+
+ del_pset(spills);
+ del_pset(reloads);
+}
+
+#define INVALID_FILE_FORMAT assert(0 && "Invalid file format.")
+#define BUFLEN 32
+#define BUFCONV " %32s "
/**
* Read in the actions performed by the external allocator.
- * Apply these transformations to the irg->
+ * Apply these transformations to the irg.
+ * @return 1 if an allocation was read in. 0 otherwise.
*/
-static void read_and_apply_results(be_raext_env_t *raenv, char *filename) {
+static int read_and_apply_results(be_raext_env_t *raenv, char *filename) {
FILE *f;
+ char buf[BUFLEN];
+ int is_allocation = 0;
if (!(f = fopen(filename, "rt"))) {
fprintf(stderr, "Could not open file %s for reading\n", filename);
- exit(1);
+ assert(0);
+ exit(0xdeadbeef);
}
raenv->f = f;
- //TODO: free pmap entries (the psets) pmap_foreach(raenv.vars, pme) del_pset(pme->value);
+ /* read the action */
+ if (fscanf(f, BUFCONV, buf) != 1)
+ INVALID_FILE_FORMAT;
+
+ /* do we spill */
+ if (!strcmp(buf, "spills")) {
+ int var_nr;
+ while (fscanf(f, " %d ", &var_nr) == 1)
+ var_add_spills_and_reloads(raenv, var_nr);
+ } else
+
+ /* or do we allocate */
+ if (!strcmp(buf, "allocs")) {
+ int var_nr, reg_nr;
+
+ is_allocation = 1;
+ while (fscanf(f, " %d %d ", &var_nr, ®_nr) == 2) {
+ ir_node *irn;
+ pset *vals = get_var_values(raenv, var_nr);
+
+ assert(vals && "Variable nr does not exist!");
+ pset_foreach(vals, irn)
+ arch_set_irn_register(raenv->aenv, irn, arch_register_for_index(raenv->cls, reg_nr));
+ }
+ } else
+ INVALID_FILE_FORMAT;
+
+ if (!feof(f))
+ INVALID_FILE_FORMAT;
fclose(f);
+
+ return is_allocation;
+}
+
+static void check_allocation(be_raext_env_t *raenv) {
+ int i, o;
+
+ for (i=0; i<raenv->n_cls_vars; ++i) {
+ var_info_t *vi1 = raenv->cls_vars[i];
+
+ if (vi1->var_nr == SET_REMOVED)
+ continue;
+
+ for (o=0; o<i; ++o) {
+ var_info_t *vi2 = raenv->cls_vars[o];
+ ir_node *irn1, *irn2;
+
+ if (vi2->var_nr == SET_REMOVED)
+ continue;
+
+ pset_foreach(vi1->values, irn1)
+ pset_foreach(vi2->values, irn2)
+ if (values_interfere(irn1, irn2)) {
+ dump_ir_block_graph_sched(raenv->irg, "ERROR");
+ ir_fprintf(stdout, "SSA values %+F and %+F interfere. They belong to varible %d and %d respectively.\n", irn1, irn2, vi1->var_nr, vi2->var_nr);
+ assert(0 && "ERROR graph dumped");
+ }
+ }
+ }
}
+/******************************************************************************
+ __ __ _
+ | \/ | (_)
+ | \ / | __ _ _ _ __
+ | |\/| |/ _` | | '_ \
+ | | | | (_| | | | | |
+ |_| |_|\__,_|_|_| |_|
+ *****************************************************************************/
+
+/**
+ * Default values for options
+ */
+static void (*ssa_destr)(be_raext_env_t*) = ssa_destr_simple;
+static char callee[128] = "\"E:/user/kimohoff/ipd-registerallocator/win32/register allocator\"";
+//static char callee[128] = "/ben/kimohoff/ipd-registerallocator/register_allocator";
+
/**
* Allocate registers with an external program using a text-file interface.
* Read in results and apply them
*
*/
-static void be_ra_extern_main(const be_main_env_t *env, ir_graph *irg) {
+static void be_ra_extern_main(const be_irg_t *bi) {
+ be_main_env_t *env = bi->main_env;
+ ir_graph *irg = bi->irg;
+
be_raext_env_t raenv;
int clsnr, clss;
+ var_info_t *vi;
- raenv.irg = irg;
- raenv.aenv = env->arch_env;
- raenv.vars = pmap_create();
- raenv.blocks = pmap_create();
+ compute_doms(irg);
- /* SSA destruction */
- be_clear_links(irg);
- irg_block_walk_graph(irg, ssa_destr_simple, NULL, &raenv);
+ raenv.irg = irg;
+ raenv.aenv = env->arch_env;
+ raenv.dom_info = be_compute_dominance_frontiers(irg);
+ raenv.vars = new_set(compare_var_infos, 64);
+ raenv.dbg = firm_dbg_register("ir.be.raextern");
+ firm_dbg_set_mask(raenv.dbg, DBG_LEVEL);
+
+ /* Insert copies for constraints */
+ handle_constraints(&raenv);
+ dump_ir_block_graph_sched(irg, "-extern-constr");
+
+ /* SSA destruction respectively transformation into "Conventional SSA" */
+ ssa_destr(&raenv);
+ dump_ir_block_graph_sched(irg, "-extern-ssadestr");
+
+ /* Mapping of SSA-Values <--> Variables */
phi_class_compute(irg);
+ be_clear_links(irg);
irg_walk_graph(irg, values_to_vars, NULL, &raenv);
- dump_ir_block_graph_sched(irg, "-extern-ssadestr");
/* For all register classes */
for(clsnr = 0, clss = arch_isa_get_n_reg_class(raenv.aenv->isa); clsnr < clss; ++clsnr) {
+ int done, round = 1;
char out[256], in[256];
raenv.cls = arch_isa_get_reg_class(raenv.aenv->isa, clsnr);
- ir_snprintf(out, sizeof(out), "%F-%s.ra", irg, raenv.cls->name);
- ir_snprintf(in, sizeof(in), "%F-%s.ra.res", irg, raenv.cls->name);
- dump_to_file(&raenv, out);
+ extract_vars_of_cls(&raenv);
+
+ do {
+ ir_snprintf(out, sizeof(out), "%F-%s-%d.ra", irg, raenv.cls->name, round);
+ ir_snprintf(in, sizeof(in), "%F-%s-%d.ra.res", irg, raenv.cls->name, round);
+
+ be_liveness(irg);
+
+ dump_to_file(&raenv, out);
+ execute(callee, out, in);
+ done = read_and_apply_results(&raenv, in);
- execute(out, in);
+ ir_snprintf(in, sizeof(in), "-extern-%s-round-%d", raenv.cls->name, round);
+ dump_ir_block_graph_sched(irg, in);
- read_and_apply_results(&raenv, in);
+ round++;
+ } while (!done);
+
+ check_allocation(&raenv);
+
+ free(raenv.cls_vars);
}
+ dump_ir_block_graph_sched(irg, "-extern-alloc");
+
/* Clean up */
- pmap_destroy(raenv.blocks);
- pmap_destroy(raenv.vars);
+ set_foreach(raenv.vars, vi)
+ del_pset(vi->values);
+ del_set(raenv.vars);
+ be_free_dominance_frontiers(raenv.dom_info);
}
+/******************************************************************************
+ ____ _ _
+ / __ \ | | (_)
+ | | | |_ __ | |_ _ ___ _ __ ___
+ | | | | '_ \| __| |/ _ \| '_ \/ __|
+ | |__| | |_) | |_| | (_) | | | \__ \
+ \____/| .__/ \__|_|\___/|_| |_|___/
+ | |
+ |_|
+ *****************************************************************************/
#ifdef WITH_LIBCORE
-static void be_ra_extern_register_options(lc_opt_entry_t *grp) {
- /* TODO */
+
+
+static const lc_opt_enum_func_ptr_items_t ssa_destr_items[] = {
+ { "simple", (int (*)()) ssa_destr_simple }, /* TODO make (void*) casts nicer */
+ { "rastello", (int (*)()) ssa_destr_rastello },
+ { NULL, NULL }
+};
+
+static lc_opt_enum_func_ptr_var_t ssa_destr_var = {
+ (int (**)()) &ssa_destr, ssa_destr_items
+};
+
+static const lc_opt_table_entry_t be_ra_extern_options[] = {
+ LC_OPT_ENT_ENUM_FUNC_PTR("ssa_destr", "SSA destruction flavor", &ssa_destr_var),
+ LC_OPT_ENT_STR("callee", "The external program to call", callee, sizeof(callee)),
+ { NULL }
+};
+
+static void be_ra_extern_register_options(lc_opt_entry_t *root) {
+ lc_opt_entry_t *grp = lc_opt_get_grp(root, "ext");
+
+ lc_opt_add_table(grp, be_ra_extern_options);
}
-#endif
+
+#endif /* WITH_LIBCORE */
const be_ra_t be_ra_external_allocator = {
#ifdef WITH_LIBCORE
projects / libfirm / blobdiff