2 * Copyright (C) 1995-2008 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 Implements several optimizations for IA32.
23 * @author Matthias Braun, Christian Wuerdig
32 #include "firm_types.h"
45 #include "../benode.h"
46 #include "../besched.h"
47 #include "../bepeephole.h"
49 #include "ia32_new_nodes.h"
50 #include "ia32_optimize.h"
51 #include "bearch_ia32_t.h"
52 #include "gen_ia32_regalloc_if.h"
53 #include "ia32_common_transform.h"
54 #include "ia32_transform.h"
55 #include "ia32_dbg_stat.h"
56 #include "ia32_architecture.h"
58 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
60 static void copy_mark(const ir_node *old, ir_node *newn)
62 if (is_ia32_is_reload(old))
63 set_ia32_is_reload(newn);
64 if (is_ia32_is_spill(old))
65 set_ia32_is_spill(newn);
66 if (is_ia32_is_remat(old))
67 set_ia32_is_remat(newn);
70 typedef enum produces_flag_t {
77 * Return which usable flag the given node produces
79 * @param node the node to check
80 * @param pn the projection number of the used result
82 static produces_flag_t produces_test_flag(ir_node *node, int pn)
85 const ia32_immediate_attr_t *imm_attr;
87 if (!is_ia32_irn(node))
88 return produces_no_flag;
90 switch (get_ia32_irn_opcode(node)) {
105 assert((int)n_ia32_ShlD_count == (int)n_ia32_ShrD_count);
106 count = get_irn_n(node, n_ia32_ShlD_count);
107 goto check_shift_amount;
112 assert((int)n_ia32_Shl_count == (int)n_ia32_Shr_count
113 && (int)n_ia32_Shl_count == (int)n_ia32_Sar_count);
114 count = get_irn_n(node, n_ia32_Shl_count);
116 /* when shift count is zero the flags are not affected, so we can only
117 * do this for constants != 0 */
118 if (!is_ia32_Immediate(count))
119 return produces_no_flag;
121 imm_attr = get_ia32_immediate_attr_const(count);
122 if (imm_attr->symconst != NULL)
123 return produces_no_flag;
124 if ((imm_attr->offset & 0x1f) == 0)
125 return produces_no_flag;
129 return pn == pn_ia32_Mul_res_high ?
130 produces_flag_carry : produces_no_flag;
133 return produces_no_flag;
136 return pn == pn_ia32_res ?
137 produces_flag_zero : produces_no_flag;
141 * Replace Cmp(x, 0) by a Test(x, x)
143 static void peephole_ia32_Cmp(ir_node *const node)
147 ia32_immediate_attr_t const *imm;
153 ia32_attr_t const *attr;
156 arch_register_t const *reg;
157 ir_edge_t const *edge;
158 ir_edge_t const *tmp;
160 if (get_ia32_op_type(node) != ia32_Normal)
163 right = get_irn_n(node, n_ia32_Cmp_right);
164 if (!is_ia32_Immediate(right))
167 imm = get_ia32_immediate_attr_const(right);
168 if (imm->symconst != NULL || imm->offset != 0)
171 dbgi = get_irn_dbg_info(node);
172 irg = get_irn_irg(node);
173 block = get_nodes_block(node);
174 noreg = ia32_new_NoReg_gp(irg);
175 nomem = get_irg_no_mem(current_ir_graph);
176 op = get_irn_n(node, n_ia32_Cmp_left);
177 attr = get_ia32_attr(node);
178 ins_permuted = attr->data.ins_permuted;
180 if (is_ia32_Cmp(node)) {
181 test = new_bd_ia32_Test(dbgi, block, noreg, noreg, nomem,
182 op, op, ins_permuted);
184 test = new_bd_ia32_Test8Bit(dbgi, block, noreg, noreg, nomem,
185 op, op, ins_permuted);
187 set_ia32_ls_mode(test, get_ia32_ls_mode(node));
189 reg = arch_irn_get_register(node, pn_ia32_Cmp_eflags);
190 arch_irn_set_register(test, pn_ia32_Test_eflags, reg);
192 foreach_out_edge_safe(node, edge, tmp) {
193 ir_node *const user = get_edge_src_irn(edge);
196 exchange(user, test);
199 sched_add_before(node, test);
200 copy_mark(node, test);
201 be_peephole_exchange(node, test);
205 * Peephole optimization for Test instructions.
206 * - Remove the Test, if an appropriate flag was produced which is still live
207 * - Change a Test(x, c) to 8Bit, if 0 <= c < 256 (3 byte shorter opcode)
209 static void peephole_ia32_Test(ir_node *node)
211 ir_node *left = get_irn_n(node, n_ia32_Test_left);
212 ir_node *right = get_irn_n(node, n_ia32_Test_right);
214 assert((int)n_ia32_Test_left == (int)n_ia32_Test8Bit_left
215 && (int)n_ia32_Test_right == (int)n_ia32_Test8Bit_right);
217 if (left == right) { /* we need a test for 0 */
218 ir_node *block = get_nodes_block(node);
219 int pn = pn_ia32_res;
224 const ir_edge_t *edge;
226 if (get_nodes_block(left) != block)
230 pn = get_Proj_proj(op);
231 op = get_Proj_pred(op);
234 /* walk schedule up and abort when we find left or some other node
235 * destroys the flags */
238 schedpoint = sched_prev(schedpoint);
239 if (schedpoint == op)
241 if (arch_irn_is(schedpoint, modify_flags))
243 if (schedpoint == block)
244 panic("couldn't find left");
247 /* make sure only Lg/Eq tests are used */
248 foreach_out_edge(node, edge) {
249 ir_node *user = get_edge_src_irn(edge);
250 ia32_condition_code_t cc = get_ia32_condcode(user);
252 if (cc != ia32_cc_equal && cc != ia32_cc_not_equal) {
257 switch (produces_test_flag(op, pn)) {
258 case produces_flag_zero:
261 case produces_flag_carry:
262 foreach_out_edge(node, edge) {
263 ir_node *user = get_edge_src_irn(edge);
264 ia32_condition_code_t cc = get_ia32_condcode(user);
267 case ia32_cc_equal: cc = ia32_cc_above_equal; break; /* CF = 0 */
268 case ia32_cc_not_equal: cc = ia32_cc_below; break; /* CF = 1 */
269 default: panic("unexpected pn");
271 set_ia32_condcode(user, cc);
279 if (get_irn_mode(op) != mode_T) {
280 set_irn_mode(op, mode_T);
282 /* If there are other users, reroute them to result proj */
283 if (get_irn_n_edges(op) != 2) {
284 ir_node *res = new_r_Proj(op, mode_Iu, pn_ia32_res);
286 edges_reroute(op, res);
287 /* Reattach the result proj to left */
288 set_Proj_pred(res, op);
291 if (get_irn_n_edges(left) == 2)
295 flags_mode = ia32_reg_classes[CLASS_ia32_flags].mode;
296 flags_proj = new_r_Proj(op, flags_mode, pn_ia32_flags);
297 arch_set_irn_register(flags_proj, &ia32_registers[REG_EFLAGS]);
299 assert(get_irn_mode(node) != mode_T);
301 be_peephole_exchange(node, flags_proj);
302 } else if (is_ia32_Immediate(right)) {
303 ia32_immediate_attr_t const *const imm = get_ia32_immediate_attr_const(right);
306 /* A test with a symconst is rather strange, but better safe than sorry */
307 if (imm->symconst != NULL)
310 offset = imm->offset;
311 if (get_ia32_op_type(node) == ia32_AddrModeS) {
312 ia32_attr_t *const attr = get_ia32_attr(node);
314 if ((offset & 0xFFFFFF00) == 0) {
315 /* attr->am_offs += 0; */
316 } else if ((offset & 0xFFFF00FF) == 0) {
317 ir_node *imm = ia32_create_Immediate(NULL, 0, offset >> 8);
318 set_irn_n(node, n_ia32_Test_right, imm);
320 } else if ((offset & 0xFF00FFFF) == 0) {
321 ir_node *imm = ia32_create_Immediate(NULL, 0, offset >> 16);
322 set_irn_n(node, n_ia32_Test_right, imm);
324 } else if ((offset & 0x00FFFFFF) == 0) {
325 ir_node *imm = ia32_create_Immediate(NULL, 0, offset >> 24);
326 set_irn_n(node, n_ia32_Test_right, imm);
331 } else if (offset < 256) {
332 arch_register_t const* const reg = arch_get_irn_register(left);
334 if (reg != &ia32_registers[REG_EAX] &&
335 reg != &ia32_registers[REG_EBX] &&
336 reg != &ia32_registers[REG_ECX] &&
337 reg != &ia32_registers[REG_EDX]) {
344 /* Technically we should build a Test8Bit because of the register
345 * constraints, but nobody changes registers at this point anymore. */
346 set_ia32_ls_mode(node, mode_Bu);
351 * AMD Athlon works faster when RET is not destination of
352 * conditional jump or directly preceded by other jump instruction.
353 * Can be avoided by placing a Rep prefix before the return.
355 static void peephole_ia32_Return(ir_node *node)
359 if (!ia32_cg_config.use_pad_return)
362 /* check if this return is the first on the block */
363 sched_foreach_reverse_from(node, irn) {
364 switch (get_irn_opcode(irn)) {
366 /* the return node itself, ignore */
370 /* ignore no code generated */
373 /* arg, IncSP 0 nodes might occur, ignore these */
374 if (be_get_IncSP_offset(irn) == 0)
384 /* ensure, that the 3 byte return is generated */
385 be_Return_set_emit_pop(node, 1);
388 /* only optimize up to 48 stores behind IncSPs */
389 #define MAXPUSH_OPTIMIZE 48
392 * Tries to create Push's from IncSP, Store combinations.
393 * The Stores are replaced by Push's, the IncSP is modified
394 * (possibly into IncSP 0, but not removed).
396 static void peephole_IncSP_Store_to_push(ir_node *irn)
402 ir_node *stores[MAXPUSH_OPTIMIZE];
407 ir_node *first_push = NULL;
408 ir_edge_t const *edge;
409 ir_edge_t const *next;
411 memset(stores, 0, sizeof(stores));
413 assert(be_is_IncSP(irn));
415 inc_ofs = be_get_IncSP_offset(irn);
420 * We first walk the schedule after the IncSP node as long as we find
421 * suitable Stores that could be transformed to a Push.
422 * We save them into the stores array which is sorted by the frame offset/4
423 * attached to the node
426 for (node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) {
431 /* it has to be a Store */
432 if (!is_ia32_Store(node))
435 /* it has to use our sp value */
436 if (get_irn_n(node, n_ia32_base) != irn)
438 /* Store has to be attached to NoMem */
439 mem = get_irn_n(node, n_ia32_mem);
443 /* unfortunately we can't support the full AMs possible for push at the
444 * moment. TODO: fix this */
445 if (!is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
448 offset = get_ia32_am_offs_int(node);
449 /* we should NEVER access uninitialized stack BELOW the current SP */
452 /* storing at half-slots is bad */
453 if ((offset & 3) != 0)
456 if (inc_ofs - 4 < offset || offset >= MAXPUSH_OPTIMIZE * 4)
458 storeslot = offset >> 2;
460 /* storing into the same slot twice is bad (and shouldn't happen...) */
461 if (stores[storeslot] != NULL)
464 stores[storeslot] = node;
465 if (storeslot > maxslot)
471 for (i = -1; i < maxslot; ++i) {
472 if (stores[i + 1] == NULL)
476 /* walk through the Stores and create Pushs for them */
477 block = get_nodes_block(irn);
478 spmode = get_irn_mode(irn);
479 irg = get_irn_irg(irn);
480 for (; i >= 0; --i) {
481 const arch_register_t *spreg;
483 ir_node *val, *mem, *mem_proj;
484 ir_node *store = stores[i];
485 ir_node *noreg = ia32_new_NoReg_gp(irg);
486 const ir_edge_t *edge;
487 const ir_edge_t *next;
489 val = get_irn_n(store, n_ia32_unary_op);
490 mem = get_irn_n(store, n_ia32_mem);
491 spreg = arch_get_irn_register(curr_sp);
493 push = new_bd_ia32_Push(get_irn_dbg_info(store), block, noreg, noreg,
495 copy_mark(store, push);
497 if (first_push == NULL)
500 sched_add_after(skip_Proj(curr_sp), push);
502 /* create stackpointer Proj */
503 curr_sp = new_r_Proj(push, spmode, pn_ia32_Push_stack);
504 arch_set_irn_register(curr_sp, spreg);
506 /* create memory Proj */
507 mem_proj = new_r_Proj(push, mode_M, pn_ia32_Push_M);
509 /* rewire Store Projs */
510 foreach_out_edge_safe(store, edge, next) {
511 ir_node *proj = get_edge_src_irn(edge);
514 switch (get_Proj_proj(proj)) {
515 case pn_ia32_Store_M:
516 exchange(proj, mem_proj);
519 panic("unexpected Proj on Store->IncSp");
523 /* use the memproj now */
524 be_peephole_exchange(store, push);
529 foreach_out_edge_safe(irn, edge, next) {
530 ir_node *const src = get_edge_src_irn(edge);
531 int const pos = get_edge_src_pos(edge);
533 if (src == first_push)
536 set_irn_n(src, pos, curr_sp);
539 be_set_IncSP_offset(irn, inc_ofs);
544 * Creates a Push instruction before the given schedule point.
546 * @param dbgi debug info
547 * @param block the block
548 * @param stack the previous stack value
549 * @param schedpoint the new node is added before this node
550 * @param reg the register to pop
552 * @return the new stack value
554 static ir_node *create_push(dbg_info *dbgi, ir_node *block,
555 ir_node *stack, ir_node *schedpoint)
557 const arch_register_t *esp = &ia32_registers[REG_ESP];
559 ir_node *val = ia32_new_NoReg_gp(cg);
560 ir_node *noreg = ia32_new_NoReg_gp(cg);
561 ir_graph *irg = get_irn_irg(block);
562 ir_node *nomem = get_irg_no_mem(irg);
563 ir_node *push = new_bd_ia32_Push(dbgi, block, noreg, noreg, nomem, val, stack);
564 sched_add_before(schedpoint, push);
566 stack = new_r_Proj(push, mode_Iu, pn_ia32_Push_stack);
567 arch_set_irn_register(stack, esp);
572 static void peephole_store_incsp(ir_node *store)
583 ir_node *am_base = get_irn_n(store, n_ia32_Store_base);
584 if (!be_is_IncSP(am_base)
585 || get_nodes_block(am_base) != get_nodes_block(store))
587 mem = get_irn_n(store, n_ia32_Store_mem);
588 if (!is_ia32_NoReg_GP(get_irn_n(store, n_ia32_Store_index))
592 int incsp_offset = be_get_IncSP_offset(am_base);
593 if (incsp_offset <= 0)
596 /* we have to be at offset 0 */
597 int my_offset = get_ia32_am_offs_int(store);
598 if (my_offset != 0) {
599 /* TODO here: find out whether there is a store with offset 0 before
600 * us and whether we can move it down to our place */
603 ir_mode *ls_mode = get_ia32_ls_mode(store);
604 int my_store_size = get_mode_size_bytes(ls_mode);
606 if (my_offset + my_store_size > incsp_offset)
609 /* correctness checking:
610 - noone else must write to that stackslot
611 (because after translation incsp won't allocate it anymore)
613 sched_foreach_reverse_from(store, node) {
619 /* make sure noone else can use the space on the stack */
620 arity = get_irn_arity(node);
621 for (i = 0; i < arity; ++i) {
622 ir_node *pred = get_irn_n(node, i);
626 if (i == n_ia32_base &&
627 (get_ia32_op_type(node) == ia32_AddrModeS
628 || get_ia32_op_type(node) == ia32_AddrModeD)) {
629 int node_offset = get_ia32_am_offs_int(node);
630 ir_mode *node_ls_mode = get_ia32_ls_mode(node);
631 int node_size = get_mode_size_bytes(node_ls_mode);
632 /* overlapping with our position? abort */
633 if (node_offset < my_offset + my_store_size
634 && node_offset + node_size >= my_offset)
636 /* otherwise it's fine */
640 /* strange use of esp: abort */
645 /* all ok, change to push */
646 dbgi = get_irn_dbg_info(store);
647 block = get_nodes_block(store);
648 noreg = ia32_new_NoReg_gp(cg);
649 val = get_irn_n(store, n_ia32_Store_val);
651 push = new_bd_ia32_Push(dbgi, block, noreg, noreg, mem,
653 create_push(dbgi, current_ir_graph, block, am_base, store);
658 * Return true if a mode can be stored in the GP register set
660 static inline int mode_needs_gp_reg(ir_mode *mode)
662 if (mode == ia32_mode_fpcw)
664 if (get_mode_size_bits(mode) > 32)
666 return mode_is_int(mode) || mode_is_reference(mode) || mode == mode_b;
670 * Tries to create Pops from Load, IncSP combinations.
671 * The Loads are replaced by Pops, the IncSP is modified
672 * (possibly into IncSP 0, but not removed).
674 static void peephole_Load_IncSP_to_pop(ir_node *irn)
676 const arch_register_t *esp = &ia32_registers[REG_ESP];
677 int i, maxslot, inc_ofs, ofs;
678 ir_node *node, *pred_sp, *block;
679 ir_node *loads[MAXPUSH_OPTIMIZE];
680 unsigned regmask = 0;
681 unsigned copymask = ~0;
683 memset(loads, 0, sizeof(loads));
684 assert(be_is_IncSP(irn));
686 inc_ofs = -be_get_IncSP_offset(irn);
691 * We first walk the schedule before the IncSP node as long as we find
692 * suitable Loads that could be transformed to a Pop.
693 * We save them into the stores array which is sorted by the frame offset/4
694 * attached to the node
697 pred_sp = be_get_IncSP_pred(irn);
698 for (node = sched_prev(irn); !sched_is_end(node); node = sched_prev(node)) {
701 const arch_register_t *sreg, *dreg;
703 /* it has to be a Load */
704 if (!is_ia32_Load(node)) {
705 if (be_is_Copy(node)) {
706 if (!mode_needs_gp_reg(get_irn_mode(node))) {
707 /* not a GP copy, ignore */
710 dreg = arch_get_irn_register(node);
711 sreg = arch_get_irn_register(be_get_Copy_op(node));
712 if (regmask & copymask & (1 << sreg->index)) {
715 if (regmask & copymask & (1 << dreg->index)) {
718 /* we CAN skip Copies if neither the destination nor the source
719 * is not in our regmask, ie none of our future Pop will overwrite it */
720 regmask |= (1 << dreg->index) | (1 << sreg->index);
721 copymask &= ~((1 << dreg->index) | (1 << sreg->index));
727 /* we can handle only GP loads */
728 if (!mode_needs_gp_reg(get_ia32_ls_mode(node)))
731 /* it has to use our predecessor sp value */
732 if (get_irn_n(node, n_ia32_base) != pred_sp) {
733 /* it would be ok if this load does not use a Pop result,
734 * but we do not check this */
738 /* should have NO index */
739 if (!is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
742 offset = get_ia32_am_offs_int(node);
743 /* we should NEVER access uninitialized stack BELOW the current SP */
746 /* storing at half-slots is bad */
747 if ((offset & 3) != 0)
750 if (offset < 0 || offset >= MAXPUSH_OPTIMIZE * 4)
752 /* ignore those outside the possible windows */
753 if (offset > inc_ofs - 4)
755 loadslot = offset >> 2;
757 /* loading from the same slot twice is bad (and shouldn't happen...) */
758 if (loads[loadslot] != NULL)
761 dreg = arch_irn_get_register(node, pn_ia32_Load_res);
762 if (regmask & (1 << dreg->index)) {
763 /* this register is already used */
766 regmask |= 1 << dreg->index;
768 loads[loadslot] = node;
769 if (loadslot > maxslot)
776 /* find the first slot */
777 for (i = maxslot; i >= 0; --i) {
778 ir_node *load = loads[i];
784 ofs = inc_ofs - (maxslot + 1) * 4;
787 /* create a new IncSP if needed */
788 block = get_nodes_block(irn);
790 pred_sp = be_new_IncSP(esp, block, pred_sp, -inc_ofs, be_get_IncSP_align(irn));
791 sched_add_before(irn, pred_sp);
794 /* walk through the Loads and create Pops for them */
795 for (++i; i <= maxslot; ++i) {
796 ir_node *load = loads[i];
798 const ir_edge_t *edge, *tmp;
799 const arch_register_t *reg;
801 mem = get_irn_n(load, n_ia32_mem);
802 reg = arch_irn_get_register(load, pn_ia32_Load_res);
804 pop = new_bd_ia32_Pop(get_irn_dbg_info(load), block, mem, pred_sp);
805 arch_irn_set_register(pop, pn_ia32_Load_res, reg);
807 copy_mark(load, pop);
809 /* create stackpointer Proj */
810 pred_sp = new_r_Proj(pop, mode_Iu, pn_ia32_Pop_stack);
811 arch_set_irn_register(pred_sp, esp);
813 sched_add_before(irn, pop);
816 foreach_out_edge_safe(load, edge, tmp) {
817 ir_node *proj = get_edge_src_irn(edge);
819 set_Proj_pred(proj, pop);
822 /* we can remove the Load now */
827 be_set_IncSP_offset(irn, -ofs);
828 be_set_IncSP_pred(irn, pred_sp);
833 * Find a free GP register if possible, else return NULL.
835 static const arch_register_t *get_free_gp_reg(ir_graph *irg)
837 be_irg_t *birg = be_birg_from_irg(irg);
840 for (i = 0; i < N_ia32_gp_REGS; ++i) {
841 const arch_register_t *reg = &ia32_reg_classes[CLASS_ia32_gp].regs[i];
842 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index))
845 if (be_peephole_get_value(CLASS_ia32_gp, i) == NULL)
853 * Creates a Pop instruction before the given schedule point.
855 * @param dbgi debug info
856 * @param block the block
857 * @param stack the previous stack value
858 * @param schedpoint the new node is added before this node
859 * @param reg the register to pop
861 * @return the new stack value
863 static ir_node *create_pop(dbg_info *dbgi, ir_node *block,
864 ir_node *stack, ir_node *schedpoint,
865 const arch_register_t *reg)
867 const arch_register_t *esp = &ia32_registers[REG_ESP];
868 ir_graph *irg = get_irn_irg(block);
874 pop = new_bd_ia32_Pop(dbgi, block, get_irg_no_mem(irg), stack);
876 stack = new_r_Proj(pop, mode_Iu, pn_ia32_Pop_stack);
877 arch_set_irn_register(stack, esp);
878 val = new_r_Proj(pop, mode_Iu, pn_ia32_Pop_res);
879 arch_set_irn_register(val, reg);
881 sched_add_before(schedpoint, pop);
884 keep = be_new_Keep(block, 1, in);
885 sched_add_before(schedpoint, keep);
891 * Optimize an IncSp by replacing it with Push/Pop.
893 static void peephole_be_IncSP(ir_node *node)
895 const arch_register_t *esp = &ia32_registers[REG_ESP];
896 const arch_register_t *reg;
902 /* first optimize incsp->incsp combinations */
903 node = be_peephole_IncSP_IncSP(node);
905 /* transform IncSP->Store combinations to Push where possible */
906 peephole_IncSP_Store_to_push(node);
908 /* transform Load->IncSP combinations to Pop where possible */
909 peephole_Load_IncSP_to_pop(node);
911 if (arch_get_irn_register(node) != esp)
914 /* replace IncSP -4 by Pop freereg when possible */
915 offset = be_get_IncSP_offset(node);
916 if ((offset != -8 || ia32_cg_config.use_add_esp_8) &&
917 (offset != -4 || ia32_cg_config.use_add_esp_4) &&
918 (offset != +4 || ia32_cg_config.use_sub_esp_4) &&
919 (offset != +8 || ia32_cg_config.use_sub_esp_8))
923 /* we need a free register for pop */
924 reg = get_free_gp_reg(get_irn_irg(node));
928 dbgi = get_irn_dbg_info(node);
929 block = get_nodes_block(node);
930 stack = be_get_IncSP_pred(node);
932 stack = create_pop(dbgi, block, stack, node, reg);
935 stack = create_pop(dbgi, block, stack, node, reg);
938 dbgi = get_irn_dbg_info(node);
939 block = get_nodes_block(node);
940 stack = be_get_IncSP_pred(node);
941 stack = new_bd_ia32_PushEax(dbgi, block, stack);
942 arch_set_irn_register(stack, esp);
943 sched_add_before(node, stack);
946 stack = new_bd_ia32_PushEax(dbgi, block, stack);
947 arch_set_irn_register(stack, esp);
948 sched_add_before(node, stack);
952 be_peephole_exchange(node, stack);
956 * Peephole optimisation for ia32_Const's
958 static void peephole_ia32_Const(ir_node *node)
960 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node);
961 const arch_register_t *reg;
966 /* try to transform a mov 0, reg to xor reg reg */
967 if (attr->offset != 0 || attr->symconst != NULL)
969 if (ia32_cg_config.use_mov_0)
971 /* xor destroys the flags, so no-one must be using them */
972 if (be_peephole_get_value(CLASS_ia32_flags, REG_FLAGS_EFLAGS) != NULL)
975 reg = arch_get_irn_register(node);
976 assert(be_peephole_get_reg_value(reg) == NULL);
978 /* create xor(produceval, produceval) */
979 block = get_nodes_block(node);
980 dbgi = get_irn_dbg_info(node);
981 xorn = new_bd_ia32_Xor0(dbgi, block);
982 arch_set_irn_register(xorn, reg);
984 sched_add_before(node, xorn);
986 copy_mark(node, xorn);
987 be_peephole_exchange(node, xorn);
990 static inline int is_noreg(const ir_node *node)
992 return is_ia32_NoReg_GP(node);
995 ir_node *ia32_immediate_from_long(long val)
997 ir_graph *irg = current_ir_graph;
998 ir_node *start_block = get_irg_start_block(irg);
1000 = new_bd_ia32_Immediate(NULL, start_block, NULL, 0, 0, val);
1001 arch_set_irn_register(immediate, &ia32_registers[REG_GP_NOREG]);
1006 static ir_node *create_immediate_from_am(const ir_node *node)
1008 ir_node *block = get_nodes_block(node);
1009 int offset = get_ia32_am_offs_int(node);
1010 int sc_sign = is_ia32_am_sc_sign(node);
1011 const ia32_attr_t *attr = get_ia32_attr_const(node);
1012 int sc_no_pic_adjust = attr->data.am_sc_no_pic_adjust;
1013 ir_entity *entity = get_ia32_am_sc(node);
1016 res = new_bd_ia32_Immediate(NULL, block, entity, sc_sign, sc_no_pic_adjust,
1018 arch_set_irn_register(res, &ia32_registers[REG_GP_NOREG]);
1022 static int is_am_one(const ir_node *node)
1024 int offset = get_ia32_am_offs_int(node);
1025 ir_entity *entity = get_ia32_am_sc(node);
1027 return offset == 1 && entity == NULL;
1030 static int is_am_minus_one(const ir_node *node)
1032 int offset = get_ia32_am_offs_int(node);
1033 ir_entity *entity = get_ia32_am_sc(node);
1035 return offset == -1 && entity == NULL;
1039 * Transforms a LEA into an Add or SHL if possible.
1041 static void peephole_ia32_Lea(ir_node *node)
1046 const arch_register_t *base_reg;
1047 const arch_register_t *index_reg;
1048 const arch_register_t *out_reg;
1059 assert(is_ia32_Lea(node));
1061 /* we can only do this if it is allowed to clobber the flags */
1062 if (be_peephole_get_value(CLASS_ia32_flags, REG_FLAGS_EFLAGS) != NULL)
1065 base = get_irn_n(node, n_ia32_Lea_base);
1066 index = get_irn_n(node, n_ia32_Lea_index);
1068 if (is_noreg(base)) {
1072 base_reg = arch_get_irn_register(base);
1074 if (is_noreg(index)) {
1078 index_reg = arch_get_irn_register(index);
1081 if (base == NULL && index == NULL) {
1082 /* we shouldn't construct these in the first place... */
1083 #ifdef DEBUG_libfirm
1084 ir_fprintf(stderr, "Optimisation warning: found immediate only lea\n");
1089 out_reg = arch_get_irn_register(node);
1090 scale = get_ia32_am_scale(node);
1091 assert(!is_ia32_need_stackent(node) || get_ia32_frame_ent(node) != NULL);
1092 /* check if we have immediates values (frame entities should already be
1093 * expressed in the offsets) */
1094 if (get_ia32_am_offs_int(node) != 0 || get_ia32_am_sc(node) != NULL) {
1100 /* we can transform leas where the out register is the same as either the
1101 * base or index register back to an Add or Shl */
1102 if (out_reg == base_reg) {
1103 if (index == NULL) {
1104 #ifdef DEBUG_libfirm
1105 if (!has_immediates) {
1106 ir_fprintf(stderr, "Optimisation warning: found lea which is "
1111 goto make_add_immediate;
1113 if (scale == 0 && !has_immediates) {
1118 /* can't create an add */
1120 } else if (out_reg == index_reg) {
1122 if (has_immediates && scale == 0) {
1124 goto make_add_immediate;
1125 } else if (!has_immediates && scale > 0) {
1127 op2 = ia32_immediate_from_long(scale);
1129 } else if (!has_immediates) {
1130 #ifdef DEBUG_libfirm
1131 ir_fprintf(stderr, "Optimisation warning: found lea which is "
1135 } else if (scale == 0 && !has_immediates) {
1140 /* can't create an add */
1143 /* can't create an add */
1148 if (ia32_cg_config.use_incdec) {
1149 if (is_am_one(node)) {
1150 dbgi = get_irn_dbg_info(node);
1151 block = get_nodes_block(node);
1152 res = new_bd_ia32_Inc(dbgi, block, op1);
1153 arch_set_irn_register(res, out_reg);
1156 if (is_am_minus_one(node)) {
1157 dbgi = get_irn_dbg_info(node);
1158 block = get_nodes_block(node);
1159 res = new_bd_ia32_Dec(dbgi, block, op1);
1160 arch_set_irn_register(res, out_reg);
1164 op2 = create_immediate_from_am(node);
1167 dbgi = get_irn_dbg_info(node);
1168 block = get_nodes_block(node);
1169 irg = get_irn_irg(node);
1170 noreg = ia32_new_NoReg_gp(irg);
1171 nomem = get_irg_no_mem(irg);
1172 res = new_bd_ia32_Add(dbgi, block, noreg, noreg, nomem, op1, op2);
1173 arch_set_irn_register(res, out_reg);
1174 set_ia32_commutative(res);
1178 dbgi = get_irn_dbg_info(node);
1179 block = get_nodes_block(node);
1180 irg = get_irn_irg(node);
1181 noreg = ia32_new_NoReg_gp(irg);
1182 nomem = get_irg_no_mem(irg);
1183 res = new_bd_ia32_Shl(dbgi, block, op1, op2);
1184 arch_set_irn_register(res, out_reg);
1188 SET_IA32_ORIG_NODE(res, node);
1190 /* add new ADD/SHL to schedule */
1191 DBG_OPT_LEA2ADD(node, res);
1193 /* exchange the Add and the LEA */
1194 sched_add_before(node, res);
1195 copy_mark(node, res);
1196 be_peephole_exchange(node, res);
1200 * Split a Imul mem, imm into a Load mem and Imul reg, imm if possible.
1202 static void peephole_ia32_Imul_split(ir_node *imul)
1204 const ir_node *right = get_irn_n(imul, n_ia32_IMul_right);
1205 const arch_register_t *reg;
1208 if (!is_ia32_Immediate(right) || get_ia32_op_type(imul) != ia32_AddrModeS) {
1209 /* no memory, imm form ignore */
1212 /* we need a free register */
1213 reg = get_free_gp_reg(get_irn_irg(imul));
1217 /* fine, we can rebuild it */
1218 res = ia32_turn_back_am(imul);
1219 arch_set_irn_register(res, reg);
1223 * Replace xorps r,r and xorpd r,r by pxor r,r
1225 static void peephole_ia32_xZero(ir_node *xorn)
1227 set_irn_op(xorn, op_ia32_xPzero);
1231 * Replace 16bit sign extension from ax to eax by shorter cwtl
1233 static void peephole_ia32_Conv_I2I(ir_node *node)
1235 const arch_register_t *eax = &ia32_registers[REG_EAX];
1236 ir_mode *smaller_mode = get_ia32_ls_mode(node);
1237 ir_node *val = get_irn_n(node, n_ia32_Conv_I2I_val);
1242 if (get_mode_size_bits(smaller_mode) != 16 ||
1243 !mode_is_signed(smaller_mode) ||
1244 eax != arch_get_irn_register(val) ||
1245 eax != arch_irn_get_register(node, pn_ia32_Conv_I2I_res))
1248 dbgi = get_irn_dbg_info(node);
1249 block = get_nodes_block(node);
1250 cwtl = new_bd_ia32_Cwtl(dbgi, block, val);
1251 arch_set_irn_register(cwtl, eax);
1252 sched_add_before(node, cwtl);
1253 be_peephole_exchange(node, cwtl);
1257 * Register a peephole optimisation function.
1259 static void register_peephole_optimisation(ir_op *op, peephole_opt_func func)
1261 assert(op->ops.generic == NULL);
1262 op->ops.generic = (op_func)func;
1265 /* Perform peephole-optimizations. */
1266 void ia32_peephole_optimization(ir_graph *irg)
1268 /* register peephole optimisations */
1269 clear_irp_opcodes_generic_func();
1270 register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
1271 register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
1272 register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
1273 register_peephole_optimisation(op_ia32_Cmp, peephole_ia32_Cmp);
1274 register_peephole_optimisation(op_ia32_Cmp8Bit, peephole_ia32_Cmp);
1275 register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test);
1276 register_peephole_optimisation(op_ia32_Test8Bit, peephole_ia32_Test);
1277 register_peephole_optimisation(op_be_Return, peephole_ia32_Return);
1278 if (! ia32_cg_config.use_imul_mem_imm32)
1279 register_peephole_optimisation(op_ia32_IMul, peephole_ia32_Imul_split);
1280 if (ia32_cg_config.use_pxor)
1281 register_peephole_optimisation(op_ia32_xZero, peephole_ia32_xZero);
1282 if (ia32_cg_config.use_short_sex_eax)
1283 register_peephole_optimisation(op_ia32_Conv_I2I, peephole_ia32_Conv_I2I);
1285 be_peephole_opt(irg);
1289 * Removes node from schedule if it is not used anymore. If irn is a mode_T node
1290 * all its Projs are removed as well.
1291 * @param irn The irn to be removed from schedule
1293 static inline void try_kill(ir_node *node)
1295 if (get_irn_mode(node) == mode_T) {
1296 const ir_edge_t *edge, *next;
1297 foreach_out_edge_safe(node, edge, next) {
1298 ir_node *proj = get_edge_src_irn(edge);
1303 if (get_irn_n_edges(node) != 0)
1306 if (sched_is_scheduled(node)) {
1313 static void optimize_conv_store(ir_node *node)
1318 ir_mode *store_mode;
1320 if (!is_ia32_Store(node) && !is_ia32_Store8Bit(node))
1323 assert((int)n_ia32_Store_val == (int)n_ia32_Store8Bit_val);
1324 pred_proj = get_irn_n(node, n_ia32_Store_val);
1325 if (is_Proj(pred_proj)) {
1326 pred = get_Proj_pred(pred_proj);
1330 if (!is_ia32_Conv_I2I(pred) && !is_ia32_Conv_I2I8Bit(pred))
1332 if (get_ia32_op_type(pred) != ia32_Normal)
1335 /* the store only stores the lower bits, so we only need the conv
1336 * it it shrinks the mode */
1337 conv_mode = get_ia32_ls_mode(pred);
1338 store_mode = get_ia32_ls_mode(node);
1339 if (get_mode_size_bits(conv_mode) < get_mode_size_bits(store_mode))
1342 set_irn_n(node, n_ia32_Store_val, get_irn_n(pred, n_ia32_Conv_I2I_val));
1343 if (get_irn_n_edges(pred_proj) == 0) {
1344 kill_node(pred_proj);
1345 if (pred != pred_proj)
1350 static void optimize_load_conv(ir_node *node)
1352 ir_node *pred, *predpred;
1356 if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node))
1359 assert((int)n_ia32_Conv_I2I_val == (int)n_ia32_Conv_I2I8Bit_val);
1360 pred = get_irn_n(node, n_ia32_Conv_I2I_val);
1364 predpred = get_Proj_pred(pred);
1365 if (!is_ia32_Load(predpred))
1368 /* the load is sign extending the upper bits, so we only need the conv
1369 * if it shrinks the mode */
1370 load_mode = get_ia32_ls_mode(predpred);
1371 conv_mode = get_ia32_ls_mode(node);
1372 if (get_mode_size_bits(conv_mode) < get_mode_size_bits(load_mode))
1375 if (get_mode_sign(conv_mode) != get_mode_sign(load_mode)) {
1376 /* change the load if it has only 1 user */
1377 if (get_irn_n_edges(pred) == 1) {
1379 if (get_mode_sign(conv_mode)) {
1380 newmode = find_signed_mode(load_mode);
1382 newmode = find_unsigned_mode(load_mode);
1384 assert(newmode != NULL);
1385 set_ia32_ls_mode(predpred, newmode);
1387 /* otherwise we have to keep the conv */
1393 exchange(node, pred);
1396 static void optimize_conv_conv(ir_node *node)
1398 ir_node *pred_proj, *pred, *result_conv;
1399 ir_mode *pred_mode, *conv_mode;
1403 if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node))
1406 assert((int)n_ia32_Conv_I2I_val == (int)n_ia32_Conv_I2I8Bit_val);
1407 pred_proj = get_irn_n(node, n_ia32_Conv_I2I_val);
1408 if (is_Proj(pred_proj))
1409 pred = get_Proj_pred(pred_proj);
1413 if (!is_ia32_Conv_I2I(pred) && !is_ia32_Conv_I2I8Bit(pred))
1416 /* we know that after a conv, the upper bits are sign extended
1417 * so we only need the 2nd conv if it shrinks the mode */
1418 conv_mode = get_ia32_ls_mode(node);
1419 conv_mode_bits = get_mode_size_bits(conv_mode);
1420 pred_mode = get_ia32_ls_mode(pred);
1421 pred_mode_bits = get_mode_size_bits(pred_mode);
1423 if (conv_mode_bits == pred_mode_bits
1424 && get_mode_sign(conv_mode) == get_mode_sign(pred_mode)) {
1425 result_conv = pred_proj;
1426 } else if (conv_mode_bits <= pred_mode_bits) {
1427 /* if 2nd conv is smaller then first conv, then we can always take the
1429 if (get_irn_n_edges(pred_proj) == 1) {
1430 result_conv = pred_proj;
1431 set_ia32_ls_mode(pred, conv_mode);
1433 /* Argh:We must change the opcode to 8bit AND copy the register constraints */
1434 if (get_mode_size_bits(conv_mode) == 8) {
1435 set_irn_op(pred, op_ia32_Conv_I2I8Bit);
1436 arch_set_in_register_reqs(pred,
1437 arch_get_in_register_reqs(node));
1440 /* we don't want to end up with 2 loads, so we better do nothing */
1441 if (get_irn_mode(pred) == mode_T) {
1445 result_conv = exact_copy(pred);
1446 set_ia32_ls_mode(result_conv, conv_mode);
1448 /* Argh:We must change the opcode to 8bit AND copy the register constraints */
1449 if (get_mode_size_bits(conv_mode) == 8) {
1450 set_irn_op(result_conv, op_ia32_Conv_I2I8Bit);
1451 arch_set_in_register_reqs(result_conv,
1452 arch_get_in_register_reqs(node));
1456 /* if both convs have the same sign, then we can take the smaller one */
1457 if (get_mode_sign(conv_mode) == get_mode_sign(pred_mode)) {
1458 result_conv = pred_proj;
1460 /* no optimisation possible if smaller conv is sign-extend */
1461 if (mode_is_signed(pred_mode)) {
1464 /* we can take the smaller conv if it is unsigned */
1465 result_conv = pred_proj;
1469 /* Some user (like Phis) won't be happy if we change the mode. */
1470 set_irn_mode(result_conv, get_irn_mode(node));
1473 exchange(node, result_conv);
1475 if (get_irn_n_edges(pred_proj) == 0) {
1476 kill_node(pred_proj);
1477 if (pred != pred_proj)
1480 optimize_conv_conv(result_conv);
1483 static void optimize_node(ir_node *node, void *env)
1487 optimize_load_conv(node);
1488 optimize_conv_store(node);
1489 optimize_conv_conv(node);
1493 * Performs conv and address mode optimization.
1495 void ia32_optimize_graph(ir_graph *irg)
1497 irg_walk_blkwise_graph(irg, NULL, optimize_node, NULL);
1500 void ia32_init_optimize(void)
1502 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.optimize");