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));
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);
259 } else if (mode_is_intb(m)) {
260 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
261 switch (get_irn_opcode(irn)) {
263 z = o = get_Const_tarval(irn);
268 ir_node* const v = get_Confirm_value(irn);
269 bitinfo* const b = get_bitinfo(v);
270 /* TODO Use bound and relation. */
277 bitinfo* const l = get_bitinfo(get_Shl_left(irn));
278 bitinfo* const r = get_bitinfo(get_Shl_right(irn));
279 ir_tarval* const rz = r->z;
281 z = tarval_shl(l->z, rz);
282 o = tarval_shl(l->o, rz);
290 bitinfo* const l = get_bitinfo(get_Shr_left(irn));
291 bitinfo* const r = get_bitinfo(get_Shr_right(irn));
292 ir_tarval* const rz = r->z;
294 z = tarval_shr(l->z, rz);
295 o = tarval_shr(l->o, rz);
303 bitinfo* const l = get_bitinfo(get_Shrs_left(irn));
304 bitinfo* const r = get_bitinfo(get_Shrs_right(irn));
305 ir_tarval* const rz = r->z;
307 z = tarval_shrs(l->z, rz);
308 o = tarval_shrs(l->o, rz);
316 bitinfo* const l = get_bitinfo(get_Rotl_left(irn));
317 bitinfo* const r = get_bitinfo(get_Rotl_right(irn));
318 ir_tarval* const rz = r->z;
320 z = tarval_rotl(l->z, rz);
321 o = tarval_rotl(l->o, rz);
329 bitinfo* const l = get_bitinfo(get_Add_left(irn));
330 bitinfo* const r = get_bitinfo(get_Add_right(irn));
331 ir_tarval* const lz = l->z;
332 ir_tarval* const lo = l->o;
333 ir_tarval* const rz = r->z;
334 ir_tarval* const ro = r->o;
335 if (lz == lo && rz == ro) {
336 z = o = tarval_add(lz, rz);
338 // TODO improve: can only do lower disjoint bits
339 /* Determine where any of the operands has zero bits, i.e. where no
340 * carry out is generated if there is not carry in */
341 ir_tarval* const no_c_in_no_c_out = tarval_and(lz, rz);
342 /* Generate a mask of the lower consecutive zeroes: x | -x. In this
343 * range the addition is disjoint and therefore Add behaves like Or.
345 ir_tarval* const low_zero_mask = tarval_or(no_c_in_no_c_out, tarval_neg(no_c_in_no_c_out));
346 ir_tarval* const low_one_mask = tarval_not(low_zero_mask);
347 z = tarval_or( tarval_or(lz, rz), low_zero_mask);
348 o = tarval_and(tarval_or(lo, ro), low_one_mask);
354 bitinfo* const l = get_bitinfo(get_Sub_left(irn));
355 bitinfo* const r = get_bitinfo(get_Sub_right(irn));
356 if (l != NULL && r != NULL) { // Sub might subtract pointers.
357 ir_tarval* const lz = l->z;
358 ir_tarval* const lo = l->o;
359 ir_tarval* const rz = r->z;
360 ir_tarval* const ro = r->o;
361 if (lz == lo && rz == ro) {
362 z = o = tarval_sub(lz, rz, NULL);
363 } else if (tarval_is_null(tarval_andnot(rz, lo))) {
364 /* Every possible one of the subtrahend is backed by a safe one of the
365 * minuend, i.e. there are no borrows. */
366 // TODO extend no-borrow like carry for Add above
367 z = tarval_andnot(lz, ro);
368 o = tarval_andnot(lo, rz);
379 bitinfo* const l = get_bitinfo(get_Mul_left(irn));
380 bitinfo* const r = get_bitinfo(get_Mul_right(irn));
381 ir_tarval* const lz = l->z;
382 ir_tarval* const lo = l->o;
383 ir_tarval* const rz = r->z;
384 ir_tarval* const ro = r->o;
385 if (lz == lo && rz == ro) {
386 z = o = tarval_mul(lz, rz);
389 // Determine safe lower zeroes: x | -x.
390 ir_tarval* const lzn = tarval_or(lz, tarval_neg(lz));
391 ir_tarval* const rzn = tarval_or(rz, tarval_neg(rz));
392 // Concatenate safe lower zeroes.
393 if (tarval_cmp(lzn, rzn) == ir_relation_less) {
394 z = tarval_mul(tarval_eor(lzn, tarval_shl(lzn, get_tarval_one(m))), rzn);
396 z = tarval_mul(tarval_eor(rzn, tarval_shl(rzn, get_tarval_one(m))), lzn);
398 o = get_tarval_null(m);
404 bitinfo* const b = get_bitinfo(get_Minus_op(irn));
406 z = o = tarval_neg(b->z);
414 bitinfo* const l = get_bitinfo(get_And_left(irn));
415 bitinfo* const r = get_bitinfo(get_And_right(irn));
416 z = tarval_and(l->z, r->z);
417 o = tarval_and(l->o, r->o);
422 bitinfo* const l = get_bitinfo(get_Or_left(irn));
423 bitinfo* const r = get_bitinfo(get_Or_right(irn));
424 z = tarval_or(l->z, r->z);
425 o = tarval_or(l->o, r->o);
430 bitinfo* const l = get_bitinfo(get_Eor_left(irn));
431 bitinfo* const r = get_bitinfo(get_Eor_right(irn));
432 ir_tarval* const lz = l->z;
433 ir_tarval* const lo = l->o;
434 ir_tarval* const rz = r->z;
435 ir_tarval* const ro = r->o;
436 z = tarval_or(tarval_andnot(lz, ro), tarval_andnot(rz, lo));
437 o = tarval_or(tarval_andnot(ro, lz), tarval_andnot(lo, rz));
442 bitinfo* const b = get_bitinfo(get_Not_op(irn));
443 z = tarval_not(b->o);
444 o = tarval_not(b->z);
449 bitinfo* const b = get_bitinfo(get_Conv_op(irn));
450 if (b == NULL) // Happens when converting from float values.
452 z = tarval_convert_to(b->z, m);
453 o = tarval_convert_to(b->o, m);
458 bitinfo* const bf = get_bitinfo(get_Mux_false(irn));
459 bitinfo* const bt = get_bitinfo(get_Mux_true(irn));
460 bitinfo* const c = get_bitinfo(get_Mux_sel(irn));
464 } else if (c->z == f) {
468 z = tarval_or( bf->z, bt->z);
469 o = tarval_and(bf->o, bt->o);
475 ir_node* const block = get_nodes_block(irn);
476 int const arity = get_Phi_n_preds(irn);
479 z = get_tarval_null(m);
480 o = get_tarval_all_one(m);
481 for (i = 0; i != arity; ++i) {
482 bitinfo* const b_cfg = get_bitinfo(get_Block_cfgpred(block, i));
483 if (b_cfg != NULL && b_cfg->z != f) {
484 bitinfo* const b = get_bitinfo(get_Phi_pred(irn, i));
485 z = tarval_or( z, b->z);
486 o = tarval_and(o, b->o);
493 bitinfo* const l = get_bitinfo(get_Cmp_left(irn));
494 bitinfo* const r = get_bitinfo(get_Cmp_right(irn));
495 if (l == NULL || r == NULL) {
496 goto result_unknown; // Cmp compares something we cannot evaluate.
498 ir_tarval* const lz = l->z;
499 ir_tarval* const lo = l->o;
500 ir_tarval* const rz = r->z;
501 ir_tarval* const ro = r->o;
502 ir_relation const relation = get_Cmp_relation(irn);
504 case ir_relation_less_greater:
505 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
506 !tarval_is_null(tarval_andnot(lo, rz))) {
507 // At least one bit differs.
509 } else if (lz == lo && rz == ro && lz == rz) {
516 case ir_relation_equal:
517 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
518 !tarval_is_null(tarval_andnot(lo, rz))) {
519 // At least one bit differs.
521 } else if (lz == lo && rz == ro && lz == rz) {
528 case ir_relation_less_equal:
529 case ir_relation_less:
530 /* TODO handle negative values */
531 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
532 tarval_is_negative(rz) || tarval_is_negative(ro))
535 if (tarval_cmp(lz, ro) & relation) {
536 /* Left upper bound is smaller(/equal) than right lower bound. */
538 } else if (!(tarval_cmp(lo, rz) & relation)) {
539 /* Left lower bound is not smaller(/equal) than right upper bound. */
546 case ir_relation_greater_equal:
547 case ir_relation_greater:
548 /* TODO handle negative values */
549 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
550 tarval_is_negative(rz) || tarval_is_negative(ro))
553 if (!(tarval_cmp(lz, ro) & relation)) {
554 /* Left upper bound is not greater(/equal) than right lower bound. */
556 } else if (tarval_cmp(lo, rz) & relation) {
557 /* Left lower bound is greater(/equal) than right upper bound. */
573 DB((dbg, LEVEL_4, "cannot analyse %+F\n", irn));
575 z = get_tarval_all_one(m);
576 o = get_tarval_null(m);
584 return set_bitinfo(irn, z, o);
587 static void first_round(ir_node* const irn, void* const env)
589 pdeq* const q = (pdeq*)env;
592 if (is_Phi(irn) || is_Block(irn)) {
593 /* Only Phis (and their users) need another round, if we did not have
594 * information about all their inputs in the first round, i.e. in loops. */
595 /* TODO inserts all Phis, should only insert Phis, which did no have all
596 * predecessors available */
601 static void apply_result(ir_node* const irn, void* ctx)
603 environment_t* env = (environment_t*)ctx;
610 bitinfo* const block_b = get_bitinfo(irn);
611 /* Trivially unreachable blocks have no info. */
612 if (block_b == NULL || block_b->z == get_tarval_b_false()) {
613 exchange(irn, get_irg_bad(get_Block_irg(irn)));
619 /* Unreachable blocks are replaced before the nodes in them. */
620 block = get_nodes_block(irn);
622 exchange(irn, block);
627 b = get_bitinfo(irn);
629 if (is_Const(irn)) return; // It cannot get any better than a Const.
633 // Only display information if we could find out anything about the value.
634 DEBUG_ONLY(if (!tarval_is_all_one(z) || !tarval_is_null(o)))
635 DB((dbg, LEVEL_2, "%+F: 0:%T 1:%T%s\n", irn, z, o, z == o ? " --- constant" : ""));
637 // Replace node with constant value by Const.
639 ir_mode* const m = get_irn_mode(irn);
641 if (mode_is_intb(m)) {
642 ir_graph *irg = get_irn_irg(irn);
643 n = new_r_Const(irg, z);
644 } else if (m == mode_X) {
645 ir_graph* const irg = get_Block_irg(block);
646 if (z == get_tarval_b_true()) {
647 // Might produce an endless loop, so keep the block.
648 add_End_keepalive(get_irg_end(irg), block);
649 n = new_r_Jmp(block);
652 /* Transferring analysis information to the bad node makes it a
653 * candidate for replacement. */
665 switch (get_irn_opcode(irn)) {
667 ir_node* const l = get_And_left(irn);
668 ir_node* const r = get_And_right(irn);
669 bitinfo const* const bl = get_bitinfo(l);
670 bitinfo const* const br = get_bitinfo(r);
671 if (bl->z == bl->o) {
672 if (tarval_is_null(tarval_andnot(br->z, bl->z))) {
673 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
677 } else if (br->z == br->o) {
678 if (tarval_is_null(tarval_andnot(bl->z, br->z))) {
679 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
688 ir_node* const l = get_Or_left(irn);
689 ir_node* const r = get_Or_right(irn);
690 bitinfo const* const bl = get_bitinfo(l);
691 bitinfo const* const br = get_bitinfo(r);
692 if (bl->z == bl->o) {
693 if (tarval_is_null(tarval_andnot(bl->o, br->o))) {
694 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
698 } else if (br->z == br->o) {
699 if (tarval_is_null(tarval_andnot(br->o, bl->o))) {
700 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
710 static void queue_users(pdeq* const q, ir_node* const n)
712 if (get_irn_mode(n) == mode_X) {
713 /* When the state of a control flow node changes, not only queue its
714 * successor blocks, but also the Phis in these blocks, because the Phis
715 * must reconsider this input path. */
717 foreach_out_edge(n, e) {
718 ir_node* const src = get_edge_src_irn(e);
720 /* should always be a block */
723 for (phi = get_Block_phis(src); phi; phi = get_Phi_next(phi))
729 foreach_out_edge(n, e) {
730 ir_node* const src = get_edge_src_irn(e);
731 if (get_irn_mode(src) == mode_T) {
740 static void clear_links(ir_node *irn, void *env)
743 set_irn_link(irn, NULL);
745 set_Block_phis(irn, NULL);
748 static void build_phi_lists(ir_node *irn, void *env)
752 add_Block_phi(get_nodes_block(irn), irn);
755 void fixpoint_vrp(ir_graph* const irg)
759 FIRM_DBG_REGISTER(dbg, "firm.opt.fp-vrp");
760 DB((dbg, LEVEL_1, "===> Performing constant propagation on %+F\n", irg));
764 /* HACK: to avoid finding dead code */
765 edges_deactivate(irg);
771 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
774 pdeq* const q = new_pdeq();
776 /* We need this extra step because the dom tree does not contain unreachable
777 blocks in Firm. Moreover build phi list. */
778 irg_walk_anchors(irg, clear_links, build_phi_lists, NULL);
780 { ir_tarval* const f = get_tarval_b_false();
781 ir_tarval* const t = get_tarval_b_true();
782 set_bitinfo(get_irg_bad(irg), f, t); /* Undefined. */
783 set_bitinfo(get_irg_end_block(irg), t, f); /* Reachable. */
786 /* TODO Improve iteration order. Best is reverse postorder in data flow
787 * direction and respecting loop nesting for fastest convergence. */
788 irg_walk_blkwise_dom_top_down(irg, NULL, first_round, q);
790 while (!pdeq_empty(q)) {
791 ir_node* const n = (ir_node*)pdeq_getl(q);
799 DB((dbg, LEVEL_2, "---> Applying analysis results\n"));
801 irg_walk_graph(irg, NULL, apply_result, &env);
804 /* control flow might changed */
805 set_irg_outs_inconsistent(irg);
806 set_irg_extblk_inconsistent(irg);
807 set_irg_doms_inconsistent(irg);
808 set_irg_loopinfo_inconsistent(irg);
809 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
812 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
814 obstack_free(&obst, NULL);
817 ir_graph_pass_t *fixpoint_vrp_irg_pass(const char *name)
819 return def_graph_pass(name ? name : "fixpoint_vrp", fixpoint_vrp);