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 is the main ia32 firm backend driver.
23 * @author Christian Wuerdig
29 #include "lc_opts_enum.h"
33 #include "pseudo_irg.h"
38 #include "iredges_t.h"
51 #include "iroptimize.h"
52 #include "instrument.h"
57 #include "../benode.h"
58 #include "../belower.h"
59 #include "../besched.h"
62 #include "../beirgmod.h"
63 #include "../be_dbgout.h"
64 #include "../beblocksched.h"
65 #include "../bemachine.h"
66 #include "../beilpsched.h"
67 #include "../bespillslots.h"
68 #include "../bemodule.h"
69 #include "../begnuas.h"
70 #include "../bestate.h"
71 #include "../beflags.h"
72 #include "../betranshlp.h"
73 #include "../belistsched.h"
75 #include "bearch_ia32_t.h"
77 #include "ia32_new_nodes.h"
78 #include "gen_ia32_regalloc_if.h"
79 #include "gen_ia32_machine.h"
80 #include "ia32_common_transform.h"
81 #include "ia32_transform.h"
82 #include "ia32_emitter.h"
83 #include "ia32_map_regs.h"
84 #include "ia32_optimize.h"
86 #include "ia32_dbg_stat.h"
87 #include "ia32_finish.h"
88 #include "ia32_util.h"
90 #include "ia32_architecture.h"
93 #include "ia32_pbqp_transform.h"
95 transformer_t be_transformer = TRANSFORMER_DEFAULT;
98 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
100 ir_mode *mode_fpcw = NULL;
101 ia32_code_gen_t *ia32_current_cg = NULL;
103 /** The current omit-fp state */
104 static unsigned ia32_curr_fp_ommitted = 0;
105 static ir_type *omit_fp_between_type = NULL;
106 static ir_type *between_type = NULL;
107 static ir_entity *old_bp_ent = NULL;
108 static ir_entity *ret_addr_ent = NULL;
109 static ir_entity *omit_fp_ret_addr_ent = NULL;
112 * The environment for the intrinsic mapping.
114 static ia32_intrinsic_env_t intrinsic_env = {
116 NULL, /* the irg, these entities belong to */
117 NULL, /* entity for __divdi3 library call */
118 NULL, /* entity for __moddi3 library call */
119 NULL, /* entity for __udivdi3 library call */
120 NULL, /* entity for __umoddi3 library call */
124 typedef ir_node *(*create_const_node_func) (dbg_info *dbg, ir_node *block);
127 * Used to create per-graph unique pseudo nodes.
129 static inline ir_node *create_const(ia32_code_gen_t *cg, ir_node **place,
130 create_const_node_func func,
131 const arch_register_t* reg)
133 ir_node *block, *res;
138 block = get_irg_start_block(cg->irg);
139 res = func(NULL, block);
140 arch_set_irn_register(res, reg);
146 /* Creates the unique per irg GP NoReg node. */
147 ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg)
149 return create_const(cg, &cg->noreg_gp, new_bd_ia32_NoReg_GP,
150 &ia32_gp_regs[REG_GP_NOREG]);
153 ir_node *ia32_new_NoReg_vfp(ia32_code_gen_t *cg)
155 return create_const(cg, &cg->noreg_vfp, new_bd_ia32_NoReg_VFP,
156 &ia32_vfp_regs[REG_VFP_NOREG]);
159 ir_node *ia32_new_NoReg_xmm(ia32_code_gen_t *cg)
161 return create_const(cg, &cg->noreg_xmm, new_bd_ia32_NoReg_XMM,
162 &ia32_xmm_regs[REG_XMM_NOREG]);
165 ir_node *ia32_new_Fpu_truncate(ia32_code_gen_t *cg)
167 return create_const(cg, &cg->fpu_trunc_mode, new_bd_ia32_ChangeCW,
168 &ia32_fp_cw_regs[REG_FPCW]);
173 * Returns the admissible noreg register node for input register pos of node irn.
175 static ir_node *ia32_get_admissible_noreg(ia32_code_gen_t *cg, ir_node *irn, int pos)
177 const arch_register_req_t *req = arch_get_register_req(irn, pos);
179 assert(req != NULL && "Missing register requirements");
180 if (req->cls == &ia32_reg_classes[CLASS_ia32_gp])
181 return ia32_new_NoReg_gp(cg);
183 if (ia32_cg_config.use_sse2) {
184 return ia32_new_NoReg_xmm(cg);
186 return ia32_new_NoReg_vfp(cg);
190 /**************************************************
193 * _ __ ___ __ _ __ _| | | ___ ___ _| |_
194 * | '__/ _ \/ _` | / _` | | |/ _ \ / __| | | _|
195 * | | | __/ (_| | | (_| | | | (_) | (__ | | |
196 * |_| \___|\__, | \__,_|_|_|\___/ \___| |_|_|
199 **************************************************/
201 static const arch_register_req_t *get_ia32_SwitchJmp_out_req(
202 const ir_node *node, int pos)
206 return arch_no_register_req;
209 static arch_irn_class_t ia32_classify(const ir_node *irn)
211 arch_irn_class_t classification = 0;
213 assert(is_ia32_irn(irn));
215 if (is_ia32_is_reload(irn))
216 classification |= arch_irn_class_reload;
218 if (is_ia32_is_spill(irn))
219 classification |= arch_irn_class_spill;
221 if (is_ia32_is_remat(irn))
222 classification |= arch_irn_class_remat;
224 return classification;
228 * The IA32 ABI callback object.
231 be_abi_call_flags_bits_t flags; /**< The call flags. */
232 const arch_env_t *aenv; /**< The architecture environment. */
233 ir_graph *irg; /**< The associated graph. */
236 static ir_entity *ia32_get_frame_entity(const ir_node *irn)
238 return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
241 static void ia32_set_frame_entity(ir_node *irn, ir_entity *ent)
243 set_ia32_frame_ent(irn, ent);
246 static void ia32_set_frame_offset(ir_node *irn, int bias)
248 if (get_ia32_frame_ent(irn) == NULL)
251 if (is_ia32_Pop(irn) || is_ia32_PopMem(irn)) {
252 ia32_code_gen_t *cg = ia32_current_cg;
253 int omit_fp = be_abi_omit_fp(cg->birg->abi);
255 /* Pop nodes modify the stack pointer before calculating the
256 * destination address, so fix this here
261 add_ia32_am_offs_int(irn, bias);
264 static int ia32_get_sp_bias(const ir_node *node)
266 if (is_ia32_Call(node))
267 return -(int)get_ia32_call_attr_const(node)->pop;
269 if (is_ia32_Push(node))
272 if (is_ia32_Pop(node) || is_ia32_PopMem(node))
279 * Generate the routine prologue.
281 * @param self The callback object.
282 * @param mem A pointer to the mem node. Update this if you define new memory.
283 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
284 * @param stack_bias Points to the current stack bias, can be modified if needed.
286 * @return The register which shall be used as a stack frame base.
288 * All nodes which define registers in @p reg_map must keep @p reg_map current.
290 static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map, int *stack_bias)
292 ia32_abi_env_t *env = self;
293 ia32_code_gen_t *cg = ia32_current_cg;
294 const arch_env_t *arch_env = env->aenv;
296 ia32_curr_fp_ommitted = env->flags.try_omit_fp;
297 if (! env->flags.try_omit_fp) {
298 ir_node *bl = get_irg_start_block(env->irg);
299 ir_node *curr_sp = be_abi_reg_map_get(reg_map, arch_env->sp);
300 ir_node *curr_bp = be_abi_reg_map_get(reg_map, arch_env->bp);
301 ir_node *noreg = ia32_new_NoReg_gp(cg);
304 /* mark bp register as ignore */
305 be_set_constr_single_reg_out(get_Proj_pred(curr_bp),
306 get_Proj_proj(curr_bp), arch_env->bp, arch_register_req_type_ignore);
309 push = new_bd_ia32_Push(NULL, bl, noreg, noreg, *mem, curr_bp, curr_sp);
310 curr_sp = new_r_Proj(push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
311 *mem = new_r_Proj(push, mode_M, pn_ia32_Push_M);
313 /* the push must have SP out register */
314 arch_set_irn_register(curr_sp, arch_env->sp);
316 /* this modifies the stack bias, because we pushed 32bit */
319 /* move esp to ebp */
320 curr_bp = be_new_Copy(arch_env->bp->reg_class, bl, curr_sp);
321 be_set_constr_single_reg_out(curr_bp, 0, arch_env->bp,
322 arch_register_req_type_ignore);
324 /* beware: the copy must be done before any other sp use */
325 curr_sp = be_new_CopyKeep_single(arch_env->sp->reg_class, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
326 be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
327 arch_register_req_type_produces_sp);
329 be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
330 be_abi_reg_map_set(reg_map, arch_env->bp, curr_bp);
339 * Generate the routine epilogue.
340 * @param self The callback object.
341 * @param bl The block for the epilog
342 * @param mem A pointer to the mem node. Update this if you define new memory.
343 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
344 * @return The register which shall be used as a stack frame base.
346 * All nodes which define registers in @p reg_map must keep @p reg_map current.
348 static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
350 ia32_abi_env_t *env = self;
351 const arch_env_t *arch_env = env->aenv;
352 ir_node *curr_sp = be_abi_reg_map_get(reg_map, arch_env->sp);
353 ir_node *curr_bp = be_abi_reg_map_get(reg_map, arch_env->bp);
355 if (env->flags.try_omit_fp) {
356 /* simply remove the stack frame here */
357 curr_sp = be_new_IncSP(arch_env->sp, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK, 0);
359 ir_mode *mode_bp = arch_env->bp->reg_class->mode;
361 if (ia32_cg_config.use_leave) {
365 leave = new_bd_ia32_Leave(NULL, bl, curr_bp);
366 curr_bp = new_r_Proj(leave, mode_bp, pn_ia32_Leave_frame);
367 curr_sp = new_r_Proj(leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
371 /* the old SP is not needed anymore (kill the proj) */
372 assert(is_Proj(curr_sp));
375 /* copy ebp to esp */
376 curr_sp = be_new_Copy(&ia32_reg_classes[CLASS_ia32_gp], bl, curr_bp);
377 arch_set_irn_register(curr_sp, arch_env->sp);
378 be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
379 arch_register_req_type_ignore);
382 pop = new_bd_ia32_PopEbp(NULL, bl, *mem, curr_sp);
383 curr_bp = new_r_Proj(pop, mode_bp, pn_ia32_Pop_res);
384 curr_sp = new_r_Proj(pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
386 *mem = new_r_Proj(pop, mode_M, pn_ia32_Pop_M);
388 arch_set_irn_register(curr_sp, arch_env->sp);
389 arch_set_irn_register(curr_bp, arch_env->bp);
392 be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
393 be_abi_reg_map_set(reg_map, arch_env->bp, curr_bp);
397 * Initialize the callback object.
398 * @param call The call object.
399 * @param aenv The architecture environment.
400 * @param irg The graph with the method.
401 * @return Some pointer. This pointer is passed to all other callback functions as self object.
403 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
405 ia32_abi_env_t *env = XMALLOC(ia32_abi_env_t);
406 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
407 env->flags = fl.bits;
414 * Destroy the callback object.
415 * @param self The callback object.
417 static void ia32_abi_done(void *self)
423 * Build the between type and entities if not already build.
425 static void ia32_build_between_type(void)
427 #define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
428 if (! between_type) {
429 ir_type *old_bp_type = new_type_primitive(mode_Iu);
430 ir_type *ret_addr_type = new_type_primitive(mode_Iu);
432 between_type = new_type_struct(IDENT("ia32_between_type"));
433 old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
434 ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type);
436 set_entity_offset(old_bp_ent, 0);
437 set_entity_offset(ret_addr_ent, get_type_size_bytes(old_bp_type));
438 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
439 set_type_state(between_type, layout_fixed);
441 omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
442 omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type);
444 set_entity_offset(omit_fp_ret_addr_ent, 0);
445 set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
446 set_type_state(omit_fp_between_type, layout_fixed);
452 * Produces the type which sits between the stack args and the locals on the stack.
453 * it will contain the return address and space to store the old base pointer.
454 * @return The Firm type modeling the ABI between type.
456 static ir_type *ia32_abi_get_between_type(void *self)
458 ia32_abi_env_t *env = self;
460 ia32_build_between_type();
461 return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
465 * Return the stack entity that contains the return address.
467 ir_entity *ia32_get_return_address_entity(void)
469 ia32_build_between_type();
470 return ia32_curr_fp_ommitted ? omit_fp_ret_addr_ent : ret_addr_ent;
474 * Return the stack entity that contains the frame address.
476 ir_entity *ia32_get_frame_address_entity(void)
478 ia32_build_between_type();
479 return ia32_curr_fp_ommitted ? NULL : old_bp_ent;
483 * Get the estimated cycle count for @p irn.
485 * @param self The this pointer.
486 * @param irn The node.
488 * @return The estimated cycle count for this operation
490 static int ia32_get_op_estimated_cost(const ir_node *irn)
493 ia32_op_type_t op_tp;
497 if (!is_ia32_irn(irn))
500 assert(is_ia32_irn(irn));
502 cost = get_ia32_latency(irn);
503 op_tp = get_ia32_op_type(irn);
505 if (is_ia32_CopyB(irn)) {
508 else if (is_ia32_CopyB_i(irn)) {
509 int size = get_ia32_copyb_size(irn);
510 cost = 20 + (int)ceil((4/3) * size);
512 /* in case of address mode operations add additional cycles */
513 else if (op_tp == ia32_AddrModeD || op_tp == ia32_AddrModeS) {
515 In case of stack access and access to fixed addresses add 5 cycles
516 (we assume they are in cache), other memory operations cost 20
519 if (is_ia32_use_frame(irn) || (
520 is_ia32_NoReg_GP(get_irn_n(irn, n_ia32_base)) &&
521 is_ia32_NoReg_GP(get_irn_n(irn, n_ia32_index))
533 * Returns the inverse operation if @p irn, recalculating the argument at position @p i.
535 * @param irn The original operation
536 * @param i Index of the argument we want the inverse operation to yield
537 * @param inverse struct to be filled with the resulting inverse op
538 * @param obstack The obstack to use for allocation of the returned nodes array
539 * @return The inverse operation or NULL if operation invertible
541 static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst)
552 ir_node *block, *noreg, *nomem;
555 /* we cannot invert non-ia32 irns */
556 if (! is_ia32_irn(irn))
559 /* operand must always be a real operand (not base, index or mem) */
560 if (i != n_ia32_binary_left && i != n_ia32_binary_right)
563 /* we don't invert address mode operations */
564 if (get_ia32_op_type(irn) != ia32_Normal)
567 /* TODO: adjust for new immediates... */
568 ir_fprintf(stderr, "TODO: fix get_inverse for new immediates (%+F)\n",
572 block = get_nodes_block(irn);
573 mode = get_irn_mode(irn);
574 irn_mode = get_irn_mode(irn);
575 noreg = get_irn_n(irn, 0);
577 dbg = get_irn_dbg_info(irn);
579 /* initialize structure */
580 inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0]));
584 switch (get_ia32_irn_opcode(irn)) {
587 if (get_ia32_immop_type(irn) == ia32_ImmConst) {
588 /* we have an add with a const here */
589 /* invers == add with negated const */
590 inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
592 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
593 set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
594 set_ia32_commutative(inverse->nodes[0]);
596 else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
597 /* we have an add with a symconst here */
598 /* invers == sub with const */
599 inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
601 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
604 /* normal add: inverse == sub */
605 inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, i ^ 1));
612 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
613 /* we have a sub with a const/symconst here */
614 /* invers == add with this const */
615 inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
616 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
617 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
621 if (i == n_ia32_binary_left) {
622 inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, 3));
625 inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, get_irn_n(irn, n_ia32_binary_left), (ir_node*) irn);
633 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
634 /* xor with const: inverse = xor */
635 inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
636 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
637 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
641 inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, (ir_node *) irn, get_irn_n(irn, i));
647 inverse->nodes[0] = new_bd_ia32_Not(dbg, block, (ir_node*) irn);
652 inverse->nodes[0] = new_bd_ia32_Neg(dbg, block, (ir_node*) irn);
657 /* inverse operation not supported */
665 static ir_mode *get_spill_mode_mode(const ir_mode *mode)
667 if (mode_is_float(mode))
674 * Get the mode that should be used for spilling value node
676 static ir_mode *get_spill_mode(const ir_node *node)
678 ir_mode *mode = get_irn_mode(node);
679 return get_spill_mode_mode(mode);
683 * Checks whether an addressmode reload for a node with mode mode is compatible
684 * with a spillslot of mode spill_mode
686 static int ia32_is_spillmode_compatible(const ir_mode *mode, const ir_mode *spillmode)
688 return !mode_is_float(mode) || mode == spillmode;
692 * Check if irn can load its operand at position i from memory (source addressmode).
693 * @param irn The irn to be checked
694 * @param i The operands position
695 * @return Non-Zero if operand can be loaded
697 static int ia32_possible_memory_operand(const ir_node *irn, unsigned int i)
699 ir_node *op = get_irn_n(irn, i);
700 const ir_mode *mode = get_irn_mode(op);
701 const ir_mode *spillmode = get_spill_mode(op);
703 if (!is_ia32_irn(irn) || /* must be an ia32 irn */
704 get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
705 !ia32_is_spillmode_compatible(mode, spillmode) ||
706 is_ia32_use_frame(irn)) /* must not already use frame */
709 switch (get_ia32_am_support(irn)) {
714 if (i != n_ia32_unary_op)
720 case n_ia32_binary_left: {
721 const arch_register_req_t *req;
722 if (!is_ia32_commutative(irn))
725 /* we can't swap left/right for limited registers
726 * (As this (currently) breaks constraint handling copies)
728 req = get_ia32_in_req(irn, n_ia32_binary_left);
729 if (req->type & arch_register_req_type_limited)
734 case n_ia32_binary_right:
743 panic("Unknown AM type");
746 /* HACK: must not already use "real" memory.
747 * This can happen for Call and Div */
748 if (!is_NoMem(get_irn_n(irn, n_ia32_mem)))
754 static void ia32_perform_memory_operand(ir_node *irn, ir_node *spill,
758 ir_mode *dest_op_mode;
760 assert(ia32_possible_memory_operand(irn, i) && "Cannot perform memory operand change");
762 set_ia32_op_type(irn, ia32_AddrModeS);
764 load_mode = get_irn_mode(get_irn_n(irn, i));
765 dest_op_mode = get_ia32_ls_mode(irn);
766 if (get_mode_size_bits(load_mode) <= get_mode_size_bits(dest_op_mode)) {
767 set_ia32_ls_mode(irn, load_mode);
769 set_ia32_use_frame(irn);
770 set_ia32_need_stackent(irn);
772 if (i == n_ia32_binary_left &&
773 get_ia32_am_support(irn) == ia32_am_binary &&
774 /* immediates are only allowed on the right side */
775 !is_ia32_Immediate(get_irn_n(irn, n_ia32_binary_right))) {
776 ia32_swap_left_right(irn);
777 i = n_ia32_binary_right;
780 assert(is_NoMem(get_irn_n(irn, n_ia32_mem)));
782 set_irn_n(irn, n_ia32_base, get_irg_frame(get_irn_irg(irn)));
783 set_irn_n(irn, n_ia32_mem, spill);
784 set_irn_n(irn, i, ia32_get_admissible_noreg(ia32_current_cg, irn, i));
785 set_ia32_is_reload(irn);
788 static const be_abi_callbacks_t ia32_abi_callbacks = {
791 ia32_abi_get_between_type,
796 /* register allocator interface */
797 static const arch_irn_ops_t ia32_irn_ops = {
800 ia32_get_frame_entity,
801 ia32_set_frame_entity,
802 ia32_set_frame_offset,
805 ia32_get_op_estimated_cost,
806 ia32_possible_memory_operand,
807 ia32_perform_memory_operand,
810 /* special register allocator interface for SwitchJmp
811 as it possibly has a WIDE range of Proj numbers.
812 We don't want to allocate output for register constraints for
814 static const arch_irn_ops_t ia32_SwitchJmp_irn_ops = {
815 /* Note: we also use SwitchJmp_out_req for the inputs too:
816 This is because the bearch API has a conceptual problem at the moment.
817 Querying for negative proj numbers which can happen for switchs
818 isn't possible and will result in inputs getting queried */
819 get_ia32_SwitchJmp_out_req,
821 ia32_get_frame_entity,
822 ia32_set_frame_entity,
823 ia32_set_frame_offset,
826 ia32_get_op_estimated_cost,
827 ia32_possible_memory_operand,
828 ia32_perform_memory_operand,
831 /**************************************************
834 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
835 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
836 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
837 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
840 **************************************************/
842 static ir_entity *mcount = NULL;
844 #define ID(s) new_id_from_chars(s, sizeof(s) - 1)
846 static void ia32_before_abi(void *self)
848 lower_mode_b_config_t lower_mode_b_config = {
849 mode_Iu, /* lowered mode */
850 mode_Bu, /* preferred mode for set */
851 0, /* don't lower direct compares */
853 ia32_code_gen_t *cg = self;
855 ir_lower_mode_b(cg->irg, &lower_mode_b_config);
857 be_dump(cg->irg, "-lower_modeb", dump_ir_block_graph_sched);
860 if (mcount == NULL) {
861 ir_type *tp = new_type_method(0, 0);
862 mcount = new_entity(get_glob_type(), ID("mcount"), tp);
863 /* FIXME: enter the right ld_ident here */
864 set_entity_ld_ident(mcount, get_entity_ident(mcount));
865 set_entity_visibility(mcount, ir_visibility_external);
867 instrument_initcall(cg->irg, mcount);
872 * Transforms the standard firm graph into
875 static void ia32_prepare_graph(void *self)
877 ia32_code_gen_t *cg = self;
880 switch (be_transformer) {
881 case TRANSFORMER_DEFAULT:
882 /* transform remaining nodes into assembler instructions */
883 ia32_transform_graph(cg);
886 case TRANSFORMER_PBQP:
887 case TRANSFORMER_RAND:
888 /* transform nodes into assembler instructions by PBQP magic */
889 ia32_transform_graph_by_pbqp(cg);
893 panic("invalid transformer");
896 ia32_transform_graph(cg);
899 /* do local optimizations (mainly CSE) */
900 optimize_graph_df(cg->irg);
903 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
905 /* optimize address mode */
906 ia32_optimize_graph(cg);
908 /* do code placement, to optimize the position of constants */
912 be_dump(cg->irg, "-place", dump_ir_block_graph_sched);
915 ir_node *turn_back_am(ir_node *node)
917 dbg_info *dbgi = get_irn_dbg_info(node);
918 ir_node *block = get_nodes_block(node);
919 ir_node *base = get_irn_n(node, n_ia32_base);
920 ir_node *index = get_irn_n(node, n_ia32_index);
921 ir_node *mem = get_irn_n(node, n_ia32_mem);
924 ir_node *load = new_bd_ia32_Load(dbgi, block, base, index, mem);
925 ir_node *load_res = new_rd_Proj(dbgi, load, mode_Iu, pn_ia32_Load_res);
927 ia32_copy_am_attrs(load, node);
928 if (is_ia32_is_reload(node))
929 set_ia32_is_reload(load);
930 set_irn_n(node, n_ia32_mem, new_NoMem());
932 switch (get_ia32_am_support(node)) {
934 set_irn_n(node, n_ia32_unary_op, load_res);
938 if (is_ia32_Immediate(get_irn_n(node, n_ia32_binary_right))) {
939 set_irn_n(node, n_ia32_binary_left, load_res);
941 set_irn_n(node, n_ia32_binary_right, load_res);
946 panic("Unknown AM type");
948 noreg = ia32_new_NoReg_gp(ia32_current_cg);
949 set_irn_n(node, n_ia32_base, noreg);
950 set_irn_n(node, n_ia32_index, noreg);
951 set_ia32_am_offs_int(node, 0);
952 set_ia32_am_sc(node, NULL);
953 set_ia32_am_scale(node, 0);
954 clear_ia32_am_sc_sign(node);
956 /* rewire mem-proj */
957 if (get_irn_mode(node) == mode_T) {
958 const ir_edge_t *edge;
959 foreach_out_edge(node, edge) {
960 ir_node *out = get_edge_src_irn(edge);
961 if (get_irn_mode(out) == mode_M) {
962 set_Proj_pred(out, load);
963 set_Proj_proj(out, pn_ia32_Load_M);
969 set_ia32_op_type(node, ia32_Normal);
970 if (sched_is_scheduled(node))
971 sched_add_before(node, load);
976 static ir_node *flags_remat(ir_node *node, ir_node *after)
978 /* we should turn back source address mode when rematerializing nodes */
983 if (is_Block(after)) {
986 block = get_nodes_block(after);
989 type = get_ia32_op_type(node);
996 /* TODO implement this later... */
997 panic("found DestAM with flag user %+F this should not happen", node);
1000 default: assert(type == ia32_Normal); break;
1003 copy = exact_copy(node);
1004 set_nodes_block(copy, block);
1005 sched_add_after(after, copy);
1011 * Called before the register allocator.
1013 static void ia32_before_ra(void *self)
1015 ia32_code_gen_t *cg = self;
1017 /* setup fpu rounding modes */
1018 ia32_setup_fpu_mode(cg);
1021 be_sched_fix_flags(cg->birg, &ia32_reg_classes[CLASS_ia32_flags],
1024 ia32_add_missing_keeps(cg);
1029 * Transforms a be_Reload into a ia32 Load.
1031 static void transform_to_Load(ia32_code_gen_t *cg, ir_node *node)
1033 ir_graph *irg = get_irn_irg(node);
1034 dbg_info *dbg = get_irn_dbg_info(node);
1035 ir_node *block = get_nodes_block(node);
1036 ir_entity *ent = be_get_frame_entity(node);
1037 ir_mode *mode = get_irn_mode(node);
1038 ir_mode *spillmode = get_spill_mode(node);
1039 ir_node *noreg = ia32_new_NoReg_gp(cg);
1040 ir_node *sched_point = NULL;
1041 ir_node *ptr = get_irg_frame(irg);
1042 ir_node *mem = get_irn_n(node, be_pos_Reload_mem);
1043 ir_node *new_op, *proj;
1044 const arch_register_t *reg;
1046 if (sched_is_scheduled(node)) {
1047 sched_point = sched_prev(node);
1050 if (mode_is_float(spillmode)) {
1051 if (ia32_cg_config.use_sse2)
1052 new_op = new_bd_ia32_xLoad(dbg, block, ptr, noreg, mem, spillmode);
1054 new_op = new_bd_ia32_vfld(dbg, block, ptr, noreg, mem, spillmode);
1056 else if (get_mode_size_bits(spillmode) == 128) {
1057 /* Reload 128 bit SSE registers */
1058 new_op = new_bd_ia32_xxLoad(dbg, block, ptr, noreg, mem);
1061 new_op = new_bd_ia32_Load(dbg, block, ptr, noreg, mem);
1063 set_ia32_op_type(new_op, ia32_AddrModeS);
1064 set_ia32_ls_mode(new_op, spillmode);
1065 set_ia32_frame_ent(new_op, ent);
1066 set_ia32_use_frame(new_op);
1067 set_ia32_is_reload(new_op);
1069 DBG_OPT_RELOAD2LD(node, new_op);
1071 proj = new_rd_Proj(dbg, new_op, mode, pn_ia32_Load_res);
1074 sched_add_after(sched_point, new_op);
1078 /* copy the register from the old node to the new Load */
1079 reg = arch_get_irn_register(node);
1080 arch_set_irn_register(proj, reg);
1082 SET_IA32_ORIG_NODE(new_op, node);
1084 exchange(node, proj);
1088 * Transforms a be_Spill node into a ia32 Store.
1090 static void transform_to_Store(ia32_code_gen_t *cg, ir_node *node)
1092 ir_graph *irg = get_irn_irg(node);
1093 dbg_info *dbg = get_irn_dbg_info(node);
1094 ir_node *block = get_nodes_block(node);
1095 ir_entity *ent = be_get_frame_entity(node);
1096 const ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
1097 ir_mode *mode = get_spill_mode(spillval);
1098 ir_node *noreg = ia32_new_NoReg_gp(cg);
1099 ir_node *nomem = new_NoMem();
1100 ir_node *ptr = get_irg_frame(irg);
1101 ir_node *val = get_irn_n(node, be_pos_Spill_val);
1103 ir_node *sched_point = NULL;
1105 if (sched_is_scheduled(node)) {
1106 sched_point = sched_prev(node);
1109 if (mode_is_float(mode)) {
1110 if (ia32_cg_config.use_sse2)
1111 store = new_bd_ia32_xStore(dbg, block, ptr, noreg, nomem, val);
1113 store = new_bd_ia32_vfst(dbg, block, ptr, noreg, nomem, val, mode);
1114 } else if (get_mode_size_bits(mode) == 128) {
1115 /* Spill 128 bit SSE registers */
1116 store = new_bd_ia32_xxStore(dbg, block, ptr, noreg, nomem, val);
1117 } else if (get_mode_size_bits(mode) == 8) {
1118 store = new_bd_ia32_Store8Bit(dbg, block, ptr, noreg, nomem, val);
1120 store = new_bd_ia32_Store(dbg, block, ptr, noreg, nomem, val);
1123 set_ia32_op_type(store, ia32_AddrModeD);
1124 set_ia32_ls_mode(store, mode);
1125 set_ia32_frame_ent(store, ent);
1126 set_ia32_use_frame(store);
1127 set_ia32_is_spill(store);
1128 SET_IA32_ORIG_NODE(store, node);
1129 DBG_OPT_SPILL2ST(node, store);
1132 sched_add_after(sched_point, store);
1136 exchange(node, store);
1139 static ir_node *create_push(ia32_code_gen_t *cg, ir_node *node, ir_node *schedpoint, ir_node *sp, ir_node *mem, ir_entity *ent)
1141 dbg_info *dbg = get_irn_dbg_info(node);
1142 ir_node *block = get_nodes_block(node);
1143 ir_node *noreg = ia32_new_NoReg_gp(cg);
1144 ir_graph *irg = get_irn_irg(node);
1145 ir_node *frame = get_irg_frame(irg);
1147 ir_node *push = new_bd_ia32_Push(dbg, block, frame, noreg, mem, noreg, sp);
1149 set_ia32_frame_ent(push, ent);
1150 set_ia32_use_frame(push);
1151 set_ia32_op_type(push, ia32_AddrModeS);
1152 set_ia32_ls_mode(push, mode_Is);
1153 set_ia32_is_spill(push);
1155 sched_add_before(schedpoint, push);
1159 static ir_node *create_pop(ia32_code_gen_t *cg, ir_node *node, ir_node *schedpoint, ir_node *sp, ir_entity *ent)
1161 dbg_info *dbg = get_irn_dbg_info(node);
1162 ir_node *block = get_nodes_block(node);
1163 ir_node *noreg = ia32_new_NoReg_gp(cg);
1164 ir_graph *irg = get_irn_irg(node);
1165 ir_node *frame = get_irg_frame(irg);
1167 ir_node *pop = new_bd_ia32_PopMem(dbg, block, frame, noreg, new_NoMem(), sp);
1169 set_ia32_frame_ent(pop, ent);
1170 set_ia32_use_frame(pop);
1171 set_ia32_op_type(pop, ia32_AddrModeD);
1172 set_ia32_ls_mode(pop, mode_Is);
1173 set_ia32_is_reload(pop);
1175 sched_add_before(schedpoint, pop);
1180 static ir_node* create_spproj(ir_node *node, ir_node *pred, int pos)
1182 dbg_info *dbg = get_irn_dbg_info(node);
1183 ir_mode *spmode = mode_Iu;
1184 const arch_register_t *spreg = &ia32_gp_regs[REG_ESP];
1187 sp = new_rd_Proj(dbg, pred, spmode, pos);
1188 arch_set_irn_register(sp, spreg);
1194 * Transform MemPerm, currently we do this the ugly way and produce
1195 * push/pop into/from memory cascades. This is possible without using
1198 static void transform_MemPerm(ia32_code_gen_t *cg, ir_node *node)
1200 ir_node *block = get_nodes_block(node);
1201 ir_node *sp = be_abi_get_ignore_irn(cg->birg->abi, &ia32_gp_regs[REG_ESP]);
1202 int arity = be_get_MemPerm_entity_arity(node);
1203 ir_node **pops = ALLOCAN(ir_node*, arity);
1207 const ir_edge_t *edge;
1208 const ir_edge_t *next;
1211 for (i = 0; i < arity; ++i) {
1212 ir_entity *inent = be_get_MemPerm_in_entity(node, i);
1213 ir_entity *outent = be_get_MemPerm_out_entity(node, i);
1214 ir_type *enttype = get_entity_type(inent);
1215 unsigned entsize = get_type_size_bytes(enttype);
1216 unsigned entsize2 = get_type_size_bytes(get_entity_type(outent));
1217 ir_node *mem = get_irn_n(node, i + 1);
1220 /* work around cases where entities have different sizes */
1221 if (entsize2 < entsize)
1223 assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
1225 push = create_push(cg, node, node, sp, mem, inent);
1226 sp = create_spproj(node, push, pn_ia32_Push_stack);
1228 /* add another push after the first one */
1229 push = create_push(cg, node, node, sp, mem, inent);
1230 add_ia32_am_offs_int(push, 4);
1231 sp = create_spproj(node, push, pn_ia32_Push_stack);
1234 set_irn_n(node, i, new_Bad());
1238 for (i = arity - 1; i >= 0; --i) {
1239 ir_entity *inent = be_get_MemPerm_in_entity(node, i);
1240 ir_entity *outent = be_get_MemPerm_out_entity(node, i);
1241 ir_type *enttype = get_entity_type(outent);
1242 unsigned entsize = get_type_size_bytes(enttype);
1243 unsigned entsize2 = get_type_size_bytes(get_entity_type(inent));
1246 /* work around cases where entities have different sizes */
1247 if (entsize2 < entsize)
1249 assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
1251 pop = create_pop(cg, node, node, sp, outent);
1252 sp = create_spproj(node, pop, pn_ia32_Pop_stack);
1254 add_ia32_am_offs_int(pop, 4);
1256 /* add another pop after the first one */
1257 pop = create_pop(cg, node, node, sp, outent);
1258 sp = create_spproj(node, pop, pn_ia32_Pop_stack);
1265 keep = be_new_Keep(block, 1, in);
1266 sched_add_before(node, keep);
1268 /* exchange memprojs */
1269 foreach_out_edge_safe(node, edge, next) {
1270 ir_node *proj = get_edge_src_irn(edge);
1271 int p = get_Proj_proj(proj);
1275 set_Proj_pred(proj, pops[p]);
1276 set_Proj_proj(proj, pn_ia32_Pop_M);
1279 /* remove memperm */
1280 arity = get_irn_arity(node);
1281 for (i = 0; i < arity; ++i) {
1282 set_irn_n(node, i, new_Bad());
1288 * Block-Walker: Calls the transform functions Spill and Reload.
1290 static void ia32_after_ra_walker(ir_node *block, void *env)
1292 ir_node *node, *prev;
1293 ia32_code_gen_t *cg = env;
1295 /* beware: the schedule is changed here */
1296 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
1297 prev = sched_prev(node);
1299 if (be_is_Reload(node)) {
1300 transform_to_Load(cg, node);
1301 } else if (be_is_Spill(node)) {
1302 transform_to_Store(cg, node);
1303 } else if (be_is_MemPerm(node)) {
1304 transform_MemPerm(cg, node);
1310 * Collects nodes that need frame entities assigned.
1312 static void ia32_collect_frame_entity_nodes(ir_node *node, void *data)
1314 be_fec_env_t *env = data;
1315 const ir_mode *mode;
1318 if (be_is_Reload(node) && be_get_frame_entity(node) == NULL) {
1319 mode = get_spill_mode_mode(get_irn_mode(node));
1320 align = get_mode_size_bytes(mode);
1321 } else if (is_ia32_irn(node) &&
1322 get_ia32_frame_ent(node) == NULL &&
1323 is_ia32_use_frame(node)) {
1324 if (is_ia32_need_stackent(node))
1327 switch (get_ia32_irn_opcode(node)) {
1329 case iro_ia32_Load: {
1330 const ia32_attr_t *attr = get_ia32_attr_const(node);
1332 if (attr->data.need_32bit_stackent) {
1334 } else if (attr->data.need_64bit_stackent) {
1337 mode = get_ia32_ls_mode(node);
1338 if (is_ia32_is_reload(node))
1339 mode = get_spill_mode_mode(mode);
1341 align = get_mode_size_bytes(mode);
1345 case iro_ia32_vfild:
1347 case iro_ia32_xLoad: {
1348 mode = get_ia32_ls_mode(node);
1353 case iro_ia32_FldCW: {
1354 /* although 2 byte would be enough 4 byte performs best */
1362 panic("unexpected frame user while collection frame entity nodes");
1364 case iro_ia32_FnstCW:
1365 case iro_ia32_Store8Bit:
1366 case iro_ia32_Store:
1369 case iro_ia32_vfist:
1370 case iro_ia32_vfisttp:
1372 case iro_ia32_xStore:
1373 case iro_ia32_xStoreSimple:
1380 be_node_needs_frame_entity(env, node, mode, align);
1384 * We transform Spill and Reload here. This needs to be done before
1385 * stack biasing otherwise we would miss the corrected offset for these nodes.
1387 static void ia32_after_ra(void *self)
1389 ia32_code_gen_t *cg = self;
1390 ir_graph *irg = cg->irg;
1391 be_fec_env_t *fec_env = be_new_frame_entity_coalescer(cg->birg);
1393 /* create and coalesce frame entities */
1394 irg_walk_graph(irg, NULL, ia32_collect_frame_entity_nodes, fec_env);
1395 be_assign_entities(fec_env);
1396 be_free_frame_entity_coalescer(fec_env);
1398 irg_block_walk_graph(irg, NULL, ia32_after_ra_walker, cg);
1402 * Last touchups for the graph before emit: x87 simulation to replace the
1403 * virtual with real x87 instructions, creating a block schedule and peephole
1406 static void ia32_finish(void *self)
1408 ia32_code_gen_t *cg = self;
1409 ir_graph *irg = cg->irg;
1411 ia32_finish_irg(irg, cg);
1413 /* we might have to rewrite x87 virtual registers */
1414 if (cg->do_x87_sim) {
1415 x87_simulate_graph(cg->birg);
1418 /* do peephole optimisations */
1419 ia32_peephole_optimization(cg);
1421 /* create block schedule, this also removes empty blocks which might
1422 * produce critical edges */
1423 cg->blk_sched = be_create_block_schedule(irg, cg->birg->exec_freq);
1427 * Emits the code, closes the output file and frees
1428 * the code generator interface.
1430 static void ia32_codegen(void *self)
1432 ia32_code_gen_t *cg = self;
1433 ir_graph *irg = cg->irg;
1435 if (ia32_cg_config.emit_machcode) {
1436 ia32_gen_binary_routine(cg, irg);
1438 ia32_gen_routine(cg, irg);
1441 /* remove it from the isa */
1444 assert(ia32_current_cg == cg);
1445 ia32_current_cg = NULL;
1447 /* de-allocate code generator */
1452 * Returns the node representing the PIC base.
1454 static ir_node *ia32_get_pic_base(void *self)
1457 ia32_code_gen_t *cg = self;
1458 ir_node *get_eip = cg->get_eip;
1459 if (get_eip != NULL)
1462 block = get_irg_start_block(cg->irg);
1463 get_eip = new_bd_ia32_GetEIP(NULL, block);
1464 cg->get_eip = get_eip;
1466 be_dep_on_frame(get_eip);
1470 static void *ia32_cg_init(be_irg_t *birg);
1472 static const arch_code_generator_if_t ia32_code_gen_if = {
1474 ia32_get_pic_base, /* return node used as base in pic code addresses */
1475 ia32_before_abi, /* before abi introduce hook */
1478 ia32_before_ra, /* before register allocation hook */
1479 ia32_after_ra, /* after register allocation hook */
1480 ia32_finish, /* called before codegen */
1481 ia32_codegen /* emit && done */
1485 * Initializes a IA32 code generator.
1487 static void *ia32_cg_init(be_irg_t *birg)
1489 ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env;
1490 ia32_code_gen_t *cg = XMALLOCZ(ia32_code_gen_t);
1492 cg->impl = &ia32_code_gen_if;
1493 cg->irg = birg->irg;
1496 cg->blk_sched = NULL;
1497 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
1498 cg->gprof = (birg->main_env->options->gprof) ? 1 : 0;
1501 /* Linux gprof implementation needs base pointer */
1502 birg->main_env->options->omit_fp = 0;
1509 if (isa->name_obst) {
1510 obstack_free(isa->name_obst, NULL);
1511 obstack_init(isa->name_obst);
1515 assert(ia32_current_cg == NULL);
1516 ia32_current_cg = cg;
1518 return (arch_code_generator_t *)cg;
1523 /*****************************************************************
1524 * ____ _ _ _____ _____
1525 * | _ \ | | | | |_ _|/ ____| /\
1526 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
1527 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
1528 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
1529 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
1531 *****************************************************************/
1534 * Set output modes for GCC
1536 static const tarval_mode_info mo_integer = {
1543 * set the tarval output mode of all integer modes to decimal
1545 static void set_tarval_output_modes(void)
1549 for (i = get_irp_n_modes() - 1; i >= 0; --i) {
1550 ir_mode *mode = get_irp_mode(i);
1552 if (mode_is_int(mode))
1553 set_tarval_mode_output_option(mode, &mo_integer);
1557 const arch_isa_if_t ia32_isa_if;
1560 * The template that generates a new ISA object.
1561 * Note that this template can be changed by command line
1564 static ia32_isa_t ia32_isa_template = {
1566 &ia32_isa_if, /* isa interface implementation */
1567 &ia32_gp_regs[REG_ESP], /* stack pointer register */
1568 &ia32_gp_regs[REG_EBP], /* base pointer register */
1569 &ia32_reg_classes[CLASS_ia32_gp], /* static link pointer register class */
1570 -1, /* stack direction */
1571 2, /* power of two stack alignment, 2^2 == 4 */
1572 NULL, /* main environment */
1573 7, /* costs for a spill instruction */
1574 5, /* costs for a reload instruction */
1576 NULL, /* 16bit register names */
1577 NULL, /* 8bit register names */
1578 NULL, /* 8bit register names high */
1581 NULL, /* current code generator */
1582 NULL, /* abstract machine */
1584 NULL, /* name obstack */
1588 static void init_asm_constraints(void)
1590 be_init_default_asm_constraint_flags();
1592 asm_constraint_flags['a'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1593 asm_constraint_flags['b'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1594 asm_constraint_flags['c'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1595 asm_constraint_flags['d'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1596 asm_constraint_flags['D'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1597 asm_constraint_flags['S'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1598 asm_constraint_flags['Q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1599 asm_constraint_flags['q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1600 asm_constraint_flags['A'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1601 asm_constraint_flags['l'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1602 asm_constraint_flags['R'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1603 asm_constraint_flags['r'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1604 asm_constraint_flags['p'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1605 asm_constraint_flags['f'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1606 asm_constraint_flags['t'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1607 asm_constraint_flags['u'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1608 asm_constraint_flags['Y'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1609 asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1610 asm_constraint_flags['n'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
1611 asm_constraint_flags['g'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
1613 /* no support for autodecrement/autoincrement */
1614 asm_constraint_flags['<'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1615 asm_constraint_flags['>'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1616 /* no float consts */
1617 asm_constraint_flags['E'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1618 asm_constraint_flags['F'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1619 /* makes no sense on x86 */
1620 asm_constraint_flags['s'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1621 /* no support for sse consts yet */
1622 asm_constraint_flags['C'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1623 /* no support for x87 consts yet */
1624 asm_constraint_flags['G'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1625 /* no support for mmx registers yet */
1626 asm_constraint_flags['y'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1627 /* not available in 32bit mode */
1628 asm_constraint_flags['Z'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1629 asm_constraint_flags['e'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1631 /* no code yet to determine register class needed... */
1632 asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1636 * Initializes the backend ISA.
1638 static arch_env_t *ia32_init(FILE *file_handle)
1640 static int inited = 0;
1648 set_tarval_output_modes();
1650 isa = XMALLOC(ia32_isa_t);
1651 memcpy(isa, &ia32_isa_template, sizeof(*isa));
1653 if (mode_fpcw == NULL) {
1654 mode_fpcw = new_ir_mode("Fpcw", irms_int_number, 16, 0, irma_none, 0);
1657 ia32_register_init();
1658 ia32_create_opcodes(&ia32_irn_ops);
1659 /* special handling for SwitchJmp */
1660 op_ia32_SwitchJmp->ops.be_ops = &ia32_SwitchJmp_irn_ops;
1662 be_emit_init(file_handle);
1663 isa->regs_16bit = pmap_create();
1664 isa->regs_8bit = pmap_create();
1665 isa->regs_8bit_high = pmap_create();
1666 isa->types = pmap_create();
1667 isa->tv_ent = pmap_create();
1668 isa->cpu = ia32_init_machine_description();
1670 ia32_build_16bit_reg_map(isa->regs_16bit);
1671 ia32_build_8bit_reg_map(isa->regs_8bit);
1672 ia32_build_8bit_reg_map_high(isa->regs_8bit_high);
1675 isa->name_obst = XMALLOC(struct obstack);
1676 obstack_init(isa->name_obst);
1679 /* enter the ISA object into the intrinsic environment */
1680 intrinsic_env.isa = isa;
1682 /* emit asm includes */
1683 n = get_irp_n_asms();
1684 for (i = 0; i < n; ++i) {
1685 be_emit_cstring("#APP\n");
1686 be_emit_ident(get_irp_asm(i));
1687 be_emit_cstring("\n#NO_APP\n");
1690 /* needed for the debug support */
1691 be_gas_emit_switch_section(GAS_SECTION_TEXT);
1692 be_emit_irprintf("%stext0:\n", be_gas_get_private_prefix());
1693 be_emit_write_line();
1695 return &isa->arch_env;
1701 * Closes the output file and frees the ISA structure.
1703 static void ia32_done(void *self)
1705 ia32_isa_t *isa = self;
1707 /* emit now all global declarations */
1708 be_gas_emit_decls(isa->arch_env.main_env);
1710 pmap_destroy(isa->regs_16bit);
1711 pmap_destroy(isa->regs_8bit);
1712 pmap_destroy(isa->regs_8bit_high);
1713 pmap_destroy(isa->tv_ent);
1714 pmap_destroy(isa->types);
1717 obstack_free(isa->name_obst, NULL);
1727 * Return the number of register classes for this architecture.
1728 * We report always these:
1729 * - the general purpose registers
1730 * - the SSE floating point register set
1731 * - the virtual floating point registers
1732 * - the SSE vector register set
1734 static unsigned ia32_get_n_reg_class(void)
1740 * Return the register class for index i.
1742 static const arch_register_class_t *ia32_get_reg_class(unsigned i)
1744 assert(i < N_CLASSES);
1745 return &ia32_reg_classes[i];
1749 * Get the register class which shall be used to store a value of a given mode.
1750 * @param self The this pointer.
1751 * @param mode The mode in question.
1752 * @return A register class which can hold values of the given mode.
1754 static const arch_register_class_t *ia32_get_reg_class_for_mode(const ir_mode *mode)
1756 if (mode_is_float(mode)) {
1757 return ia32_cg_config.use_sse2 ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1760 return &ia32_reg_classes[CLASS_ia32_gp];
1764 * Returns the register for parameter nr.
1766 static const arch_register_t *ia32_get_RegParam_reg(unsigned cc, unsigned nr,
1767 const ir_mode *mode)
1769 static const arch_register_t *gpreg_param_reg_fastcall[] = {
1770 &ia32_gp_regs[REG_ECX],
1771 &ia32_gp_regs[REG_EDX],
1774 static const unsigned MAXNUM_GPREG_ARGS = 3;
1776 static const arch_register_t *gpreg_param_reg_regparam[] = {
1777 &ia32_gp_regs[REG_EAX],
1778 &ia32_gp_regs[REG_EDX],
1779 &ia32_gp_regs[REG_ECX]
1782 static const arch_register_t *gpreg_param_reg_this[] = {
1783 &ia32_gp_regs[REG_ECX],
1788 static const arch_register_t *fpreg_sse_param_reg_std[] = {
1789 &ia32_xmm_regs[REG_XMM0],
1790 &ia32_xmm_regs[REG_XMM1],
1791 &ia32_xmm_regs[REG_XMM2],
1792 &ia32_xmm_regs[REG_XMM3],
1793 &ia32_xmm_regs[REG_XMM4],
1794 &ia32_xmm_regs[REG_XMM5],
1795 &ia32_xmm_regs[REG_XMM6],
1796 &ia32_xmm_regs[REG_XMM7]
1799 static const arch_register_t *fpreg_sse_param_reg_this[] = {
1800 NULL, /* in case of a "this" pointer, the first parameter must not be a float */
1802 static const unsigned MAXNUM_SSE_ARGS = 8;
1804 if ((cc & cc_this_call) && nr == 0)
1805 return gpreg_param_reg_this[0];
1807 if (! (cc & cc_reg_param))
1810 if (mode_is_float(mode)) {
1811 if (!ia32_cg_config.use_sse2 || (cc & cc_fpreg_param) == 0)
1813 if (nr >= MAXNUM_SSE_ARGS)
1816 if (cc & cc_this_call) {
1817 return fpreg_sse_param_reg_this[nr];
1819 return fpreg_sse_param_reg_std[nr];
1820 } else if (mode_is_int(mode) || mode_is_reference(mode)) {
1821 unsigned num_regparam;
1823 if (get_mode_size_bits(mode) > 32)
1826 if (nr >= MAXNUM_GPREG_ARGS)
1829 if (cc & cc_this_call) {
1830 return gpreg_param_reg_this[nr];
1832 num_regparam = cc & ~cc_bits;
1833 if (num_regparam == 0) {
1834 /* default fastcall */
1835 return gpreg_param_reg_fastcall[nr];
1837 if (nr < num_regparam)
1838 return gpreg_param_reg_regparam[nr];
1842 panic("unknown argument mode");
1846 * Get the ABI restrictions for procedure calls.
1847 * @param self The this pointer.
1848 * @param method_type The type of the method (procedure) in question.
1849 * @param abi The abi object to be modified
1851 static void ia32_get_call_abi(const void *self, ir_type *method_type,
1859 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
1863 /* set abi flags for calls */
1864 call_flags.bits.left_to_right = 0; /* always last arg first on stack */
1865 call_flags.bits.store_args_sequential = 0;
1866 /* call_flags.bits.try_omit_fp not changed: can handle both settings */
1867 call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
1868 call_flags.bits.call_has_imm = 0; /* No call immediate, we handle this by ourselves */
1870 /* set parameter passing style */
1871 be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
1873 cc = get_method_calling_convention(method_type);
1874 if (get_method_variadicity(method_type) == variadicity_variadic) {
1875 /* pass all parameters of a variadic function on the stack */
1876 cc = cc_cdecl_set | (cc & cc_this_call);
1878 if (get_method_additional_properties(method_type) & mtp_property_private &&
1879 ia32_cg_config.optimize_cc) {
1880 /* set the fast calling conventions (allowing up to 3) */
1881 cc = SET_FASTCALL(cc) | 3;
1885 /* we have to pop the shadow parameter ourself for compound calls */
1886 if ( (get_method_calling_convention(method_type) & cc_compound_ret)
1887 && !(cc & cc_reg_param)) {
1888 pop_amount += get_mode_size_bytes(mode_P_data);
1891 n = get_method_n_params(method_type);
1892 for (i = regnum = 0; i < n; i++) {
1894 const arch_register_t *reg = NULL;
1896 tp = get_method_param_type(method_type, i);
1897 mode = get_type_mode(tp);
1899 reg = ia32_get_RegParam_reg(cc, regnum, mode);
1902 be_abi_call_param_reg(abi, i, reg, ABI_CONTEXT_BOTH);
1905 /* Micro optimisation: if the mode is shorter than 4 bytes, load 4 bytes.
1906 * movl has a shorter opcode than mov[sz][bw]l */
1907 ir_mode *load_mode = mode;
1910 unsigned size = get_mode_size_bytes(mode);
1912 if (cc & cc_callee_clear_stk) {
1913 pop_amount += (size + 3U) & ~3U;
1916 if (size < 4) load_mode = mode_Iu;
1919 be_abi_call_param_stack(abi, i, load_mode, 4, 0, 0, ABI_CONTEXT_BOTH);
1923 be_abi_call_set_pop(abi, pop_amount);
1925 /* set return registers */
1926 n = get_method_n_ress(method_type);
1928 assert(n <= 2 && "more than two results not supported");
1930 /* In case of 64bit returns, we will have two 32bit values */
1932 tp = get_method_res_type(method_type, 0);
1933 mode = get_type_mode(tp);
1935 assert(!mode_is_float(mode) && "two FP results not supported");
1937 tp = get_method_res_type(method_type, 1);
1938 mode = get_type_mode(tp);
1940 assert(!mode_is_float(mode) && "mixed INT, FP results not supported");
1942 be_abi_call_res_reg(abi, 0, &ia32_gp_regs[REG_EAX], ABI_CONTEXT_BOTH);
1943 be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX], ABI_CONTEXT_BOTH);
1946 const arch_register_t *reg;
1948 tp = get_method_res_type(method_type, 0);
1949 assert(is_atomic_type(tp));
1950 mode = get_type_mode(tp);
1952 reg = mode_is_float(mode) ? &ia32_vfp_regs[REG_VF0] : &ia32_gp_regs[REG_EAX];
1954 be_abi_call_res_reg(abi, 0, reg, ABI_CONTEXT_BOTH);
1958 static int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn)
1962 if (!is_ia32_irn(irn)) {
1966 if (is_ia32_NoReg_GP(irn) || is_ia32_NoReg_VFP(irn) || is_ia32_NoReg_XMM(irn)
1967 || is_ia32_ChangeCW(irn) || is_ia32_Immediate(irn))
1974 * Initializes the code generator interface.
1976 static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self)
1979 return &ia32_code_gen_if;
1983 * Returns the estimated execution time of an ia32 irn.
1985 static sched_timestep_t ia32_sched_exectime(void *env, const ir_node *irn)
1988 return is_ia32_irn(irn) ? ia32_get_op_estimated_cost(irn) : 1;
1991 list_sched_selector_t ia32_sched_selector;
1994 * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
1996 static const list_sched_selector_t *ia32_get_list_sched_selector(
1997 const void *self, list_sched_selector_t *selector)
2000 memcpy(&ia32_sched_selector, selector, sizeof(ia32_sched_selector));
2001 ia32_sched_selector.exectime = ia32_sched_exectime;
2002 ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
2003 return &ia32_sched_selector;
2006 static const ilp_sched_selector_t *ia32_get_ilp_sched_selector(const void *self)
2013 * Returns the necessary byte alignment for storing a register of given class.
2015 static int ia32_get_reg_class_alignment(const arch_register_class_t *cls)
2017 ir_mode *mode = arch_register_class_mode(cls);
2018 int bytes = get_mode_size_bytes(mode);
2020 if (mode_is_float(mode) && bytes > 8)
2025 static const be_execution_unit_t ***ia32_get_allowed_execution_units(
2028 static const be_execution_unit_t *_allowed_units_BRANCH[] = {
2029 &ia32_execution_units_BRANCH[IA32_EXECUNIT_TP_BRANCH_BRANCH1],
2030 &ia32_execution_units_BRANCH[IA32_EXECUNIT_TP_BRANCH_BRANCH2],
2033 static const be_execution_unit_t *_allowed_units_GP[] = {
2034 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EAX],
2035 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EBX],
2036 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_ECX],
2037 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EDX],
2038 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_ESI],
2039 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EDI],
2040 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EBP],
2043 static const be_execution_unit_t *_allowed_units_DUMMY[] = {
2044 &be_machine_execution_units_DUMMY[0],
2047 static const be_execution_unit_t **_units_callret[] = {
2048 _allowed_units_BRANCH,
2051 static const be_execution_unit_t **_units_other[] = {
2055 static const be_execution_unit_t **_units_dummy[] = {
2056 _allowed_units_DUMMY,
2059 const be_execution_unit_t ***ret;
2061 if (is_ia32_irn(irn)) {
2062 ret = get_ia32_exec_units(irn);
2063 } else if (is_be_node(irn)) {
2064 if (be_is_Return(irn)) {
2065 ret = _units_callret;
2066 } else if (be_is_Barrier(irn)) {
2080 * Return the abstract ia32 machine.
2082 static const be_machine_t *ia32_get_machine(const void *self)
2084 const ia32_isa_t *isa = self;
2089 * Return irp irgs in the desired order.
2091 static ir_graph **ia32_get_irg_list(const void *self, ir_graph ***irg_list)
2098 static void ia32_mark_remat(ir_node *node)
2100 if (is_ia32_irn(node)) {
2101 set_ia32_is_remat(node);
2106 * Check if Mux(sel, t, f) would represent an Abs (or -Abs).
2108 static bool mux_is_abs(ir_node *sel, ir_node *mux_true, ir_node *mux_false)
2117 cmp = get_Proj_pred(sel);
2121 /* must be <, <=, >=, > */
2122 pnc = get_Proj_proj(sel);
2137 if (!is_negated_value(mux_true, mux_false))
2140 /* must be x cmp 0 */
2141 cmp_right = get_Cmp_right(cmp);
2142 if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
2145 cmp_left = get_Cmp_left(cmp);
2146 if (cmp_left != mux_true && cmp_left != mux_false)
2153 * Check if Mux(sel, mux_true, mux_false) would represent a Max or Min operation
2155 static bool mux_is_float_min_max(ir_node *sel, ir_node *mux_true,
2165 cmp = get_Proj_pred(sel);
2169 cmp_l = get_Cmp_left(cmp);
2170 cmp_r = get_Cmp_right(cmp);
2171 if (!mode_is_float(get_irn_mode(cmp_l)))
2174 /* check for min/max. They're defined as (C-Semantik):
2175 * min(a, b) = a < b ? a : b
2176 * or min(a, b) = a <= b ? a : b
2177 * max(a, b) = a > b ? a : b
2178 * or max(a, b) = a >= b ? a : b
2179 * (Note we only handle float min/max here)
2181 pnc = get_Proj_proj(sel);
2186 if (cmp_l == mux_true && cmp_r == mux_false)
2192 if (cmp_l == mux_true && cmp_r == mux_false)
2198 if (cmp_l == mux_false && cmp_r == mux_true)
2204 if (cmp_l == mux_false && cmp_r == mux_true)
2215 static bool mux_is_set(ir_node *sel, ir_node *mux_true, ir_node *mux_false)
2217 ir_mode *mode = get_irn_mode(mux_true);
2220 if (!mode_is_int(mode) && !mode_is_reference(mode)
2224 if (is_Const(mux_true) && is_Const(mux_false)) {
2225 /* we can create a set plus up two 3 instructions for any combination of constants */
2232 static bool mux_is_float_const_const(ir_node *sel, ir_node *mux_true,
2237 if (!mode_is_float(get_irn_mode(mux_true)))
2240 return is_Const(mux_true) && is_Const(mux_false);
2243 static bool mux_is_doz(ir_node *sel, ir_node *mux_true, ir_node *mux_false)
2254 cmp = get_Proj_pred(sel);
2258 mode = get_irn_mode(mux_true);
2259 if (mode_is_signed(mode) || mode_is_float(mode))
2262 pn = get_Proj_proj(sel);
2263 cmp_left = get_Cmp_left(cmp);
2264 cmp_right = get_Cmp_right(cmp);
2265 if ((pn & pn_Cmp_Gt) &&
2266 is_Const(mux_false) && is_Const_null(mux_false) && is_Sub(mux_true) &&
2267 get_Sub_left(mux_true) == cmp_left &&
2268 get_Sub_right(mux_true) == cmp_right) {
2269 /* Mux(a >=u b, a - b, 0) unsigned Doz */
2272 if ((pn & pn_Cmp_Lt) &&
2273 is_Const(mux_true) && is_Const_null(mux_true) && is_Sub(mux_false) &&
2274 get_Sub_left(mux_false) == cmp_left &&
2275 get_Sub_right(mux_false) == cmp_right) {
2276 /* Mux(a <=u b, 0, a - b) unsigned Doz */
2283 static int ia32_is_mux_allowed(ir_node *sel, ir_node *mux_false,
2288 /* we can handle Abs for all modes and compares */
2289 if (mux_is_abs(sel, mux_true, mux_false))
2291 /* we can handle Set for all modes and compares */
2292 if (mux_is_set(sel, mux_true, mux_false))
2294 /* SSE has own min/max operations */
2295 if (ia32_cg_config.use_sse2
2296 && mux_is_float_min_max(sel, mux_true, mux_false))
2298 /* we can handle Mux(?, Const[f], Const[f]) */
2299 if (mux_is_float_const_const(sel, mux_true, mux_false)) {
2300 #ifdef FIRM_GRGEN_BE
2301 /* well, some code selectors can't handle it */
2302 if (be_transformer != TRANSFORMER_PBQP
2303 || be_transformer != TRANSFORMER_RAND)
2310 /* no support for 64bit inputs to cmov */
2311 mode = get_irn_mode(mux_true);
2312 if (get_mode_size_bits(mode) > 32)
2314 /* we can't handle MuxF yet */
2315 if (mode_is_float(mode))
2318 if (mux_is_doz(sel, mux_true, mux_false))
2321 /* Check Cmp before the node */
2323 ir_node *cmp = get_Proj_pred(sel);
2325 ir_mode *cmp_mode = get_irn_mode(get_Cmp_left(cmp));
2327 /* we can't handle 64bit compares */
2328 if (get_mode_size_bits(cmp_mode) > 32)
2331 /* we can't handle float compares */
2332 if (mode_is_float(cmp_mode))
2337 /* did we disable cmov generation? */
2338 if (!ia32_cg_config.use_cmov)
2341 /* we can use a cmov */
2345 static asm_constraint_flags_t ia32_parse_asm_constraint(const char **c)
2349 /* we already added all our simple flags to the flags modifier list in
2350 * init, so this flag we don't know. */
2351 return ASM_CONSTRAINT_FLAG_INVALID;
2354 static int ia32_is_valid_clobber(const char *clobber)
2356 return ia32_get_clobber_register(clobber) != NULL;
2360 * Create the trampoline code.
2362 static ir_node *ia32_create_trampoline_fkt(ir_node *block, ir_node *mem, ir_node *trampoline, ir_node *env, ir_node *callee)
2364 ir_node *st, *p = trampoline;
2365 ir_mode *mode = get_irn_mode(p);
2368 st = new_r_Store(block, mem, p, new_Const_long(mode_Bu, 0xb9), 0);
2369 mem = new_r_Proj(st, mode_M, pn_Store_M);
2370 p = new_r_Add(block, p, new_Const_long(mode_Iu, 1), mode);
2371 st = new_r_Store(block, mem, p, env, 0);
2372 mem = new_r_Proj(st, mode_M, pn_Store_M);
2373 p = new_r_Add(block, p, new_Const_long(mode_Iu, 4), mode);
2375 st = new_r_Store(block, mem, p, new_Const_long(mode_Bu, 0xe9), 0);
2376 mem = new_r_Proj(st, mode_M, pn_Store_M);
2377 p = new_r_Add(block, p, new_Const_long(mode_Iu, 1), mode);
2378 st = new_r_Store(block, mem, p, callee, 0);
2379 mem = new_r_Proj(st, mode_M, pn_Store_M);
2380 p = new_r_Add(block, p, new_Const_long(mode_Iu, 4), mode);
2386 * Returns the libFirm configuration parameter for this backend.
2388 static const backend_params *ia32_get_libfirm_params(void)
2390 static const ir_settings_if_conv_t ifconv = {
2391 4, /* maxdepth, doesn't matter for Mux-conversion */
2392 ia32_is_mux_allowed /* allows or disallows Mux creation for given selector */
2394 static const ir_settings_arch_dep_t ad = {
2395 1, /* also use subs */
2396 4, /* maximum shifts */
2397 31, /* maximum shift amount */
2398 ia32_evaluate_insn, /* evaluate the instruction sequence */
2400 1, /* allow Mulhs */
2401 1, /* allow Mulus */
2402 32, /* Mulh allowed up to 32 bit */
2404 static backend_params p = {
2405 1, /* need dword lowering */
2406 1, /* support inline assembly */
2407 NULL, /* will be set later */
2408 ia32_create_intrinsic_fkt,
2409 &intrinsic_env, /* context for ia32_create_intrinsic_fkt */
2410 NULL, /* ifconv info will be set below */
2411 NULL, /* float arithmetic mode, will be set below */
2412 12, /* size of trampoline code */
2413 4, /* alignment of trampoline code */
2414 ia32_create_trampoline_fkt,
2415 4 /* alignment of stack parameter */
2418 ia32_setup_cg_config();
2420 /* doesn't really belong here, but this is the earliest place the backend
2422 init_asm_constraints();
2425 p.if_conv_info = &ifconv;
2426 if (! ia32_cg_config.use_sse2)
2427 p.mode_float_arithmetic = mode_E;
2431 static const lc_opt_enum_int_items_t gas_items[] = {
2432 { "elf", OBJECT_FILE_FORMAT_ELF },
2433 { "mingw", OBJECT_FILE_FORMAT_COFF },
2434 { "macho", OBJECT_FILE_FORMAT_MACH_O },
2438 static lc_opt_enum_int_var_t gas_var = {
2439 (int*) &be_gas_object_file_format, gas_items
2442 #ifdef FIRM_GRGEN_BE
2443 static const lc_opt_enum_int_items_t transformer_items[] = {
2444 { "default", TRANSFORMER_DEFAULT },
2445 { "pbqp", TRANSFORMER_PBQP },
2446 { "random", TRANSFORMER_RAND },
2450 static lc_opt_enum_int_var_t transformer_var = {
2451 (int*)&be_transformer, transformer_items
2455 static const lc_opt_table_entry_t ia32_options[] = {
2456 LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
2457 #ifdef FIRM_GRGEN_BE
2458 LC_OPT_ENT_ENUM_INT("transformer", "the transformer used for code selection", &transformer_var),
2460 LC_OPT_ENT_INT("stackalign", "set power of two stack alignment for calls",
2461 &ia32_isa_template.arch_env.stack_alignment),
2465 const arch_isa_if_t ia32_isa_if = {
2468 ia32_handle_intrinsics,
2469 ia32_get_n_reg_class,
2471 ia32_get_reg_class_for_mode,
2473 ia32_get_code_generator_if,
2474 ia32_get_list_sched_selector,
2475 ia32_get_ilp_sched_selector,
2476 ia32_get_reg_class_alignment,
2477 ia32_get_libfirm_params,
2478 ia32_get_allowed_execution_units,
2482 ia32_parse_asm_constraint,
2483 ia32_is_valid_clobber
2486 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_ia32);
2487 void be_init_arch_ia32(void)
2489 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2490 lc_opt_entry_t *ia32_grp = lc_opt_get_grp(be_grp, "ia32");
2492 lc_opt_add_table(ia32_grp, ia32_options);
2493 be_register_isa_if("ia32", &ia32_isa_if);
2495 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.cg");
2497 ia32_init_emitter();
2499 ia32_init_optimize();
2500 ia32_init_transform();
2502 ia32_init_architecture();