fix bugs when exchanging nodes to projs in bepeephole
[libfirm] / ir / be / bechordal.c
index f4638e0..09c03a7 100644 (file)
 #include "irdump.h"
 #include "irdom.h"
 #include "irtools.h"
+#include "irbitset.h"
 #include "debug.h"
 #include "xmalloc.h"
+#include "iredges.h"
 
 #include "beutil.h"
 #include "besched.h"
@@ -60,6 +62,7 @@
 #include "beinsn_t.h"
 #include "bestatevent.h"
 #include "beirg_t.h"
+#include "beintlive_t.h"
 #include "bera.h"
 #include "bechordal_t.h"
 #include "bechordal_draw.h"
@@ -154,7 +157,7 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head
        else {
                b = get_irn_link(irn);
 
-               assert(b && b->magic == BORDER_FOURCC && "Illegal border encountered");
+               DEBUG_ONLY(assert(b && b->magic == BORDER_FOURCC && "Illegal border encountered"));
        }
 
        b->pressure = pressure;
@@ -177,8 +180,7 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head
  */
 static INLINE int has_reg_class(const be_chordal_env_t *env, const ir_node *irn)
 {
-       return arch_irn_has_reg_class(env->birg->main_env->arch_env, irn, -1, env->cls);
-       // return arch_irn_consider_in_reg_alloc(env->birg->main_env->arch_env, env->cls, irn);
+       return arch_irn_consider_in_reg_alloc(env->birg->main_env->arch_env, env->cls, irn);
 }
 
 #define has_limited_constr(req, irn) \
@@ -231,7 +233,8 @@ static be_insn_t *chordal_scan_insn(be_chordal_env_t *env, ir_node *irn)
 
 static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn)
 {
-       const arch_env_t *aenv = env->birg->main_env->arch_env;
+       const be_irg_t *birg   = env->birg;
+       const arch_env_t *aenv = birg->main_env->arch_env;
        bitset_t *tmp          = bitset_alloca(env->cls->n_regs);
        bitset_t *def_constr   = bitset_alloca(env->cls->n_regs);
        ir_node *bl            = get_nodes_block(irn);
@@ -290,6 +293,14 @@ static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn)
                        if(a_op->carrier != op->carrier || !a_op->has_constraints)
                                continue;
 
+                       /* if the constraint is the same, no copy is necessary
+                        * TODO generalise unequal but overlapping constraints */
+                       if (a_op->req == op->req)
+                               continue;
+
+                       if (be_is_Copy(get_irn_n(insn->irn, a_op->pos)))
+                               continue;
+
                        copy = be_new_Copy(env->cls, env->irg, bl, op->carrier);
                        be_stat_ev("constr_copy", 1);
 
@@ -324,17 +335,17 @@ static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn)
                        3) is constrained to a register occuring in out constraints.
                */
                if(!op->has_constraints ||
-                               !values_interfere(lv, insn->irn, op->carrier) ||
-                               bitset_popcnt(tmp) == 0)
+                  !values_interfere(birg, insn->irn, op->carrier) ||
+                  bitset_popcnt(tmp) == 0)
                        continue;
 
                /*
                   only create the copy if the operand is no copy.
                   this is necessary since the assure constraints phase inserts
-                  Copies and Keeps for operands which must be different from the results.
-                  Additional copies here would destroy this.
+                  Copies and Keeps for operands which must be different from the
+                  results. Additional copies here would destroy this.
                 */
-               if(be_is_Copy(op->carrier))
+               if (be_is_Copy(get_irn_n(insn->irn, op->pos)))
                        continue;
 
                copy = be_new_Copy(env->cls, env->irg, bl, op->carrier);
@@ -371,7 +382,6 @@ static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t
        int n_defs   = be_insn_n_defs(insn);
        bitset_t *bs = bitset_alloca(env->cls->n_regs);
        int *pairing = alloca(MAX(n_defs, n_uses) * sizeof(pairing[0]));
-       be_lv_t *lv  = env->birg->lv;
 
        int i, j;
 
@@ -391,7 +401,7 @@ static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t
 
                        if (op->partner != NULL)
                                continue;
-                       if (values_interfere(lv, op->irn, op->carrier))
+                       if (values_interfere(env->birg, op->irn, op->carrier))
                                continue;
 
                        bitset_clear_all(bs);
@@ -407,6 +417,11 @@ static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t
 
                if (smallest >= 0) {
                        be_operand_t *partner = &insn->ops[smallest];
+                       for(i = insn->use_start; i < insn->n_ops; ++i) {
+                               if(insn->ops[i].carrier == partner->carrier)
+                                       insn->ops[i].partner = out_op;
+                       }
+
                        out_op->partner  = partner;
                        partner->partner = out_op;
                }
@@ -460,7 +475,6 @@ static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env,
                the Perm. Recomputing liveness is also a good idea if a Perm is inserted, since
                the live sets may change.
        */
-       // be_liveness_recompute(lv);
        obstack_free(env->obst, insn);
        *the_insn = insn = chordal_scan_insn(env, insn->irn);
 
@@ -483,26 +497,28 @@ static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env,
        return perm;
 }
 
-static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *irn, int *silent)
+static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env,
+                                   ir_node *irn, int *silent)
 {
        const arch_env_t *aenv;
        int n_regs;
        bitset_t *bs;
        ir_node **alloc_nodes;
-       hungarian_problem_t *bp;
+       //hungarian_problem_t *bp;
        int *assignment;
        pmap *partners;
        int i, n_alloc;
-       long col;
+       bitset_pos_t col;
        const ir_edge_t *edge;
        ir_node *perm = NULL;
-       int match_res, cost;
+       //int match_res, cost;
        be_chordal_env_t *env  = alloc_env->chordal_env;
        void *base             = obstack_base(env->obst);
        be_insn_t *insn        = chordal_scan_insn(env, irn);
        ir_node *res           = insn->next_insn;
        int be_silent          = *silent;
-       be_lv_t *lv            = env->birg->lv;
+       be_irg_t *birg         = env->birg;
+       bipartite_t *bp;
 
        if(insn->pre_colored) {
                int i;
@@ -534,8 +550,8 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
        n_regs      = env->cls->n_regs;
        bs          = bitset_alloca(n_regs);
        alloc_nodes = alloca(n_regs * sizeof(alloc_nodes[0]));
-       bp          = hungarian_new(n_regs, n_regs, 2, HUNGARIAN_MATCH_PERFECT);
-       // bipartite_t *bp        = bipartite_new(n_regs, n_regs);
+       //bp          = hungarian_new(n_regs, n_regs, 2, HUNGARIAN_MATCH_PERFECT);
+       bp          = bipartite_new(n_regs, n_regs);
        assignment  = alloca(n_regs * sizeof(assignment[0]));
        partners    = pmap_create();
 
@@ -563,20 +579,36 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                        set of admissible registers via a bipartite graph.
                */
                if(!op->partner || !pmap_contains(partners, op->partner->carrier)) {
+                       ir_node *partner = op->partner ? op->partner->carrier : NULL;
+                       int i;
+
+                       pmap_insert(partners, op->carrier, partner);
+                       if(partner != NULL)
+                               pmap_insert(partners, partner, op->carrier);
+
+                       /* don't insert a node twice */
+                       for(i = 0; i < n_alloc; ++i) {
+                               if(alloc_nodes[i] == op->carrier) {
+                                       break;
+                               }
+                       }
+                       if(i < n_alloc)
+                               continue;
 
-                       pmap_insert(partners, op->carrier, op->partner ? op->partner->carrier : NULL);
                        alloc_nodes[n_alloc] = op->carrier;
 
-                       DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier, op->partner ? op->partner->carrier : NULL));
+                       DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier,
+                            partner));
 
                        bitset_clear_all(bs);
                        get_decisive_partner_regs(bs, op, op->partner);
 
-                       DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier, bs));
+                       DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier,
+                            bs));
 
                        bitset_foreach(bs, col) {
-                               hungarian_add(bp, n_alloc, col, 1);
-                               // bipartite_add(bp, n_alloc, col);
+                               //hungarian_add(bp, n_alloc, col, 1);
+                               bipartite_add(bp, n_alloc, col);
                        }
 
                        n_alloc++;
@@ -589,14 +621,26 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
        */
        if(perm != NULL) {
                foreach_out_edge(perm, edge) {
+                       int i;
                        ir_node *proj = get_edge_src_irn(edge);
 
                        assert(is_Proj(proj));
 
-                       if(!values_interfere(lv, proj, irn) || pmap_contains(partners, proj))
+                       if(!values_interfere(birg, proj, irn) || pmap_contains(partners, proj))
+                               continue;
+
+                       /* don't insert a node twice */
+                       for(i = 0; i < n_alloc; ++i) {
+                               if(alloc_nodes[i] == proj) {
+                                       break;
+                               }
+                       }
+                       if(i < n_alloc)
                                continue;
 
+
                        assert(n_alloc < n_regs);
+
                        alloc_nodes[n_alloc] = proj;
                        pmap_insert(partners, proj, NULL);
 
@@ -604,8 +648,8 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                        arch_put_non_ignore_regs(aenv, env->cls, bs);
                        bitset_andnot(bs, env->ignore_colors);
                        bitset_foreach(bs, col) {
-                               hungarian_add(bp, n_alloc, col, 1);
-                               // bipartite_add(bp, n_alloc, col);
+                               //hungarian_add(bp, n_alloc, col, 1);
+                               bipartite_add(bp, n_alloc, col);
                        }
 
                        n_alloc++;
@@ -613,30 +657,35 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
        }
 
        /* Compute a valid register allocation. */
+#if 0
        hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
        match_res = hungarian_solve(bp, assignment, &cost, 1);
        assert(match_res == 0 && "matching failed");
-       //bipartite_matching(bp, assignment);
+#else
+       bipartite_matching(bp, assignment);
+#endif
 
        /* Assign colors obtained from the matching. */
        for(i = 0; i < n_alloc; ++i) {
                const arch_register_t *reg;
-               ir_node *nodes[2];
-               int j;
+               ir_node *irn;
 
                assert(assignment[i] >= 0 && "there must have been a register assigned");
                reg = arch_register_for_index(env->cls, assignment[i]);
+               assert(! (reg->type & arch_register_type_ignore));
 
-               nodes[0] = alloc_nodes[i];
-               nodes[1] = pmap_get(partners, alloc_nodes[i]);
-
-               for(j = 0; j < 2; ++j) {
-                       if(!nodes[j])
-                               continue;
+               irn = alloc_nodes[i];
+               if (irn != NULL) {
+                       arch_set_irn_register(aenv, irn, reg);
+                       (void) pset_hinsert_ptr(alloc_env->pre_colored, irn);
+                       DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", irn, reg->name));
+               }
 
-                       arch_set_irn_register(aenv, nodes[j], reg);
-                       (void) pset_hinsert_ptr(alloc_env->pre_colored, nodes[j]);
-                       DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", nodes[j], reg->name));
+               irn = pmap_get(partners, alloc_nodes[i]);
+               if (irn != NULL) {
+                       arch_set_irn_register(aenv, irn, reg);
+                       (void) pset_hinsert_ptr(alloc_env->pre_colored, irn);
+                       DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", irn, reg->name));
                }
        }
 
@@ -671,8 +720,8 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                }
        }
 
-       //bipartite_free(bp);
-       hungarian_free(bp);
+       bipartite_free(bp);
+       //hungarian_free(bp);
        pmap_destroy(partners);
 
 end:
@@ -732,11 +781,10 @@ static void pressure(ir_node *block, void *env_ptr)
        be_lv_t *lv                       = env->birg->lv;
 
        int i, n;
+       bitset_pos_t elm;
        unsigned step = 0;
        unsigned pressure = 0;
        struct list_head *head;
-       pset *live_in  = be_lv_pset_put_in(lv, block, pset_new_ptr_default());
-       pset *live_end = be_lv_pset_put_end(lv, block, pset_new_ptr_default());
 
        DBG((dbg, LEVEL_1, "Computing pressure in block %+F\n", block));
        bitset_clear_all(live);
@@ -751,7 +799,8 @@ static void pressure(ir_node *block, void *env_ptr)
         * Make final uses of all values live out of the block.
         * They are necessary to build up real intervals.
         */
-       foreach_pset(live_end, irn) {
+       be_lv_foreach(lv, block, be_lv_state_end, i) {
+               ir_node *irn = be_lv_get_irn(lv, block, i);
                if(has_reg_class(env, irn)) {
                        DBG((dbg, LEVEL_3, "\tMaking live: %+F/%d\n", irn, get_irn_idx(irn)));
                        bitset_set(live, get_irn_idx(irn));
@@ -768,6 +817,25 @@ static void pressure(ir_node *block, void *env_ptr)
                DBG((dbg, LEVEL_1, "\tinsn: %+F, pressure: %d\n", irn, pressure));
                DBG((dbg, LEVEL_2, "\tlive: %B\n", live));
 
+               if (get_irn_mode(irn) == mode_T) {
+                       const ir_edge_t *edge;
+
+                       foreach_out_edge(irn, edge) {
+                               ir_node *proj = get_edge_src_irn(edge);
+
+                               /*
+                                * If the node defines some value, which can put into a
+                                * register of the current class, make a border for it.
+                                */
+                               if(has_reg_class(env, proj)) {
+                                       int nr = get_irn_idx(proj);
+
+                                       bitset_clear(live, nr);
+                                       border_def(proj, step, 1);
+                               }
+                       }
+               }
+
                /*
                 * If the node defines some value, which can put into a
                 * register of the current class, make a border for it.
@@ -803,22 +871,11 @@ static void pressure(ir_node *block, void *env_ptr)
                ++step;
        }
 
-       /*
-        * Add initial defs for all values live in.
-        */
-       foreach_pset(live_in, irn) {
-               if(has_reg_class(env, irn)) {
-
-                       /* Mark the value live in. */
-                       bitset_set(live, get_irn_idx(irn));
-
-                       /* Add the def */
+       bitset_foreach(live, elm) {
+               ir_node *irn = get_idx_irn(env->irg, elm);
+               if (be_is_live_in(lv, block, irn))
                        border_def(irn, step, 0);
-               }
        }
-
-       del_pset(live_in);
-       del_pset(live_end);
 }
 
 static void assign(ir_node *block, void *env_ptr)
@@ -831,10 +888,10 @@ static void assign(ir_node *block, void *env_ptr)
        const arch_env_t *arch_env  = env->birg->main_env->arch_env;
        struct list_head *head      = get_block_border_head(env, block);
        be_lv_t *lv                 = env->birg->lv;
-       pset *live_in               = be_lv_pset_put_in(lv, block, pset_new_ptr_default());
 
        const ir_node *irn;
        border_t *b;
+       int idx;
 
        bitset_clear_all(colors);
        bitset_clear_all(live);
@@ -852,7 +909,8 @@ static void assign(ir_node *block, void *env_ptr)
         * Since their colors have already been assigned (The dominators were
         * allocated before), we have to mark their colors as used also.
         */
-       foreach_pset(live_in, irn) {
+       be_lv_foreach(lv, block, be_lv_state_in, idx) {
+               irn = be_lv_get_irn(lv, block, idx);
                if(has_reg_class(env, irn)) {
                        const arch_register_t *reg = arch_get_irn_register(arch_env, irn);
                        int col;
@@ -927,27 +985,24 @@ static void assign(ir_node *block, void *env_ptr)
                        bitset_clear(live, nr);
                }
        }
-
-       del_pset(live_in);
 }
 
 void be_ra_chordal_color(be_chordal_env_t *chordal_env)
 {
        be_chordal_alloc_env_t env;
        char buf[256];
+       be_lv_t *lv;
        be_irg_t *birg = chordal_env->birg;
        const arch_register_class_t *cls = chordal_env->cls;
 
        int colors_n          = arch_register_class_n_regs(cls);
        ir_graph *irg         = chordal_env->irg;
-       int allocatable_regs  = colors_n - be_put_ignore_regs(birg, cls, NULL);
-
-       /* some special classes contain only ignore regs, no work to be done */
-       if(allocatable_regs == 0)
-               return;
 
        be_assure_dom_front(birg);
-       be_assure_liveness(birg);
+       lv = be_assure_liveness(birg);
+       be_liveness_assure_sets(lv);
+       be_liveness_assure_chk(lv);
+
        assure_doms(irg);
 
        env.chordal_env   = chordal_env;
@@ -987,7 +1042,7 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env)
 
 void be_init_chordal(void)
 {
-       FIRM_DBG_REGISTER(dbg, "firm.be.chordal.constr");
+       FIRM_DBG_REGISTER(dbg, "firm.be.chordal");
 }
 
 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_chordal);