2 * Copyright (C) 1995-2010 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Data-flow driven minimal fixpoint value range propagation
23 * @author Christoph Mallon
30 #include "adt/xmalloc.h"
39 #include "iroptimize.h"
46 * - Implement cleared/set bit calculation for Add, Sub, Minus, Mul, Div, Mod, Shl, Shr, Shrs, Rotl
47 * - Implement min/max calculation for And, Eor, Or, Not, Conv, Shl, Shr, Shrs, Rotl, Mux
48 * - Implement min/max calculation for Add, Sub, Minus, Mul, Div, Mod, Conv, Shl, Shr, Shrs, Rotl, Mux
51 /* Tables of the cleared/set bit lattice
53 * Encoding of the lattice
56 * 01 - impossible state, is zero /and/ one
57 * 10 T top, may be either zero or one
71 * Half adder, half subtractor, and, xor, or, Mux
83 * Full adder, full subtractor
114 * Assume: Xmin <= Xmax and no overflow
115 * A + B = (Amin + Bmin, Amax + Bmax)
116 * -A = (-Amax, -Amin)
117 * A - B = A + -B = (Amin (-B)min, Amax + (-B)max) = (Amin - Bmax, Amax - Bmin)
120 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
122 static struct obstack obst;
124 typedef struct bitinfo
126 ir_tarval* z; // safe zeroes, 0 = bit is zero, 1 = bit maybe is 1
127 ir_tarval* o; // safe ones, 0 = bit maybe is zero, 1 = bit is 1
130 typedef struct environment_t {
131 unsigned modified:1; /**< Set, if the graph was modified. */
134 static inline bitinfo* get_bitinfo(ir_node const* const irn)
136 return (bitinfo*)get_irn_link(irn);
139 static int set_bitinfo(ir_node* const irn, ir_tarval* const z, ir_tarval* const o)
141 bitinfo* b = get_bitinfo(irn);
143 b = OALLOCZ(&obst, bitinfo);
144 set_irn_link(irn, b);
145 } else if (z == b->z && o == b->o) {
150 DB((dbg, LEVEL_3, "%+F: 0:%T 1:%T\n", irn, z, o));
154 static int mode_is_intb(ir_mode const* const m)
156 return mode_is_int(m) || m == mode_b;
159 static int transfer(ir_node* const irn)
161 ir_tarval* const f = get_tarval_b_false();
162 ir_tarval* const t = get_tarval_b_true();
163 ir_mode* const m = get_irn_mode(irn);
168 bitinfo* const b = get_bitinfo(get_nodes_block(irn));
170 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
172 if (b->z == f && b->o == f) {
175 } else switch (get_irn_opcode(irn)) {
177 ir_node* const pred = get_Proj_pred(irn);
178 if (is_Start(pred)) {
179 goto result_unknown_X;
180 } else if (is_Cond(pred)) {
181 ir_node* const selector = get_Cond_selector(pred);
182 bitinfo* const b = get_bitinfo(selector);
183 ir_tarval* const bz = b->z;
184 ir_tarval* const bo = b->o;
185 if (get_irn_mode(selector) == mode_b) {
187 if ((bz == t) == get_Proj_proj(irn)) {
193 goto result_unknown_X;
196 long const val = get_Proj_proj(irn);
197 if (val != get_Cond_default_proj(pred)) {
198 ir_tarval* const tv = new_tarval_from_long(val, get_irn_mode(selector));
199 if (!tarval_is_null(tarval_andnot(tv, bz)) ||
200 !tarval_is_null(tarval_andnot(bo, tv))) {
201 // At least one bit differs.
203 #if 0 // TODO must handle default Proj
204 } else if (bz == bo && bz == tv) {
208 goto result_unknown_X;
211 goto cannot_analyse_X;
215 goto cannot_analyse_X;
221 goto result_unknown_X;
225 DB((dbg, LEVEL_4, "cannot analyse %+F\n", irn));
231 } else if (is_Block(irn)) {
233 int const arity = get_Block_n_cfgpreds(irn);
236 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
237 for (i = 0; i != arity; ++i) {
238 bitinfo* const b = get_bitinfo(get_Block_cfgpred(irn, i));
239 if (b != NULL && b->z == t) {
246 ir_graph *const irg = get_Block_irg(irn);
248 irn == get_irg_start_block(irg) ||
249 irn == get_irg_end_block(irg);
253 z = reachable ? t : f;
254 } else if (mode_is_intb(m)) {
255 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
256 switch (get_irn_opcode(irn)) {
258 z = o = get_Const_tarval(irn);
263 ir_node* const v = get_Confirm_value(irn);
264 bitinfo* const b = get_bitinfo(v);
265 /* TODO Use bound and relation. */
272 bitinfo* const l = get_bitinfo(get_Shl_left(irn));
273 bitinfo* const r = get_bitinfo(get_Shl_right(irn));
274 ir_tarval* const rz = r->z;
276 z = tarval_shl(l->z, rz);
277 o = tarval_shl(l->o, rz);
285 bitinfo* const l = get_bitinfo(get_Shr_left(irn));
286 bitinfo* const r = get_bitinfo(get_Shr_right(irn));
287 ir_tarval* const rz = r->z;
289 z = tarval_shr(l->z, rz);
290 o = tarval_shr(l->o, rz);
298 bitinfo* const l = get_bitinfo(get_Shrs_left(irn));
299 bitinfo* const r = get_bitinfo(get_Shrs_right(irn));
300 ir_tarval* const rz = r->z;
302 z = tarval_shrs(l->z, rz);
303 o = tarval_shrs(l->o, rz);
311 bitinfo* const l = get_bitinfo(get_Rotl_left(irn));
312 bitinfo* const r = get_bitinfo(get_Rotl_right(irn));
313 ir_tarval* const rz = r->z;
315 z = tarval_rotl(l->z, rz);
316 o = tarval_rotl(l->o, rz);
324 bitinfo* const l = get_bitinfo(get_Add_left(irn));
325 bitinfo* const r = get_bitinfo(get_Add_right(irn));
326 ir_tarval* const lz = l->z;
327 ir_tarval* const lo = l->o;
328 ir_tarval* const rz = r->z;
329 ir_tarval* const ro = r->o;
330 if (lz == lo && rz == ro) {
331 z = o = tarval_add(lz, rz);
333 // TODO improve: can only do lower disjoint bits
334 /* Determine where any of the operands has zero bits, i.e. where no
335 * carry out is generated if there is not carry in */
336 ir_tarval* const no_c_in_no_c_out = tarval_and(lz, rz);
337 /* Generate a mask of the lower consecutive zeroes: x | -x. In this
338 * range the addition is disjoint and therefore Add behaves like Or.
340 ir_tarval* const low_zero_mask = tarval_or(no_c_in_no_c_out, tarval_neg(no_c_in_no_c_out));
341 ir_tarval* const low_one_mask = tarval_not(low_zero_mask);
342 z = tarval_or( tarval_or(lz, rz), low_zero_mask);
343 o = tarval_and(tarval_or(lo, ro), low_one_mask);
349 bitinfo* const l = get_bitinfo(get_Sub_left(irn));
350 bitinfo* const r = get_bitinfo(get_Sub_right(irn));
351 if (l != NULL && r != NULL) { // Sub might subtract pointers.
352 ir_tarval* const lz = l->z;
353 ir_tarval* const lo = l->o;
354 ir_tarval* const rz = r->z;
355 ir_tarval* const ro = r->o;
356 if (lz == lo && rz == ro) {
357 z = o = tarval_sub(lz, rz, NULL);
358 } else if (tarval_is_null(tarval_andnot(rz, lo))) {
359 /* Every possible one of the subtrahend is backed by a safe one of the
360 * minuend, i.e. there are no borrows. */
361 // TODO extend no-borrow like carry for Add above
362 z = tarval_andnot(lz, ro);
363 o = tarval_andnot(lo, rz);
374 bitinfo* const l = get_bitinfo(get_Mul_left(irn));
375 bitinfo* const r = get_bitinfo(get_Mul_right(irn));
376 ir_tarval* const lz = l->z;
377 ir_tarval* const lo = l->o;
378 ir_tarval* const rz = r->z;
379 ir_tarval* const ro = r->o;
380 if (lz == lo && rz == ro) {
381 z = o = tarval_mul(lz, rz);
384 // Determine safe lower zeroes: x | -x.
385 ir_tarval* const lzn = tarval_or(lz, tarval_neg(lz));
386 ir_tarval* const rzn = tarval_or(rz, tarval_neg(rz));
387 // Concatenate safe lower zeroes.
388 if (tarval_cmp(lzn, rzn) == ir_relation_less) {
389 z = tarval_mul(tarval_eor(lzn, tarval_shl(lzn, get_tarval_one(m))), rzn);
391 z = tarval_mul(tarval_eor(rzn, tarval_shl(rzn, get_tarval_one(m))), lzn);
393 o = get_tarval_null(m);
399 bitinfo* const b = get_bitinfo(get_Minus_op(irn));
401 z = o = tarval_neg(b->z);
409 bitinfo* const l = get_bitinfo(get_And_left(irn));
410 bitinfo* const r = get_bitinfo(get_And_right(irn));
411 z = tarval_and(l->z, r->z);
412 o = tarval_and(l->o, r->o);
417 bitinfo* const l = get_bitinfo(get_Or_left(irn));
418 bitinfo* const r = get_bitinfo(get_Or_right(irn));
419 z = tarval_or(l->z, r->z);
420 o = tarval_or(l->o, r->o);
425 bitinfo* const l = get_bitinfo(get_Eor_left(irn));
426 bitinfo* const r = get_bitinfo(get_Eor_right(irn));
427 ir_tarval* const lz = l->z;
428 ir_tarval* const lo = l->o;
429 ir_tarval* const rz = r->z;
430 ir_tarval* const ro = r->o;
431 z = tarval_or(tarval_andnot(lz, ro), tarval_andnot(rz, lo));
432 o = tarval_or(tarval_andnot(ro, lz), tarval_andnot(lo, rz));
437 bitinfo* const b = get_bitinfo(get_Not_op(irn));
438 z = tarval_not(b->o);
439 o = tarval_not(b->z);
444 bitinfo* const b = get_bitinfo(get_Conv_op(irn));
445 if (b == NULL) // Happens when converting from float values.
447 z = tarval_convert_to(b->z, m);
448 o = tarval_convert_to(b->o, m);
453 bitinfo* const bf = get_bitinfo(get_Mux_false(irn));
454 bitinfo* const bt = get_bitinfo(get_Mux_true(irn));
455 bitinfo* const c = get_bitinfo(get_Mux_sel(irn));
459 } else if (c->z == f) {
463 z = tarval_or( bf->z, bt->z);
464 o = tarval_and(bf->o, bt->o);
470 ir_node* const block = get_nodes_block(irn);
471 int const arity = get_Phi_n_preds(irn);
474 z = get_tarval_null(m);
475 o = get_tarval_all_one(m);
476 for (i = 0; i != arity; ++i) {
477 bitinfo* const b_cfg = get_bitinfo(get_Block_cfgpred(block, i));
478 if (b_cfg != NULL && b_cfg->z != f) {
479 bitinfo* const b = get_bitinfo(get_Phi_pred(irn, i));
480 z = tarval_or( z, b->z);
481 o = tarval_and(o, b->o);
488 bitinfo* const l = get_bitinfo(get_Cmp_left(irn));
489 bitinfo* const r = get_bitinfo(get_Cmp_right(irn));
490 if (l == NULL || r == NULL) {
491 goto result_unknown; // Cmp compares something we cannot evaluate.
493 ir_tarval* const lz = l->z;
494 ir_tarval* const lo = l->o;
495 ir_tarval* const rz = r->z;
496 ir_tarval* const ro = r->o;
497 ir_relation const relation = get_Cmp_relation(irn);
499 case ir_relation_less_greater:
500 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
501 !tarval_is_null(tarval_andnot(lo, rz))) {
502 // At least one bit differs.
504 } else if (lz == lo && rz == ro && lz == rz) {
511 case ir_relation_equal:
512 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
513 !tarval_is_null(tarval_andnot(lo, rz))) {
514 // At least one bit differs.
516 } else if (lz == lo && rz == ro && lz == rz) {
523 case ir_relation_less_equal:
524 case ir_relation_less:
525 /* TODO handle negative values */
526 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
527 tarval_is_negative(rz) || tarval_is_negative(ro))
530 if (tarval_cmp(lz, ro) & relation) {
531 /* Left upper bound is smaller(/equal) than right lower bound. */
533 } else if (!(tarval_cmp(lo, rz) & relation)) {
534 /* Left lower bound is not smaller(/equal) than right upper bound. */
541 case ir_relation_greater_equal:
542 case ir_relation_greater:
543 /* TODO handle negative values */
544 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
545 tarval_is_negative(rz) || tarval_is_negative(ro))
548 if (!(tarval_cmp(lz, ro) & relation)) {
549 /* Left upper bound is not greater(/equal) than right lower bound. */
551 } else if (tarval_cmp(lo, rz) & relation) {
552 /* Left lower bound is greater(/equal) than right upper bound. */
568 DB((dbg, LEVEL_4, "cannot analyse %+F\n", irn));
570 z = get_tarval_all_one(m);
571 o = get_tarval_null(m);
579 return set_bitinfo(irn, z, o);
582 static void first_round(ir_node* const irn, void* const env)
584 pdeq* const q = (pdeq*)env;
587 if (is_Phi(irn) || is_Block(irn)) {
588 /* Only Phis (and their users) need another round, if we did not have
589 * information about all their inputs in the first round, i.e. in loops. */
590 /* TODO inserts all Phis, should only insert Phis, which did no have all
591 * predecessors available */
596 static void apply_result(ir_node* const irn, void* ctx)
598 environment_t* env = (environment_t*)ctx;
605 bitinfo* const block_b = get_bitinfo(irn);
606 /* Trivially unreachable blocks have no info. */
607 if (block_b == NULL || block_b->z == get_tarval_b_false()) {
608 exchange(irn, get_irg_bad(get_Block_irg(irn)));
614 /* Unreachable blocks are replaced before the nodes in them. */
615 block = get_nodes_block(irn);
617 exchange(irn, block);
622 b = get_bitinfo(irn);
624 if (is_Const(irn)) return; // It cannot get any better than a Const.
628 // Only display information if we could find out anything about the value.
629 DEBUG_ONLY(if (!tarval_is_all_one(z) || !tarval_is_null(o)))
630 DB((dbg, LEVEL_2, "%+F: 0:%T 1:%T%s\n", irn, z, o, z == o ? " --- constant" : ""));
632 // Replace node with constant value by Const.
634 ir_mode* const m = get_irn_mode(irn);
636 if (mode_is_intb(m)) {
637 ir_graph *irg = get_irn_irg(irn);
638 n = new_r_Const(irg, z);
639 } else if (m == mode_X) {
640 ir_graph* const irg = get_Block_irg(block);
641 if (z == get_tarval_b_true()) {
642 // Might produce an endless loop, so keep the block.
643 add_End_keepalive(get_irg_end(irg), block);
644 n = new_r_Jmp(block);
647 /* Transferring analysis information to the bad node makes it a
648 * candidate for replacement. */
660 switch (get_irn_opcode(irn)) {
662 ir_node* const l = get_And_left(irn);
663 ir_node* const r = get_And_right(irn);
664 bitinfo const* const bl = get_bitinfo(l);
665 bitinfo const* const br = get_bitinfo(r);
666 if (bl->z == bl->o) {
667 if (tarval_is_null(tarval_andnot(br->z, bl->z))) {
668 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
672 } else if (br->z == br->o) {
673 if (tarval_is_null(tarval_andnot(bl->z, br->z))) {
674 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
683 ir_node* const l = get_Or_left(irn);
684 ir_node* const r = get_Or_right(irn);
685 bitinfo const* const bl = get_bitinfo(l);
686 bitinfo const* const br = get_bitinfo(r);
687 if (bl->z == bl->o) {
688 if (tarval_is_null(tarval_andnot(bl->o, br->o))) {
689 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
693 } else if (br->z == br->o) {
694 if (tarval_is_null(tarval_andnot(br->o, bl->o))) {
695 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
705 static void queue_users(pdeq* const q, ir_node* const n)
707 if (get_irn_mode(n) == mode_X) {
708 /* When the state of a control flow node changes, not only queue its
709 * successor blocks, but also the Phis in these blocks, because the Phis
710 * must reconsider this input path. */
712 foreach_out_edge(n, e) {
713 ir_node* const src = get_edge_src_irn(e);
715 /* should always be a block */
718 for (phi = get_Block_phis(src); phi; phi = get_Phi_next(phi))
724 foreach_out_edge(n, e) {
725 ir_node* const src = get_edge_src_irn(e);
726 if (get_irn_mode(src) == mode_T) {
735 static void clear_links(ir_node *irn, void *env)
738 set_irn_link(irn, NULL);
740 set_Block_phis(irn, NULL);
743 static void build_phi_lists(ir_node *irn, void *env)
747 add_Block_phi(get_nodes_block(irn), irn);
750 void fixpoint_vrp(ir_graph* const irg)
754 FIRM_DBG_REGISTER(dbg, "firm.opt.fp-vrp");
755 DB((dbg, LEVEL_1, "===> Performing constant propagation on %+F\n", irg));
759 /* HACK: to avoid finding dead code */
760 edges_deactivate(irg);
766 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
769 pdeq* const q = new_pdeq();
771 /* We need this extra step because the dom tree does not contain unreachable
772 blocks in Firm. Moreover build phi list. */
773 irg_walk_anchors(irg, clear_links, build_phi_lists, NULL);
775 /* TODO Improve iteration order. Best is reverse postorder in data flow
776 * direction and respecting loop nesting for fastest convergence. */
777 irg_walk_blkwise_dom_top_down(irg, NULL, first_round, q);
779 while (!pdeq_empty(q)) {
780 ir_node* const n = (ir_node*)pdeq_getl(q);
788 DB((dbg, LEVEL_2, "---> Applying analysis results\n"));
790 irg_walk_graph(irg, NULL, apply_result, &env);
793 /* control flow might changed */
794 set_irg_outs_inconsistent(irg);
795 set_irg_extblk_inconsistent(irg);
796 set_irg_doms_inconsistent(irg);
797 set_irg_loopinfo_inconsistent(irg);
798 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
801 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
803 obstack_free(&obst, NULL);
806 ir_graph_pass_t *fixpoint_vrp_irg_pass(const char *name)
808 return def_graph_pass(name ? name : "fixpoint_vrp", fixpoint_vrp);