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
32 #include "adt/xmalloc.h"
38 #include "irgraph_t.h"
41 #include "iroptimize.h"
46 #include "opt_manage.h"
49 * - Implement cleared/set bit calculation for Add, Sub, Minus, Mul, Div, Mod, Shl, Shr, Shrs, Rotl
50 * - Implement min/max calculation for And, Eor, Or, Not, Conv, Shl, Shr, Shrs, Rotl, Mux
51 * - Implement min/max calculation for Add, Sub, Minus, Mul, Div, Mod, Conv, Shl, Shr, Shrs, Rotl, Mux
54 /* Tables of the cleared/set bit lattice
56 * Encoding of the lattice
59 * 01 - impossible state, is zero /and/ one
60 * 10 T top, may be either zero or one
74 * Half adder, half subtractor, and, xor, or, Mux
86 * Full adder, full subtractor
117 * Assume: Xmin <= Xmax and no overflow
118 * A + B = (Amin + Bmin, Amax + Bmax)
119 * -A = (-Amax, -Amin)
120 * A - B = A + -B = (Amin (-B)min, Amax + (-B)max) = (Amin - Bmax, Amax - Bmin)
123 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
125 static struct obstack obst;
127 typedef struct bitinfo
129 ir_tarval* z; // safe zeroes, 0 = bit is zero, 1 = bit maybe is 1
130 ir_tarval* o; // safe ones, 0 = bit maybe is zero, 1 = bit is 1
133 typedef struct environment_t {
134 unsigned modified:1; /**< Set, if the graph was modified. */
137 static bool is_undefined(bitinfo const* const b)
139 return tarval_is_null(b->z) && tarval_is_all_one(b->o);
142 static inline bitinfo* get_bitinfo(ir_node const* const irn)
144 return (bitinfo*)get_irn_link(irn);
147 static int set_bitinfo(ir_node* const irn, ir_tarval* const z, ir_tarval* const o)
149 bitinfo* b = get_bitinfo(irn);
151 b = OALLOCZ(&obst, bitinfo);
152 set_irn_link(irn, b);
153 } else if (z == b->z && o == b->o) {
156 /* Assert monotonicity. */
157 assert(tarval_is_null(tarval_andnot(b->z, z)));
158 assert(tarval_is_null(tarval_andnot(o, b->o)));
162 DB((dbg, LEVEL_3, "%+F: 0:%T 1:%T\n", irn, z, o));
166 static int mode_is_intb(ir_mode const* const m)
168 return mode_is_int(m) || m == mode_b;
171 static int transfer(ir_node* const irn)
173 ir_tarval* const f = get_tarval_b_false();
174 ir_tarval* const t = get_tarval_b_true();
175 ir_mode* const m = get_irn_mode(irn);
179 if (is_Bad(irn)) return 0;
182 bitinfo* const b = get_bitinfo(get_nodes_block(irn));
184 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
186 /* Unreachble blocks might have no bitinfo. */
187 if (b == NULL || b->z == f) {
191 } else switch (get_irn_opcode(irn)) {
193 ir_node* const pred = get_Proj_pred(irn);
194 if (is_Start(pred)) {
195 goto result_unknown_X;
196 } else if (is_Cond(pred)) {
197 ir_node* const selector = get_Cond_selector(pred);
198 bitinfo* const b = get_bitinfo(selector);
202 if ((b->z == t) == get_Proj_proj(irn)) {
208 goto result_unknown_X;
210 } else if (is_Switch(pred)) {
211 ir_node* const selector = get_Switch_selector(pred);
212 bitinfo* const b = get_bitinfo(selector);
216 goto cannot_analyse_X;
218 goto cannot_analyse_X;
224 goto result_unknown_X;
228 DB((dbg, LEVEL_4, "cannot analyse %+F\n", irn));
234 } else if (is_Block(irn)) {
236 int const arity = get_Block_n_cfgpreds(irn);
239 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
240 for (i = 0; i != arity; ++i) {
241 bitinfo* const b = get_bitinfo(get_Block_cfgpred(irn, i));
242 if (b != NULL && b->z == t) {
249 ir_graph *const irg = get_Block_irg(irn);
251 irn == get_irg_start_block(irg) ||
252 irn == get_irg_end_block(irg);
262 } else if (mode_is_intb(m)) {
263 bitinfo* const b = get_bitinfo(get_nodes_block(irn));
265 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
267 if (b == NULL || b->z == f) {
269 z = get_tarval_null(m);
270 o = get_tarval_all_one(m);
271 } else if (is_Phi(irn)) {
272 ir_node* const block = get_nodes_block(irn);
273 int const arity = get_Phi_n_preds(irn);
276 z = get_tarval_null(m);
277 o = get_tarval_all_one(m);
278 for (i = 0; i != arity; ++i) {
279 bitinfo* const b_cfg = get_bitinfo(get_Block_cfgpred(block, i));
280 if (b_cfg != NULL && b_cfg->z != f) {
281 bitinfo* const b = get_bitinfo(get_Phi_pred(irn, i));
282 /* Only use input if it's not undefined. */
283 if (!is_undefined(b)) {
284 z = tarval_or( z, b->z);
285 o = tarval_and(o, b->o);
290 int const arity = get_irn_arity(irn);
293 /* Undefined if any input is undefined. */
294 for (i = 0; i != arity; ++i) {
295 ir_node* const pred = get_irn_n(irn, i);
296 bitinfo* const pred_b = get_bitinfo(pred);
297 if (pred_b != NULL && is_undefined(pred_b))
301 switch (get_irn_opcode(irn)) {
303 z = o = get_Const_tarval(irn);
308 ir_node* const v = get_Confirm_value(irn);
309 bitinfo* const b = get_bitinfo(v);
310 /* TODO Use bound and relation. */
313 if ((get_Confirm_relation(irn) & ~ir_relation_unordered) == ir_relation_equal) {
314 bitinfo* const bound_b = get_bitinfo(get_Confirm_bound(irn));
315 z = tarval_and(z, bound_b->z);
316 o = tarval_or( o, bound_b->o);
322 bitinfo* const l = get_bitinfo(get_Shl_left(irn));
323 bitinfo* const r = get_bitinfo(get_Shl_right(irn));
324 ir_tarval* const rz = r->z;
326 z = tarval_shl(l->z, rz);
327 o = tarval_shl(l->o, rz);
335 bitinfo* const l = get_bitinfo(get_Shr_left(irn));
336 bitinfo* const r = get_bitinfo(get_Shr_right(irn));
337 ir_tarval* const rz = r->z;
339 z = tarval_shr(l->z, rz);
340 o = tarval_shr(l->o, rz);
348 bitinfo* const l = get_bitinfo(get_Shrs_left(irn));
349 bitinfo* const r = get_bitinfo(get_Shrs_right(irn));
350 ir_tarval* const rz = r->z;
352 z = tarval_shrs(l->z, rz);
353 o = tarval_shrs(l->o, rz);
361 bitinfo* const l = get_bitinfo(get_Rotl_left(irn));
362 bitinfo* const r = get_bitinfo(get_Rotl_right(irn));
363 ir_tarval* const rz = r->z;
365 z = tarval_rotl(l->z, rz);
366 o = tarval_rotl(l->o, rz);
374 bitinfo* const l = get_bitinfo(get_Add_left(irn));
375 bitinfo* const r = get_bitinfo(get_Add_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_add(lz, rz);
383 // TODO improve: can only do lower disjoint bits
384 /* Determine where any of the operands has zero bits, i.e. where no
385 * carry out is generated if there is not carry in */
386 ir_tarval* const no_c_in_no_c_out = tarval_and(lz, rz);
387 /* Generate a mask of the lower consecutive zeroes: x | -x. In this
388 * range the addition is disjoint and therefore Add behaves like Or.
390 ir_tarval* const low_zero_mask = tarval_or(no_c_in_no_c_out, tarval_neg(no_c_in_no_c_out));
391 ir_tarval* const low_one_mask = tarval_not(low_zero_mask);
392 z = tarval_or( tarval_or(lz, rz), low_zero_mask);
393 o = tarval_and(tarval_or(lo, ro), low_one_mask);
399 bitinfo* const l = get_bitinfo(get_Sub_left(irn));
400 bitinfo* const r = get_bitinfo(get_Sub_right(irn));
401 if (l != NULL && r != NULL) { // Sub might subtract pointers.
402 ir_tarval* const lz = l->z;
403 ir_tarval* const lo = l->o;
404 ir_tarval* const rz = r->z;
405 ir_tarval* const ro = r->o;
406 if (lz == lo && rz == ro) {
407 z = o = tarval_sub(lz, rz, NULL);
408 } else if (tarval_is_null(tarval_andnot(rz, lo))) {
409 /* Every possible one of the subtrahend is backed by a safe one of the
410 * minuend, i.e. there are no borrows. */
411 // TODO extend no-borrow like carry for Add above
412 z = tarval_andnot(lz, ro);
413 o = tarval_andnot(lo, rz);
424 bitinfo* const l = get_bitinfo(get_Mul_left(irn));
425 bitinfo* const r = get_bitinfo(get_Mul_right(irn));
426 ir_tarval* const lz = l->z;
427 ir_tarval* const lo = l->o;
428 ir_tarval* const rz = r->z;
429 ir_tarval* const ro = r->o;
430 if (lz == lo && rz == ro) {
431 z = o = tarval_mul(lz, rz);
434 // Determine safe lower zeroes: x | -x.
435 ir_tarval* const lzn = tarval_or(lz, tarval_neg(lz));
436 ir_tarval* const rzn = tarval_or(rz, tarval_neg(rz));
437 // Concatenate safe lower zeroes.
438 if (tarval_cmp(lzn, rzn) == ir_relation_less) {
439 z = tarval_mul(tarval_eor(lzn, tarval_shl(lzn, get_tarval_one(m))), rzn);
441 z = tarval_mul(tarval_eor(rzn, tarval_shl(rzn, get_tarval_one(m))), lzn);
443 o = get_tarval_null(m);
449 bitinfo* const b = get_bitinfo(get_Minus_op(irn));
451 z = o = tarval_neg(b->z);
459 bitinfo* const l = get_bitinfo(get_And_left(irn));
460 bitinfo* const r = get_bitinfo(get_And_right(irn));
461 z = tarval_and(l->z, r->z);
462 o = tarval_and(l->o, r->o);
467 bitinfo* const l = get_bitinfo(get_Or_left(irn));
468 bitinfo* const r = get_bitinfo(get_Or_right(irn));
469 z = tarval_or(l->z, r->z);
470 o = tarval_or(l->o, r->o);
475 bitinfo* const l = get_bitinfo(get_Eor_left(irn));
476 bitinfo* const r = get_bitinfo(get_Eor_right(irn));
477 ir_tarval* const lz = l->z;
478 ir_tarval* const lo = l->o;
479 ir_tarval* const rz = r->z;
480 ir_tarval* const ro = r->o;
481 z = tarval_or(tarval_andnot(lz, ro), tarval_andnot(rz, lo));
482 o = tarval_or(tarval_andnot(ro, lz), tarval_andnot(lo, rz));
487 bitinfo* const b = get_bitinfo(get_Not_op(irn));
488 z = tarval_not(b->o);
489 o = tarval_not(b->z);
494 bitinfo* const b = get_bitinfo(get_Conv_op(irn));
495 if (b == NULL) // Happens when converting from float values.
497 z = tarval_convert_to(b->z, m);
498 o = tarval_convert_to(b->o, m);
503 bitinfo* const bf = get_bitinfo(get_Mux_false(irn));
504 bitinfo* const bt = get_bitinfo(get_Mux_true(irn));
505 bitinfo* const c = get_bitinfo(get_Mux_sel(irn));
509 } else if (c->z == f) {
513 z = tarval_or( bf->z, bt->z);
514 o = tarval_and(bf->o, bt->o);
520 bitinfo* const l = get_bitinfo(get_Cmp_left(irn));
521 bitinfo* const r = get_bitinfo(get_Cmp_right(irn));
522 if (l == NULL || r == NULL) {
523 goto result_unknown; // Cmp compares something we cannot evaluate.
525 ir_tarval* const lz = l->z;
526 ir_tarval* const lo = l->o;
527 ir_tarval* const rz = r->z;
528 ir_tarval* const ro = r->o;
529 ir_relation const relation = get_Cmp_relation(irn);
531 case ir_relation_less_greater:
532 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
533 !tarval_is_null(tarval_andnot(lo, rz))) {
534 // At least one bit differs.
536 } else if (lz == lo && rz == ro && lz == rz) {
543 case ir_relation_equal:
544 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
545 !tarval_is_null(tarval_andnot(lo, rz))) {
546 // At least one bit differs.
548 } else if (lz == lo && rz == ro && lz == rz) {
555 case ir_relation_less_equal:
556 case ir_relation_less:
557 /* TODO handle negative values */
558 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
559 tarval_is_negative(rz) || tarval_is_negative(ro))
562 if (tarval_cmp(lz, ro) & relation) {
563 /* Left upper bound is smaller(/equal) than right lower bound. */
565 } else if (!(tarval_cmp(lo, rz) & relation)) {
566 /* Left lower bound is not smaller(/equal) than right upper bound. */
573 case ir_relation_greater_equal:
574 case ir_relation_greater:
575 /* TODO handle negative values */
576 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
577 tarval_is_negative(rz) || tarval_is_negative(ro))
580 if (!(tarval_cmp(lz, ro) & relation)) {
581 /* Left upper bound is not greater(/equal) than right lower bound. */
583 } else if (tarval_cmp(lo, rz) & relation) {
584 /* Left lower bound is greater(/equal) than right upper bound. */
600 DB((dbg, LEVEL_4, "cannot analyse %+F\n", irn));
602 z = get_tarval_all_one(m);
603 o = get_tarval_null(m);
612 return set_bitinfo(irn, z, o);
615 static void first_round(ir_node* const irn, void* const env)
617 pdeq* const q = (pdeq*)env;
620 if (is_Phi(irn) || is_Block(irn)) {
621 /* Only Phis (and their users) need another round, if we did not have
622 * information about all their inputs in the first round, i.e. in loops. */
623 /* TODO inserts all Phis, should only insert Phis, which did no have all
624 * predecessors available */
629 static ir_node *make_bad_block(ir_graph *irg)
631 ir_node *bad = new_r_Bad(irg, mode_BB);
632 bitinfo *bb = get_bitinfo(bad);
634 ir_tarval* const f = get_tarval_b_false();
635 ir_tarval* const t = get_tarval_b_true();
636 set_bitinfo(bad, f, t); /* Undefined. */
641 static void apply_result(ir_node* const irn, void* ctx)
643 environment_t* env = (environment_t*)ctx;
651 block_b = get_bitinfo(irn);
652 /* Trivially unreachable blocks have no info. */
653 if (block_b == NULL || block_b->z == get_tarval_b_false()) {
654 ir_node *bad = make_bad_block(get_irn_irg(irn));
661 block = get_nodes_block(irn);
662 block_b = get_bitinfo(block);
663 /* Trivially unreachable blocks have no info. */
664 if (block_b == NULL || block_b->z == get_tarval_b_false()) {
665 /* Unreachable blocks might be replaced before the nodes in them. */
666 ir_mode *mode = get_irn_mode(irn);
667 ir_graph *irg = get_irn_irg(irn);
668 ir_node *bad = new_r_Bad(irg, mode);
674 b = get_bitinfo(irn);
676 if (is_Const(irn)) return; // It cannot get any better than a Const.
680 // Only display information if we could find out anything about the value.
681 DEBUG_ONLY(if (!tarval_is_all_one(z) || !tarval_is_null(o)))
682 DB((dbg, LEVEL_2, "%+F: 0:%T 1:%T%s\n", irn, z, o, z == o ? " --- constant" : ""));
684 // Replace node with constant value by Const.
686 ir_mode* const m = get_irn_mode(irn);
688 if (mode_is_intb(m)) {
689 ir_graph *irg = get_irn_irg(irn);
690 n = new_r_Const(irg, z);
691 } else if (m == mode_X) {
692 ir_graph* const irg = get_Block_irg(block);
693 if (z == get_tarval_b_true()) {
694 // Might produce an endless loop, so keep the block.
695 add_End_keepalive(get_irg_end(irg), block);
696 n = new_r_Jmp(block);
698 n = new_r_Bad(irg, mode_X);
699 /* Transferring analysis information to the bad node makes it a
700 * candidate for replacement. */
712 switch (get_irn_opcode(irn)) {
714 ir_node* const l = get_And_left(irn);
715 ir_node* const r = get_And_right(irn);
716 bitinfo const* const bl = get_bitinfo(l);
717 bitinfo const* const br = get_bitinfo(r);
718 if (bl->z == bl->o) {
719 if (tarval_is_null(tarval_andnot(br->z, bl->z))) {
720 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
724 } else if (br->z == br->o) {
725 if (tarval_is_null(tarval_andnot(bl->z, br->z))) {
726 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
735 ir_node* const l = get_Eor_left(irn);
736 ir_node* const r = get_Eor_right(irn);
737 bitinfo const* const bl = get_bitinfo(l);
738 bitinfo const* const br = get_bitinfo(r);
739 /* if each bit is guaranteed to be zero on either the left or right
740 * then an Add will have the same effect as the Eor. Change it for
742 if (tarval_is_null(tarval_and(bl->z, br->z))) {
743 dbg_info *dbgi = get_irn_dbg_info(irn);
744 ir_node *block = get_nodes_block(irn);
745 ir_mode *mode = get_irn_mode(irn);
746 ir_node *new_node = new_rd_Add(dbgi, block, l, r, mode);
747 bitinfo const *bi = get_bitinfo(irn);
748 DB((dbg, LEVEL_2, "%+F(%+F, %+F) normalised to Add\n", irn, l, r));
749 set_bitinfo(new_node, bi->z, bi->o);
750 exchange(irn, new_node);
757 ir_node* const l = get_Or_left(irn);
758 ir_node* const r = get_Or_right(irn);
759 bitinfo const* const bl = get_bitinfo(l);
760 bitinfo const* const br = get_bitinfo(r);
761 if (bl->z == bl->o) {
762 if (tarval_is_null(tarval_andnot(bl->o, br->o))) {
763 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
767 } else if (br->z == br->o) {
768 if (tarval_is_null(tarval_andnot(br->o, bl->o))) {
769 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
775 /* if each bit is guaranteed to be zero on either the left or right
776 * then an Add will have the same effect as the Or. Change it for
778 if (tarval_is_null(tarval_and(bl->z, br->z))) {
779 dbg_info *dbgi = get_irn_dbg_info(irn);
780 ir_node *block = get_nodes_block(irn);
781 ir_mode *mode = get_irn_mode(irn);
782 ir_node *new_node = new_rd_Add(dbgi, block, l, r, mode);
783 bitinfo const *bi = get_bitinfo(irn);
784 DB((dbg, LEVEL_2, "%+F(%+F, %+F) normalised to Add\n", irn, l, r));
785 set_bitinfo(new_node, bi->z, bi->o);
786 exchange(irn, new_node);
795 static void queue_users(pdeq* const q, ir_node* const n)
797 if (get_irn_mode(n) == mode_X) {
798 /* When the state of a control flow node changes, not only queue its
799 * successor blocks, but also the Phis in these blocks, because the Phis
800 * must reconsider this input path. */
802 foreach_out_edge(n, e) {
803 ir_node* const src = get_edge_src_irn(e);
805 /* should always be a block */
808 for (phi = get_Block_phis(src); phi; phi = get_Phi_next(phi))
814 foreach_out_edge(n, e) {
815 ir_node* const src = get_edge_src_irn(e);
816 if (get_irn_mode(src) == mode_T) {
825 static void clear_links(ir_node *irn, void *env)
828 set_irn_link(irn, NULL);
830 set_Block_phis(irn, NULL);
833 static void build_phi_lists(ir_node *irn, void *env)
837 add_Block_phi(get_nodes_block(irn), irn);
840 static ir_graph_state_t do_fixpoint_vrp(ir_graph* const irg)
843 ir_graph_state_t res = 0;
845 FIRM_DBG_REGISTER(dbg, "firm.opt.fp-vrp");
846 DB((dbg, LEVEL_1, "===> Performing constant propagation on %+F\n", irg));
850 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
853 pdeq* const q = new_pdeq();
855 /* We need this extra step because the dom tree does not contain
856 * unreachable blocks in Firm. Moreover build phi list. */
857 irg_walk_anchors(irg, clear_links, build_phi_lists, NULL);
860 ir_tarval* const f = get_tarval_b_false();
861 ir_tarval* const t = get_tarval_b_true();
862 set_bitinfo(get_irg_end_block(irg), t, f); /* Reachable. */
865 /* TODO Improve iteration order. Best is reverse postorder in data flow
866 * direction and respecting loop nesting for fastest convergence. */
867 irg_walk_blkwise_dom_top_down(irg, NULL, first_round, q);
869 while (!pdeq_empty(q)) {
870 ir_node* const n = (ir_node*)pdeq_getl(q);
878 DB((dbg, LEVEL_2, "---> Applying analysis results\n"));
880 irg_walk_graph(irg, NULL, apply_result, &env);
882 if (! env.modified) {
883 res |= IR_GRAPH_STATE_CONSISTENT_DOMINANCE | IR_GRAPH_STATE_CONSISTENT_ENTITY_USAGE;
886 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
888 obstack_free(&obst, NULL);
893 static optdesc_t opt_fpvrp = {
895 IR_GRAPH_STATE_NO_BADS | IR_GRAPH_STATE_NO_UNREACHABLE_CODE | IR_GRAPH_STATE_CONSISTENT_DOMINANCE | IR_GRAPH_STATE_CONSISTENT_OUT_EDGES,
899 void fixpoint_vrp(ir_graph* const irg)
901 perform_irg_optimization(irg, &opt_fpvrp);
904 ir_graph_pass_t *fixpoint_vrp_irg_pass(const char *name)
906 return def_graph_pass(name ? name : "fixpoint_vrp", fixpoint_vrp);