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 This file contains functions for matching firm graphs for
23 * nodes that can be used as address mode for x86 instructions
24 * @author Matthias Braun
29 #include "ia32_address_mode.h"
30 #include "ia32_transform.h"
36 #include "iredges_t.h"
39 #include "../benode.h"
40 #include "../belive.h"
44 /* gas/ld don't support negative symconsts :-( */
45 #undef SUPPORT_NEGATIVE_SYMCONSTS
47 static bitset_t *non_address_mode_nodes;
50 * Recursive worker for checking if a DAG with root node can be represented as a simple immediate,
52 * @param node the node
53 * @param symconsts number of symconsts found so far
54 * @param negate if set, the immediate must be negated
56 * @return non-zero if the DAG represents an immediate, 0 else
58 static bool do_is_immediate(const ir_node *node, int *symconsts, bool negate)
63 switch (get_irn_opcode(node)) {
65 /* Consts are typically immediates */
66 if (!tarval_is_long(get_Const_tarval(node))) {
69 "Optimisation warning tarval of %+F(%+F) is not a long.\n",
70 node, current_ir_graph);
76 /* the first SymConst of a DAG can be fold into an immediate */
77 #ifndef SUPPORT_NEGATIVE_SYMCONSTS
78 /* unfortunately the assembler/linker doesn't support -symconst */
82 if (get_SymConst_kind(node) != symconst_addr_ent)
89 /* we can use '0' for Unknowns */
93 /* Add's and Sub's are typically supported as long as both operands are
95 if (ia32_is_non_address_mode_node(node))
98 left = get_binop_left(node);
99 if (!do_is_immediate(left, symconsts, negate))
101 right = get_binop_right(node);
102 if (!do_is_immediate(right, symconsts, is_Sub(node) ? !negate : negate))
107 /* all other nodes are NO immediates */
113 * Check if a DAG starting with root node can be folded into an address mode
116 * @param addr the address mode data so far
117 * @param node the node
118 * @param negate if set, the immediate must be negated
120 static int is_immediate(ia32_address_t *addr, const ir_node *node, bool negate)
122 int symconsts = (addr->symconst_ent != NULL);
123 return do_is_immediate(node, &symconsts, negate);
127 * Place a DAG with root node into an address mode.
129 * @param addr the address mode data so far
130 * @param node the node
131 * @param negate if set, the immediate must be negated
133 static void eat_immediate(ia32_address_t *addr, ir_node *node, bool negate)
140 switch (get_irn_opcode(node)) {
142 /* simply add the value to the offset */
143 tv = get_Const_tarval(node);
144 val = get_tarval_long(tv);
152 /* place the entity into the symconst */
153 if (addr->symconst_ent != NULL) {
154 panic("Internal error: more than 1 symconst in address calculation");
156 addr->symconst_ent = get_SymConst_entity(node);
157 if (get_entity_owner(addr->symconst_ent) == get_tls_type())
158 addr->tls_segment = true;
159 #ifndef SUPPORT_NEGATIVE_SYMCONSTS
162 addr->symconst_sign = negate;
167 assert(!ia32_is_non_address_mode_node(node));
168 left = get_Add_left(node);
169 eat_immediate(addr, left, negate);
170 right = get_Add_right(node);
171 eat_immediate(addr, right, negate);
174 assert(!ia32_is_non_address_mode_node(node));
175 left = get_Sub_left(node);
176 eat_immediate(addr, left, negate);
177 right = get_Sub_right(node);
178 eat_immediate(addr, right, !negate);
181 panic("Internal error in immediate address calculation");
186 * Place operands of node into an address mode.
188 * @param addr the address mode data so far
189 * @param node the node
190 * @param flags the flags
192 * @return the folded node
194 static ir_node *eat_immediates(ia32_address_t *addr, ir_node *node,
195 ia32_create_am_flags_t flags)
197 if (!(flags & ia32_create_am_force) &&
198 ia32_is_non_address_mode_node(node) &&
199 (!(flags & ia32_create_am_double_use) || get_irn_n_edges(node) > 2))
203 ir_node *left = get_Add_left(node);
204 ir_node *right = get_Add_right(node);
206 if (is_immediate(addr, left, 0)) {
207 eat_immediate(addr, left, 0);
208 return eat_immediates(addr, right, ia32_create_am_normal);
210 if (is_immediate(addr, right, 0)) {
211 eat_immediate(addr, right, 0);
212 return eat_immediates(addr, left, ia32_create_am_normal);
214 } else if (is_Sub(node)) {
215 ir_node *left = get_Sub_left(node);
216 ir_node *right = get_Sub_right(node);
218 if (is_immediate(addr, right, 1)) {
219 eat_immediate(addr, right, 1);
220 return eat_immediates(addr, left, ia32_create_am_normal);
228 * Try to place a Shl into an address mode.
230 * @param addr the address mode data so far
231 * @param node the node to place
233 * @return non-zero on success
235 static int eat_shl(ia32_address_t *addr, ir_node *node)
237 ir_node *shifted_val;
241 ir_node *right = get_Shl_right(node);
244 /* we can use shl with 1, 2 or 3 shift */
245 if (!is_Const(right))
247 tv = get_Const_tarval(right);
248 if (!tarval_is_long(tv))
251 val = get_tarval_long(tv);
252 if (val < 0 || val > 3)
255 ir_fprintf(stderr, "Optimisation warning: unoptimized Shl(,0) found\n");
258 shifted_val = get_Shl_left(node);
259 } else if (is_Add(node)) {
260 /* might be an add x, x */
261 ir_node *left = get_Add_left(node);
262 ir_node *right = get_Add_right(node);
275 /* we can only eat a shl if we don't have a scale or index set yet */
276 if (addr->scale != 0 || addr->index != NULL)
278 if (ia32_is_non_address_mode_node(node))
281 #ifndef AGGRESSIVE_AM
282 if (get_irn_n_edges(node) > 1)
287 addr->index = shifted_val;
291 /* Create an address mode for a given node. */
292 void ia32_create_address_mode(ia32_address_t *addr, ir_node *node, ia32_create_am_flags_t flags)
296 if (is_immediate(addr, node, 0)) {
297 eat_immediate(addr, node, 0);
301 #ifndef AGGRESSIVE_AM
302 if (!(flags & ia32_create_am_force) && get_irn_n_edges(node) > 1) {
308 if (!(flags & ia32_create_am_force) &&
309 ia32_is_non_address_mode_node(node) &&
310 (!(flags & ia32_create_am_double_use) || get_irn_n_edges(node) > 2)) {
315 eat_imms = eat_immediates(addr, node, flags);
316 if (eat_imms != node) {
317 if (flags & ia32_create_am_force) {
318 eat_imms = ia32_skip_downconv(eat_imms);
322 #ifndef AGGRESSIVE_AM
323 if (get_irn_n_edges(node) > 1) {
328 if (ia32_is_non_address_mode_node(node)) {
334 /* starting point Add, Sub or Shl, FrameAddr */
336 /* We don't want to eat add x, x as shl here, so only test for real Shl
337 * instructions, because we want the former as Lea x, x, not Shl x, 1 */
338 if (eat_shl(addr, node))
340 } else if (is_immediate(addr, node, 0)) {
341 eat_immediate(addr, node, 0);
343 } else if (be_is_FrameAddr(node)) {
344 assert(addr->base == NULL);
345 assert(addr->frame_entity == NULL);
346 addr->base = be_get_FrameAddr_frame(node);
348 addr->frame_entity = be_get_FrameAddr_entity(node);
350 } else if (is_Add(node)) {
351 ir_node *left = get_Add_left(node);
352 ir_node *right = get_Add_right(node);
354 if (flags & ia32_create_am_force) {
355 left = ia32_skip_downconv(left);
356 right = ia32_skip_downconv(right);
359 assert(flags & ia32_create_am_force || !is_immediate(addr, left, 0));
360 assert(flags & ia32_create_am_force || !is_immediate(addr, right, 0));
362 if (eat_shl(addr, left)) {
364 } else if (eat_shl(addr, right)) {
368 be_is_FrameAddr(left) &&
369 !ia32_is_non_address_mode_node(left)) {
370 assert(addr->base == NULL);
371 assert(addr->frame_entity == NULL);
372 addr->base = be_get_FrameAddr_frame(left);
374 addr->frame_entity = be_get_FrameAddr_entity(left);
376 } else if (right != NULL &&
377 be_is_FrameAddr(right) &&
378 !ia32_is_non_address_mode_node(right)) {
379 assert(addr->base == NULL);
380 assert(addr->frame_entity == NULL);
381 addr->base = be_get_FrameAddr_frame(right);
383 addr->frame_entity = be_get_FrameAddr_entity(right);
388 if (addr->base != NULL) {
389 assert(addr->index == NULL && addr->scale == 0);
390 assert(right == NULL);
397 if (addr->base == NULL) {
400 assert(addr->index == NULL && addr->scale == 0);
410 void ia32_mark_non_am(ir_node *node)
412 bitset_set(non_address_mode_nodes, get_irn_idx(node));
415 int ia32_is_non_address_mode_node(ir_node const *node)
417 return bitset_is_set(non_address_mode_nodes, get_irn_idx(node));
421 * Check if a given value is last used (i.e. die after) the block of some
424 static int value_last_used_here(be_lv_t *lv, ir_node *here, ir_node *value)
426 ir_node *block = get_nodes_block(here);
427 const ir_edge_t *edge;
429 /* If the value is live end it is for sure it does not die here */
430 if (be_is_live_end(lv, block, value)) return 0;
432 /* if multiple nodes in this block use the value, then we cannot decide
433 * whether the value will die here (because there is no schedule yet).
434 * Assume it does not die in this case. */
435 foreach_out_edge(value, edge) {
436 ir_node *user = get_edge_src_irn(edge);
437 if (user != here && get_nodes_block(user) == block) {
445 static bool simple_is_immediate(const ir_node *node)
448 return do_is_immediate(node, &symconsts, false);
452 * Walker: mark those nodes that cannot be part of an address mode because
453 * their value must be accessed through a register
455 static void mark_non_address_nodes(ir_node *node, void *env)
457 be_lv_t *lv = (be_lv_t*)env;
465 mode = get_irn_mode(node);
466 if (!mode_is_int(mode) && !mode_is_reference(mode) && mode != mode_b)
469 switch (get_irn_opcode(node)) {
471 /* Nothing to do. especially do not mark the pointer, because we want to
472 * turn it into AM. */
476 /* Do not mark the pointer, because we want to turn it into AM. */
477 val = get_Store_value(node);
478 ia32_mark_non_am(val);
483 /* only 1 user: AM folding is always beneficial */
484 if (get_irn_n_edges(node) <= 1)
487 /* for adds and shls with multiple users we use this heuristic:
488 * we do not fold them into address mode if their operands don't live
489 * out of the block, because in this case we will reduce register
490 * pressure. Otherwise we fold them in aggressively in the hope, that
491 * the node itself doesn't exist anymore and we were able to save the
492 * register for the result */
493 left = get_binop_left(node);
494 right = get_binop_right(node);
496 /* if any of the operands is an immediate then this will not
497 * increase register pressure */
498 if (simple_is_immediate(left) || simple_is_immediate(right))
501 /* Fold AM if any of the two operands does not die here. This duplicates
502 * an addition and has the same register pressure for the case that only
503 * one operand dies, but is faster (on Pentium 4).
504 * && instead of || only folds AM if both operands do not die here */
505 if (!value_last_used_here(lv, node, left) ||
506 !value_last_used_here(lv, node, right)) {
510 /* At least one of left and right are not used by anyone else, so it is
511 * beneficial for the register pressure (if both are unused otherwise,
512 * else neutral) and ALU use to not fold AM. */
513 ia32_mark_non_am(node);
517 arity = get_irn_arity(node);
519 for (i = 0; i < arity; ++i) {
520 ir_node *in = get_irn_n(node, i);
521 ia32_mark_non_am(in);
527 void ia32_calculate_non_address_mode_nodes(ir_graph *irg)
529 be_lv_t *lv = be_assure_liveness(irg);
531 non_address_mode_nodes = bitset_malloc(get_irg_last_idx(irg));
533 irg_walk_graph(irg, NULL, mark_non_address_nodes, lv);
536 void ia32_free_non_address_mode_nodes(void)
538 bitset_free(non_address_mode_nodes);