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"
41 #include "iroptimize.h"
48 * - Implement cleared/set bit calculation for Add, Sub, Minus, Mul, Div, Mod, Shl, Shr, Shrs, Rotl
49 * - Implement min/max calculation for And, Eor, Or, Not, Conv, Shl, Shr, Shrs, Rotl, Mux
50 * - Implement min/max calculation for Add, Sub, Minus, Mul, Div, Mod, Conv, Shl, Shr, Shrs, Rotl, Mux
53 /* Tables of the cleared/set bit lattice
55 * Encoding of the lattice
58 * 01 - impossible state, is zero /and/ one
59 * 10 T top, may be either zero or one
73 * Half adder, half subtractor, and, xor, or, Mux
85 * Full adder, full subtractor
116 * Assume: Xmin <= Xmax and no overflow
117 * A + B = (Amin + Bmin, Amax + Bmax)
118 * -A = (-Amax, -Amin)
119 * A - B = A + -B = (Amin (-B)min, Amax + (-B)max) = (Amin - Bmax, Amax - Bmin)
122 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
124 static struct obstack obst;
126 typedef struct bitinfo
128 ir_tarval* z; // safe zeroes, 0 = bit is zero, 1 = bit maybe is 1
129 ir_tarval* o; // safe ones, 0 = bit maybe is zero, 1 = bit is 1
132 typedef struct environment_t {
133 unsigned modified:1; /**< Set, if the graph was modified. */
136 static bool is_undefined(bitinfo const* const b)
138 return tarval_is_null(b->z) && tarval_is_all_one(b->o);
141 static inline bitinfo* get_bitinfo(ir_node const* const irn)
143 return (bitinfo*)get_irn_link(irn);
146 static int set_bitinfo(ir_node* const irn, ir_tarval* const z, ir_tarval* const o)
148 bitinfo* b = get_bitinfo(irn);
150 b = OALLOCZ(&obst, bitinfo);
151 set_irn_link(irn, b);
152 } else if (z == b->z && o == b->o) {
157 DB((dbg, LEVEL_3, "%+F: 0:%T 1:%T\n", irn, z, o));
161 static int mode_is_intb(ir_mode const* const m)
163 return mode_is_int(m) || m == mode_b;
166 static int transfer(ir_node* const irn)
168 ir_tarval* const f = get_tarval_b_false();
169 ir_tarval* const t = get_tarval_b_true();
170 ir_mode* const m = get_irn_mode(irn);
175 bitinfo* const b = get_bitinfo(get_nodes_block(irn));
177 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
179 /* Unreachble blocks might have no bitinfo. */
180 if (b == NULL || b->z == f) {
184 } else switch (get_irn_opcode(irn)) {
186 ir_node* const pred = get_Proj_pred(irn);
187 if (is_Start(pred)) {
188 goto result_unknown_X;
189 } else if (is_Cond(pred)) {
190 ir_node* const selector = get_Cond_selector(pred);
191 bitinfo* const b = get_bitinfo(selector);
192 if (is_undefined(b)) {
194 } else if (get_irn_mode(selector) == mode_b) {
196 if ((b->z == t) == get_Proj_proj(irn)) {
202 goto result_unknown_X;
205 long const val = get_Proj_proj(irn);
206 if (val != get_Cond_default_proj(pred)) {
207 ir_tarval* const tv = new_tarval_from_long(val, get_irn_mode(selector));
208 if (!tarval_is_null(tarval_andnot(tv, b->z)) ||
209 !tarval_is_null(tarval_andnot(b->o, tv))) {
210 // At least one bit differs.
212 #if 0 // TODO must handle default Proj
213 } else if (b->z == b->o && b->z == tv) {
217 goto result_unknown_X;
220 goto cannot_analyse_X;
224 goto cannot_analyse_X;
230 goto result_unknown_X;
234 DB((dbg, LEVEL_4, "cannot analyse %+F\n", irn));
240 } else if (is_Block(irn)) {
242 int const arity = get_Block_n_cfgpreds(irn);
245 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
246 for (i = 0; i != arity; ++i) {
247 bitinfo* const b = get_bitinfo(get_Block_cfgpred(irn, i));
248 if (b != NULL && b->z == t) {
255 ir_graph *const irg = get_Block_irg(irn);
257 irn == get_irg_start_block(irg) ||
258 irn == get_irg_end_block(irg);
268 } else if (mode_is_intb(m)) {
269 bitinfo* const b = get_bitinfo(get_nodes_block(irn));
271 DB((dbg, LEVEL_3, "transfer %+F\n", irn));
275 z = get_tarval_null(m);
276 o = get_tarval_all_one(m);
277 } else if (is_Phi(irn)) {
278 ir_node* const block = get_nodes_block(irn);
279 int const arity = get_Phi_n_preds(irn);
282 z = get_tarval_null(m);
283 o = get_tarval_all_one(m);
284 for (i = 0; i != arity; ++i) {
285 bitinfo* const b_cfg = get_bitinfo(get_Block_cfgpred(block, i));
286 if (b_cfg != NULL && b_cfg->z != f) {
287 bitinfo* const b = get_bitinfo(get_Phi_pred(irn, i));
288 /* Only use input if it's not undefined. */
289 if (!is_undefined(b)) {
290 z = tarval_or( z, b->z);
291 o = tarval_and(o, b->o);
296 int const arity = get_irn_arity(irn);
299 /* Undefined if any input is undefined. */
300 for (i = 0; i != arity; ++i) {
301 ir_node* const pred = get_irn_n(irn, i);
302 bitinfo* const pred_b = get_bitinfo(pred);
303 if (pred_b != NULL && is_undefined(pred_b))
307 switch (get_irn_opcode(irn)) {
309 z = o = get_Const_tarval(irn);
314 ir_node* const v = get_Confirm_value(irn);
315 bitinfo* const b = get_bitinfo(v);
316 /* TODO Use bound and relation. */
323 bitinfo* const l = get_bitinfo(get_Shl_left(irn));
324 bitinfo* const r = get_bitinfo(get_Shl_right(irn));
325 ir_tarval* const rz = r->z;
327 z = tarval_shl(l->z, rz);
328 o = tarval_shl(l->o, rz);
336 bitinfo* const l = get_bitinfo(get_Shr_left(irn));
337 bitinfo* const r = get_bitinfo(get_Shr_right(irn));
338 ir_tarval* const rz = r->z;
340 z = tarval_shr(l->z, rz);
341 o = tarval_shr(l->o, rz);
349 bitinfo* const l = get_bitinfo(get_Shrs_left(irn));
350 bitinfo* const r = get_bitinfo(get_Shrs_right(irn));
351 ir_tarval* const rz = r->z;
353 z = tarval_shrs(l->z, rz);
354 o = tarval_shrs(l->o, rz);
362 bitinfo* const l = get_bitinfo(get_Rotl_left(irn));
363 bitinfo* const r = get_bitinfo(get_Rotl_right(irn));
364 ir_tarval* const rz = r->z;
366 z = tarval_rotl(l->z, rz);
367 o = tarval_rotl(l->o, rz);
375 bitinfo* const l = get_bitinfo(get_Add_left(irn));
376 bitinfo* const r = get_bitinfo(get_Add_right(irn));
377 ir_tarval* const lz = l->z;
378 ir_tarval* const lo = l->o;
379 ir_tarval* const rz = r->z;
380 ir_tarval* const ro = r->o;
381 if (lz == lo && rz == ro) {
382 z = o = tarval_add(lz, rz);
384 // TODO improve: can only do lower disjoint bits
385 /* Determine where any of the operands has zero bits, i.e. where no
386 * carry out is generated if there is not carry in */
387 ir_tarval* const no_c_in_no_c_out = tarval_and(lz, rz);
388 /* Generate a mask of the lower consecutive zeroes: x | -x. In this
389 * range the addition is disjoint and therefore Add behaves like Or.
391 ir_tarval* const low_zero_mask = tarval_or(no_c_in_no_c_out, tarval_neg(no_c_in_no_c_out));
392 ir_tarval* const low_one_mask = tarval_not(low_zero_mask);
393 z = tarval_or( tarval_or(lz, rz), low_zero_mask);
394 o = tarval_and(tarval_or(lo, ro), low_one_mask);
400 bitinfo* const l = get_bitinfo(get_Sub_left(irn));
401 bitinfo* const r = get_bitinfo(get_Sub_right(irn));
402 if (l != NULL && r != NULL) { // Sub might subtract pointers.
403 ir_tarval* const lz = l->z;
404 ir_tarval* const lo = l->o;
405 ir_tarval* const rz = r->z;
406 ir_tarval* const ro = r->o;
407 if (lz == lo && rz == ro) {
408 z = o = tarval_sub(lz, rz, NULL);
409 } else if (tarval_is_null(tarval_andnot(rz, lo))) {
410 /* Every possible one of the subtrahend is backed by a safe one of the
411 * minuend, i.e. there are no borrows. */
412 // TODO extend no-borrow like carry for Add above
413 z = tarval_andnot(lz, ro);
414 o = tarval_andnot(lo, rz);
425 bitinfo* const l = get_bitinfo(get_Mul_left(irn));
426 bitinfo* const r = get_bitinfo(get_Mul_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 if (lz == lo && rz == ro) {
432 z = o = tarval_mul(lz, rz);
435 // Determine safe lower zeroes: x | -x.
436 ir_tarval* const lzn = tarval_or(lz, tarval_neg(lz));
437 ir_tarval* const rzn = tarval_or(rz, tarval_neg(rz));
438 // Concatenate safe lower zeroes.
439 if (tarval_cmp(lzn, rzn) == ir_relation_less) {
440 z = tarval_mul(tarval_eor(lzn, tarval_shl(lzn, get_tarval_one(m))), rzn);
442 z = tarval_mul(tarval_eor(rzn, tarval_shl(rzn, get_tarval_one(m))), lzn);
444 o = get_tarval_null(m);
450 bitinfo* const b = get_bitinfo(get_Minus_op(irn));
452 z = o = tarval_neg(b->z);
460 bitinfo* const l = get_bitinfo(get_And_left(irn));
461 bitinfo* const r = get_bitinfo(get_And_right(irn));
462 z = tarval_and(l->z, r->z);
463 o = tarval_and(l->o, r->o);
468 bitinfo* const l = get_bitinfo(get_Or_left(irn));
469 bitinfo* const r = get_bitinfo(get_Or_right(irn));
470 z = tarval_or(l->z, r->z);
471 o = tarval_or(l->o, r->o);
476 bitinfo* const l = get_bitinfo(get_Eor_left(irn));
477 bitinfo* const r = get_bitinfo(get_Eor_right(irn));
478 ir_tarval* const lz = l->z;
479 ir_tarval* const lo = l->o;
480 ir_tarval* const rz = r->z;
481 ir_tarval* const ro = r->o;
482 z = tarval_or(tarval_andnot(lz, ro), tarval_andnot(rz, lo));
483 o = tarval_or(tarval_andnot(ro, lz), tarval_andnot(lo, rz));
488 bitinfo* const b = get_bitinfo(get_Not_op(irn));
489 z = tarval_not(b->o);
490 o = tarval_not(b->z);
495 bitinfo* const b = get_bitinfo(get_Conv_op(irn));
496 if (b == NULL) // Happens when converting from float values.
498 z = tarval_convert_to(b->z, m);
499 o = tarval_convert_to(b->o, m);
504 bitinfo* const bf = get_bitinfo(get_Mux_false(irn));
505 bitinfo* const bt = get_bitinfo(get_Mux_true(irn));
506 bitinfo* const c = get_bitinfo(get_Mux_sel(irn));
510 } else if (c->z == f) {
514 z = tarval_or( bf->z, bt->z);
515 o = tarval_and(bf->o, bt->o);
521 bitinfo* const l = get_bitinfo(get_Cmp_left(irn));
522 bitinfo* const r = get_bitinfo(get_Cmp_right(irn));
523 if (l == NULL || r == NULL) {
524 goto result_unknown; // Cmp compares something we cannot evaluate.
526 ir_tarval* const lz = l->z;
527 ir_tarval* const lo = l->o;
528 ir_tarval* const rz = r->z;
529 ir_tarval* const ro = r->o;
530 ir_relation const relation = get_Cmp_relation(irn);
532 case ir_relation_less_greater:
533 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
534 !tarval_is_null(tarval_andnot(lo, rz))) {
535 // At least one bit differs.
537 } else if (lz == lo && rz == ro && lz == rz) {
544 case ir_relation_equal:
545 if (!tarval_is_null(tarval_andnot(ro, lz)) ||
546 !tarval_is_null(tarval_andnot(lo, rz))) {
547 // At least one bit differs.
549 } else if (lz == lo && rz == ro && lz == rz) {
556 case ir_relation_less_equal:
557 case ir_relation_less:
558 /* TODO handle negative values */
559 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
560 tarval_is_negative(rz) || tarval_is_negative(ro))
563 if (tarval_cmp(lz, ro) & relation) {
564 /* Left upper bound is smaller(/equal) than right lower bound. */
566 } else if (!(tarval_cmp(lo, rz) & relation)) {
567 /* Left lower bound is not smaller(/equal) than right upper bound. */
574 case ir_relation_greater_equal:
575 case ir_relation_greater:
576 /* TODO handle negative values */
577 if (tarval_is_negative(lz) || tarval_is_negative(lo) ||
578 tarval_is_negative(rz) || tarval_is_negative(ro))
581 if (!(tarval_cmp(lz, ro) & relation)) {
582 /* Left upper bound is not greater(/equal) than right lower bound. */
584 } else if (tarval_cmp(lo, rz) & relation) {
585 /* Left lower bound is greater(/equal) than right upper bound. */
601 DB((dbg, LEVEL_4, "cannot analyse %+F\n", irn));
603 z = get_tarval_all_one(m);
604 o = get_tarval_null(m);
613 return set_bitinfo(irn, z, o);
616 static void first_round(ir_node* const irn, void* const env)
618 pdeq* const q = (pdeq*)env;
621 if (is_Phi(irn) || is_Block(irn)) {
622 /* Only Phis (and their users) need another round, if we did not have
623 * information about all their inputs in the first round, i.e. in loops. */
624 /* TODO inserts all Phis, should only insert Phis, which did no have all
625 * predecessors available */
630 static void apply_result(ir_node* const irn, void* ctx)
632 environment_t* env = (environment_t*)ctx;
640 block_b = get_bitinfo(irn);
641 /* Trivially unreachable blocks have no info. */
642 if (block_b == NULL || block_b->z == get_tarval_b_false()) {
643 exchange(irn, get_irg_bad(get_Block_irg(irn)));
649 block = get_nodes_block(irn);
650 block_b = get_bitinfo(block);
651 /* Trivially unreachable blocks have no info. */
652 if (block_b == NULL || block_b->z == get_tarval_b_false()) {
653 /* Unreachable blocks might be replaced before the nodes in them. */
654 exchange(irn, is_Bad(block) ? block : get_irg_bad(get_Block_irg(block)));
659 b = get_bitinfo(irn);
661 if (is_Const(irn)) return; // It cannot get any better than a Const.
665 // Only display information if we could find out anything about the value.
666 DEBUG_ONLY(if (!tarval_is_all_one(z) || !tarval_is_null(o)))
667 DB((dbg, LEVEL_2, "%+F: 0:%T 1:%T%s\n", irn, z, o, z == o ? " --- constant" : ""));
669 // Replace node with constant value by Const.
671 ir_mode* const m = get_irn_mode(irn);
673 if (mode_is_intb(m)) {
674 ir_graph *irg = get_irn_irg(irn);
675 n = new_r_Const(irg, z);
676 } else if (m == mode_X) {
677 ir_graph* const irg = get_Block_irg(block);
678 if (z == get_tarval_b_true()) {
679 // Might produce an endless loop, so keep the block.
680 add_End_keepalive(get_irg_end(irg), block);
681 n = new_r_Jmp(block);
684 /* Transferring analysis information to the bad node makes it a
685 * candidate for replacement. */
697 switch (get_irn_opcode(irn)) {
699 ir_node* const l = get_And_left(irn);
700 ir_node* const r = get_And_right(irn);
701 bitinfo const* const bl = get_bitinfo(l);
702 bitinfo const* const br = get_bitinfo(r);
703 if (bl->z == bl->o) {
704 if (tarval_is_null(tarval_andnot(br->z, bl->z))) {
705 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
709 } else if (br->z == br->o) {
710 if (tarval_is_null(tarval_andnot(bl->z, br->z))) {
711 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
720 ir_node* const l = get_Or_left(irn);
721 ir_node* const r = get_Or_right(irn);
722 bitinfo const* const bl = get_bitinfo(l);
723 bitinfo const* const br = get_bitinfo(r);
724 if (bl->z == bl->o) {
725 if (tarval_is_null(tarval_andnot(bl->o, br->o))) {
726 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
730 } else if (br->z == br->o) {
731 if (tarval_is_null(tarval_andnot(br->o, bl->o))) {
732 DB((dbg, LEVEL_2, "%+F(%+F, %+F) is superfluous\n", irn, l, r));
742 static void queue_users(pdeq* const q, ir_node* const n)
744 if (get_irn_mode(n) == mode_X) {
745 /* When the state of a control flow node changes, not only queue its
746 * successor blocks, but also the Phis in these blocks, because the Phis
747 * must reconsider this input path. */
749 foreach_out_edge(n, e) {
750 ir_node* const src = get_edge_src_irn(e);
752 /* should always be a block */
755 for (phi = get_Block_phis(src); phi; phi = get_Phi_next(phi))
761 foreach_out_edge(n, e) {
762 ir_node* const src = get_edge_src_irn(e);
763 if (get_irn_mode(src) == mode_T) {
772 static void clear_links(ir_node *irn, void *env)
775 set_irn_link(irn, NULL);
777 set_Block_phis(irn, NULL);
780 static void build_phi_lists(ir_node *irn, void *env)
784 add_Block_phi(get_nodes_block(irn), irn);
787 void fixpoint_vrp(ir_graph* const irg)
791 FIRM_DBG_REGISTER(dbg, "firm.opt.fp-vrp");
792 DB((dbg, LEVEL_1, "===> Performing constant propagation on %+F\n", irg));
796 /* HACK: to avoid finding dead code */
797 edges_deactivate(irg);
803 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
806 pdeq* const q = new_pdeq();
808 /* We need this extra step because the dom tree does not contain unreachable
809 blocks in Firm. Moreover build phi list. */
810 irg_walk_anchors(irg, clear_links, build_phi_lists, NULL);
812 { ir_tarval* const f = get_tarval_b_false();
813 ir_tarval* const t = get_tarval_b_true();
814 set_bitinfo(get_irg_bad(irg), f, t); /* Undefined. */
815 set_bitinfo(get_irg_end_block(irg), t, f); /* Reachable. */
818 /* TODO Improve iteration order. Best is reverse postorder in data flow
819 * direction and respecting loop nesting for fastest convergence. */
820 irg_walk_blkwise_dom_top_down(irg, NULL, first_round, q);
822 while (!pdeq_empty(q)) {
823 ir_node* const n = (ir_node*)pdeq_getl(q);
831 DB((dbg, LEVEL_2, "---> Applying analysis results\n"));
833 irg_walk_graph(irg, NULL, apply_result, &env);
836 /* control flow might changed */
837 set_irg_outs_inconsistent(irg);
838 set_irg_extblk_inconsistent(irg);
839 set_irg_doms_inconsistent(irg);
840 set_irg_loopinfo_inconsistent(irg);
841 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
844 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
846 obstack_free(&obst, NULL);
849 ir_graph_pass_t *fixpoint_vrp_irg_pass(const char *name)
851 return def_graph_pass(name ? name : "fixpoint_vrp", fixpoint_vrp);