1 /* Static Single Assignment conversion routines for the GNU compiler.
2 Copyright (C) 2000, 2001 Free Software Foundation, Inc.
4 This file is part of GNU CC.
6 GNU CC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 2, or (at your option) any
11 GNU CC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to the Free
18 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 Building an Optimizing Compiler
25 Butterworth-Heinemann, 1998
27 Static Single Assignment Construction
28 Preston Briggs, Tim Harvey, Taylor Simpson
29 Technical Report, Rice University, 1995
30 ftp://ftp.cs.rice.edu/public/preston/optimizer/SSA.ps.gz. */
38 #include "partition.h"
42 #include "hard-reg-set.h"
46 #include "insn-config.h"
48 #include "basic-block.h"
54 Handle subregs better, maybe. For now, if a reg that's set in a
55 subreg expression is duplicated going into SSA form, an extra copy
56 is inserted first that copies the entire reg into the duplicate, so
57 that the other bits are preserved. This isn't strictly SSA, since
58 at least part of the reg is assigned in more than one place (though
61 ??? What to do about strict_low_part. Probably I'll have to split
62 them out of their current instructions first thing.
64 Actually the best solution may be to have a kind of "mid-level rtl"
65 in which the RTL encodes exactly what we want, without exposing a
66 lot of niggling processor details. At some later point we lower
67 the representation, calling back into optabs to finish any necessary
70 /* All pseudo-registers and select hard registers are converted to SSA
71 form. When converting out of SSA, these select hard registers are
72 guaranteed to be mapped to their original register number. Each
73 machine's .h file should define CONVERT_HARD_REGISTER_TO_SSA_P
74 indicating which hard registers should be converted.
76 When converting out of SSA, temporaries for all registers are
77 partitioned. The partition is checked to ensure that all uses of
78 the same hard register in the same machine mode are in the same
81 /* If conservative_reg_partition is non-zero, use a conservative
82 register partitioning algorithm (which leaves more regs after
83 emerging from SSA) instead of the coalescing one. This is being
84 left in for a limited time only, as a debugging tool until the
85 coalescing algorithm is validated. */
87 static int conservative_reg_partition;
89 /* This flag is set when the CFG is in SSA form. */
92 /* Element I is the single instruction that sets register I. */
93 varray_type ssa_definition;
95 /* Element I-PSEUDO is the normal register that originated the ssa
96 register in question. */
97 varray_type ssa_rename_from;
99 /* Element I is the normal register that originated the ssa
100 register in question.
102 A hash table stores the (register, rtl) pairs. These are each
103 xmalloc'ed and deleted when the hash table is destroyed. */
104 htab_t ssa_rename_from_ht;
106 /* The running target ssa register for a given pseudo register.
107 (Pseudo registers appear in only one mode.) */
108 static rtx *ssa_rename_to_pseudo;
109 /* Similar, but for hard registers. A hard register can appear in
110 many modes, so we store an equivalent pseudo for each of the
112 static rtx ssa_rename_to_hard[FIRST_PSEUDO_REGISTER][NUM_MACHINE_MODES];
114 /* ssa_rename_from maps pseudo registers to the original corresponding
115 RTL. It is implemented as using a hash table. */
120 } ssa_rename_from_pair;
122 struct ssa_rename_from_hash_table_data {
123 sbitmap canonical_elements;
124 partition reg_partition;
127 static void ssa_rename_from_initialize
129 static rtx ssa_rename_from_lookup
131 static unsigned int original_register
132 PARAMS ((unsigned int regno));
133 static void ssa_rename_from_insert
134 PARAMS ((unsigned int reg, rtx r));
135 static void ssa_rename_from_free
137 typedef int (*srf_trav) PARAMS ((int regno, rtx r, sbitmap canonical_elements, partition reg_partition));
138 static void ssa_rename_from_traverse
139 PARAMS ((htab_trav callback_function, sbitmap canonical_elements, partition reg_partition));
140 /*static Avoid warnign message. */ void ssa_rename_from_print
142 static int ssa_rename_from_print_1
143 PARAMS ((void **slot, void *data));
144 static hashval_t ssa_rename_from_hash_function
145 PARAMS ((const void * srfp));
146 static int ssa_rename_from_equal
147 PARAMS ((const void *srfp1, const void *srfp2));
148 static void ssa_rename_from_delete
149 PARAMS ((void *srfp));
151 static rtx ssa_rename_to_lookup
153 static void ssa_rename_to_insert
154 PARAMS ((rtx reg, rtx r));
156 /* The number of registers that were live on entry to the SSA routines. */
157 static unsigned int ssa_max_reg_num;
159 /* Local function prototypes. */
161 struct rename_context;
163 static inline rtx * phi_alternative
165 static rtx first_insn_after_basic_block_note
166 PARAMS ((basic_block));
167 static void compute_dominance_frontiers_1
168 PARAMS ((sbitmap *frontiers, int *idom, int bb, sbitmap done));
169 static void find_evaluations_1
170 PARAMS ((rtx dest, rtx set, void *data));
171 static void find_evaluations
172 PARAMS ((sbitmap *evals, int nregs));
173 static void compute_iterated_dominance_frontiers
174 PARAMS ((sbitmap *idfs, sbitmap *frontiers, sbitmap *evals, int nregs));
175 static void insert_phi_node
176 PARAMS ((int regno, int b));
177 static void insert_phi_nodes
178 PARAMS ((sbitmap *idfs, sbitmap *evals, int nregs));
179 static void create_delayed_rename
180 PARAMS ((struct rename_context *, rtx *));
181 static void apply_delayed_renames
182 PARAMS ((struct rename_context *));
183 static int rename_insn_1
184 PARAMS ((rtx *ptr, void *data));
185 static void rename_block
186 PARAMS ((int b, int *idom));
187 static void rename_registers
188 PARAMS ((int nregs, int *idom));
190 static inline int ephi_add_node
191 PARAMS ((rtx reg, rtx *nodes, int *n_nodes));
192 static int * ephi_forward
193 PARAMS ((int t, sbitmap visited, sbitmap *succ, int *tstack));
194 static void ephi_backward
195 PARAMS ((int t, sbitmap visited, sbitmap *pred, rtx *nodes));
196 static void ephi_create
197 PARAMS ((int t, sbitmap visited, sbitmap *pred, sbitmap *succ, rtx *nodes));
198 static void eliminate_phi
199 PARAMS ((edge e, partition reg_partition));
200 static int make_regs_equivalent_over_bad_edges
201 PARAMS ((int bb, partition reg_partition));
203 /* These are used only in the conservative register partitioning
205 static int make_equivalent_phi_alternatives_equivalent
206 PARAMS ((int bb, partition reg_partition));
207 static partition compute_conservative_reg_partition
209 static int record_canonical_element_1
210 PARAMS ((void **srfp, void *data));
211 static int check_hard_regs_in_partition
212 PARAMS ((partition reg_partition));
213 static int rename_equivalent_regs_in_insn
214 PARAMS ((rtx *ptr, void *data));
216 /* These are used in the register coalescing algorithm. */
217 static int coalesce_if_unconflicting
218 PARAMS ((partition p, conflict_graph conflicts, int reg1, int reg2));
219 static int coalesce_regs_in_copies
220 PARAMS ((basic_block bb, partition p, conflict_graph conflicts));
221 static int coalesce_reg_in_phi
222 PARAMS ((rtx, int dest_regno, int src_regno, void *data));
223 static int coalesce_regs_in_successor_phi_nodes
224 PARAMS ((basic_block bb, partition p, conflict_graph conflicts));
225 static partition compute_coalesced_reg_partition
227 static int mark_reg_in_phi
228 PARAMS ((rtx *ptr, void *data));
229 static void mark_phi_and_copy_regs
230 PARAMS ((regset phi_set));
232 static int rename_equivalent_regs_in_insn
233 PARAMS ((rtx *ptr, void *data));
234 static void rename_equivalent_regs
235 PARAMS ((partition reg_partition));
237 /* Deal with hard registers. */
238 static int conflicting_hard_regs_p
239 PARAMS ((int reg1, int reg2));
241 /* ssa_rename_to maps registers and machine modes to SSA pseudo registers. */
243 /* Find the register associated with REG in the indicated mode. */
246 ssa_rename_to_lookup (reg)
249 if (!HARD_REGISTER_P (reg))
250 return ssa_rename_to_pseudo[REGNO (reg) - FIRST_PSEUDO_REGISTER];
252 return ssa_rename_to_hard[REGNO (reg)][GET_MODE (reg)];
255 /* Store a new value mapping REG to R in ssa_rename_to. */
258 ssa_rename_to_insert(reg, r)
262 if (!HARD_REGISTER_P (reg))
263 ssa_rename_to_pseudo[REGNO (reg) - FIRST_PSEUDO_REGISTER] = r;
265 ssa_rename_to_hard[REGNO (reg)][GET_MODE (reg)] = r;
268 /* Prepare ssa_rename_from for use. */
271 ssa_rename_from_initialize ()
273 /* We use an arbitrary initial hash table size of 64. */
274 ssa_rename_from_ht = htab_create (64,
275 &ssa_rename_from_hash_function,
276 &ssa_rename_from_equal,
277 &ssa_rename_from_delete);
280 /* Find the REG entry in ssa_rename_from. Return NULL_RTX if no entry is
284 ssa_rename_from_lookup (reg)
287 ssa_rename_from_pair srfp;
288 ssa_rename_from_pair *answer;
290 srfp.original = NULL_RTX;
291 answer = (ssa_rename_from_pair *)
292 htab_find_with_hash (ssa_rename_from_ht, (void *) &srfp, reg);
293 return (answer == 0 ? NULL_RTX : answer->original);
296 /* Find the number of the original register specified by REGNO. If
297 the register is a pseudo, return the original register's number.
298 Otherwise, return this register number REGNO. */
301 original_register (regno)
304 rtx original_rtx = ssa_rename_from_lookup (regno);
305 return original_rtx != NULL_RTX ? REGNO (original_rtx) : regno;
308 /* Add mapping from R to REG to ssa_rename_from even if already present. */
311 ssa_rename_from_insert (reg, r)
316 ssa_rename_from_pair *srfp = xmalloc (sizeof (ssa_rename_from_pair));
319 slot = htab_find_slot_with_hash (ssa_rename_from_ht, (const void *) srfp,
322 free ((void *) *slot);
326 /* Apply the CALLBACK_FUNCTION to each element in ssa_rename_from.
327 CANONICAL_ELEMENTS and REG_PARTITION pass data needed by the only
328 current use of this function. */
331 ssa_rename_from_traverse (callback_function,
332 canonical_elements, reg_partition)
333 htab_trav callback_function;
334 sbitmap canonical_elements;
335 partition reg_partition;
337 struct ssa_rename_from_hash_table_data srfhd;
338 srfhd.canonical_elements = canonical_elements;
339 srfhd.reg_partition = reg_partition;
340 htab_traverse (ssa_rename_from_ht, callback_function, (void *) &srfhd);
343 /* Destroy ssa_rename_from. */
346 ssa_rename_from_free ()
348 htab_delete (ssa_rename_from_ht);
351 /* Print the contents of ssa_rename_from. */
353 /* static Avoid erroneous error message. */
355 ssa_rename_from_print ()
357 printf ("ssa_rename_from's hash table contents:\n");
358 htab_traverse (ssa_rename_from_ht, &ssa_rename_from_print_1, NULL);
361 /* Print the contents of the hash table entry SLOT, passing the unused
362 sttribute DATA. Used as a callback function with htab_traverse (). */
365 ssa_rename_from_print_1 (slot, data)
367 void *data ATTRIBUTE_UNUSED;
369 ssa_rename_from_pair * p = *slot;
370 printf ("ssa_rename_from maps pseudo %i to original %i.\n",
371 p->reg, REGNO (p->original));
375 /* Given a hash entry SRFP, yield a hash value. */
378 ssa_rename_from_hash_function (srfp)
381 return ((const ssa_rename_from_pair *) srfp)->reg;
384 /* Test whether two hash table entries SRFP1 and SRFP2 are equal. */
387 ssa_rename_from_equal (srfp1, srfp2)
391 return ssa_rename_from_hash_function (srfp1) ==
392 ssa_rename_from_hash_function (srfp2);
395 /* Delete the hash table entry SRFP. */
398 ssa_rename_from_delete (srfp)
404 /* Given the SET of a PHI node, return the address of the alternative
405 for predecessor block C. */
408 phi_alternative (set, c)
412 rtvec phi_vec = XVEC (SET_SRC (set), 0);
415 for (v = GET_NUM_ELEM (phi_vec) - 2; v >= 0; v -= 2)
416 if (INTVAL (RTVEC_ELT (phi_vec, v + 1)) == c)
417 return &RTVEC_ELT (phi_vec, v);
422 /* Given the SET of a phi node, remove the alternative for predecessor
423 block C. Return non-zero on success, or zero if no alternative is
427 remove_phi_alternative (set, block)
431 rtvec phi_vec = XVEC (SET_SRC (set), 0);
432 int num_elem = GET_NUM_ELEM (phi_vec);
436 for (v = num_elem - 2; v >= 0; v -= 2)
437 if (INTVAL (RTVEC_ELT (phi_vec, v + 1)) == c)
439 if (v < num_elem - 2)
441 RTVEC_ELT (phi_vec, v) = RTVEC_ELT (phi_vec, num_elem - 2);
442 RTVEC_ELT (phi_vec, v + 1) = RTVEC_ELT (phi_vec, num_elem - 1);
444 PUT_NUM_ELEM (phi_vec, num_elem - 2);
451 /* For all registers, find all blocks in which they are set.
453 This is the transform of what would be local kill information that
454 we ought to be getting from flow. */
456 static sbitmap *fe_evals;
457 static int fe_current_bb;
460 find_evaluations_1 (dest, set, data)
462 rtx set ATTRIBUTE_UNUSED;
463 void *data ATTRIBUTE_UNUSED;
465 if (GET_CODE (dest) == REG
466 && CONVERT_REGISTER_TO_SSA_P (REGNO (dest)))
467 SET_BIT (fe_evals[REGNO (dest)], fe_current_bb);
471 find_evaluations (evals, nregs)
477 sbitmap_vector_zero (evals, nregs);
480 for (bb = n_basic_blocks; --bb >= 0; )
486 last = BLOCK_END (bb);
490 note_stores (PATTERN (p), find_evaluations_1, NULL);
499 /* Computing the Dominance Frontier:
501 As decribed in Morgan, section 3.5, this may be done simply by
502 walking the dominator tree bottom-up, computing the frontier for
503 the children before the parent. When considering a block B,
506 (1) A flow graph edge leaving B that does not lead to a child
507 of B in the dominator tree must be a block that is either equal
508 to B or not dominated by B. Such blocks belong in the frontier
511 (2) Consider a block X in the frontier of one of the children C
512 of B. If X is not equal to B and is not dominated by B, it
513 is in the frontier of B.
517 compute_dominance_frontiers_1 (frontiers, idom, bb, done)
523 basic_block b = BASIC_BLOCK (bb);
528 sbitmap_zero (frontiers[bb]);
530 /* Do the frontier of the children first. Not all children in the
531 dominator tree (blocks dominated by this one) are children in the
532 CFG, so check all blocks. */
533 for (c = 0; c < n_basic_blocks; ++c)
534 if (idom[c] == bb && ! TEST_BIT (done, c))
535 compute_dominance_frontiers_1 (frontiers, idom, c, done);
537 /* Find blocks conforming to rule (1) above. */
538 for (e = b->succ; e; e = e->succ_next)
540 if (e->dest == EXIT_BLOCK_PTR)
542 if (idom[e->dest->index] != bb)
543 SET_BIT (frontiers[bb], e->dest->index);
546 /* Find blocks conforming to rule (2). */
547 for (c = 0; c < n_basic_blocks; ++c)
551 EXECUTE_IF_SET_IN_SBITMAP (frontiers[c], 0, x,
554 SET_BIT (frontiers[bb], x);
560 compute_dominance_frontiers (frontiers, idom)
564 sbitmap done = sbitmap_alloc (n_basic_blocks);
567 compute_dominance_frontiers_1 (frontiers, idom, 0, done);
572 /* Computing the Iterated Dominance Frontier:
574 This is the set of merge points for a given register.
576 This is not particularly intuitive. See section 7.1 of Morgan, in
577 particular figures 7.3 and 7.4 and the immediately surrounding text.
581 compute_iterated_dominance_frontiers (idfs, frontiers, evals, nregs)
590 worklist = sbitmap_alloc (n_basic_blocks);
592 for (reg = 0; reg < nregs; ++reg)
594 sbitmap idf = idfs[reg];
597 /* Start the iterative process by considering those blocks that
598 evaluate REG. We'll add their dominance frontiers to the
599 IDF, and then consider the blocks we just added. */
600 sbitmap_copy (worklist, evals[reg]);
602 /* Morgan's algorithm is incorrect here. Blocks that evaluate
603 REG aren't necessarily in REG's IDF. Start with an empty IDF. */
606 /* Iterate until the worklist is empty. */
611 EXECUTE_IF_SET_IN_SBITMAP (worklist, 0, b,
613 RESET_BIT (worklist, b);
614 /* For each block on the worklist, add to the IDF all
615 blocks on its dominance frontier that aren't already
616 on the IDF. Every block that's added is also added
618 sbitmap_union_of_diff (worklist, worklist, frontiers[b], idf);
619 sbitmap_a_or_b (idf, idf, frontiers[b]);
626 sbitmap_free (worklist);
630 fprintf(rtl_dump_file,
631 "Iterated dominance frontier: %d passes on %d regs.\n",
636 /* Return the INSN immediately following the NOTE_INSN_BASIC_BLOCK
637 note associated with the BLOCK. */
640 first_insn_after_basic_block_note (block)
645 /* Get the first instruction in the block. */
648 if (insn == NULL_RTX)
650 if (GET_CODE (insn) == CODE_LABEL)
651 insn = NEXT_INSN (insn);
652 if (!NOTE_INSN_BASIC_BLOCK_P (insn))
655 return NEXT_INSN (insn);
658 /* Insert the phi nodes. */
661 insert_phi_node (regno, bb)
664 basic_block b = BASIC_BLOCK (bb);
672 /* Find out how many predecessors there are. */
673 for (e = b->pred, npred = 0; e; e = e->pred_next)
674 if (e->src != ENTRY_BLOCK_PTR)
677 /* If this block has no "interesting" preds, then there is nothing to
678 do. Consider a block that only has the entry block as a pred. */
682 /* This is the register to which the phi function will be assigned. */
683 reg = regno_reg_rtx[regno];
685 /* Construct the arguments to the PHI node. The use of pc_rtx is just
686 a placeholder; we'll insert the proper value in rename_registers. */
687 vec = rtvec_alloc (npred * 2);
688 for (e = b->pred, i = 0; e ; e = e->pred_next, i += 2)
689 if (e->src != ENTRY_BLOCK_PTR)
691 RTVEC_ELT (vec, i + 0) = pc_rtx;
692 RTVEC_ELT (vec, i + 1) = GEN_INT (e->src->index);
695 phi = gen_rtx_PHI (VOIDmode, vec);
696 phi = gen_rtx_SET (VOIDmode, reg, phi);
698 insn = first_insn_after_basic_block_note (b);
699 end_p = PREV_INSN (insn) == b->end;
700 emit_insn_before (phi, insn);
702 b->end = PREV_INSN (insn);
706 insert_phi_nodes (idfs, evals, nregs)
708 sbitmap *evals ATTRIBUTE_UNUSED;
713 for (reg = 0; reg < nregs; ++reg)
714 if (CONVERT_REGISTER_TO_SSA_P (reg))
717 EXECUTE_IF_SET_IN_SBITMAP (idfs[reg], 0, b,
719 if (REGNO_REG_SET_P (BASIC_BLOCK (b)->global_live_at_start, reg))
720 insert_phi_node (reg, b);
725 /* Rename the registers to conform to SSA.
727 This is essentially the algorithm presented in Figure 7.8 of Morgan,
728 with a few changes to reduce pattern search time in favour of a bit
729 more memory usage. */
731 /* One of these is created for each set. It will live in a list local
732 to its basic block for the duration of that block's processing. */
733 struct rename_set_data
735 struct rename_set_data *next;
736 /* This is the SET_DEST of the (first) SET that sets the REG. */
738 /* This is what used to be at *REG_LOC. */
740 /* This is the REG that will replace OLD_REG. It's set only
741 when the rename data is moved onto the DONE_RENAMES queue. */
743 /* This is what to restore ssa_rename_to_lookup (old_reg) to. It is
744 usually the previous contents of ssa_rename_to_lookup (old_reg). */
746 /* This is the insn that contains all the SETs of the REG. */
750 /* This struct is used to pass information to callback functions while
751 renaming registers. */
752 struct rename_context
754 struct rename_set_data *new_renames;
755 struct rename_set_data *done_renames;
759 /* Queue the rename of *REG_LOC. */
761 create_delayed_rename (c, reg_loc)
762 struct rename_context *c;
765 struct rename_set_data *r;
766 r = (struct rename_set_data *) xmalloc (sizeof(*r));
768 if (GET_CODE (*reg_loc) != REG
769 || !CONVERT_REGISTER_TO_SSA_P (REGNO (*reg_loc)))
772 r->reg_loc = reg_loc;
773 r->old_reg = *reg_loc;
774 r->prev_reg = ssa_rename_to_lookup(r->old_reg);
775 r->set_insn = c->current_insn;
776 r->next = c->new_renames;
780 /* This is part of a rather ugly hack to allow the pre-ssa regno to be
781 reused. If, during processing, a register has not yet been touched,
782 ssa_rename_to[regno][machno] will be NULL. Now, in the course of pushing
783 and popping values from ssa_rename_to, when we would ordinarily
784 pop NULL back in, we pop RENAME_NO_RTX. We treat this exactly the
785 same as NULL, except that it signals that the original regno has
786 already been reused. */
787 #define RENAME_NO_RTX pc_rtx
789 /* Move all the entries from NEW_RENAMES onto DONE_RENAMES by
790 applying all the renames on NEW_RENAMES. */
793 apply_delayed_renames (c)
794 struct rename_context *c;
796 struct rename_set_data *r;
797 struct rename_set_data *last_r = NULL;
799 for (r = c->new_renames; r != NULL; r = r->next)
803 /* Failure here means that someone has a PARALLEL that sets
804 a register twice (bad!). */
805 if (ssa_rename_to_lookup (r->old_reg) != r->prev_reg)
807 /* Failure here means we have changed REG_LOC before applying
809 /* For the first set we come across, reuse the original regno. */
810 if (r->prev_reg == NULL_RTX && !HARD_REGISTER_P (r->old_reg))
812 r->new_reg = r->old_reg;
813 /* We want to restore RENAME_NO_RTX rather than NULL_RTX. */
814 r->prev_reg = RENAME_NO_RTX;
817 r->new_reg = gen_reg_rtx (GET_MODE (r->old_reg));
818 new_regno = REGNO (r->new_reg);
819 ssa_rename_to_insert (r->old_reg, r->new_reg);
821 if (new_regno >= (int) ssa_definition->num_elements)
823 int new_limit = new_regno * 5 / 4;
824 VARRAY_GROW (ssa_definition, new_limit);
827 VARRAY_RTX (ssa_definition, new_regno) = r->set_insn;
828 ssa_rename_from_insert (new_regno, r->old_reg);
833 last_r->next = c->done_renames;
834 c->done_renames = c->new_renames;
835 c->new_renames = NULL;
839 /* Part one of the first step of rename_block, called through for_each_rtx.
840 Mark pseudos that are set for later update. Transform uses of pseudos. */
843 rename_insn_1 (ptr, data)
848 struct rename_context *context = data;
853 switch (GET_CODE (x))
857 rtx *destp = &SET_DEST (x);
858 rtx dest = SET_DEST (x);
860 /* Some SETs also use the REG specified in their LHS.
861 These can be detected by the presence of
862 STRICT_LOW_PART, SUBREG, SIGN_EXTRACT, and ZERO_EXTRACT
863 in the LHS. Handle these by changing
864 (set (subreg (reg foo)) ...)
866 (sequence [(set (reg foo_1) (reg foo))
867 (set (subreg (reg foo_1)) ...)])
869 FIXME: Much of the time this is too much. For many libcalls,
870 paradoxical SUBREGs, etc., the input register is dead. We should
871 recognise this in rename_block or here and not make a false
874 if (GET_CODE (dest) == STRICT_LOW_PART
875 || GET_CODE (dest) == SUBREG
876 || GET_CODE (dest) == SIGN_EXTRACT
877 || GET_CODE (dest) == ZERO_EXTRACT)
882 while (GET_CODE (reg) == STRICT_LOW_PART
883 || GET_CODE (reg) == SUBREG
884 || GET_CODE (reg) == SIGN_EXTRACT
885 || GET_CODE (reg) == ZERO_EXTRACT)
888 if (GET_CODE (reg) == REG
889 && CONVERT_REGISTER_TO_SSA_P (REGNO (reg)))
891 /* Generate (set reg reg), and do renaming on it so
892 that it becomes (set reg_1 reg_0), and we will
893 replace reg with reg_1 in the SUBREG. */
895 struct rename_set_data *saved_new_renames;
896 saved_new_renames = context->new_renames;
897 context->new_renames = NULL;
898 i = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
899 for_each_rtx (&i, rename_insn_1, data);
900 apply_delayed_renames (context);
901 context->new_renames = saved_new_renames;
904 else if (GET_CODE (dest) == REG &&
905 CONVERT_REGISTER_TO_SSA_P (REGNO (dest)))
907 /* We found a genuine set of an interesting register. Tag
908 it so that we can create a new name for it after we finish
909 processing this insn. */
911 create_delayed_rename (context, destp);
913 /* Since we do not wish to (directly) traverse the
914 SET_DEST, recurse through for_each_rtx for the SET_SRC
916 if (GET_CODE (x) == SET)
917 for_each_rtx (&SET_SRC (x), rename_insn_1, data);
921 /* Otherwise, this was not an interesting destination. Continue
922 on, marking uses as normal. */
927 if (CONVERT_REGISTER_TO_SSA_P (REGNO (x)) &&
928 REGNO (x) < ssa_max_reg_num)
930 rtx new_reg = ssa_rename_to_lookup (x);
932 if (new_reg != NULL_RTX && new_reg != RENAME_NO_RTX)
934 if (GET_MODE (x) != GET_MODE (new_reg))
938 /* Else this is a use before a set. Warn? */
943 /* There is considerable debate on how CLOBBERs ought to be
944 handled in SSA. For now, we're keeping the CLOBBERs, which
945 means that we don't really have SSA form. There are a couple
946 of proposals for how to fix this problem, but neither is
949 rtx dest = XCEXP (x, 0, CLOBBER);
952 if (CONVERT_REGISTER_TO_SSA_P (REGNO (dest))
953 && REGNO (dest) < ssa_max_reg_num)
955 rtx new_reg = ssa_rename_to_lookup (dest);
956 if (new_reg != NULL_RTX && new_reg != RENAME_NO_RTX)
957 XCEXP (x, 0, CLOBBER) = new_reg;
959 /* Stop traversing. */
963 /* Continue traversing. */
968 /* Never muck with the phi. We do that elsewhere, special-like. */
972 /* Anything else, continue traversing. */
978 rename_block (bb, idom)
982 basic_block b = BASIC_BLOCK (bb);
984 rtx insn, next, last;
985 struct rename_set_data *set_data = NULL;
988 /* Step One: Walk the basic block, adding new names for sets and
998 struct rename_context context;
999 context.done_renames = set_data;
1000 context.new_renames = NULL;
1001 context.current_insn = insn;
1004 for_each_rtx (&PATTERN (insn), rename_insn_1, &context);
1005 for_each_rtx (®_NOTES (insn), rename_insn_1, &context);
1007 /* Sometimes, we end up with a sequence of insns that
1008 SSA needs to treat as a single insn. Wrap these in a
1009 SEQUENCE. (Any notes now get attached to the SEQUENCE,
1010 not to the old version inner insn.) */
1011 if (get_insns () != NULL_RTX)
1016 emit (PATTERN (insn));
1017 seq = gen_sequence ();
1018 /* We really want a SEQUENCE of SETs, not a SEQUENCE
1020 for (i = 0; i < XVECLEN (seq, 0); i++)
1021 XVECEXP (seq, 0, i) = PATTERN (XVECEXP (seq, 0, i));
1022 PATTERN (insn) = seq;
1026 apply_delayed_renames (&context);
1027 set_data = context.done_renames;
1030 next = NEXT_INSN (insn);
1032 while (insn != last);
1034 /* Step Two: Update the phi nodes of this block's successors. */
1036 for (e = b->succ; e; e = e->succ_next)
1038 if (e->dest == EXIT_BLOCK_PTR)
1041 insn = first_insn_after_basic_block_note (e->dest);
1043 while (PHI_NODE_P (insn))
1045 rtx phi = PATTERN (insn);
1048 /* Find out which of our outgoing registers this node is
1049 intended to replace. Note that if this is not the first PHI
1050 node to have been created for this register, we have to
1051 jump through rename links to figure out which register
1052 we're talking about. This can easily be recognized by
1053 noting that the regno is new to this pass. */
1054 reg = SET_DEST (phi);
1055 if (REGNO (reg) >= ssa_max_reg_num)
1056 reg = ssa_rename_from_lookup (REGNO (reg));
1057 if (reg == NULL_RTX)
1059 reg = ssa_rename_to_lookup (reg);
1061 /* It is possible for the variable to be uninitialized on
1062 edges in. Reduce the arity of the PHI so that we don't
1063 consider those edges. */
1064 if (reg == NULL || reg == RENAME_NO_RTX)
1066 if (! remove_phi_alternative (phi, b))
1071 /* When we created the PHI nodes, we did not know what mode
1072 the register should be. Now that we've found an original,
1073 we can fill that in. */
1074 if (GET_MODE (SET_DEST (phi)) == VOIDmode)
1075 PUT_MODE (SET_DEST (phi), GET_MODE (reg));
1076 else if (GET_MODE (SET_DEST (phi)) != GET_MODE (reg))
1079 *phi_alternative (phi, bb) = reg;
1082 insn = NEXT_INSN (insn);
1086 /* Step Three: Do the same to the children of this block in
1089 for (c = 0; c < n_basic_blocks; ++c)
1091 rename_block (c, idom);
1093 /* Step Four: Update the sets to refer to their new register,
1094 and restore ssa_rename_to to its previous state. */
1098 struct rename_set_data *next;
1099 rtx old_reg = *set_data->reg_loc;
1101 if (*set_data->reg_loc != set_data->old_reg)
1103 *set_data->reg_loc = set_data->new_reg;
1105 ssa_rename_to_insert (old_reg, set_data->prev_reg);
1107 next = set_data->next;
1114 rename_registers (nregs, idom)
1118 VARRAY_RTX_INIT (ssa_definition, nregs * 3, "ssa_definition");
1119 ssa_rename_from_initialize ();
1121 ssa_rename_to_pseudo = (rtx *) alloca (nregs * sizeof(rtx));
1122 memset ((char *) ssa_rename_to_pseudo, 0, nregs * sizeof(rtx));
1123 memset ((char *) ssa_rename_to_hard, 0,
1124 FIRST_PSEUDO_REGISTER * NUM_MACHINE_MODES * sizeof (rtx));
1126 rename_block (0, idom);
1128 /* ??? Update basic_block_live_at_start, and other flow info
1131 ssa_rename_to_pseudo = NULL;
1134 /* The main entry point for moving to SSA. */
1139 /* Element I is the set of blocks that set register I. */
1142 /* Dominator bitmaps. */
1146 /* Element I is the immediate dominator of block I. */
1151 /* Don't do it twice. */
1155 /* Need global_live_at_{start,end} up to date. Do not remove any
1156 dead code. We'll let the SSA optimizers do that. */
1157 life_analysis (get_insns (), NULL, 0);
1159 idom = (int *) alloca (n_basic_blocks * sizeof (int));
1160 memset ((void *)idom, -1, (size_t)n_basic_blocks * sizeof (int));
1161 calculate_dominance_info (idom, NULL, CDI_DOMINATORS);
1166 fputs (";; Immediate Dominators:\n", rtl_dump_file);
1167 for (i = 0; i < n_basic_blocks; ++i)
1168 fprintf (rtl_dump_file, ";\t%3d = %3d\n", i, idom[i]);
1169 fflush (rtl_dump_file);
1172 /* Compute dominance frontiers. */
1174 dfs = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
1175 compute_dominance_frontiers (dfs, idom);
1179 dump_sbitmap_vector (rtl_dump_file, ";; Dominance Frontiers:",
1180 "; Basic Block", dfs, n_basic_blocks);
1181 fflush (rtl_dump_file);
1184 /* Compute register evaluations. */
1186 ssa_max_reg_num = max_reg_num();
1187 nregs = ssa_max_reg_num;
1188 evals = sbitmap_vector_alloc (nregs, n_basic_blocks);
1189 find_evaluations (evals, nregs);
1191 /* Compute the iterated dominance frontier for each register. */
1193 idfs = sbitmap_vector_alloc (nregs, n_basic_blocks);
1194 compute_iterated_dominance_frontiers (idfs, dfs, evals, nregs);
1198 dump_sbitmap_vector (rtl_dump_file, ";; Iterated Dominance Frontiers:",
1199 "; Register", idfs, nregs);
1200 fflush (rtl_dump_file);
1203 /* Insert the phi nodes. */
1205 insert_phi_nodes (idfs, evals, nregs);
1207 /* Rename the registers to satisfy SSA. */
1209 rename_registers (nregs, idom);
1211 /* All done! Clean up and go home. */
1213 sbitmap_vector_free (dfs);
1214 sbitmap_vector_free (evals);
1215 sbitmap_vector_free (idfs);
1218 reg_scan (get_insns (), max_reg_num (), 1);
1221 /* REG is the representative temporary of its partition. Add it to the
1222 set of nodes to be processed, if it hasn't been already. Return the
1223 index of this register in the node set. */
1226 ephi_add_node (reg, nodes, n_nodes)
1231 for (i = *n_nodes - 1; i >= 0; --i)
1232 if (REGNO (reg) == REGNO (nodes[i]))
1235 nodes[i = (*n_nodes)++] = reg;
1239 /* Part one of the topological sort. This is a forward (downward) search
1240 through the graph collecting a stack of nodes to process. Assuming no
1241 cycles, the nodes at top of the stack when we are finished will have
1242 no other dependancies. */
1245 ephi_forward (t, visited, succ, tstack)
1253 SET_BIT (visited, t);
1255 EXECUTE_IF_SET_IN_SBITMAP (succ[t], 0, s,
1257 if (! TEST_BIT (visited, s))
1258 tstack = ephi_forward (s, visited, succ, tstack);
1265 /* Part two of the topological sort. The is a backward search through
1266 a cycle in the graph, copying the data forward as we go. */
1269 ephi_backward (t, visited, pred, nodes)
1271 sbitmap visited, *pred;
1276 SET_BIT (visited, t);
1278 EXECUTE_IF_SET_IN_SBITMAP (pred[t], 0, p,
1280 if (! TEST_BIT (visited, p))
1282 ephi_backward (p, visited, pred, nodes);
1283 emit_move_insn (nodes[p], nodes[t]);
1288 /* Part two of the topological sort. Create the copy for a register
1289 and any cycle of which it is a member. */
1292 ephi_create (t, visited, pred, succ, nodes)
1294 sbitmap visited, *pred, *succ;
1297 rtx reg_u = NULL_RTX;
1298 int unvisited_predecessors = 0;
1301 /* Iterate through the predecessor list looking for unvisited nodes.
1302 If there are any, we have a cycle, and must deal with that. At
1303 the same time, look for a visited predecessor. If there is one,
1304 we won't need to create a temporary. */
1306 EXECUTE_IF_SET_IN_SBITMAP (pred[t], 0, p,
1308 if (! TEST_BIT (visited, p))
1309 unvisited_predecessors = 1;
1314 if (unvisited_predecessors)
1316 /* We found a cycle. Copy out one element of the ring (if necessary),
1317 then traverse the ring copying as we go. */
1321 reg_u = gen_reg_rtx (GET_MODE (nodes[t]));
1322 emit_move_insn (reg_u, nodes[t]);
1325 EXECUTE_IF_SET_IN_SBITMAP (pred[t], 0, p,
1327 if (! TEST_BIT (visited, p))
1329 ephi_backward (p, visited, pred, nodes);
1330 emit_move_insn (nodes[p], reg_u);
1336 /* No cycle. Just copy the value from a successor. */
1339 EXECUTE_IF_SET_IN_SBITMAP (succ[t], 0, s,
1341 SET_BIT (visited, t);
1342 emit_move_insn (nodes[t], nodes[s]);
1348 /* Convert the edge to normal form. */
1351 eliminate_phi (e, reg_partition)
1353 partition reg_partition;
1356 sbitmap *pred, *succ;
1359 int *stack, *tstack;
1363 /* Collect an upper bound on the number of registers needing processing. */
1365 insn = first_insn_after_basic_block_note (e->dest);
1368 while (PHI_NODE_P (insn))
1370 insn = next_nonnote_insn (insn);
1377 /* Build the auxilliary graph R(B).
1379 The nodes of the graph are the members of the register partition
1380 present in Phi(B). There is an edge from FIND(T0)->FIND(T1) for
1381 each T0 = PHI(...,T1,...), where T1 is for the edge from block C. */
1383 nodes = (rtx *) alloca (n_nodes * sizeof(rtx));
1384 pred = sbitmap_vector_alloc (n_nodes, n_nodes);
1385 succ = sbitmap_vector_alloc (n_nodes, n_nodes);
1386 sbitmap_vector_zero (pred, n_nodes);
1387 sbitmap_vector_zero (succ, n_nodes);
1389 insn = first_insn_after_basic_block_note (e->dest);
1392 for (; PHI_NODE_P (insn); insn = next_nonnote_insn (insn))
1394 rtx* preg = phi_alternative (PATTERN (insn), e->src->index);
1395 rtx tgt = SET_DEST (PATTERN (insn));
1398 /* There may be no phi alternative corresponding to this edge.
1399 This indicates that the phi variable is undefined along this
1405 if (GET_CODE (reg) != REG || GET_CODE (tgt) != REG)
1408 reg = regno_reg_rtx[partition_find (reg_partition, REGNO (reg))];
1409 tgt = regno_reg_rtx[partition_find (reg_partition, REGNO (tgt))];
1410 /* If the two registers are already in the same partition,
1411 nothing will need to be done. */
1416 ireg = ephi_add_node (reg, nodes, &n_nodes);
1417 itgt = ephi_add_node (tgt, nodes, &n_nodes);
1419 SET_BIT (pred[ireg], itgt);
1420 SET_BIT (succ[itgt], ireg);
1427 /* Begin a topological sort of the graph. */
1429 visited = sbitmap_alloc (n_nodes);
1430 sbitmap_zero (visited);
1432 tstack = stack = (int *) alloca (n_nodes * sizeof (int));
1434 for (i = 0; i < n_nodes; ++i)
1435 if (! TEST_BIT (visited, i))
1436 tstack = ephi_forward (i, visited, succ, tstack);
1438 sbitmap_zero (visited);
1440 /* As we find a solution to the tsort, collect the implementation
1441 insns in a sequence. */
1444 while (tstack != stack)
1447 if (! TEST_BIT (visited, i))
1448 ephi_create (i, visited, pred, succ, nodes);
1451 insn = gen_sequence ();
1453 insert_insn_on_edge (insn, e);
1455 fprintf (rtl_dump_file, "Emitting copy on edge (%d,%d)\n",
1456 e->src->index, e->dest->index);
1458 sbitmap_free (visited);
1460 sbitmap_vector_free (pred);
1461 sbitmap_vector_free (succ);
1464 /* For basic block B, consider all phi insns which provide an
1465 alternative corresponding to an incoming abnormal critical edge.
1466 Place the phi alternative corresponding to that abnormal critical
1467 edge in the same register class as the destination of the set.
1469 From Morgan, p. 178:
1471 For each abnormal critical edge (C, B),
1472 if T0 = phi (T1, ..., Ti, ..., Tm) is a phi node in B,
1473 and C is the ith predecessor of B,
1474 then T0 and Ti must be equivalent.
1476 Return non-zero iff any such cases were found for which the two
1477 regs were not already in the same class. */
1480 make_regs_equivalent_over_bad_edges (bb, reg_partition)
1482 partition reg_partition;
1485 basic_block b = BASIC_BLOCK (bb);
1488 /* Advance to the first phi node. */
1489 phi = first_insn_after_basic_block_note (b);
1491 /* Scan all the phi nodes. */
1494 phi = next_nonnote_insn (phi))
1498 rtx set = PATTERN (phi);
1499 rtx tgt = SET_DEST (set);
1501 /* The set target is expected to be an SSA register. */
1502 if (GET_CODE (tgt) != REG
1503 || !CONVERT_REGISTER_TO_SSA_P (REGNO (tgt)))
1505 tgt_regno = REGNO (tgt);
1507 /* Scan incoming abnormal critical edges. */
1508 for (e = b->pred; e; e = e->pred_next)
1509 if ((e->flags & (EDGE_ABNORMAL | EDGE_CRITICAL))
1510 == (EDGE_ABNORMAL | EDGE_CRITICAL))
1512 rtx *alt = phi_alternative (set, e->src->index);
1515 /* If there is no alternative corresponding to this edge,
1516 the value is undefined along the edge, so just go on. */
1520 /* The phi alternative is expected to be an SSA register. */
1521 if (GET_CODE (*alt) != REG
1522 || !CONVERT_REGISTER_TO_SSA_P (REGNO (*alt)))
1524 alt_regno = REGNO (*alt);
1526 /* If the set destination and the phi alternative aren't
1527 already in the same class... */
1528 if (partition_find (reg_partition, tgt_regno)
1529 != partition_find (reg_partition, alt_regno))
1531 /* ... make them such. */
1532 if (conflicting_hard_regs_p (tgt_regno, alt_regno))
1533 /* It is illegal to unify a hard register with a
1534 different register. */
1537 partition_union (reg_partition,
1538 tgt_regno, alt_regno);
1547 /* Consider phi insns in basic block BB pairwise. If the set target
1548 of both isns are equivalent pseudos, make the corresponding phi
1549 alternatives in each phi corresponding equivalent.
1551 Return nonzero if any new register classes were unioned. */
1554 make_equivalent_phi_alternatives_equivalent (bb, reg_partition)
1556 partition reg_partition;
1559 basic_block b = BASIC_BLOCK (bb);
1562 /* Advance to the first phi node. */
1563 phi = first_insn_after_basic_block_note (b);
1565 /* Scan all the phi nodes. */
1568 phi = next_nonnote_insn (phi))
1570 rtx set = PATTERN (phi);
1571 /* The regno of the destination of the set. */
1572 int tgt_regno = REGNO (SET_DEST (PATTERN (phi)));
1574 rtx phi2 = next_nonnote_insn (phi);
1576 /* Scan all phi nodes following this one. */
1579 phi2 = next_nonnote_insn (phi2))
1581 rtx set2 = PATTERN (phi2);
1582 /* The regno of the destination of the set. */
1583 int tgt2_regno = REGNO (SET_DEST (set2));
1585 /* Are the set destinations equivalent regs? */
1586 if (partition_find (reg_partition, tgt_regno) ==
1587 partition_find (reg_partition, tgt2_regno))
1590 /* Scan over edges. */
1591 for (e = b->pred; e; e = e->pred_next)
1593 int pred_block = e->src->index;
1594 /* Identify the phi alternatives from both phi
1595 nodes corresponding to this edge. */
1596 rtx *alt = phi_alternative (set, pred_block);
1597 rtx *alt2 = phi_alternative (set2, pred_block);
1599 /* If one of the phi nodes doesn't have a
1600 corresponding alternative, just skip it. */
1601 if (alt == 0 || alt2 == 0)
1604 /* Both alternatives should be SSA registers. */
1605 if (GET_CODE (*alt) != REG
1606 || !CONVERT_REGISTER_TO_SSA_P (REGNO (*alt)))
1608 if (GET_CODE (*alt2) != REG
1609 || !CONVERT_REGISTER_TO_SSA_P (REGNO (*alt2)))
1612 /* If the alternatives aren't already in the same
1614 if (partition_find (reg_partition, REGNO (*alt))
1615 != partition_find (reg_partition, REGNO (*alt2)))
1617 /* ... make them so. */
1618 if (conflicting_hard_regs_p (REGNO (*alt), REGNO (*alt2)))
1619 /* It is illegal to unify a hard register with
1620 a different register. */
1623 partition_union (reg_partition,
1624 REGNO (*alt), REGNO (*alt2));
1635 /* Compute a conservative partition of outstanding pseudo registers.
1636 See Morgan 7.3.1. */
1639 compute_conservative_reg_partition ()
1644 /* We don't actually work with hard registers, but it's easier to
1645 carry them around anyway rather than constantly doing register
1646 number arithmetic. */
1648 partition_new (ssa_definition->num_elements);
1650 /* The first priority is to make sure registers that might have to
1651 be copied on abnormal critical edges are placed in the same
1652 partition. This saves us from having to split abnormal critical
1654 for (bb = n_basic_blocks; --bb >= 0; )
1655 changed += make_regs_equivalent_over_bad_edges (bb, p);
1657 /* Now we have to insure that corresponding arguments of phi nodes
1658 assigning to corresponding regs are equivalent. Iterate until
1663 for (bb = n_basic_blocks; --bb >= 0; )
1664 changed += make_equivalent_phi_alternatives_equivalent (bb, p);
1670 /* The following functions compute a register partition that attempts
1671 to eliminate as many reg copies and phi node copies as possible by
1672 coalescing registers. This is the strategy:
1674 1. As in the conservative case, the top priority is to coalesce
1675 registers that otherwise would cause copies to be placed on
1676 abnormal critical edges (which isn't possible).
1678 2. Figure out which regs are involved (in the LHS or RHS) of
1679 copies and phi nodes. Compute conflicts among these regs.
1681 3. Walk around the instruction stream, placing two regs in the
1682 same class of the partition if one appears on the LHS and the
1683 other on the RHS of a copy or phi node and the two regs don't
1684 conflict. The conflict information of course needs to be
1687 4. If anything has changed, there may be new opportunities to
1688 coalesce regs, so go back to 2.
1691 /* If REG1 and REG2 don't conflict in CONFLICTS, place them in the
1692 same class of partition P, if they aren't already. Update
1693 CONFLICTS appropriately.
1695 Returns one if REG1 and REG2 were placed in the same class but were
1696 not previously; zero otherwise.
1698 See Morgan figure 11.15. */
1701 coalesce_if_unconflicting (p, conflicts, reg1, reg2)
1703 conflict_graph conflicts;
1709 /* Work only on SSA registers. */
1710 if (!CONVERT_REGISTER_TO_SSA_P (reg1) || !CONVERT_REGISTER_TO_SSA_P (reg2))
1713 /* Find the canonical regs for the classes containing REG1 and
1715 reg1 = partition_find (p, reg1);
1716 reg2 = partition_find (p, reg2);
1718 /* If they're already in the same class, there's nothing to do. */
1722 /* If the regs conflict, our hands are tied. */
1723 if (conflicting_hard_regs_p (reg1, reg2) ||
1724 conflict_graph_conflict_p (conflicts, reg1, reg2))
1727 /* We're good to go. Put the regs in the same partition. */
1728 partition_union (p, reg1, reg2);
1730 /* Find the new canonical reg for the merged class. */
1731 reg = partition_find (p, reg1);
1733 /* Merge conflicts from the two previous classes. */
1734 conflict_graph_merge_regs (conflicts, reg, reg1);
1735 conflict_graph_merge_regs (conflicts, reg, reg2);
1740 /* For each register copy insn in basic block BB, place the LHS and
1741 RHS regs in the same class in partition P if they do not conflict
1742 according to CONFLICTS.
1744 Returns the number of changes that were made to P.
1746 See Morgan figure 11.14. */
1749 coalesce_regs_in_copies (bb, p, conflicts)
1752 conflict_graph conflicts;
1758 /* Scan the instruction stream of the block. */
1759 for (insn = bb->head; insn != end; insn = NEXT_INSN (insn))
1765 /* If this isn't a set insn, go to the next insn. */
1766 if (GET_CODE (insn) != INSN)
1768 pattern = PATTERN (insn);
1769 if (GET_CODE (pattern) != SET)
1772 src = SET_SRC (pattern);
1773 dest = SET_DEST (pattern);
1775 /* We're only looking for copies. */
1776 if (GET_CODE (src) != REG || GET_CODE (dest) != REG)
1779 /* Coalesce only if the reg modes are the same. As long as
1780 each reg's rtx is unique, it can have only one mode, so two
1781 pseudos of different modes can't be coalesced into one.
1783 FIXME: We can probably get around this by inserting SUBREGs
1784 where appropriate, but for now we don't bother. */
1785 if (GET_MODE (src) != GET_MODE (dest))
1788 /* Found a copy; see if we can use the same reg for both the
1789 source and destination (and thus eliminate the copy,
1791 changed += coalesce_if_unconflicting (p, conflicts,
1792 REGNO (src), REGNO (dest));
1798 struct phi_coalesce_context
1801 conflict_graph conflicts;
1805 /* Callback function for for_each_successor_phi. If the set
1806 destination and the phi alternative regs do not conflict, place
1807 them in the same paritition class. DATA is a pointer to a
1808 phi_coalesce_context struct. */
1811 coalesce_reg_in_phi (insn, dest_regno, src_regno, data)
1812 rtx insn ATTRIBUTE_UNUSED;
1817 struct phi_coalesce_context *context =
1818 (struct phi_coalesce_context *) data;
1820 /* Attempt to use the same reg, if they don't conflict. */
1822 += coalesce_if_unconflicting (context->p, context->conflicts,
1823 dest_regno, src_regno);
1827 /* For each alternative in a phi function corresponding to basic block
1828 BB (in phi nodes in successor block to BB), place the reg in the
1829 phi alternative and the reg to which the phi value is set into the
1830 same class in partition P, if allowed by CONFLICTS.
1832 Return the number of changes that were made to P.
1834 See Morgan figure 11.14. */
1837 coalesce_regs_in_successor_phi_nodes (bb, p, conflicts)
1840 conflict_graph conflicts;
1842 struct phi_coalesce_context context;
1844 context.conflicts = conflicts;
1845 context.changed = 0;
1847 for_each_successor_phi (bb, &coalesce_reg_in_phi, &context);
1849 return context.changed;
1852 /* Compute and return a partition of pseudos. Where possible,
1853 non-conflicting pseudos are placed in the same class.
1855 The caller is responsible for deallocating the returned partition. */
1858 compute_coalesced_reg_partition ()
1864 partition_new (ssa_definition->num_elements);
1866 /* The first priority is to make sure registers that might have to
1867 be copied on abnormal critical edges are placed in the same
1868 partition. This saves us from having to split abnormal critical
1869 edges (which can't be done). */
1870 for (bb = n_basic_blocks; --bb >= 0; )
1871 make_regs_equivalent_over_bad_edges (bb, p);
1875 regset_head phi_set;
1876 conflict_graph conflicts;
1880 /* Build the set of registers involved in phi nodes, either as
1881 arguments to the phi function or as the target of a set. */
1882 INITIALIZE_REG_SET (phi_set);
1883 mark_phi_and_copy_regs (&phi_set);
1885 /* Compute conflicts. */
1886 conflicts = conflict_graph_compute (&phi_set, p);
1888 /* FIXME: Better would be to process most frequently executed
1889 blocks first, so that most frequently executed copies would
1890 be more likely to be removed by register coalescing. But any
1891 order will generate correct, if non-optimal, results. */
1892 for (bb = n_basic_blocks; --bb >= 0; )
1894 basic_block block = BASIC_BLOCK (bb);
1895 changed += coalesce_regs_in_copies (block, p, conflicts);
1897 coalesce_regs_in_successor_phi_nodes (block, p, conflicts);
1900 conflict_graph_delete (conflicts);
1902 while (changed > 0);
1907 /* Mark the regs in a phi node. PTR is a phi expression or one of its
1908 components (a REG or a CONST_INT). DATA is a reg set in which to
1909 set all regs. Called from for_each_rtx. */
1912 mark_reg_in_phi (ptr, data)
1917 regset set = (regset) data;
1919 switch (GET_CODE (expr))
1922 SET_REGNO_REG_SET (set, REGNO (expr));
1932 /* Mark in PHI_SET all pseudos that are used in a phi node -- either
1933 set from a phi expression, or used as an argument in one. Also
1934 mark regs that are the source or target of a reg copy. Uses
1938 mark_phi_and_copy_regs (phi_set)
1943 /* Scan the definitions of all regs. */
1944 for (reg = 0; reg < VARRAY_SIZE (ssa_definition); ++reg)
1945 if (CONVERT_REGISTER_TO_SSA_P (reg))
1947 rtx insn = VARRAY_RTX (ssa_definition, reg);
1953 pattern = PATTERN (insn);
1954 /* Sometimes we get PARALLEL insns. These aren't phi nodes or
1956 if (GET_CODE (pattern) != SET)
1958 src = SET_SRC (pattern);
1960 if (GET_CODE (src) == REG)
1962 /* It's a reg copy. */
1963 SET_REGNO_REG_SET (phi_set, reg);
1964 SET_REGNO_REG_SET (phi_set, REGNO (src));
1966 else if (GET_CODE (src) == PHI)
1968 /* It's a phi node. Mark the reg being set. */
1969 SET_REGNO_REG_SET (phi_set, reg);
1970 /* Mark the regs used in the phi function. */
1971 for_each_rtx (&src, mark_reg_in_phi, phi_set);
1973 /* ... else nothing to do. */
1977 /* Rename regs in insn PTR that are equivalent. DATA is the register
1978 partition which specifies equivalences. */
1981 rename_equivalent_regs_in_insn (ptr, data)
1986 partition reg_partition = (partition) data;
1991 switch (GET_CODE (x))
1994 if (CONVERT_REGISTER_TO_SSA_P (REGNO (x)))
1996 unsigned int regno = REGNO (x);
1997 unsigned int new_regno = partition_find (reg_partition, regno);
1998 rtx canonical_element_rtx = ssa_rename_from_lookup (new_regno);
2000 if (canonical_element_rtx != NULL_RTX &&
2001 HARD_REGISTER_P (canonical_element_rtx))
2003 if (REGNO (canonical_element_rtx) != regno)
2004 *ptr = canonical_element_rtx;
2006 else if (regno != new_regno)
2008 rtx new_reg = regno_reg_rtx[new_regno];
2009 if (GET_MODE (x) != GET_MODE (new_reg))
2017 /* No need to rename the phi nodes. We'll check equivalence
2018 when inserting copies. */
2022 /* Anything else, continue traversing. */
2027 /* Record the register's canonical element stored in SRFP in the
2028 canonical_elements sbitmap packaged in DATA. This function is used
2029 as a callback function for traversing ssa_rename_from. */
2032 record_canonical_element_1 (srfp, data)
2036 unsigned int reg = ((ssa_rename_from_pair *) *srfp)->reg;
2037 sbitmap canonical_elements =
2038 ((struct ssa_rename_from_hash_table_data *) data)->canonical_elements;
2039 partition reg_partition =
2040 ((struct ssa_rename_from_hash_table_data *) data)->reg_partition;
2042 SET_BIT (canonical_elements, partition_find (reg_partition, reg));
2046 /* For each class in the REG_PARTITION corresponding to a particular
2047 hard register and machine mode, check that there are no other
2048 classes with the same hard register and machine mode. Returns
2049 nonzero if this is the case, i.e., the partition is acceptable. */
2052 check_hard_regs_in_partition (reg_partition)
2053 partition reg_partition;
2055 /* CANONICAL_ELEMENTS has a nonzero bit if a class with the given register
2056 number and machine mode has already been seen. This is a
2057 problem with the partition. */
2058 sbitmap canonical_elements;
2060 int already_seen[FIRST_PSEUDO_REGISTER][NUM_MACHINE_MODES];
2064 /* Collect a list of canonical elements. */
2065 canonical_elements = sbitmap_alloc (max_reg_num ());
2066 sbitmap_zero (canonical_elements);
2067 ssa_rename_from_traverse (&record_canonical_element_1,
2068 canonical_elements, reg_partition);
2070 /* We have not seen any hard register uses. */
2071 for (reg = 0; reg < FIRST_PSEUDO_REGISTER; ++reg)
2072 for (mach_mode = 0; mach_mode < NUM_MACHINE_MODES; ++mach_mode)
2073 already_seen[reg][mach_mode] = 0;
2075 /* Check for classes with the same hard register and machine mode. */
2076 EXECUTE_IF_SET_IN_SBITMAP (canonical_elements, 0, element_index,
2078 rtx hard_reg_rtx = ssa_rename_from_lookup (element_index);
2079 if (hard_reg_rtx != NULL_RTX &&
2080 HARD_REGISTER_P (hard_reg_rtx) &&
2081 already_seen[REGNO (hard_reg_rtx)][GET_MODE (hard_reg_rtx)] != 0)
2082 /* Two distinct partition classes should be mapped to the same
2087 sbitmap_free (canonical_elements);
2092 /* Rename regs that are equivalent in REG_PARTITION. Also collapse
2093 any SEQUENCE insns. */
2096 rename_equivalent_regs (reg_partition)
2097 partition reg_partition;
2101 for (bb = n_basic_blocks; --bb >= 0; )
2103 basic_block b = BASIC_BLOCK (bb);
2113 for_each_rtx (&PATTERN (insn),
2114 rename_equivalent_regs_in_insn,
2116 for_each_rtx (®_NOTES (insn),
2117 rename_equivalent_regs_in_insn,
2120 if (GET_CODE (PATTERN (insn)) == SEQUENCE)
2122 rtx s = PATTERN (insn);
2123 int slen = XVECLEN (s, 0);
2129 PATTERN (insn) = XVECEXP (s, 0, slen-1);
2130 for (i = 0; i < slen - 1; i++)
2131 emit_block_insn_before (XVECEXP (s, 0, i), insn, b);
2135 next = NEXT_INSN (insn);
2137 while (insn != last);
2141 /* The main entry point for moving from SSA. */
2147 partition reg_partition;
2148 rtx insns = get_insns ();
2150 /* Need global_live_at_{start,end} up to date. There should not be
2151 any significant dead code at this point, except perhaps dead
2152 stores. So do not take the time to perform dead code elimination.
2154 Register coalescing needs death notes, so generate them. */
2155 life_analysis (insns, NULL, PROP_DEATH_NOTES);
2157 /* Figure out which regs in copies and phi nodes don't conflict and
2158 therefore can be coalesced. */
2159 if (conservative_reg_partition)
2160 reg_partition = compute_conservative_reg_partition ();
2162 reg_partition = compute_coalesced_reg_partition ();
2164 if (!check_hard_regs_in_partition (reg_partition))
2165 /* Two separate partitions should correspond to the same hard
2166 register but do not. */
2169 rename_equivalent_regs (reg_partition);
2171 /* Eliminate the PHI nodes. */
2172 for (bb = n_basic_blocks; --bb >= 0; )
2174 basic_block b = BASIC_BLOCK (bb);
2177 for (e = b->pred; e; e = e->pred_next)
2178 if (e->src != ENTRY_BLOCK_PTR)
2179 eliminate_phi (e, reg_partition);
2182 partition_delete (reg_partition);
2184 /* Actually delete the PHI nodes. */
2185 for (bb = n_basic_blocks; --bb >= 0; )
2187 rtx insn = BLOCK_HEAD (bb);
2191 /* If this is a PHI node delete it. */
2192 if (PHI_NODE_P (insn))
2194 if (insn == BLOCK_END (bb))
2195 BLOCK_END (bb) = PREV_INSN (insn);
2196 insn = delete_insn (insn);
2198 /* Since all the phi nodes come at the beginning of the
2199 block, if we find an ordinary insn, we can stop looking
2200 for more phi nodes. */
2201 else if (INSN_P (insn))
2203 /* If we've reached the end of the block, stop. */
2204 else if (insn == BLOCK_END (bb))
2207 insn = NEXT_INSN (insn);
2211 /* Commit all the copy nodes needed to convert out of SSA form. */
2212 commit_edge_insertions ();
2216 count_or_remove_death_notes (NULL, 1);
2218 /* Deallocate the data structures. */
2219 VARRAY_FREE (ssa_definition);
2220 ssa_rename_from_free ();
2223 /* Scan phi nodes in successors to BB. For each such phi node that
2224 has a phi alternative value corresponding to BB, invoke FN. FN
2225 is passed the entire phi node insn, the regno of the set
2226 destination, the regno of the phi argument corresponding to BB,
2229 If FN ever returns non-zero, stops immediately and returns this
2230 value. Otherwise, returns zero. */
2233 for_each_successor_phi (bb, fn, data)
2235 successor_phi_fn fn;
2240 if (bb == EXIT_BLOCK_PTR)
2243 /* Scan outgoing edges. */
2244 for (e = bb->succ; e != NULL; e = e->succ_next)
2248 basic_block successor = e->dest;
2249 if (successor == ENTRY_BLOCK_PTR
2250 || successor == EXIT_BLOCK_PTR)
2253 /* Advance to the first non-label insn of the successor block. */
2254 insn = first_insn_after_basic_block_note (successor);
2259 /* Scan phi nodes in the successor. */
2260 for ( ; PHI_NODE_P (insn); insn = NEXT_INSN (insn))
2263 rtx phi_set = PATTERN (insn);
2264 rtx *alternative = phi_alternative (phi_set, bb->index);
2267 /* This phi function may not have an alternative
2268 corresponding to the incoming edge, indicating the
2269 assigned variable is not defined along the edge. */
2270 if (alternative == NULL)
2272 phi_src = *alternative;
2274 /* Invoke the callback. */
2275 result = (*fn) (insn, REGNO (SET_DEST (phi_set)),
2276 REGNO (phi_src), data);
2278 /* Terminate if requested. */
2287 /* Assuming the ssa_rename_from mapping has been established, yields
2288 nonzero if 1) only one SSA register of REG1 and REG2 comes from a
2289 hard register or 2) both SSA registers REG1 and REG2 come from
2290 different hard registers. */
2293 conflicting_hard_regs_p (reg1, reg2)
2297 int orig_reg1 = original_register (reg1);
2298 int orig_reg2 = original_register (reg2);
2299 if (HARD_REGISTER_NUM_P (orig_reg1) && HARD_REGISTER_NUM_P (orig_reg2)
2300 && orig_reg1 != orig_reg2)
2302 if (HARD_REGISTER_NUM_P (orig_reg1) && !HARD_REGISTER_NUM_P (orig_reg2))
2304 if (!HARD_REGISTER_NUM_P (orig_reg1) && HARD_REGISTER_NUM_P (orig_reg2))