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 This file implements the x87 support and virtual to stack
23 * register translation for the ia32 backend.
24 * @author Michael Beck
34 #include "iredges_t.h"
46 #include "../belive_t.h"
47 #include "../besched.h"
48 #include "../benode.h"
49 #include "bearch_ia32_t.h"
50 #include "ia32_new_nodes.h"
51 #include "gen_ia32_new_nodes.h"
52 #include "gen_ia32_regalloc_if.h"
54 #include "ia32_architecture.h"
56 #define MASK_TOS(x) ((x) & (N_ia32_st_REGS - 1))
58 /** the debug handle */
59 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
61 /* Forward declaration. */
62 typedef struct x87_simulator x87_simulator;
65 * An exchange template.
66 * Note that our virtual functions have the same inputs
67 * and attributes as the real ones, so we can simple exchange
69 * Further, x87 supports inverse instructions, so we can handle them.
71 typedef struct exchange_tmpl {
72 ir_op *normal_op; /**< the normal one */
73 ir_op *reverse_op; /**< the reverse one if exists */
74 ir_op *normal_pop_op; /**< the normal one with tos pop */
75 ir_op *reverse_pop_op; /**< the reverse one with tos pop */
79 * An entry on the simulated x87 stack.
81 typedef struct st_entry {
82 int reg_idx; /**< the virtual register index of this stack value */
83 ir_node *node; /**< the node that produced this value */
89 typedef struct x87_state {
90 st_entry st[N_ia32_st_REGS]; /**< the register stack */
91 int depth; /**< the current stack depth */
92 int tos; /**< position of the tos */
93 x87_simulator *sim; /**< The simulator. */
96 /** An empty state, used for blocks without fp instructions. */
97 static x87_state _empty = { { {0, NULL}, }, 0, 0, NULL };
98 static x87_state *empty = (x87_state *)&_empty;
101 * Return values of the instruction simulator functions.
104 NO_NODE_ADDED = 0, /**< No node that needs simulation was added. */
105 NODE_ADDED = 1 /**< A node that must be simulated was added by the simulator
106 in the schedule AFTER the current node. */
110 * The type of an instruction simulator function.
112 * @param state the x87 state
113 * @param n the node to be simulated
115 * @return NODE_ADDED if a node was added AFTER n in schedule that MUST be
117 * NO_NODE_ADDED otherwise
119 typedef int (*sim_func)(x87_state *state, ir_node *n);
122 * A block state: Every block has a x87 state at the beginning and at the end.
124 typedef struct blk_state {
125 x87_state *begin; /**< state at the begin or NULL if not assigned */
126 x87_state *end; /**< state at the end or NULL if not assigned */
129 #define PTR_TO_BLKSTATE(p) ((blk_state *)(p))
131 /** liveness bitset for vfp registers. */
132 typedef unsigned char vfp_liveness;
137 struct x87_simulator {
138 struct obstack obst; /**< An obstack for fast allocating. */
139 pmap *blk_states; /**< Map blocks to states. */
140 be_lv_t *lv; /**< intrablock liveness. */
141 vfp_liveness *live; /**< Liveness information. */
142 unsigned n_idx; /**< The cached get_irg_last_idx() result. */
143 waitq *worklist; /**< Worklist of blocks that must be processed. */
144 ia32_isa_t *isa; /**< the ISA object */
148 * Returns the current stack depth.
150 * @param state the x87 state
152 * @return the x87 stack depth
154 static int x87_get_depth(const x87_state *state)
157 } /* x87_get_depth */
160 * Return the virtual register index at st(pos).
162 * @param state the x87 state
163 * @param pos a stack position
165 * @return the vfp register index that produced the value at st(pos)
167 static int x87_get_st_reg(const x87_state *state, int pos)
169 assert(pos < state->depth);
170 return state->st[MASK_TOS(state->tos + pos)].reg_idx;
171 } /* x87_get_st_reg */
175 * Return the node at st(pos).
177 * @param state the x87 state
178 * @param pos a stack position
180 * @return the IR node that produced the value at st(pos)
182 static ir_node *x87_get_st_node(const x87_state *state, int pos)
184 assert(pos < state->depth);
185 return state->st[MASK_TOS(state->tos + pos)].node;
186 } /* x87_get_st_node */
189 * Dump the stack for debugging.
191 * @param state the x87 state
193 static void x87_dump_stack(const x87_state *state)
197 for (i = state->depth - 1; i >= 0; --i) {
198 DB((dbg, LEVEL_2, "vf%d(%+F) ", x87_get_st_reg(state, i),
199 x87_get_st_node(state, i)));
201 DB((dbg, LEVEL_2, "<-- TOS\n"));
202 } /* x87_dump_stack */
203 #endif /* DEBUG_libfirm */
206 * Set a virtual register to st(pos).
208 * @param state the x87 state
209 * @param reg_idx the vfp register index that should be set
210 * @param node the IR node that produces the value of the vfp register
211 * @param pos the stack position where the new value should be entered
213 static void x87_set_st(x87_state *state, int reg_idx, ir_node *node, int pos)
215 assert(0 < state->depth);
216 state->st[MASK_TOS(state->tos + pos)].reg_idx = reg_idx;
217 state->st[MASK_TOS(state->tos + pos)].node = node;
219 DB((dbg, LEVEL_2, "After SET_REG: "));
220 DEBUG_ONLY(x87_dump_stack(state));
224 * Set the tos virtual register.
226 * @param state the x87 state
227 * @param reg_idx the vfp register index that should be set
228 * @param node the IR node that produces the value of the vfp register
230 static void x87_set_tos(x87_state *state, int reg_idx, ir_node *node)
232 x87_set_st(state, reg_idx, node, 0);
236 * Swap st(0) with st(pos).
238 * @param state the x87 state
239 * @param pos the stack position to change the tos with
241 static void x87_fxch(x87_state *state, int pos)
244 assert(pos < state->depth);
246 entry = state->st[MASK_TOS(state->tos + pos)];
247 state->st[MASK_TOS(state->tos + pos)] = state->st[MASK_TOS(state->tos)];
248 state->st[MASK_TOS(state->tos)] = entry;
250 DB((dbg, LEVEL_2, "After FXCH: "));
251 DEBUG_ONLY(x87_dump_stack(state));
255 * Convert a virtual register to the stack index.
257 * @param state the x87 state
258 * @param reg_idx the register vfp index
260 * @return the stack position where the register is stacked
261 * or -1 if the virtual register was not found
263 static int x87_on_stack(const x87_state *state, int reg_idx)
265 int i, tos = state->tos;
267 for (i = 0; i < state->depth; ++i)
268 if (state->st[MASK_TOS(tos + i)].reg_idx == reg_idx)
274 * Push a virtual Register onto the stack, double pushed allowed.
276 * @param state the x87 state
277 * @param reg_idx the register vfp index
278 * @param node the node that produces the value of the vfp register
280 static void x87_push_dbl(x87_state *state, int reg_idx, ir_node *node)
282 assert(state->depth < N_ia32_st_REGS && "stack overrun");
285 state->tos = MASK_TOS(state->tos - 1);
286 state->st[state->tos].reg_idx = reg_idx;
287 state->st[state->tos].node = node;
289 DB((dbg, LEVEL_2, "After PUSH: ")); DEBUG_ONLY(x87_dump_stack(state));
293 * Push a virtual Register onto the stack, double pushes are NOT allowed.
295 * @param state the x87 state
296 * @param reg_idx the register vfp index
297 * @param node the node that produces the value of the vfp register
298 * @param dbl_push if != 0 double pushes are allowed
300 static void x87_push(x87_state *state, int reg_idx, ir_node *node)
302 assert(x87_on_stack(state, reg_idx) == -1 && "double push");
304 x87_push_dbl(state, reg_idx, node);
308 * Pop a virtual Register from the stack.
310 * @param state the x87 state
312 static void x87_pop(x87_state *state)
314 assert(state->depth > 0 && "stack underrun");
317 state->tos = MASK_TOS(state->tos + 1);
319 DB((dbg, LEVEL_2, "After POP: ")); DEBUG_ONLY(x87_dump_stack(state));
323 * Empty the fpu stack
325 * @param state the x87 state
327 static void x87_emms(x87_state *state)
334 * Returns the block state of a block.
336 * @param sim the x87 simulator handle
337 * @param block the current block
339 * @return the block state
341 static blk_state *x87_get_bl_state(x87_simulator *sim, ir_node *block)
343 pmap_entry *entry = pmap_find(sim->blk_states, block);
346 blk_state *bl_state = obstack_alloc(&sim->obst, sizeof(*bl_state));
347 bl_state->begin = NULL;
348 bl_state->end = NULL;
350 pmap_insert(sim->blk_states, block, bl_state);
354 return PTR_TO_BLKSTATE(entry->value);
355 } /* x87_get_bl_state */
358 * Creates a new x87 state.
360 * @param sim the x87 simulator handle
362 * @return a new x87 state
364 static x87_state *x87_alloc_state(x87_simulator *sim)
366 x87_state *res = obstack_alloc(&sim->obst, sizeof(*res));
370 } /* x87_alloc_state */
375 * @param sim the x87 simulator handle
376 * @param src the x87 state that will be cloned
378 * @return a cloned copy of the src state
380 static x87_state *x87_clone_state(x87_simulator *sim, const x87_state *src)
382 x87_state *res = x87_alloc_state(sim);
386 } /* x87_clone_state */
389 * Patch a virtual instruction into a x87 one and return
390 * the node representing the result value.
392 * @param n the IR node to patch
393 * @param op the x87 opcode to patch in
395 static ir_node *x87_patch_insn(ir_node *n, ir_op *op)
397 ir_mode *mode = get_irn_mode(n);
402 if (mode == mode_T) {
403 /* patch all Proj's */
404 const ir_edge_t *edge;
406 foreach_out_edge(n, edge) {
407 ir_node *proj = get_edge_src_irn(edge);
409 mode = get_irn_mode(proj);
410 if (mode_is_float(mode)) {
412 set_irn_mode(proj, ia32_reg_classes[CLASS_ia32_st].mode);
416 } else if (mode_is_float(mode))
417 set_irn_mode(n, ia32_reg_classes[CLASS_ia32_st].mode);
419 } /* x87_patch_insn */
422 * Returns the first Proj of a mode_T node having a given mode.
424 * @param n the mode_T node
425 * @param m the desired mode of the Proj
426 * @return The first Proj of mode @p m found or NULL.
428 static ir_node *get_irn_Proj_for_mode(ir_node *n, ir_mode *m)
430 const ir_edge_t *edge;
432 assert(get_irn_mode(n) == mode_T && "Need mode_T node");
434 foreach_out_edge(n, edge) {
435 ir_node *proj = get_edge_src_irn(edge);
436 if (get_irn_mode(proj) == m)
441 } /* get_irn_Proj_for_mode */
444 * Wrap the arch_* function here so we can check for errors.
446 static inline const arch_register_t *x87_get_irn_register(const ir_node *irn)
448 const arch_register_t *res = arch_get_irn_register(irn);
450 assert(res->reg_class == &ia32_reg_classes[CLASS_ia32_vfp]);
452 } /* x87_get_irn_register */
454 static inline const arch_register_t *x87_irn_get_register(const ir_node *irn,
457 const arch_register_t *res = arch_irn_get_register(irn, pos);
459 assert(res->reg_class == &ia32_reg_classes[CLASS_ia32_vfp]);
461 } /* x87_irn_get_register */
463 static inline const arch_register_t *get_st_reg(int index)
465 return &ia32_registers[REG_ST0 + index];
468 /* -------------- x87 perm --------------- */
471 * Creates a fxch for shuffle.
473 * @param state the x87 state
474 * @param pos parameter for fxch
475 * @param block the block were fxch is inserted
477 * Creates a new fxch node and reroute the user of the old node
480 * @return the fxch node
482 static ir_node *x87_fxch_shuffle(x87_state *state, int pos, ir_node *block)
485 ia32_x87_attr_t *attr;
487 fxch = new_bd_ia32_fxch(NULL, block);
488 attr = get_ia32_x87_attr(fxch);
489 attr->x87[0] = get_st_reg(pos);
490 attr->x87[2] = get_st_reg(0);
494 x87_fxch(state, pos);
496 } /* x87_fxch_shuffle */
499 * Calculate the necessary permutations to reach dst_state.
501 * These permutations are done with fxch instructions and placed
502 * at the end of the block.
504 * Note that critical edges are removed here, so we need only
505 * a shuffle if the current block has only one successor.
507 * @param sim the simulator handle
508 * @param block the current block
509 * @param state the current x87 stack state, might be modified
510 * @param dst_block the destination block
511 * @param dst_state destination state
515 static x87_state *x87_shuffle(x87_simulator *sim, ir_node *block,
516 x87_state *state, ir_node *dst_block,
517 const x87_state *dst_state)
519 int i, n_cycles, k, ri;
520 unsigned cycles[4], all_mask;
521 char cycle_idx[4][8];
522 ir_node *fxch, *before, *after;
526 assert(state->depth == dst_state->depth);
528 /* Some mathematics here:
529 If we have a cycle of length n that includes the tos,
530 we need n-1 exchange operations.
531 We can always add the tos and restore it, so we need
532 n+1 exchange operations for a cycle not containing the tos.
533 So, the maximum of needed operations is for a cycle of 7
534 not including the tos == 8.
535 This is the same number of ops we would need for using stores,
536 so exchange is cheaper (we save the loads).
537 On the other hand, we might need an additional exchange
538 in the next block to bring one operand on top, so the
539 number of ops in the first case is identical.
540 Further, no more than 4 cycles can exists (4 x 2).
542 all_mask = (1 << (state->depth)) - 1;
544 for (n_cycles = 0; all_mask; ++n_cycles) {
545 int src_idx, dst_idx;
547 /* find the first free slot */
548 for (i = 0; i < state->depth; ++i) {
549 if (all_mask & (1 << i)) {
550 all_mask &= ~(1 << i);
552 /* check if there are differences here */
553 if (x87_get_st_reg(state, i) != x87_get_st_reg(dst_state, i))
559 /* no more cycles found */
564 cycles[n_cycles] = (1 << i);
565 cycle_idx[n_cycles][k++] = i;
566 for (src_idx = i; ; src_idx = dst_idx) {
567 dst_idx = x87_on_stack(dst_state, x87_get_st_reg(state, src_idx));
569 if ((all_mask & (1 << dst_idx)) == 0)
572 cycle_idx[n_cycles][k++] = dst_idx;
573 cycles[n_cycles] |= (1 << dst_idx);
574 all_mask &= ~(1 << dst_idx);
576 cycle_idx[n_cycles][k] = -1;
580 /* no permutation needed */
584 /* Hmm: permutation needed */
585 DB((dbg, LEVEL_2, "\n%+F needs permutation: from\n", block));
586 DEBUG_ONLY(x87_dump_stack(state));
587 DB((dbg, LEVEL_2, " to\n"));
588 DEBUG_ONLY(x87_dump_stack(dst_state));
592 DB((dbg, LEVEL_2, "Need %d cycles\n", n_cycles));
593 for (ri = 0; ri < n_cycles; ++ri) {
594 DB((dbg, LEVEL_2, " Ring %d:\n ", ri));
595 for (k = 0; cycle_idx[ri][k] != -1; ++k)
596 DB((dbg, LEVEL_2, " st%d ->", cycle_idx[ri][k]));
597 DB((dbg, LEVEL_2, "\n"));
604 * Find the place node must be insert.
605 * We have only one successor block, so the last instruction should
608 before = sched_last(block);
609 assert(is_cfop(before));
611 /* now do the permutations */
612 for (ri = 0; ri < n_cycles; ++ri) {
613 if ((cycles[ri] & 1) == 0) {
614 /* this cycle does not include the tos */
615 fxch = x87_fxch_shuffle(state, cycle_idx[ri][0], block);
617 sched_add_after(after, fxch);
619 sched_add_before(before, fxch);
622 for (k = 1; cycle_idx[ri][k] != -1; ++k) {
623 fxch = x87_fxch_shuffle(state, cycle_idx[ri][k], block);
625 sched_add_after(after, fxch);
627 sched_add_before(before, fxch);
630 if ((cycles[ri] & 1) == 0) {
631 /* this cycle does not include the tos */
632 fxch = x87_fxch_shuffle(state, cycle_idx[ri][0], block);
633 sched_add_after(after, fxch);
640 * Create a fxch node before another node.
642 * @param state the x87 state
643 * @param n the node after the fxch
644 * @param pos exchange st(pos) with st(0)
648 static ir_node *x87_create_fxch(x87_state *state, ir_node *n, int pos)
651 ia32_x87_attr_t *attr;
652 ir_node *block = get_nodes_block(n);
654 x87_fxch(state, pos);
656 fxch = new_bd_ia32_fxch(NULL, block);
657 attr = get_ia32_x87_attr(fxch);
658 attr->x87[0] = get_st_reg(pos);
659 attr->x87[2] = get_st_reg(0);
663 sched_add_before(n, fxch);
664 DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(fxch), attr->x87[0]->name, attr->x87[2]->name));
666 } /* x87_create_fxch */
669 * Create a fpush before node n.
671 * @param state the x87 state
672 * @param n the node after the fpush
673 * @param pos push st(pos) on stack
674 * @param op_idx replace input op_idx of n with the fpush result
676 static void x87_create_fpush(x87_state *state, ir_node *n, int pos, int op_idx)
678 ir_node *fpush, *pred = get_irn_n(n, op_idx);
679 ia32_x87_attr_t *attr;
680 const arch_register_t *out = x87_get_irn_register(pred);
682 x87_push_dbl(state, arch_register_get_index(out), pred);
684 fpush = new_bd_ia32_fpush(NULL, get_nodes_block(n));
685 attr = get_ia32_x87_attr(fpush);
686 attr->x87[0] = get_st_reg(pos);
687 attr->x87[2] = get_st_reg(0);
690 sched_add_before(n, fpush);
692 DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(fpush), attr->x87[0]->name, attr->x87[2]->name));
693 } /* x87_create_fpush */
696 * Create a fpop before node n.
698 * @param state the x87 state
699 * @param n the node after the fpop
700 * @param num pop 1 or 2 values
702 * @return the fpop node
704 static ir_node *x87_create_fpop(x87_state *state, ir_node *n, int num)
706 ir_node *fpop = NULL;
707 ia32_x87_attr_t *attr;
712 if (ia32_cg_config.use_ffreep)
713 fpop = new_bd_ia32_ffreep(NULL, get_nodes_block(n));
715 fpop = new_bd_ia32_fpop(NULL, get_nodes_block(n));
716 attr = get_ia32_x87_attr(fpop);
717 attr->x87[0] = get_st_reg(0);
718 attr->x87[1] = get_st_reg(0);
719 attr->x87[2] = get_st_reg(0);
722 sched_add_before(n, fpop);
723 DB((dbg, LEVEL_1, "<<< %s %s\n", get_irn_opname(fpop), attr->x87[0]->name));
728 } /* x87_create_fpop */
730 /* --------------------------------- liveness ------------------------------------------ */
733 * The liveness transfer function.
734 * Updates a live set over a single step from a given node to its predecessor.
735 * Everything defined at the node is removed from the set, the uses of the node get inserted.
737 * @param irn The node at which liveness should be computed.
738 * @param live The bitset of registers live before @p irn. This set gets modified by updating it to
739 * the registers live after irn.
741 * @return The live bitset.
743 static vfp_liveness vfp_liveness_transfer(ir_node *irn, vfp_liveness live)
746 const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
748 if (get_irn_mode(irn) == mode_T) {
749 const ir_edge_t *edge;
751 foreach_out_edge(irn, edge) {
752 ir_node *proj = get_edge_src_irn(edge);
754 if (arch_irn_consider_in_reg_alloc(cls, proj)) {
755 const arch_register_t *reg = x87_get_irn_register(proj);
756 live &= ~(1 << arch_register_get_index(reg));
759 } else if (arch_irn_consider_in_reg_alloc(cls, irn)) {
760 const arch_register_t *reg = x87_get_irn_register(irn);
761 live &= ~(1 << arch_register_get_index(reg));
764 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
765 ir_node *op = get_irn_n(irn, i);
767 if (mode_is_float(get_irn_mode(op)) &&
768 arch_irn_consider_in_reg_alloc(cls, op)) {
769 const arch_register_t *reg = x87_get_irn_register(op);
770 live |= 1 << arch_register_get_index(reg);
774 } /* vfp_liveness_transfer */
777 * Put all live virtual registers at the end of a block into a bitset.
779 * @param sim the simulator handle
780 * @param lv the liveness information
781 * @param bl the block
783 * @return The live bitset at the end of this block
785 static vfp_liveness vfp_liveness_end_of_block(x87_simulator *sim, const ir_node *block)
788 vfp_liveness live = 0;
789 const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
790 const be_lv_t *lv = sim->lv;
792 be_lv_foreach(lv, block, be_lv_state_end, i) {
793 const arch_register_t *reg;
794 const ir_node *node = be_lv_get_irn(lv, block, i);
795 if (!arch_irn_consider_in_reg_alloc(cls, node))
798 reg = x87_get_irn_register(node);
799 live |= 1 << arch_register_get_index(reg);
803 } /* vfp_liveness_end_of_block */
805 /** get the register mask from an arch_register */
806 #define REGMASK(reg) (1 << (arch_register_get_index(reg)))
809 * Return a bitset of argument registers which are live at the end of a node.
811 * @param sim the simulator handle
812 * @param pos the node
813 * @param kill kill mask for the output registers
815 * @return The live bitset.
817 static unsigned vfp_live_args_after(x87_simulator *sim, const ir_node *pos, unsigned kill)
819 unsigned idx = get_irn_idx(pos);
821 assert(idx < sim->n_idx);
822 return sim->live[idx] & ~kill;
823 } /* vfp_live_args_after */
826 * Calculate the liveness for a whole block and cache it.
828 * @param sim the simulator handle
829 * @param lv the liveness handle
830 * @param block the block
832 static void update_liveness(x87_simulator *sim, ir_node *block)
834 vfp_liveness live = vfp_liveness_end_of_block(sim, block);
838 /* now iterate through the block backward and cache the results */
839 sched_foreach_reverse(block, irn) {
840 /* stop at the first Phi: this produces the live-in */
844 idx = get_irn_idx(irn);
845 sim->live[idx] = live;
847 live = vfp_liveness_transfer(irn, live);
849 idx = get_irn_idx(block);
850 sim->live[idx] = live;
851 } /* update_liveness */
854 * Returns true if a register is live in a set.
856 * @param reg_idx the vfp register index
857 * @param live a live bitset
859 #define is_vfp_live(reg_idx, live) ((live) & (1 << (reg_idx)))
863 * Dump liveness info.
865 * @param live the live bitset
867 static void vfp_dump_live(vfp_liveness live)
871 DB((dbg, LEVEL_2, "Live after: "));
872 for (i = 0; i < 8; ++i) {
873 if (live & (1 << i)) {
874 DB((dbg, LEVEL_2, "vf%d ", i));
877 DB((dbg, LEVEL_2, "\n"));
878 } /* vfp_dump_live */
879 #endif /* DEBUG_libfirm */
881 /* --------------------------------- simulators ---------------------------------------- */
884 * Simulate a virtual binop.
886 * @param state the x87 state
887 * @param n the node that should be simulated (and patched)
888 * @param tmpl the template containing the 4 possible x87 opcodes
890 * @return NO_NODE_ADDED
892 static int sim_binop(x87_state *state, ir_node *n, const exchange_tmpl *tmpl)
894 int op2_idx = 0, op1_idx;
895 int out_idx, do_pop = 0;
896 ia32_x87_attr_t *attr;
898 ir_node *patched_insn;
900 x87_simulator *sim = state->sim;
901 ir_node *op1 = get_irn_n(n, n_ia32_binary_left);
902 ir_node *op2 = get_irn_n(n, n_ia32_binary_right);
903 const arch_register_t *op1_reg = x87_get_irn_register(op1);
904 const arch_register_t *op2_reg = x87_get_irn_register(op2);
905 const arch_register_t *out = x87_irn_get_register(n, pn_ia32_res);
906 int reg_index_1 = arch_register_get_index(op1_reg);
907 int reg_index_2 = arch_register_get_index(op2_reg);
908 vfp_liveness live = vfp_live_args_after(sim, n, REGMASK(out));
912 DB((dbg, LEVEL_1, ">>> %+F %s, %s -> %s\n", n,
913 arch_register_get_name(op1_reg), arch_register_get_name(op2_reg),
914 arch_register_get_name(out)));
915 DEBUG_ONLY(vfp_dump_live(live));
916 DB((dbg, LEVEL_1, "Stack before: "));
917 DEBUG_ONLY(x87_dump_stack(state));
919 op1_idx = x87_on_stack(state, reg_index_1);
920 assert(op1_idx >= 0);
921 op1_live_after = is_vfp_live(reg_index_1, live);
923 attr = get_ia32_x87_attr(n);
924 permuted = attr->attr.data.ins_permuted;
926 if (reg_index_2 != REG_VFP_VFP_NOREG) {
929 /* second operand is a vfp register */
930 op2_idx = x87_on_stack(state, reg_index_2);
931 assert(op2_idx >= 0);
932 op2_live_after = is_vfp_live(reg_index_2, live);
934 if (op2_live_after) {
935 /* Second operand is live. */
937 if (op1_live_after) {
938 /* Both operands are live: push the first one.
939 This works even for op1 == op2. */
940 x87_create_fpush(state, n, op1_idx, n_ia32_binary_right);
941 /* now do fxxx (tos=tos X op) */
945 dst = tmpl->normal_op;
947 /* Second live, first operand is dead here, bring it to tos. */
949 x87_create_fxch(state, n, op1_idx);
954 /* now do fxxx (tos=tos X op) */
956 dst = tmpl->normal_op;
959 /* Second operand is dead. */
960 if (op1_live_after) {
961 /* First operand is live: bring second to tos. */
963 x87_create_fxch(state, n, op2_idx);
968 /* now do fxxxr (tos = op X tos) */
970 dst = tmpl->reverse_op;
972 /* Both operands are dead here, pop them from the stack. */
975 /* Both are identically and on tos, no pop needed. */
976 /* here fxxx (tos = tos X tos) */
977 dst = tmpl->normal_op;
980 /* now do fxxxp (op = op X tos, pop) */
981 dst = tmpl->normal_pop_op;
985 } else if (op1_idx == 0) {
986 assert(op1_idx != op2_idx);
987 /* now do fxxxrp (op = tos X op, pop) */
988 dst = tmpl->reverse_pop_op;
992 /* Bring the second on top. */
993 x87_create_fxch(state, n, op2_idx);
994 if (op1_idx == op2_idx) {
995 /* Both are identically and on tos now, no pop needed. */
998 /* use fxxx (tos = tos X tos) */
999 dst = tmpl->normal_op;
1002 /* op2 is on tos now */
1004 /* use fxxxp (op = op X tos, pop) */
1005 dst = tmpl->normal_pop_op;
1013 /* second operand is an address mode */
1014 if (op1_live_after) {
1015 /* first operand is live: push it here */
1016 x87_create_fpush(state, n, op1_idx, n_ia32_binary_left);
1019 /* first operand is dead: bring it to tos */
1021 x87_create_fxch(state, n, op1_idx);
1026 /* use fxxx (tos = tos X mem) */
1027 dst = permuted ? tmpl->reverse_op : tmpl->normal_op;
1031 patched_insn = x87_patch_insn(n, dst);
1032 x87_set_st(state, arch_register_get_index(out), patched_insn, out_idx);
1037 /* patch the operation */
1038 attr->x87[0] = op1_reg = get_st_reg(op1_idx);
1039 if (reg_index_2 != REG_VFP_VFP_NOREG) {
1040 attr->x87[1] = op2_reg = get_st_reg(op2_idx);
1042 attr->x87[2] = out = get_st_reg(out_idx);
1044 if (reg_index_2 != REG_VFP_VFP_NOREG) {
1045 DB((dbg, LEVEL_1, "<<< %s %s, %s -> %s\n", get_irn_opname(n),
1046 arch_register_get_name(op1_reg), arch_register_get_name(op2_reg),
1047 arch_register_get_name(out)));
1049 DB((dbg, LEVEL_1, "<<< %s %s, [AM] -> %s\n", get_irn_opname(n),
1050 arch_register_get_name(op1_reg),
1051 arch_register_get_name(out)));
1054 return NO_NODE_ADDED;
1058 * Simulate a virtual Unop.
1060 * @param state the x87 state
1061 * @param n the node that should be simulated (and patched)
1062 * @param op the x87 opcode that will replace n's opcode
1064 * @return NO_NODE_ADDED
1066 static int sim_unop(x87_state *state, ir_node *n, ir_op *op)
1069 x87_simulator *sim = state->sim;
1070 const arch_register_t *op1 = x87_get_irn_register(get_irn_n(n, 0));
1071 const arch_register_t *out = x87_get_irn_register(n);
1072 ia32_x87_attr_t *attr;
1073 unsigned live = vfp_live_args_after(sim, n, REGMASK(out));
1075 DB((dbg, LEVEL_1, ">>> %+F -> %s\n", n, out->name));
1076 DEBUG_ONLY(vfp_dump_live(live));
1078 op1_idx = x87_on_stack(state, arch_register_get_index(op1));
1080 if (is_vfp_live(arch_register_get_index(op1), live)) {
1081 /* push the operand here */
1082 x87_create_fpush(state, n, op1_idx, 0);
1086 /* operand is dead, bring it to tos */
1088 x87_create_fxch(state, n, op1_idx);
1093 x87_set_tos(state, arch_register_get_index(out), x87_patch_insn(n, op));
1094 attr = get_ia32_x87_attr(n);
1095 attr->x87[0] = op1 = get_st_reg(0);
1096 attr->x87[2] = out = get_st_reg(0);
1097 DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), out->name));
1099 return NO_NODE_ADDED;
1103 * Simulate a virtual Load instruction.
1105 * @param state the x87 state
1106 * @param n the node that should be simulated (and patched)
1107 * @param op the x87 opcode that will replace n's opcode
1109 * @return NO_NODE_ADDED
1111 static int sim_load(x87_state *state, ir_node *n, ir_op *op, int res_pos)
1113 const arch_register_t *out = x87_irn_get_register(n, res_pos);
1114 ia32_x87_attr_t *attr;
1116 DB((dbg, LEVEL_1, ">>> %+F -> %s\n", n, arch_register_get_name(out)));
1117 x87_push(state, arch_register_get_index(out), x87_patch_insn(n, op));
1118 assert(out == x87_irn_get_register(n, res_pos));
1119 attr = get_ia32_x87_attr(n);
1120 attr->x87[2] = out = get_st_reg(0);
1121 DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), arch_register_get_name(out)));
1123 return NO_NODE_ADDED;
1127 * Rewire all users of @p old_val to @new_val iff they are scheduled after @p store.
1129 * @param store The store
1130 * @param old_val The former value
1131 * @param new_val The new value
1133 static void collect_and_rewire_users(ir_node *store, ir_node *old_val, ir_node *new_val)
1135 const ir_edge_t *edge, *ne;
1137 foreach_out_edge_safe(old_val, edge, ne) {
1138 ir_node *user = get_edge_src_irn(edge);
1140 if (! user || user == store)
1143 /* if the user is scheduled after the store: rewire */
1144 if (sched_is_scheduled(user) && sched_comes_after(store, user)) {
1146 /* find the input of the user pointing to the old value */
1147 for (i = get_irn_arity(user) - 1; i >= 0; i--) {
1148 if (get_irn_n(user, i) == old_val)
1149 set_irn_n(user, i, new_val);
1153 } /* collect_and_rewire_users */
1156 * Simulate a virtual Store.
1158 * @param state the x87 state
1159 * @param n the node that should be simulated (and patched)
1160 * @param op the x87 store opcode
1161 * @param op_p the x87 store and pop opcode
1163 static int sim_store(x87_state *state, ir_node *n, ir_op *op, ir_op *op_p)
1165 ir_node *val = get_irn_n(n, n_ia32_vfst_val);
1166 const arch_register_t *op2 = x87_get_irn_register(val);
1167 unsigned live = vfp_live_args_after(state->sim, n, 0);
1168 int insn = NO_NODE_ADDED;
1169 ia32_x87_attr_t *attr;
1170 int op2_reg_idx, op2_idx, depth;
1171 int live_after_node;
1174 op2_reg_idx = arch_register_get_index(op2);
1175 op2_idx = x87_on_stack(state, op2_reg_idx);
1176 live_after_node = is_vfp_live(arch_register_get_index(op2), live);
1177 DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
1178 assert(op2_idx >= 0);
1180 mode = get_ia32_ls_mode(n);
1181 depth = x87_get_depth(state);
1183 if (live_after_node) {
1185 Problem: fst doesn't support 96bit modes (spills), only fstp does
1186 fist doesn't support 64bit mode, only fistp
1188 - stack not full: push value and fstp
1189 - stack full: fstp value and load again
1190 Note that we cannot test on mode_E, because floats might be 96bit ...
1192 if (get_mode_size_bits(mode) > 64 || (mode_is_int(mode) && get_mode_size_bits(mode) > 32)) {
1193 if (depth < N_ia32_st_REGS) {
1194 /* ok, we have a free register: push + fstp */
1195 x87_create_fpush(state, n, op2_idx, n_ia32_vfst_val);
1197 x87_patch_insn(n, op_p);
1199 ir_node *vfld, *mem, *block, *rproj, *mproj;
1200 ir_graph *irg = get_irn_irg(n);
1202 /* stack full here: need fstp + load */
1204 x87_patch_insn(n, op_p);
1206 block = get_nodes_block(n);
1207 vfld = new_bd_ia32_vfld(NULL, block, get_irn_n(n, 0), get_irn_n(n, 1), new_r_NoMem(irg), get_ia32_ls_mode(n));
1209 /* copy all attributes */
1210 set_ia32_frame_ent(vfld, get_ia32_frame_ent(n));
1211 if (is_ia32_use_frame(n))
1212 set_ia32_use_frame(vfld);
1213 set_ia32_op_type(vfld, ia32_AddrModeS);
1214 add_ia32_am_offs_int(vfld, get_ia32_am_offs_int(n));
1215 set_ia32_am_sc(vfld, get_ia32_am_sc(n));
1216 set_ia32_ls_mode(vfld, get_ia32_ls_mode(n));
1218 rproj = new_r_Proj(vfld, get_ia32_ls_mode(vfld), pn_ia32_vfld_res);
1219 mproj = new_r_Proj(vfld, mode_M, pn_ia32_vfld_M);
1220 mem = get_irn_Proj_for_mode(n, mode_M);
1222 assert(mem && "Store memory not found");
1224 arch_set_irn_register(rproj, op2);
1226 /* reroute all former users of the store memory to the load memory */
1227 edges_reroute(mem, mproj, irg);
1228 /* set the memory input of the load to the store memory */
1229 set_irn_n(vfld, n_ia32_vfld_mem, mem);
1231 sched_add_after(n, vfld);
1232 sched_add_after(vfld, rproj);
1234 /* rewire all users, scheduled after the store, to the loaded value */
1235 collect_and_rewire_users(n, val, rproj);
1240 /* we can only store the tos to memory */
1242 x87_create_fxch(state, n, op2_idx);
1244 /* mode size 64 or smaller -> use normal fst */
1245 x87_patch_insn(n, op);
1248 /* we can only store the tos to memory */
1250 x87_create_fxch(state, n, op2_idx);
1253 x87_patch_insn(n, op_p);
1256 attr = get_ia32_x87_attr(n);
1257 attr->x87[1] = op2 = get_st_reg(0);
1258 DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2)));
1263 #define _GEN_BINOP(op, rev) \
1264 static int sim_##op(x87_state *state, ir_node *n) { \
1265 exchange_tmpl tmpl = { op_ia32_##op, op_ia32_##rev, op_ia32_##op##p, op_ia32_##rev##p }; \
1266 return sim_binop(state, n, &tmpl); \
1269 #define GEN_BINOP(op) _GEN_BINOP(op, op)
1270 #define GEN_BINOPR(op) _GEN_BINOP(op, op##r)
1272 #define GEN_LOAD(op) \
1273 static int sim_##op(x87_state *state, ir_node *n) { \
1274 return sim_load(state, n, op_ia32_##op, pn_ia32_v##op##_res); \
1277 #define GEN_UNOP(op) \
1278 static int sim_##op(x87_state *state, ir_node *n) { \
1279 return sim_unop(state, n, op_ia32_##op); \
1282 #define GEN_STORE(op) \
1283 static int sim_##op(x87_state *state, ir_node *n) { \
1284 return sim_store(state, n, op_ia32_##op, op_ia32_##op##p); \
1306 * Simulate a virtual fisttp.
1308 * @param state the x87 state
1309 * @param n the node that should be simulated (and patched)
1311 * @return NO_NODE_ADDED
1313 static int sim_fisttp(x87_state *state, ir_node *n)
1315 ir_node *val = get_irn_n(n, n_ia32_vfst_val);
1316 const arch_register_t *op2 = x87_get_irn_register(val);
1317 ia32_x87_attr_t *attr;
1318 int op2_reg_idx, op2_idx;
1320 op2_reg_idx = arch_register_get_index(op2);
1321 op2_idx = x87_on_stack(state, op2_reg_idx);
1322 DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
1323 assert(op2_idx >= 0);
1325 /* Note: although the value is still live here, it is destroyed because
1326 of the pop. The register allocator is aware of that and introduced a copy
1327 if the value must be alive. */
1329 /* we can only store the tos to memory */
1331 x87_create_fxch(state, n, op2_idx);
1334 x87_patch_insn(n, op_ia32_fisttp);
1336 attr = get_ia32_x87_attr(n);
1337 attr->x87[1] = op2 = get_st_reg(0);
1338 DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2)));
1340 return NO_NODE_ADDED;
1344 * Simulate a virtual FtstFnstsw.
1346 * @param state the x87 state
1347 * @param n the node that should be simulated (and patched)
1349 * @return NO_NODE_ADDED
1351 static int sim_FtstFnstsw(x87_state *state, ir_node *n)
1353 x87_simulator *sim = state->sim;
1354 ia32_x87_attr_t *attr = get_ia32_x87_attr(n);
1355 ir_node *op1_node = get_irn_n(n, n_ia32_vFtstFnstsw_left);
1356 const arch_register_t *reg1 = x87_get_irn_register(op1_node);
1357 int reg_index_1 = arch_register_get_index(reg1);
1358 int op1_idx = x87_on_stack(state, reg_index_1);
1359 unsigned live = vfp_live_args_after(sim, n, 0);
1361 DB((dbg, LEVEL_1, ">>> %+F %s\n", n, arch_register_get_name(reg1)));
1362 DEBUG_ONLY(vfp_dump_live(live));
1363 DB((dbg, LEVEL_1, "Stack before: "));
1364 DEBUG_ONLY(x87_dump_stack(state));
1365 assert(op1_idx >= 0);
1368 /* bring the value to tos */
1369 x87_create_fxch(state, n, op1_idx);
1373 /* patch the operation */
1374 x87_patch_insn(n, op_ia32_FtstFnstsw);
1375 reg1 = get_st_reg(op1_idx);
1376 attr->x87[0] = reg1;
1377 attr->x87[1] = NULL;
1378 attr->x87[2] = NULL;
1380 if (!is_vfp_live(reg_index_1, live))
1381 x87_create_fpop(state, sched_next(n), 1);
1383 return NO_NODE_ADDED;
1384 } /* sim_FtstFnstsw */
1389 * @param state the x87 state
1390 * @param n the node that should be simulated (and patched)
1392 * @return NO_NODE_ADDED
1394 static int sim_Fucom(x87_state *state, ir_node *n)
1398 ia32_x87_attr_t *attr = get_ia32_x87_attr(n);
1400 x87_simulator *sim = state->sim;
1401 ir_node *op1_node = get_irn_n(n, n_ia32_vFucomFnstsw_left);
1402 ir_node *op2_node = get_irn_n(n, n_ia32_vFucomFnstsw_right);
1403 const arch_register_t *op1 = x87_get_irn_register(op1_node);
1404 const arch_register_t *op2 = x87_get_irn_register(op2_node);
1405 int reg_index_1 = arch_register_get_index(op1);
1406 int reg_index_2 = arch_register_get_index(op2);
1407 unsigned live = vfp_live_args_after(sim, n, 0);
1408 bool permuted = attr->attr.data.ins_permuted;
1412 DB((dbg, LEVEL_1, ">>> %+F %s, %s\n", n,
1413 arch_register_get_name(op1), arch_register_get_name(op2)));
1414 DEBUG_ONLY(vfp_dump_live(live));
1415 DB((dbg, LEVEL_1, "Stack before: "));
1416 DEBUG_ONLY(x87_dump_stack(state));
1418 op1_idx = x87_on_stack(state, reg_index_1);
1419 assert(op1_idx >= 0);
1421 /* BEWARE: check for comp a,a cases, they might happen */
1422 if (reg_index_2 != REG_VFP_VFP_NOREG) {
1423 /* second operand is a vfp register */
1424 op2_idx = x87_on_stack(state, reg_index_2);
1425 assert(op2_idx >= 0);
1427 if (is_vfp_live(reg_index_2, live)) {
1428 /* second operand is live */
1430 if (is_vfp_live(reg_index_1, live)) {
1431 /* both operands are live */
1434 /* res = tos X op */
1435 } else if (op2_idx == 0) {
1436 /* res = op X tos */
1437 permuted = !permuted;
1440 /* bring the first one to tos */
1441 x87_create_fxch(state, n, op1_idx);
1442 if (op1_idx == op2_idx) {
1444 } else if (op2_idx == 0) {
1448 /* res = tos X op */
1451 /* second live, first operand is dead here, bring it to tos.
1452 This means further, op1_idx != op2_idx. */
1453 assert(op1_idx != op2_idx);
1455 x87_create_fxch(state, n, op1_idx);
1460 /* res = tos X op, pop */
1464 /* second operand is dead */
1465 if (is_vfp_live(reg_index_1, live)) {
1466 /* first operand is live: bring second to tos.
1467 This means further, op1_idx != op2_idx. */
1468 assert(op1_idx != op2_idx);
1470 x87_create_fxch(state, n, op2_idx);
1475 /* res = op X tos, pop */
1477 permuted = !permuted;
1480 /* both operands are dead here, check first for identity. */
1481 if (op1_idx == op2_idx) {
1482 /* identically, one pop needed */
1484 x87_create_fxch(state, n, op1_idx);
1488 /* res = tos X op, pop */
1491 /* different, move them to st and st(1) and pop both.
1492 The tricky part is to get one into st(1).*/
1493 else if (op2_idx == 1) {
1494 /* good, second operand is already in the right place, move the first */
1496 /* bring the first on top */
1497 x87_create_fxch(state, n, op1_idx);
1498 assert(op2_idx != 0);
1501 /* res = tos X op, pop, pop */
1503 } else if (op1_idx == 1) {
1504 /* good, first operand is already in the right place, move the second */
1506 /* bring the first on top */
1507 x87_create_fxch(state, n, op2_idx);
1508 assert(op1_idx != 0);
1511 /* res = op X tos, pop, pop */
1512 permuted = !permuted;
1516 /* if one is already the TOS, we need two fxch */
1518 /* first one is TOS, move to st(1) */
1519 x87_create_fxch(state, n, 1);
1520 assert(op2_idx != 1);
1522 x87_create_fxch(state, n, op2_idx);
1524 /* res = op X tos, pop, pop */
1526 permuted = !permuted;
1528 } else if (op2_idx == 0) {
1529 /* second one is TOS, move to st(1) */
1530 x87_create_fxch(state, n, 1);
1531 assert(op1_idx != 1);
1533 x87_create_fxch(state, n, op1_idx);
1535 /* res = tos X op, pop, pop */
1538 /* none of them is either TOS or st(1), 3 fxch needed */
1539 x87_create_fxch(state, n, op2_idx);
1540 assert(op1_idx != 0);
1541 x87_create_fxch(state, n, 1);
1543 x87_create_fxch(state, n, op1_idx);
1545 /* res = tos X op, pop, pop */
1552 /* second operand is an address mode */
1553 if (is_vfp_live(reg_index_1, live)) {
1554 /* first operand is live: bring it to TOS */
1556 x87_create_fxch(state, n, op1_idx);
1560 /* first operand is dead: bring it to tos */
1562 x87_create_fxch(state, n, op1_idx);
1569 /* patch the operation */
1570 if (is_ia32_vFucomFnstsw(n)) {
1574 case 0: dst = op_ia32_FucomFnstsw; break;
1575 case 1: dst = op_ia32_FucompFnstsw; break;
1576 case 2: dst = op_ia32_FucomppFnstsw; break;
1577 default: panic("invalid popcount in sim_Fucom");
1580 for (i = 0; i < pops; ++i) {
1583 } else if (is_ia32_vFucomi(n)) {
1585 case 0: dst = op_ia32_Fucomi; break;
1586 case 1: dst = op_ia32_Fucompi; x87_pop(state); break;
1588 dst = op_ia32_Fucompi;
1590 x87_create_fpop(state, sched_next(n), 1);
1592 default: panic("invalid popcount in sim_Fucom");
1595 panic("invalid operation %+F in sim_FucomFnstsw", n);
1598 x87_patch_insn(n, dst);
1605 op1 = get_st_reg(op1_idx);
1608 op2 = get_st_reg(op2_idx);
1611 attr->x87[2] = NULL;
1612 attr->attr.data.ins_permuted = permuted;
1615 DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(n),
1616 arch_register_get_name(op1), arch_register_get_name(op2)));
1618 DB((dbg, LEVEL_1, "<<< %s %s, [AM]\n", get_irn_opname(n),
1619 arch_register_get_name(op1)));
1622 return NO_NODE_ADDED;
1628 * @param state the x87 state
1629 * @param n the node that should be simulated (and patched)
1631 * @return NO_NODE_ADDED
1633 static int sim_Keep(x87_state *state, ir_node *node)
1636 const arch_register_t *op_reg;
1642 DB((dbg, LEVEL_1, ">>> %+F\n", node));
1644 arity = get_irn_arity(node);
1645 for (i = 0; i < arity; ++i) {
1646 op = get_irn_n(node, i);
1647 op_reg = arch_get_irn_register(op);
1648 if (arch_register_get_class(op_reg) != &ia32_reg_classes[CLASS_ia32_vfp])
1651 reg_id = arch_register_get_index(op_reg);
1652 live = vfp_live_args_after(state->sim, node, 0);
1654 op_stack_idx = x87_on_stack(state, reg_id);
1655 if (op_stack_idx >= 0 && !is_vfp_live(reg_id, live))
1656 x87_create_fpop(state, sched_next(node), 1);
1659 DB((dbg, LEVEL_1, "Stack after: "));
1660 DEBUG_ONLY(x87_dump_stack(state));
1662 return NO_NODE_ADDED;
1666 * Keep the given node alive by adding a be_Keep.
1668 * @param node the node to kept alive
1670 static void keep_float_node_alive(ir_node *node)
1672 ir_node *block = get_nodes_block(node);
1673 ir_node *keep = be_new_Keep(block, 1, &node);
1675 assert(sched_is_scheduled(node));
1676 sched_add_after(node, keep);
1680 * Create a copy of a node. Recreate the node if it's a constant.
1682 * @param state the x87 state
1683 * @param n the node to be copied
1685 * @return the copy of n
1687 static ir_node *create_Copy(x87_state *state, ir_node *n)
1689 dbg_info *n_dbg = get_irn_dbg_info(n);
1690 ir_mode *mode = get_irn_mode(n);
1691 ir_node *block = get_nodes_block(n);
1692 ir_node *pred = get_irn_n(n, 0);
1693 ir_node *(*cnstr)(dbg_info *, ir_node *, ir_mode *) = NULL;
1695 const arch_register_t *out;
1696 const arch_register_t *op1;
1697 ia32_x87_attr_t *attr;
1699 /* Do not copy constants, recreate them. */
1700 switch (get_ia32_irn_opcode(pred)) {
1702 cnstr = new_bd_ia32_fldz;
1705 cnstr = new_bd_ia32_fld1;
1707 case iro_ia32_fldpi:
1708 cnstr = new_bd_ia32_fldpi;
1710 case iro_ia32_fldl2e:
1711 cnstr = new_bd_ia32_fldl2e;
1713 case iro_ia32_fldl2t:
1714 cnstr = new_bd_ia32_fldl2t;
1716 case iro_ia32_fldlg2:
1717 cnstr = new_bd_ia32_fldlg2;
1719 case iro_ia32_fldln2:
1720 cnstr = new_bd_ia32_fldln2;
1726 out = x87_get_irn_register(n);
1727 op1 = x87_get_irn_register(pred);
1729 if (cnstr != NULL) {
1730 /* copy a constant */
1731 res = (*cnstr)(n_dbg, block, mode);
1733 x87_push(state, arch_register_get_index(out), res);
1735 attr = get_ia32_x87_attr(res);
1736 attr->x87[2] = get_st_reg(0);
1738 int op1_idx = x87_on_stack(state, arch_register_get_index(op1));
1740 res = new_bd_ia32_fpushCopy(n_dbg, block, pred, mode);
1742 x87_push(state, arch_register_get_index(out), res);
1744 attr = get_ia32_x87_attr(res);
1745 attr->x87[0] = get_st_reg(op1_idx);
1746 attr->x87[2] = get_st_reg(0);
1748 arch_set_irn_register(res, out);
1754 * Simulate a be_Copy.
1756 * @param state the x87 state
1757 * @param n the node that should be simulated (and patched)
1759 * @return NO_NODE_ADDED
1761 static int sim_Copy(x87_state *state, ir_node *n)
1764 const arch_register_t *out;
1765 const arch_register_t *op1;
1766 const arch_register_class_t *cls;
1767 ir_node *node, *next;
1768 int op1_idx, out_idx;
1771 cls = arch_get_irn_reg_class_out(n);
1772 if (cls != &ia32_reg_classes[CLASS_ia32_vfp])
1775 pred = get_irn_n(n, 0);
1776 out = x87_get_irn_register(n);
1777 op1 = x87_get_irn_register(pred);
1778 live = vfp_live_args_after(state->sim, n, REGMASK(out));
1780 DB((dbg, LEVEL_1, ">>> %+F %s -> %s\n", n,
1781 arch_register_get_name(op1), arch_register_get_name(out)));
1782 DEBUG_ONLY(vfp_dump_live(live));
1784 op1_idx = x87_on_stack(state, arch_register_get_index(op1));
1786 if (is_vfp_live(arch_register_get_index(op1), live)) {
1787 ir_node *pred = get_irn_n(n, 0);
1789 /* Operand is still live, a real copy. We need here an fpush that can
1790 hold a a register, so use the fpushCopy or recreate constants */
1791 node = create_Copy(state, n);
1793 /* We have to make sure the old value doesn't go dead (which can happen
1794 * when we recreate constants). As the simulator expected that value in
1795 * the pred blocks. This is unfortunate as removing it would save us 1
1796 * instruction, but we would have to rerun all the simulation to get
1799 next = sched_next(n);
1802 sched_add_before(next, node);
1804 if (get_irn_n_edges(pred) == 0) {
1805 keep_float_node_alive(pred);
1808 DB((dbg, LEVEL_1, "<<< %+F %s -> ?\n", node, op1->name));
1810 out_idx = x87_on_stack(state, arch_register_get_index(out));
1812 if (out_idx >= 0 && out_idx != op1_idx) {
1813 /* Matze: out already on stack? how can this happen? */
1814 panic("invalid stack state in x87 simulator");
1817 /* op1 must be killed and placed where out is */
1819 ia32_x87_attr_t *attr;
1820 /* best case, simple remove and rename */
1821 x87_patch_insn(n, op_ia32_Pop);
1822 attr = get_ia32_x87_attr(n);
1823 attr->x87[0] = op1 = get_st_reg(0);
1826 x87_set_st(state, arch_register_get_index(out), n, op1_idx - 1);
1828 ia32_x87_attr_t *attr;
1829 /* move op1 to tos, store and pop it */
1831 x87_create_fxch(state, n, op1_idx);
1834 x87_patch_insn(n, op_ia32_Pop);
1835 attr = get_ia32_x87_attr(n);
1836 attr->x87[0] = op1 = get_st_reg(out_idx);
1839 x87_set_st(state, arch_register_get_index(out), n, out_idx - 1);
1841 DB((dbg, LEVEL_1, "<<< %+F %s\n", n, op1->name));
1844 /* just a virtual copy */
1845 x87_set_st(state, arch_register_get_index(out), get_unop_op(n), op1_idx);
1846 /* don't remove the node to keep the verifier quiet :),
1847 the emitter won't emit any code for the node */
1850 DB((dbg, LEVEL_1, "<<< KILLED %s\n", get_irn_opname(n)));
1851 exchange(n, get_unop_op(n));
1855 return NO_NODE_ADDED;
1859 * Returns the vf0 result Proj of a Call.
1861 * @para call the Call node
1863 static ir_node *get_call_result_proj(ir_node *call)
1865 const ir_edge_t *edge;
1867 /* search the result proj */
1868 foreach_out_edge(call, edge) {
1869 ir_node *proj = get_edge_src_irn(edge);
1870 long pn = get_Proj_proj(proj);
1872 if (pn == pn_ia32_Call_vf0)
1877 } /* get_call_result_proj */
1880 * Simulate a ia32_Call.
1882 * @param state the x87 state
1883 * @param n the node that should be simulated (and patched)
1885 * @return NO_NODE_ADDED
1887 static int sim_Call(x87_state *state, ir_node *n)
1889 ir_type *call_tp = get_ia32_call_attr_const(n)->call_tp;
1893 const arch_register_t *reg;
1895 DB((dbg, LEVEL_1, ">>> %+F\n", n));
1897 /* at the begin of a call the x87 state should be empty */
1898 assert(state->depth == 0 && "stack not empty before call");
1900 if (get_method_n_ress(call_tp) <= 0)
1904 * If the called function returns a float, it is returned in st(0).
1905 * This even happens if the return value is NOT used.
1906 * Moreover, only one return result is supported.
1908 res_type = get_method_res_type(call_tp, 0);
1909 mode = get_type_mode(res_type);
1911 if (mode == NULL || !mode_is_float(mode))
1914 resproj = get_call_result_proj(n);
1915 assert(resproj != NULL);
1917 reg = x87_get_irn_register(resproj);
1918 x87_push(state, arch_register_get_index(reg), resproj);
1921 DB((dbg, LEVEL_1, "Stack after: "));
1922 DEBUG_ONLY(x87_dump_stack(state));
1924 return NO_NODE_ADDED;
1928 * Simulate a be_Return.
1930 * @param state the x87 state
1931 * @param n the node that should be simulated (and patched)
1933 * @return NO_NODE_ADDED
1935 static int sim_Return(x87_state *state, ir_node *n)
1937 int n_res = be_Return_get_n_rets(n);
1938 int i, n_float_res = 0;
1940 /* only floating point return values must reside on stack */
1941 for (i = 0; i < n_res; ++i) {
1942 ir_node *res = get_irn_n(n, be_pos_Return_val + i);
1944 if (mode_is_float(get_irn_mode(res)))
1947 assert(x87_get_depth(state) == n_float_res);
1949 /* pop them virtually */
1950 for (i = n_float_res - 1; i >= 0; --i)
1953 return NO_NODE_ADDED;
1956 typedef struct perm_data_t {
1957 const arch_register_t *in;
1958 const arch_register_t *out;
1962 * Simulate a be_Perm.
1964 * @param state the x87 state
1965 * @param irn the node that should be simulated (and patched)
1967 * @return NO_NODE_ADDED
1969 static int sim_Perm(x87_state *state, ir_node *irn)
1972 ir_node *pred = get_irn_n(irn, 0);
1974 const ir_edge_t *edge;
1976 /* handle only floating point Perms */
1977 if (! mode_is_float(get_irn_mode(pred)))
1978 return NO_NODE_ADDED;
1980 DB((dbg, LEVEL_1, ">>> %+F\n", irn));
1982 /* Perm is a pure virtual instruction on x87.
1983 All inputs must be on the FPU stack and are pairwise
1984 different from each other.
1985 So, all we need to do is to permutate the stack state. */
1986 n = get_irn_arity(irn);
1987 NEW_ARR_A(int, stack_pos, n);
1989 /* collect old stack positions */
1990 for (i = 0; i < n; ++i) {
1991 const arch_register_t *inreg = x87_get_irn_register(get_irn_n(irn, i));
1992 int idx = x87_on_stack(state, arch_register_get_index(inreg));
1994 assert(idx >= 0 && "Perm argument not on x87 stack");
1998 /* now do the permutation */
1999 foreach_out_edge(irn, edge) {
2000 ir_node *proj = get_edge_src_irn(edge);
2001 const arch_register_t *out = x87_get_irn_register(proj);
2002 long num = get_Proj_proj(proj);
2004 assert(0 <= num && num < n && "More Proj's than Perm inputs");
2005 x87_set_st(state, arch_register_get_index(out), proj, stack_pos[(unsigned)num]);
2007 DB((dbg, LEVEL_1, "<<< %+F\n", irn));
2009 return NO_NODE_ADDED;
2013 * Kill any dead registers at block start by popping them from the stack.
2015 * @param sim the simulator handle
2016 * @param block the current block
2017 * @param start_state the x87 state at the begin of the block
2019 * @return the x87 state after dead register killed
2021 static x87_state *x87_kill_deads(x87_simulator *sim, ir_node *block, x87_state *start_state)
2023 x87_state *state = start_state;
2024 ir_node *first_insn = sched_first(block);
2025 ir_node *keep = NULL;
2026 unsigned live = vfp_live_args_after(sim, block, 0);
2028 int i, depth, num_pop;
2031 depth = x87_get_depth(state);
2032 for (i = depth - 1; i >= 0; --i) {
2033 int reg = x87_get_st_reg(state, i);
2035 if (! is_vfp_live(reg, live))
2036 kill_mask |= (1 << i);
2040 /* create a new state, will be changed */
2041 state = x87_clone_state(sim, state);
2043 DB((dbg, LEVEL_1, "Killing deads:\n"));
2044 DEBUG_ONLY(vfp_dump_live(live));
2045 DEBUG_ONLY(x87_dump_stack(state));
2047 if (kill_mask != 0 && live == 0) {
2048 /* special case: kill all registers */
2049 if (ia32_cg_config.use_femms || ia32_cg_config.use_emms) {
2050 if (ia32_cg_config.use_femms) {
2051 /* use FEMMS on AMD processors to clear all */
2052 keep = new_bd_ia32_femms(NULL, block);
2054 /* use EMMS to clear all */
2055 keep = new_bd_ia32_emms(NULL, block);
2057 sched_add_before(first_insn, keep);
2063 /* now kill registers */
2065 /* we can only kill from TOS, so bring them up */
2066 if (! (kill_mask & 1)) {
2067 /* search from behind, because we can to a double-pop */
2068 for (i = depth - 1; i >= 0; --i) {
2069 if (kill_mask & (1 << i)) {
2070 kill_mask &= ~(1 << i);
2077 x87_set_st(state, -1, keep, i);
2078 x87_create_fxch(state, first_insn, i);
2081 if ((kill_mask & 3) == 3) {
2082 /* we can do a double-pop */
2086 /* only a single pop */
2091 kill_mask >>= num_pop;
2092 keep = x87_create_fpop(state, first_insn, num_pop);
2097 } /* x87_kill_deads */
2100 * Run a simulation and fix all virtual instructions for a block.
2102 * @param sim the simulator handle
2103 * @param block the current block
2105 static void x87_simulate_block(x87_simulator *sim, ir_node *block)
2108 blk_state *bl_state = x87_get_bl_state(sim, block);
2109 x87_state *state = bl_state->begin;
2110 const ir_edge_t *edge;
2111 ir_node *start_block;
2113 assert(state != NULL);
2114 /* already processed? */
2115 if (bl_state->end != NULL)
2118 DB((dbg, LEVEL_1, "Simulate %+F\n", block));
2119 DB((dbg, LEVEL_2, "State at Block begin:\n "));
2120 DEBUG_ONLY(x87_dump_stack(state));
2122 /* at block begin, kill all dead registers */
2123 state = x87_kill_deads(sim, block, state);
2124 /* create a new state, will be changed */
2125 state = x87_clone_state(sim, state);
2127 /* beware, n might change */
2128 for (n = sched_first(block); !sched_is_end(n); n = next) {
2131 ir_op *op = get_irn_op(n);
2134 * get the next node to be simulated here.
2135 * n might be completely removed from the schedule-
2137 next = sched_next(n);
2138 if (op->ops.generic != NULL) {
2139 func = (sim_func)op->ops.generic;
2142 node_inserted = (*func)(state, n);
2145 * sim_func might have added an additional node after n,
2146 * so update next node
2147 * beware: n must not be changed by sim_func
2148 * (i.e. removed from schedule) in this case
2150 if (node_inserted != NO_NODE_ADDED)
2151 next = sched_next(n);
2155 start_block = get_irg_start_block(get_irn_irg(block));
2157 DB((dbg, LEVEL_2, "State at Block end:\n ")); DEBUG_ONLY(x87_dump_stack(state));
2159 /* check if the state must be shuffled */
2160 foreach_block_succ(block, edge) {
2161 ir_node *succ = get_edge_src_irn(edge);
2162 blk_state *succ_state;
2164 if (succ == start_block)
2167 succ_state = x87_get_bl_state(sim, succ);
2169 if (succ_state->begin == NULL) {
2170 DB((dbg, LEVEL_2, "Set begin state for succ %+F:\n", succ));
2171 DEBUG_ONLY(x87_dump_stack(state));
2172 succ_state->begin = state;
2174 waitq_put(sim->worklist, succ);
2176 DB((dbg, LEVEL_2, "succ %+F already has a state, shuffling\n", succ));
2177 /* There is already a begin state for the successor, bad.
2178 Do the necessary permutations.
2179 Note that critical edges are removed, so this is always possible:
2180 If the successor has more than one possible input, then it must
2183 x87_shuffle(sim, block, state, succ, succ_state->begin);
2186 bl_state->end = state;
2187 } /* x87_simulate_block */
2190 * Register a simulator function.
2192 * @param op the opcode to simulate
2193 * @param func the simulator function for the opcode
2195 static void register_sim(ir_op *op, sim_func func)
2197 assert(op->ops.generic == NULL);
2198 op->ops.generic = (op_func) func;
2199 } /* register_sim */
2202 * Create a new x87 simulator.
2204 * @param sim a simulator handle, will be initialized
2205 * @param irg the current graph
2207 static void x87_init_simulator(x87_simulator *sim, ir_graph *irg)
2209 obstack_init(&sim->obst);
2210 sim->blk_states = pmap_create();
2211 sim->n_idx = get_irg_last_idx(irg);
2212 sim->live = obstack_alloc(&sim->obst, sizeof(*sim->live) * sim->n_idx);
2214 DB((dbg, LEVEL_1, "--------------------------------\n"
2215 "x87 Simulator started for %+F\n", irg));
2217 /* set the generic function pointer of instruction we must simulate */
2218 clear_irp_opcodes_generic_func();
2220 register_sim(op_ia32_Call, sim_Call);
2221 register_sim(op_ia32_vfld, sim_fld);
2222 register_sim(op_ia32_vfild, sim_fild);
2223 register_sim(op_ia32_vfld1, sim_fld1);
2224 register_sim(op_ia32_vfldz, sim_fldz);
2225 register_sim(op_ia32_vfadd, sim_fadd);
2226 register_sim(op_ia32_vfsub, sim_fsub);
2227 register_sim(op_ia32_vfmul, sim_fmul);
2228 register_sim(op_ia32_vfdiv, sim_fdiv);
2229 register_sim(op_ia32_vfprem, sim_fprem);
2230 register_sim(op_ia32_vfabs, sim_fabs);
2231 register_sim(op_ia32_vfchs, sim_fchs);
2232 register_sim(op_ia32_vfist, sim_fist);
2233 register_sim(op_ia32_vfisttp, sim_fisttp);
2234 register_sim(op_ia32_vfst, sim_fst);
2235 register_sim(op_ia32_vFtstFnstsw, sim_FtstFnstsw);
2236 register_sim(op_ia32_vFucomFnstsw, sim_Fucom);
2237 register_sim(op_ia32_vFucomi, sim_Fucom);
2238 register_sim(op_be_Copy, sim_Copy);
2239 register_sim(op_be_Return, sim_Return);
2240 register_sim(op_be_Perm, sim_Perm);
2241 register_sim(op_be_Keep, sim_Keep);
2242 } /* x87_init_simulator */
2245 * Destroy a x87 simulator.
2247 * @param sim the simulator handle
2249 static void x87_destroy_simulator(x87_simulator *sim)
2251 pmap_destroy(sim->blk_states);
2252 obstack_free(&sim->obst, NULL);
2253 DB((dbg, LEVEL_1, "x87 Simulator stopped\n\n"));
2254 } /* x87_destroy_simulator */
2257 * Pre-block walker: calculate the liveness information for the block
2258 * and store it into the sim->live cache.
2260 static void update_liveness_walker(ir_node *block, void *data)
2262 x87_simulator *sim = data;
2263 update_liveness(sim, block);
2264 } /* update_liveness_walker */
2267 * Run a simulation and fix all virtual instructions for a graph.
2268 * Replaces all virtual floating point instructions and registers
2271 void x87_simulate_graph(ir_graph *irg)
2273 /* TODO improve code quality (less executed fxch) by using execfreqs */
2275 ir_node *block, *start_block;
2276 blk_state *bl_state;
2279 /* create the simulator */
2280 x87_init_simulator(&sim, irg);
2282 start_block = get_irg_start_block(irg);
2283 bl_state = x87_get_bl_state(&sim, start_block);
2285 /* start with the empty state */
2286 bl_state->begin = empty;
2289 sim.worklist = new_waitq();
2290 waitq_put(sim.worklist, start_block);
2292 be_assure_liveness(irg);
2293 sim.lv = be_get_irg_liveness(irg);
2294 be_liveness_assure_sets(sim.lv);
2296 /* Calculate the liveness for all nodes. We must precalculate this info,
2297 * because the simulator adds new nodes (possible before Phi nodes) which
2298 * would let a lazy calculation fail.
2299 * On the other hand we reduce the computation amount due to
2300 * precaching from O(n^2) to O(n) at the expense of O(n) cache memory.
2302 irg_block_walk_graph(irg, update_liveness_walker, NULL, &sim);
2306 block = waitq_get(sim.worklist);
2307 x87_simulate_block(&sim, block);
2308 } while (! waitq_empty(sim.worklist));
2311 del_waitq(sim.worklist);
2312 x87_destroy_simulator(&sim);
2313 } /* x87_simulate_graph */
2315 /* Initializes the x87 simulator. */
2316 void ia32_init_x87(void)
2318 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.x87");
2319 } /* ia32_init_x87 */