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"
52 #include "iroptimize.h"
53 #include "instrument.h"
58 #include "../benode.h"
59 #include "../belower.h"
60 #include "../besched.h"
63 #include "../beirgmod.h"
64 #include "../be_dbgout.h"
65 #include "../beblocksched.h"
66 #include "../bemachine.h"
67 #include "../beilpsched.h"
68 #include "../bespillslots.h"
69 #include "../bemodule.h"
70 #include "../begnuas.h"
71 #include "../bestate.h"
72 #include "../beflags.h"
73 #include "../betranshlp.h"
74 #include "../belistsched.h"
76 #include "bearch_ia32_t.h"
78 #include "ia32_new_nodes.h"
79 #include "gen_ia32_regalloc_if.h"
80 #include "gen_ia32_machine.h"
81 #include "ia32_common_transform.h"
82 #include "ia32_transform.h"
83 #include "ia32_emitter.h"
84 #include "ia32_map_regs.h"
85 #include "ia32_optimize.h"
87 #include "ia32_dbg_stat.h"
88 #include "ia32_finish.h"
89 #include "ia32_util.h"
91 #include "ia32_architecture.h"
94 #include "ia32_pbqp_transform.h"
96 transformer_t be_transformer = TRANSFORMER_DEFAULT;
99 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
101 ir_mode *mode_fpcw = NULL;
102 ia32_code_gen_t *ia32_current_cg = NULL;
104 /** The current omit-fp state */
105 static unsigned ia32_curr_fp_ommitted = 0;
106 static ir_type *omit_fp_between_type = NULL;
107 static ir_type *between_type = NULL;
108 static ir_entity *old_bp_ent = NULL;
109 static ir_entity *ret_addr_ent = NULL;
110 static ir_entity *omit_fp_ret_addr_ent = NULL;
113 * The environment for the intrinsic mapping.
115 static ia32_intrinsic_env_t intrinsic_env = {
117 NULL, /* the irg, these entities belong to */
118 NULL, /* entity for __divdi3 library call */
119 NULL, /* entity for __moddi3 library call */
120 NULL, /* entity for __udivdi3 library call */
121 NULL, /* entity for __umoddi3 library call */
125 typedef ir_node *(*create_const_node_func) (dbg_info *dbg, ir_node *block);
128 * Used to create per-graph unique pseudo nodes.
130 static inline ir_node *create_const(ia32_code_gen_t *cg, ir_node **place,
131 create_const_node_func func,
132 const arch_register_t* reg)
134 ir_node *block, *res;
139 block = get_irg_start_block(cg->irg);
140 res = func(NULL, block);
141 arch_set_irn_register(res, reg);
147 /* Creates the unique per irg GP NoReg node. */
148 ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg)
150 return create_const(cg, &cg->noreg_gp, new_bd_ia32_NoReg_GP,
151 &ia32_gp_regs[REG_GP_NOREG]);
154 ir_node *ia32_new_NoReg_vfp(ia32_code_gen_t *cg)
156 return create_const(cg, &cg->noreg_vfp, new_bd_ia32_NoReg_VFP,
157 &ia32_vfp_regs[REG_VFP_NOREG]);
160 ir_node *ia32_new_NoReg_xmm(ia32_code_gen_t *cg)
162 return create_const(cg, &cg->noreg_xmm, new_bd_ia32_NoReg_XMM,
163 &ia32_xmm_regs[REG_XMM_NOREG]);
166 ir_node *ia32_new_Fpu_truncate(ia32_code_gen_t *cg)
168 return create_const(cg, &cg->fpu_trunc_mode, new_bd_ia32_ChangeCW,
169 &ia32_fp_cw_regs[REG_FPCW]);
174 * Returns the admissible noreg register node for input register pos of node irn.
176 static ir_node *ia32_get_admissible_noreg(ia32_code_gen_t *cg, ir_node *irn, int pos)
178 const arch_register_req_t *req = arch_get_register_req(irn, pos);
180 assert(req != NULL && "Missing register requirements");
181 if (req->cls == &ia32_reg_classes[CLASS_ia32_gp])
182 return ia32_new_NoReg_gp(cg);
184 if (ia32_cg_config.use_sse2) {
185 return ia32_new_NoReg_xmm(cg);
187 return ia32_new_NoReg_vfp(cg);
192 static const arch_register_req_t *get_ia32_SwitchJmp_out_req(
193 const ir_node *node, int pos)
197 return arch_no_register_req;
200 static arch_irn_class_t ia32_classify(const ir_node *irn)
202 arch_irn_class_t classification = 0;
204 assert(is_ia32_irn(irn));
206 if (is_ia32_is_reload(irn))
207 classification |= arch_irn_class_reload;
209 if (is_ia32_is_spill(irn))
210 classification |= arch_irn_class_spill;
212 if (is_ia32_is_remat(irn))
213 classification |= arch_irn_class_remat;
215 return classification;
219 * The IA32 ABI callback object.
222 be_abi_call_flags_bits_t flags; /**< The call flags. */
223 const arch_env_t *aenv; /**< The architecture environment. */
224 ir_graph *irg; /**< The associated graph. */
227 static ir_entity *ia32_get_frame_entity(const ir_node *irn)
229 return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
232 static void ia32_set_frame_entity(ir_node *node, ir_entity *entity)
234 if (is_be_node(node))
235 be_node_set_frame_entity(node, entity);
237 set_ia32_frame_ent(node, entity);
240 static void ia32_set_frame_offset(ir_node *irn, int bias)
242 if (get_ia32_frame_ent(irn) == NULL)
245 if (is_ia32_Pop(irn) || is_ia32_PopMem(irn)) {
246 ir_graph *irg = get_irn_irg(irn);
247 be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
248 if (layout->sp_relative) {
249 /* Pop nodes modify the stack pointer before calculating the
250 * destination address, so fix this here
255 add_ia32_am_offs_int(irn, bias);
258 static int ia32_get_sp_bias(const ir_node *node)
260 if (is_ia32_Call(node))
261 return -(int)get_ia32_call_attr_const(node)->pop;
263 if (is_ia32_Push(node))
266 if (is_ia32_Pop(node) || is_ia32_PopMem(node))
273 * Generate the routine prologue.
275 * @param self The callback object.
276 * @param mem A pointer to the mem node. Update this if you define new memory.
277 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
278 * @param stack_bias Points to the current stack bias, can be modified if needed.
280 * @return The register which shall be used as a stack frame base.
282 * All nodes which define registers in @p reg_map must keep @p reg_map current.
284 static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map, int *stack_bias)
286 ia32_abi_env_t *env = self;
287 ia32_code_gen_t *cg = ia32_current_cg;
288 const arch_env_t *arch_env = env->aenv;
290 ia32_curr_fp_ommitted = env->flags.try_omit_fp;
291 if (! env->flags.try_omit_fp) {
292 ir_node *bl = get_irg_start_block(env->irg);
293 ir_node *curr_sp = be_abi_reg_map_get(reg_map, arch_env->sp);
294 ir_node *curr_bp = be_abi_reg_map_get(reg_map, arch_env->bp);
295 ir_node *noreg = ia32_new_NoReg_gp(cg);
298 /* mark bp register as ignore */
299 be_set_constr_single_reg_out(get_Proj_pred(curr_bp),
300 get_Proj_proj(curr_bp), arch_env->bp, arch_register_req_type_ignore);
303 push = new_bd_ia32_Push(NULL, bl, noreg, noreg, *mem, curr_bp, curr_sp);
304 curr_sp = new_r_Proj(push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
305 *mem = new_r_Proj(push, mode_M, pn_ia32_Push_M);
307 /* the push must have SP out register */
308 arch_set_irn_register(curr_sp, arch_env->sp);
310 /* this modifies the stack bias, because we pushed 32bit */
313 /* move esp to ebp */
314 curr_bp = be_new_Copy(arch_env->bp->reg_class, bl, curr_sp);
315 be_set_constr_single_reg_out(curr_bp, 0, arch_env->bp,
316 arch_register_req_type_ignore);
318 /* beware: the copy must be done before any other sp use */
319 curr_sp = be_new_CopyKeep_single(arch_env->sp->reg_class, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
320 be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
321 arch_register_req_type_produces_sp);
323 be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
324 be_abi_reg_map_set(reg_map, arch_env->bp, curr_bp);
333 * Generate the routine epilogue.
334 * @param self The callback object.
335 * @param bl The block for the epilog
336 * @param mem A pointer to the mem node. Update this if you define new memory.
337 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
338 * @return The register which shall be used as a stack frame base.
340 * All nodes which define registers in @p reg_map must keep @p reg_map current.
342 static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
344 ia32_abi_env_t *env = self;
345 const arch_env_t *arch_env = env->aenv;
346 ir_node *curr_sp = be_abi_reg_map_get(reg_map, arch_env->sp);
347 ir_node *curr_bp = be_abi_reg_map_get(reg_map, arch_env->bp);
349 if (env->flags.try_omit_fp) {
350 /* simply remove the stack frame here */
351 curr_sp = be_new_IncSP(arch_env->sp, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK, 0);
353 ir_mode *mode_bp = arch_env->bp->reg_class->mode;
355 if (ia32_cg_config.use_leave) {
359 leave = new_bd_ia32_Leave(NULL, bl, curr_bp);
360 curr_bp = new_r_Proj(leave, mode_bp, pn_ia32_Leave_frame);
361 curr_sp = new_r_Proj(leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
365 /* the old SP is not needed anymore (kill the proj) */
366 assert(is_Proj(curr_sp));
369 /* copy ebp to esp */
370 curr_sp = be_new_Copy(&ia32_reg_classes[CLASS_ia32_gp], bl, curr_bp);
371 arch_set_irn_register(curr_sp, arch_env->sp);
372 be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
373 arch_register_req_type_ignore);
376 pop = new_bd_ia32_PopEbp(NULL, bl, *mem, curr_sp);
377 curr_bp = new_r_Proj(pop, mode_bp, pn_ia32_Pop_res);
378 curr_sp = new_r_Proj(pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
380 *mem = new_r_Proj(pop, mode_M, pn_ia32_Pop_M);
382 arch_set_irn_register(curr_sp, arch_env->sp);
383 arch_set_irn_register(curr_bp, arch_env->bp);
386 be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
387 be_abi_reg_map_set(reg_map, arch_env->bp, curr_bp);
391 * Initialize the callback object.
392 * @param call The call object.
393 * @param aenv The architecture environment.
394 * @param irg The graph with the method.
395 * @return Some pointer. This pointer is passed to all other callback functions as self object.
397 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
399 ia32_abi_env_t *env = XMALLOC(ia32_abi_env_t);
400 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
401 env->flags = fl.bits;
408 * Destroy the callback object.
409 * @param self The callback object.
411 static void ia32_abi_done(void *self)
417 * Build the between type and entities if not already build.
419 static void ia32_build_between_type(void)
421 #define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
422 if (! between_type) {
423 ir_type *old_bp_type = new_type_primitive(mode_Iu);
424 ir_type *ret_addr_type = new_type_primitive(mode_Iu);
426 between_type = new_type_struct(IDENT("ia32_between_type"));
427 old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
428 ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type);
430 set_entity_offset(old_bp_ent, 0);
431 set_entity_offset(ret_addr_ent, get_type_size_bytes(old_bp_type));
432 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
433 set_type_state(between_type, layout_fixed);
435 omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
436 omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type);
438 set_entity_offset(omit_fp_ret_addr_ent, 0);
439 set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
440 set_type_state(omit_fp_between_type, layout_fixed);
446 * Produces the type which sits between the stack args and the locals on the stack.
447 * it will contain the return address and space to store the old base pointer.
448 * @return The Firm type modeling the ABI between type.
450 static ir_type *ia32_abi_get_between_type(void *self)
452 ia32_abi_env_t *env = self;
454 ia32_build_between_type();
455 return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
459 * Return the stack entity that contains the return address.
461 ir_entity *ia32_get_return_address_entity(void)
463 ia32_build_between_type();
464 return ia32_curr_fp_ommitted ? omit_fp_ret_addr_ent : ret_addr_ent;
468 * Return the stack entity that contains the frame address.
470 ir_entity *ia32_get_frame_address_entity(void)
472 ia32_build_between_type();
473 return ia32_curr_fp_ommitted ? NULL : old_bp_ent;
477 * Get the estimated cycle count for @p irn.
479 * @param self The this pointer.
480 * @param irn The node.
482 * @return The estimated cycle count for this operation
484 static int ia32_get_op_estimated_cost(const ir_node *irn)
487 ia32_op_type_t op_tp;
491 if (!is_ia32_irn(irn))
494 assert(is_ia32_irn(irn));
496 cost = get_ia32_latency(irn);
497 op_tp = get_ia32_op_type(irn);
499 if (is_ia32_CopyB(irn)) {
502 else if (is_ia32_CopyB_i(irn)) {
503 int size = get_ia32_copyb_size(irn);
504 cost = 20 + (int)ceil((4/3) * size);
506 /* in case of address mode operations add additional cycles */
507 else if (op_tp == ia32_AddrModeD || op_tp == ia32_AddrModeS) {
509 In case of stack access and access to fixed addresses add 5 cycles
510 (we assume they are in cache), other memory operations cost 20
513 if (is_ia32_use_frame(irn) || (
514 is_ia32_NoReg_GP(get_irn_n(irn, n_ia32_base)) &&
515 is_ia32_NoReg_GP(get_irn_n(irn, n_ia32_index))
527 * Returns the inverse operation if @p irn, recalculating the argument at position @p i.
529 * @param irn The original operation
530 * @param i Index of the argument we want the inverse operation to yield
531 * @param inverse struct to be filled with the resulting inverse op
532 * @param obstack The obstack to use for allocation of the returned nodes array
533 * @return The inverse operation or NULL if operation invertible
535 static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst)
546 ir_node *block, *noreg, *nomem;
549 /* we cannot invert non-ia32 irns */
550 if (! is_ia32_irn(irn))
553 /* operand must always be a real operand (not base, index or mem) */
554 if (i != n_ia32_binary_left && i != n_ia32_binary_right)
557 /* we don't invert address mode operations */
558 if (get_ia32_op_type(irn) != ia32_Normal)
561 /* TODO: adjust for new immediates... */
562 ir_fprintf(stderr, "TODO: fix get_inverse for new immediates (%+F)\n",
566 block = get_nodes_block(irn);
567 mode = get_irn_mode(irn);
568 irn_mode = get_irn_mode(irn);
569 noreg = get_irn_n(irn, 0);
571 dbg = get_irn_dbg_info(irn);
573 /* initialize structure */
574 inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0]));
578 switch (get_ia32_irn_opcode(irn)) {
581 if (get_ia32_immop_type(irn) == ia32_ImmConst) {
582 /* we have an add with a const here */
583 /* invers == add with negated const */
584 inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
586 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
587 set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
588 set_ia32_commutative(inverse->nodes[0]);
590 else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
591 /* we have an add with a symconst here */
592 /* invers == sub with const */
593 inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
595 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
598 /* normal add: inverse == sub */
599 inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, i ^ 1));
606 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
607 /* we have a sub with a const/symconst here */
608 /* invers == add with this const */
609 inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
610 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
611 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
615 if (i == n_ia32_binary_left) {
616 inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, 3));
619 inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, get_irn_n(irn, n_ia32_binary_left), (ir_node*) irn);
627 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
628 /* xor with const: inverse = xor */
629 inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
630 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
631 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
635 inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, (ir_node *) irn, get_irn_n(irn, i));
641 inverse->nodes[0] = new_bd_ia32_Not(dbg, block, (ir_node*) irn);
646 inverse->nodes[0] = new_bd_ia32_Neg(dbg, block, (ir_node*) irn);
651 /* inverse operation not supported */
659 static ir_mode *get_spill_mode_mode(const ir_mode *mode)
661 if (mode_is_float(mode))
668 * Get the mode that should be used for spilling value node
670 static ir_mode *get_spill_mode(const ir_node *node)
672 ir_mode *mode = get_irn_mode(node);
673 return get_spill_mode_mode(mode);
677 * Checks whether an addressmode reload for a node with mode mode is compatible
678 * with a spillslot of mode spill_mode
680 static int ia32_is_spillmode_compatible(const ir_mode *mode, const ir_mode *spillmode)
682 return !mode_is_float(mode) || mode == spillmode;
686 * Check if irn can load its operand at position i from memory (source addressmode).
687 * @param irn The irn to be checked
688 * @param i The operands position
689 * @return Non-Zero if operand can be loaded
691 static int ia32_possible_memory_operand(const ir_node *irn, unsigned int i)
693 ir_node *op = get_irn_n(irn, i);
694 const ir_mode *mode = get_irn_mode(op);
695 const ir_mode *spillmode = get_spill_mode(op);
697 if (!is_ia32_irn(irn) || /* must be an ia32 irn */
698 get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
699 !ia32_is_spillmode_compatible(mode, spillmode) ||
700 is_ia32_use_frame(irn)) /* must not already use frame */
703 switch (get_ia32_am_support(irn)) {
708 if (i != n_ia32_unary_op)
714 case n_ia32_binary_left: {
715 const arch_register_req_t *req;
716 if (!is_ia32_commutative(irn))
719 /* we can't swap left/right for limited registers
720 * (As this (currently) breaks constraint handling copies)
722 req = get_ia32_in_req(irn, n_ia32_binary_left);
723 if (req->type & arch_register_req_type_limited)
728 case n_ia32_binary_right:
737 panic("Unknown AM type");
740 /* HACK: must not already use "real" memory.
741 * This can happen for Call and Div */
742 if (!is_NoMem(get_irn_n(irn, n_ia32_mem)))
748 static void ia32_perform_memory_operand(ir_node *irn, ir_node *spill,
752 ir_mode *dest_op_mode;
754 assert(ia32_possible_memory_operand(irn, i) && "Cannot perform memory operand change");
756 set_ia32_op_type(irn, ia32_AddrModeS);
758 load_mode = get_irn_mode(get_irn_n(irn, i));
759 dest_op_mode = get_ia32_ls_mode(irn);
760 if (get_mode_size_bits(load_mode) <= get_mode_size_bits(dest_op_mode)) {
761 set_ia32_ls_mode(irn, load_mode);
763 set_ia32_use_frame(irn);
764 set_ia32_need_stackent(irn);
766 if (i == n_ia32_binary_left &&
767 get_ia32_am_support(irn) == ia32_am_binary &&
768 /* immediates are only allowed on the right side */
769 !is_ia32_Immediate(get_irn_n(irn, n_ia32_binary_right))) {
770 ia32_swap_left_right(irn);
771 i = n_ia32_binary_right;
774 assert(is_NoMem(get_irn_n(irn, n_ia32_mem)));
776 set_irn_n(irn, n_ia32_base, get_irg_frame(get_irn_irg(irn)));
777 set_irn_n(irn, n_ia32_mem, spill);
778 set_irn_n(irn, i, ia32_get_admissible_noreg(ia32_current_cg, irn, i));
779 set_ia32_is_reload(irn);
782 static const be_abi_callbacks_t ia32_abi_callbacks = {
785 ia32_abi_get_between_type,
790 /* register allocator interface */
791 static const arch_irn_ops_t ia32_irn_ops = {
794 ia32_get_frame_entity,
795 ia32_set_frame_offset,
798 ia32_get_op_estimated_cost,
799 ia32_possible_memory_operand,
800 ia32_perform_memory_operand,
803 /* special register allocator interface for SwitchJmp
804 as it possibly has a WIDE range of Proj numbers.
805 We don't want to allocate output for register constraints for
807 static const arch_irn_ops_t ia32_SwitchJmp_irn_ops = {
808 /* Note: we also use SwitchJmp_out_req for the inputs too:
809 This is because the bearch API has a conceptual problem at the moment.
810 Querying for negative proj numbers which can happen for switchs
811 isn't possible and will result in inputs getting queried */
812 get_ia32_SwitchJmp_out_req,
814 ia32_get_frame_entity,
815 ia32_set_frame_offset,
818 ia32_get_op_estimated_cost,
819 ia32_possible_memory_operand,
820 ia32_perform_memory_operand,
824 static ir_entity *mcount = NULL;
826 #define ID(s) new_id_from_chars(s, sizeof(s) - 1)
828 static void ia32_before_abi(void *self)
830 lower_mode_b_config_t lower_mode_b_config = {
831 mode_Iu, /* lowered mode */
832 mode_Bu, /* preferred mode for set */
833 0, /* don't lower direct compares */
835 ia32_code_gen_t *cg = self;
837 ir_lower_mode_b(cg->irg, &lower_mode_b_config);
839 dump_ir_graph(cg->irg, "lower_modeb");
842 if (mcount == NULL) {
843 ir_type *tp = new_type_method(0, 0);
844 mcount = new_entity(get_glob_type(), ID("mcount"), tp);
845 /* FIXME: enter the right ld_ident here */
846 set_entity_ld_ident(mcount, get_entity_ident(mcount));
847 set_entity_visibility(mcount, ir_visibility_external);
849 instrument_initcall(cg->irg, mcount);
854 * Transforms the standard firm graph into
857 static void ia32_prepare_graph(void *self)
859 ia32_code_gen_t *cg = self;
862 switch (be_transformer) {
863 case TRANSFORMER_DEFAULT:
864 /* transform remaining nodes into assembler instructions */
865 ia32_transform_graph(cg);
868 case TRANSFORMER_PBQP:
869 case TRANSFORMER_RAND:
870 /* transform nodes into assembler instructions by PBQP magic */
871 ia32_transform_graph_by_pbqp(cg);
875 panic("invalid transformer");
878 ia32_transform_graph(cg);
881 /* do local optimizations (mainly CSE) */
882 optimize_graph_df(cg->irg);
885 dump_ir_graph(cg->irg, "transformed");
887 /* optimize address mode */
888 ia32_optimize_graph(cg);
890 /* do code placement, to optimize the position of constants */
894 dump_ir_graph(cg->irg, "place");
897 ir_node *turn_back_am(ir_node *node)
899 dbg_info *dbgi = get_irn_dbg_info(node);
900 ir_node *block = get_nodes_block(node);
901 ir_node *base = get_irn_n(node, n_ia32_base);
902 ir_node *index = get_irn_n(node, n_ia32_index);
903 ir_node *mem = get_irn_n(node, n_ia32_mem);
906 ir_node *load = new_bd_ia32_Load(dbgi, block, base, index, mem);
907 ir_node *load_res = new_rd_Proj(dbgi, load, mode_Iu, pn_ia32_Load_res);
909 ia32_copy_am_attrs(load, node);
910 if (is_ia32_is_reload(node))
911 set_ia32_is_reload(load);
912 set_irn_n(node, n_ia32_mem, new_NoMem());
914 switch (get_ia32_am_support(node)) {
916 set_irn_n(node, n_ia32_unary_op, load_res);
920 if (is_ia32_Immediate(get_irn_n(node, n_ia32_binary_right))) {
921 set_irn_n(node, n_ia32_binary_left, load_res);
923 set_irn_n(node, n_ia32_binary_right, load_res);
928 panic("Unknown AM type");
930 noreg = ia32_new_NoReg_gp(ia32_current_cg);
931 set_irn_n(node, n_ia32_base, noreg);
932 set_irn_n(node, n_ia32_index, noreg);
933 set_ia32_am_offs_int(node, 0);
934 set_ia32_am_sc(node, NULL);
935 set_ia32_am_scale(node, 0);
936 clear_ia32_am_sc_sign(node);
938 /* rewire mem-proj */
939 if (get_irn_mode(node) == mode_T) {
940 const ir_edge_t *edge;
941 foreach_out_edge(node, edge) {
942 ir_node *out = get_edge_src_irn(edge);
943 if (get_irn_mode(out) == mode_M) {
944 set_Proj_pred(out, load);
945 set_Proj_proj(out, pn_ia32_Load_M);
951 set_ia32_op_type(node, ia32_Normal);
952 if (sched_is_scheduled(node))
953 sched_add_before(node, load);
958 static ir_node *flags_remat(ir_node *node, ir_node *after)
960 /* we should turn back source address mode when rematerializing nodes */
965 if (is_Block(after)) {
968 block = get_nodes_block(after);
971 type = get_ia32_op_type(node);
978 /* TODO implement this later... */
979 panic("found DestAM with flag user %+F this should not happen", node);
982 default: assert(type == ia32_Normal); break;
985 copy = exact_copy(node);
986 set_nodes_block(copy, block);
987 sched_add_after(after, copy);
993 * Called before the register allocator.
995 static void ia32_before_ra(void *self)
997 ia32_code_gen_t *cg = self;
999 /* setup fpu rounding modes */
1000 ia32_setup_fpu_mode(cg);
1003 be_sched_fix_flags(cg->irg, &ia32_reg_classes[CLASS_ia32_flags],
1006 ia32_add_missing_keeps(cg);
1011 * Transforms a be_Reload into a ia32 Load.
1013 static void transform_to_Load(ia32_code_gen_t *cg, ir_node *node)
1015 ir_graph *irg = get_irn_irg(node);
1016 dbg_info *dbg = get_irn_dbg_info(node);
1017 ir_node *block = get_nodes_block(node);
1018 ir_entity *ent = be_get_frame_entity(node);
1019 ir_mode *mode = get_irn_mode(node);
1020 ir_mode *spillmode = get_spill_mode(node);
1021 ir_node *noreg = ia32_new_NoReg_gp(cg);
1022 ir_node *sched_point = NULL;
1023 ir_node *ptr = get_irg_frame(irg);
1024 ir_node *mem = get_irn_n(node, be_pos_Reload_mem);
1025 ir_node *new_op, *proj;
1026 const arch_register_t *reg;
1028 if (sched_is_scheduled(node)) {
1029 sched_point = sched_prev(node);
1032 if (mode_is_float(spillmode)) {
1033 if (ia32_cg_config.use_sse2)
1034 new_op = new_bd_ia32_xLoad(dbg, block, ptr, noreg, mem, spillmode);
1036 new_op = new_bd_ia32_vfld(dbg, block, ptr, noreg, mem, spillmode);
1038 else if (get_mode_size_bits(spillmode) == 128) {
1039 /* Reload 128 bit SSE registers */
1040 new_op = new_bd_ia32_xxLoad(dbg, block, ptr, noreg, mem);
1043 new_op = new_bd_ia32_Load(dbg, block, ptr, noreg, mem);
1045 set_ia32_op_type(new_op, ia32_AddrModeS);
1046 set_ia32_ls_mode(new_op, spillmode);
1047 set_ia32_frame_ent(new_op, ent);
1048 set_ia32_use_frame(new_op);
1049 set_ia32_is_reload(new_op);
1051 DBG_OPT_RELOAD2LD(node, new_op);
1053 proj = new_rd_Proj(dbg, new_op, mode, pn_ia32_Load_res);
1056 sched_add_after(sched_point, new_op);
1060 /* copy the register from the old node to the new Load */
1061 reg = arch_get_irn_register(node);
1062 arch_set_irn_register(proj, reg);
1064 SET_IA32_ORIG_NODE(new_op, node);
1066 exchange(node, proj);
1070 * Transforms a be_Spill node into a ia32 Store.
1072 static void transform_to_Store(ia32_code_gen_t *cg, ir_node *node)
1074 ir_graph *irg = get_irn_irg(node);
1075 dbg_info *dbg = get_irn_dbg_info(node);
1076 ir_node *block = get_nodes_block(node);
1077 ir_entity *ent = be_get_frame_entity(node);
1078 const ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
1079 ir_mode *mode = get_spill_mode(spillval);
1080 ir_node *noreg = ia32_new_NoReg_gp(cg);
1081 ir_node *nomem = new_NoMem();
1082 ir_node *ptr = get_irg_frame(irg);
1083 ir_node *val = get_irn_n(node, be_pos_Spill_val);
1085 ir_node *sched_point = NULL;
1087 if (sched_is_scheduled(node)) {
1088 sched_point = sched_prev(node);
1091 if (mode_is_float(mode)) {
1092 if (ia32_cg_config.use_sse2)
1093 store = new_bd_ia32_xStore(dbg, block, ptr, noreg, nomem, val);
1095 store = new_bd_ia32_vfst(dbg, block, ptr, noreg, nomem, val, mode);
1096 } else if (get_mode_size_bits(mode) == 128) {
1097 /* Spill 128 bit SSE registers */
1098 store = new_bd_ia32_xxStore(dbg, block, ptr, noreg, nomem, val);
1099 } else if (get_mode_size_bits(mode) == 8) {
1100 store = new_bd_ia32_Store8Bit(dbg, block, ptr, noreg, nomem, val);
1102 store = new_bd_ia32_Store(dbg, block, ptr, noreg, nomem, val);
1105 set_ia32_op_type(store, ia32_AddrModeD);
1106 set_ia32_ls_mode(store, mode);
1107 set_ia32_frame_ent(store, ent);
1108 set_ia32_use_frame(store);
1109 set_ia32_is_spill(store);
1110 SET_IA32_ORIG_NODE(store, node);
1111 DBG_OPT_SPILL2ST(node, store);
1114 sched_add_after(sched_point, store);
1118 exchange(node, store);
1121 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)
1123 dbg_info *dbg = get_irn_dbg_info(node);
1124 ir_node *block = get_nodes_block(node);
1125 ir_node *noreg = ia32_new_NoReg_gp(cg);
1126 ir_graph *irg = get_irn_irg(node);
1127 ir_node *frame = get_irg_frame(irg);
1129 ir_node *push = new_bd_ia32_Push(dbg, block, frame, noreg, mem, noreg, sp);
1131 set_ia32_frame_ent(push, ent);
1132 set_ia32_use_frame(push);
1133 set_ia32_op_type(push, ia32_AddrModeS);
1134 set_ia32_ls_mode(push, mode_Is);
1135 set_ia32_is_spill(push);
1137 sched_add_before(schedpoint, push);
1141 static ir_node *create_pop(ia32_code_gen_t *cg, ir_node *node, ir_node *schedpoint, ir_node *sp, ir_entity *ent)
1143 dbg_info *dbg = get_irn_dbg_info(node);
1144 ir_node *block = get_nodes_block(node);
1145 ir_node *noreg = ia32_new_NoReg_gp(cg);
1146 ir_graph *irg = get_irn_irg(node);
1147 ir_node *frame = get_irg_frame(irg);
1149 ir_node *pop = new_bd_ia32_PopMem(dbg, block, frame, noreg, new_NoMem(), sp);
1151 set_ia32_frame_ent(pop, ent);
1152 set_ia32_use_frame(pop);
1153 set_ia32_op_type(pop, ia32_AddrModeD);
1154 set_ia32_ls_mode(pop, mode_Is);
1155 set_ia32_is_reload(pop);
1157 sched_add_before(schedpoint, pop);
1162 static ir_node* create_spproj(ir_node *node, ir_node *pred, int pos)
1164 dbg_info *dbg = get_irn_dbg_info(node);
1165 ir_mode *spmode = mode_Iu;
1166 const arch_register_t *spreg = &ia32_gp_regs[REG_ESP];
1169 sp = new_rd_Proj(dbg, pred, spmode, pos);
1170 arch_set_irn_register(sp, spreg);
1176 * Transform MemPerm, currently we do this the ugly way and produce
1177 * push/pop into/from memory cascades. This is possible without using
1180 static void transform_MemPerm(ia32_code_gen_t *cg, ir_node *node)
1182 ir_node *block = get_nodes_block(node);
1183 ir_node *sp = be_abi_get_ignore_irn(be_get_irg_abi(cg->irg), &ia32_gp_regs[REG_ESP]);
1184 int arity = be_get_MemPerm_entity_arity(node);
1185 ir_node **pops = ALLOCAN(ir_node*, arity);
1189 const ir_edge_t *edge;
1190 const ir_edge_t *next;
1193 for (i = 0; i < arity; ++i) {
1194 ir_entity *inent = be_get_MemPerm_in_entity(node, i);
1195 ir_entity *outent = be_get_MemPerm_out_entity(node, i);
1196 ir_type *enttype = get_entity_type(inent);
1197 unsigned entsize = get_type_size_bytes(enttype);
1198 unsigned entsize2 = get_type_size_bytes(get_entity_type(outent));
1199 ir_node *mem = get_irn_n(node, i + 1);
1202 /* work around cases where entities have different sizes */
1203 if (entsize2 < entsize)
1205 assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
1207 push = create_push(cg, node, node, sp, mem, inent);
1208 sp = create_spproj(node, push, pn_ia32_Push_stack);
1210 /* add another push after the first one */
1211 push = create_push(cg, node, node, sp, mem, inent);
1212 add_ia32_am_offs_int(push, 4);
1213 sp = create_spproj(node, push, pn_ia32_Push_stack);
1216 set_irn_n(node, i, new_Bad());
1220 for (i = arity - 1; i >= 0; --i) {
1221 ir_entity *inent = be_get_MemPerm_in_entity(node, i);
1222 ir_entity *outent = be_get_MemPerm_out_entity(node, i);
1223 ir_type *enttype = get_entity_type(outent);
1224 unsigned entsize = get_type_size_bytes(enttype);
1225 unsigned entsize2 = get_type_size_bytes(get_entity_type(inent));
1228 /* work around cases where entities have different sizes */
1229 if (entsize2 < entsize)
1231 assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
1233 pop = create_pop(cg, node, node, sp, outent);
1234 sp = create_spproj(node, pop, pn_ia32_Pop_stack);
1236 add_ia32_am_offs_int(pop, 4);
1238 /* add another pop after the first one */
1239 pop = create_pop(cg, node, node, sp, outent);
1240 sp = create_spproj(node, pop, pn_ia32_Pop_stack);
1247 keep = be_new_Keep(block, 1, in);
1248 sched_add_before(node, keep);
1250 /* exchange memprojs */
1251 foreach_out_edge_safe(node, edge, next) {
1252 ir_node *proj = get_edge_src_irn(edge);
1253 int p = get_Proj_proj(proj);
1257 set_Proj_pred(proj, pops[p]);
1258 set_Proj_proj(proj, pn_ia32_Pop_M);
1261 /* remove memperm */
1262 arity = get_irn_arity(node);
1263 for (i = 0; i < arity; ++i) {
1264 set_irn_n(node, i, new_Bad());
1270 * Block-Walker: Calls the transform functions Spill and Reload.
1272 static void ia32_after_ra_walker(ir_node *block, void *env)
1274 ir_node *node, *prev;
1275 ia32_code_gen_t *cg = env;
1277 /* beware: the schedule is changed here */
1278 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
1279 prev = sched_prev(node);
1281 if (be_is_Reload(node)) {
1282 transform_to_Load(cg, node);
1283 } else if (be_is_Spill(node)) {
1284 transform_to_Store(cg, node);
1285 } else if (be_is_MemPerm(node)) {
1286 transform_MemPerm(cg, node);
1292 * Collects nodes that need frame entities assigned.
1294 static void ia32_collect_frame_entity_nodes(ir_node *node, void *data)
1296 be_fec_env_t *env = data;
1297 const ir_mode *mode;
1300 if (be_is_Reload(node) && be_get_frame_entity(node) == NULL) {
1301 mode = get_spill_mode_mode(get_irn_mode(node));
1302 align = get_mode_size_bytes(mode);
1303 } else if (is_ia32_irn(node) &&
1304 get_ia32_frame_ent(node) == NULL &&
1305 is_ia32_use_frame(node)) {
1306 if (is_ia32_need_stackent(node))
1309 switch (get_ia32_irn_opcode(node)) {
1311 case iro_ia32_Load: {
1312 const ia32_attr_t *attr = get_ia32_attr_const(node);
1314 if (attr->data.need_32bit_stackent) {
1316 } else if (attr->data.need_64bit_stackent) {
1319 mode = get_ia32_ls_mode(node);
1320 if (is_ia32_is_reload(node))
1321 mode = get_spill_mode_mode(mode);
1323 align = get_mode_size_bytes(mode);
1327 case iro_ia32_vfild:
1329 case iro_ia32_xLoad: {
1330 mode = get_ia32_ls_mode(node);
1335 case iro_ia32_FldCW: {
1336 /* although 2 byte would be enough 4 byte performs best */
1344 panic("unexpected frame user while collection frame entity nodes");
1346 case iro_ia32_FnstCW:
1347 case iro_ia32_Store8Bit:
1348 case iro_ia32_Store:
1351 case iro_ia32_vfist:
1352 case iro_ia32_vfisttp:
1354 case iro_ia32_xStore:
1355 case iro_ia32_xStoreSimple:
1362 be_node_needs_frame_entity(env, node, mode, align);
1366 * We transform Spill and Reload here. This needs to be done before
1367 * stack biasing otherwise we would miss the corrected offset for these nodes.
1369 static void ia32_after_ra(void *self)
1371 ia32_code_gen_t *cg = self;
1372 ir_graph *irg = cg->irg;
1373 be_fec_env_t *fec_env = be_new_frame_entity_coalescer(cg->irg);
1375 /* create and coalesce frame entities */
1376 irg_walk_graph(irg, NULL, ia32_collect_frame_entity_nodes, fec_env);
1377 be_assign_entities(fec_env, ia32_set_frame_entity);
1378 be_free_frame_entity_coalescer(fec_env);
1380 irg_block_walk_graph(irg, NULL, ia32_after_ra_walker, cg);
1384 * Last touchups for the graph before emit: x87 simulation to replace the
1385 * virtual with real x87 instructions, creating a block schedule and peephole
1388 static void ia32_finish(void *self)
1390 ia32_code_gen_t *cg = self;
1391 ir_graph *irg = cg->irg;
1393 ia32_finish_irg(irg, cg);
1395 /* we might have to rewrite x87 virtual registers */
1396 if (cg->do_x87_sim) {
1397 x87_simulate_graph(cg->irg);
1400 /* do peephole optimisations */
1401 ia32_peephole_optimization(cg);
1403 /* create block schedule, this also removes empty blocks which might
1404 * produce critical edges */
1405 cg->blk_sched = be_create_block_schedule(irg);
1409 * Emits the code, closes the output file and frees
1410 * the code generator interface.
1412 static void ia32_codegen(void *self)
1414 ia32_code_gen_t *cg = self;
1415 ir_graph *irg = cg->irg;
1417 if (ia32_cg_config.emit_machcode) {
1418 ia32_gen_binary_routine(cg, irg);
1420 ia32_gen_routine(cg, irg);
1423 /* remove it from the isa */
1426 assert(ia32_current_cg == cg);
1427 ia32_current_cg = NULL;
1429 /* de-allocate code generator */
1434 * Returns the node representing the PIC base.
1436 static ir_node *ia32_get_pic_base(void *self)
1439 ia32_code_gen_t *cg = self;
1440 ir_node *get_eip = cg->get_eip;
1441 if (get_eip != NULL)
1444 block = get_irg_start_block(cg->irg);
1445 get_eip = new_bd_ia32_GetEIP(NULL, block);
1446 cg->get_eip = get_eip;
1448 be_dep_on_frame(get_eip);
1452 static void *ia32_cg_init(ir_graph *irg);
1454 static const arch_code_generator_if_t ia32_code_gen_if = {
1456 ia32_get_pic_base, /* return node used as base in pic code addresses */
1457 ia32_before_abi, /* before abi introduce hook */
1460 ia32_before_ra, /* before register allocation hook */
1461 ia32_after_ra, /* after register allocation hook */
1462 ia32_finish, /* called before codegen */
1463 ia32_codegen /* emit && done */
1467 * Initializes a IA32 code generator.
1469 static void *ia32_cg_init(ir_graph *irg)
1471 ia32_isa_t *isa = (ia32_isa_t *)be_get_irg_arch_env(irg);
1472 ia32_code_gen_t *cg = XMALLOCZ(ia32_code_gen_t);
1474 cg->impl = &ia32_code_gen_if;
1477 cg->blk_sched = NULL;
1478 cg->dump = (be_get_irg_options(irg)->dump_flags & DUMP_BE) ? 1 : 0;
1479 cg->gprof = (be_get_irg_options(irg)->gprof) ? 1 : 0;
1482 /* Linux gprof implementation needs base pointer */
1483 be_get_irg_options(irg)->omit_fp = 0;
1490 if (isa->name_obst) {
1491 obstack_free(isa->name_obst, NULL);
1492 obstack_init(isa->name_obst);
1496 assert(ia32_current_cg == NULL);
1497 ia32_current_cg = cg;
1499 return (arch_code_generator_t *)cg;
1504 * Set output modes for GCC
1506 static const tarval_mode_info mo_integer = {
1513 * set the tarval output mode of all integer modes to decimal
1515 static void set_tarval_output_modes(void)
1519 for (i = get_irp_n_modes() - 1; i >= 0; --i) {
1520 ir_mode *mode = get_irp_mode(i);
1522 if (mode_is_int(mode))
1523 set_tarval_mode_output_option(mode, &mo_integer);
1527 const arch_isa_if_t ia32_isa_if;
1530 * The template that generates a new ISA object.
1531 * Note that this template can be changed by command line
1534 static ia32_isa_t ia32_isa_template = {
1536 &ia32_isa_if, /* isa interface implementation */
1537 &ia32_gp_regs[REG_ESP], /* stack pointer register */
1538 &ia32_gp_regs[REG_EBP], /* base pointer register */
1539 &ia32_reg_classes[CLASS_ia32_gp], /* static link pointer register class */
1540 -1, /* stack direction */
1541 2, /* power of two stack alignment, 2^2 == 4 */
1542 NULL, /* main environment */
1543 7, /* costs for a spill instruction */
1544 5, /* costs for a reload instruction */
1545 false, /* no custom abi handling */
1547 NULL, /* 16bit register names */
1548 NULL, /* 8bit register names */
1549 NULL, /* 8bit register names high */
1552 NULL, /* current code generator */
1553 NULL, /* abstract machine */
1555 NULL, /* name obstack */
1559 static void init_asm_constraints(void)
1561 be_init_default_asm_constraint_flags();
1563 asm_constraint_flags['a'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1564 asm_constraint_flags['b'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1565 asm_constraint_flags['c'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1566 asm_constraint_flags['d'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1567 asm_constraint_flags['D'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1568 asm_constraint_flags['S'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1569 asm_constraint_flags['Q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1570 asm_constraint_flags['q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1571 asm_constraint_flags['A'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1572 asm_constraint_flags['l'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1573 asm_constraint_flags['R'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1574 asm_constraint_flags['r'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1575 asm_constraint_flags['p'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1576 asm_constraint_flags['f'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1577 asm_constraint_flags['t'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1578 asm_constraint_flags['u'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1579 asm_constraint_flags['Y'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1580 asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
1581 asm_constraint_flags['n'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
1582 asm_constraint_flags['g'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
1584 /* no support for autodecrement/autoincrement */
1585 asm_constraint_flags['<'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1586 asm_constraint_flags['>'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1587 /* no float consts */
1588 asm_constraint_flags['E'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1589 asm_constraint_flags['F'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1590 /* makes no sense on x86 */
1591 asm_constraint_flags['s'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1592 /* no support for sse consts yet */
1593 asm_constraint_flags['C'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1594 /* no support for x87 consts yet */
1595 asm_constraint_flags['G'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1596 /* no support for mmx registers yet */
1597 asm_constraint_flags['y'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1598 /* not available in 32bit mode */
1599 asm_constraint_flags['Z'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1600 asm_constraint_flags['e'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1602 /* no code yet to determine register class needed... */
1603 asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
1607 * Initializes the backend ISA.
1609 static arch_env_t *ia32_init(FILE *file_handle)
1611 static int inited = 0;
1619 set_tarval_output_modes();
1621 isa = XMALLOC(ia32_isa_t);
1622 memcpy(isa, &ia32_isa_template, sizeof(*isa));
1624 if (mode_fpcw == NULL) {
1625 mode_fpcw = new_ir_mode("Fpcw", irms_int_number, 16, 0, irma_none, 0);
1628 ia32_register_init();
1629 ia32_create_opcodes(&ia32_irn_ops);
1630 /* special handling for SwitchJmp */
1631 op_ia32_SwitchJmp->ops.be_ops = &ia32_SwitchJmp_irn_ops;
1633 be_emit_init(file_handle);
1634 isa->regs_16bit = pmap_create();
1635 isa->regs_8bit = pmap_create();
1636 isa->regs_8bit_high = pmap_create();
1637 isa->types = pmap_create();
1638 isa->tv_ent = pmap_create();
1639 isa->cpu = ia32_init_machine_description();
1641 ia32_build_16bit_reg_map(isa->regs_16bit);
1642 ia32_build_8bit_reg_map(isa->regs_8bit);
1643 ia32_build_8bit_reg_map_high(isa->regs_8bit_high);
1646 isa->name_obst = XMALLOC(struct obstack);
1647 obstack_init(isa->name_obst);
1650 /* enter the ISA object into the intrinsic environment */
1651 intrinsic_env.isa = isa;
1653 /* emit asm includes */
1654 n = get_irp_n_asms();
1655 for (i = 0; i < n; ++i) {
1656 be_emit_cstring("#APP\n");
1657 be_emit_ident(get_irp_asm(i));
1658 be_emit_cstring("\n#NO_APP\n");
1661 /* needed for the debug support */
1662 be_gas_emit_switch_section(GAS_SECTION_TEXT);
1663 be_emit_irprintf("%stext0:\n", be_gas_get_private_prefix());
1664 be_emit_write_line();
1666 return &isa->arch_env;
1672 * Closes the output file and frees the ISA structure.
1674 static void ia32_done(void *self)
1676 ia32_isa_t *isa = self;
1678 /* emit now all global declarations */
1679 be_gas_emit_decls(isa->arch_env.main_env);
1681 pmap_destroy(isa->regs_16bit);
1682 pmap_destroy(isa->regs_8bit);
1683 pmap_destroy(isa->regs_8bit_high);
1684 pmap_destroy(isa->tv_ent);
1685 pmap_destroy(isa->types);
1688 obstack_free(isa->name_obst, NULL);
1698 * Return the number of register classes for this architecture.
1699 * We report always these:
1700 * - the general purpose registers
1701 * - the SSE floating point register set
1702 * - the virtual floating point registers
1703 * - the SSE vector register set
1705 static unsigned ia32_get_n_reg_class(void)
1711 * Return the register class for index i.
1713 static const arch_register_class_t *ia32_get_reg_class(unsigned i)
1715 assert(i < N_CLASSES);
1716 return &ia32_reg_classes[i];
1720 * Get the register class which shall be used to store a value of a given mode.
1721 * @param self The this pointer.
1722 * @param mode The mode in question.
1723 * @return A register class which can hold values of the given mode.
1725 static const arch_register_class_t *ia32_get_reg_class_for_mode(const ir_mode *mode)
1727 if (mode_is_float(mode)) {
1728 return ia32_cg_config.use_sse2 ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1731 return &ia32_reg_classes[CLASS_ia32_gp];
1735 * Returns the register for parameter nr.
1737 static const arch_register_t *ia32_get_RegParam_reg(unsigned cc, unsigned nr,
1738 const ir_mode *mode)
1740 static const arch_register_t *gpreg_param_reg_fastcall[] = {
1741 &ia32_gp_regs[REG_ECX],
1742 &ia32_gp_regs[REG_EDX],
1745 static const unsigned MAXNUM_GPREG_ARGS = 3;
1747 static const arch_register_t *gpreg_param_reg_regparam[] = {
1748 &ia32_gp_regs[REG_EAX],
1749 &ia32_gp_regs[REG_EDX],
1750 &ia32_gp_regs[REG_ECX]
1753 static const arch_register_t *gpreg_param_reg_this[] = {
1754 &ia32_gp_regs[REG_ECX],
1759 static const arch_register_t *fpreg_sse_param_reg_std[] = {
1760 &ia32_xmm_regs[REG_XMM0],
1761 &ia32_xmm_regs[REG_XMM1],
1762 &ia32_xmm_regs[REG_XMM2],
1763 &ia32_xmm_regs[REG_XMM3],
1764 &ia32_xmm_regs[REG_XMM4],
1765 &ia32_xmm_regs[REG_XMM5],
1766 &ia32_xmm_regs[REG_XMM6],
1767 &ia32_xmm_regs[REG_XMM7]
1770 static const arch_register_t *fpreg_sse_param_reg_this[] = {
1771 NULL, /* in case of a "this" pointer, the first parameter must not be a float */
1773 static const unsigned MAXNUM_SSE_ARGS = 8;
1775 if ((cc & cc_this_call) && nr == 0)
1776 return gpreg_param_reg_this[0];
1778 if (! (cc & cc_reg_param))
1781 if (mode_is_float(mode)) {
1782 if (!ia32_cg_config.use_sse2 || (cc & cc_fpreg_param) == 0)
1784 if (nr >= MAXNUM_SSE_ARGS)
1787 if (cc & cc_this_call) {
1788 return fpreg_sse_param_reg_this[nr];
1790 return fpreg_sse_param_reg_std[nr];
1791 } else if (mode_is_int(mode) || mode_is_reference(mode)) {
1792 unsigned num_regparam;
1794 if (get_mode_size_bits(mode) > 32)
1797 if (nr >= MAXNUM_GPREG_ARGS)
1800 if (cc & cc_this_call) {
1801 return gpreg_param_reg_this[nr];
1803 num_regparam = cc & ~cc_bits;
1804 if (num_regparam == 0) {
1805 /* default fastcall */
1806 return gpreg_param_reg_fastcall[nr];
1808 if (nr < num_regparam)
1809 return gpreg_param_reg_regparam[nr];
1813 panic("unknown argument mode");
1817 * Get the ABI restrictions for procedure calls.
1818 * @param self The this pointer.
1819 * @param method_type The type of the method (procedure) in question.
1820 * @param abi The abi object to be modified
1822 static void ia32_get_call_abi(const void *self, ir_type *method_type,
1830 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
1834 /* set abi flags for calls */
1835 call_flags.bits.left_to_right = 0; /* always last arg first on stack */
1836 call_flags.bits.store_args_sequential = 0;
1837 /* call_flags.bits.try_omit_fp not changed: can handle both settings */
1838 call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
1839 call_flags.bits.call_has_imm = 0; /* No call immediate, we handle this by ourselves */
1841 /* set parameter passing style */
1842 be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
1844 cc = get_method_calling_convention(method_type);
1845 if (get_method_variadicity(method_type) == variadicity_variadic) {
1846 /* pass all parameters of a variadic function on the stack */
1847 cc = cc_cdecl_set | (cc & cc_this_call);
1849 if (get_method_additional_properties(method_type) & mtp_property_private &&
1850 ia32_cg_config.optimize_cc) {
1851 /* set the fast calling conventions (allowing up to 3) */
1852 cc = SET_FASTCALL(cc) | 3;
1856 /* we have to pop the shadow parameter ourself for compound calls */
1857 if ( (get_method_calling_convention(method_type) & cc_compound_ret)
1858 && !(cc & cc_reg_param)) {
1859 pop_amount += get_mode_size_bytes(mode_P_data);
1862 n = get_method_n_params(method_type);
1863 for (i = regnum = 0; i < n; i++) {
1865 const arch_register_t *reg = NULL;
1867 tp = get_method_param_type(method_type, i);
1868 mode = get_type_mode(tp);
1870 reg = ia32_get_RegParam_reg(cc, regnum, mode);
1873 be_abi_call_param_reg(abi, i, reg, ABI_CONTEXT_BOTH);
1876 /* Micro optimisation: if the mode is shorter than 4 bytes, load 4 bytes.
1877 * movl has a shorter opcode than mov[sz][bw]l */
1878 ir_mode *load_mode = mode;
1881 unsigned size = get_mode_size_bytes(mode);
1883 if (cc & cc_callee_clear_stk) {
1884 pop_amount += (size + 3U) & ~3U;
1887 if (size < 4) load_mode = mode_Iu;
1890 be_abi_call_param_stack(abi, i, load_mode, 4, 0, 0, ABI_CONTEXT_BOTH);
1894 be_abi_call_set_pop(abi, pop_amount);
1896 /* set return registers */
1897 n = get_method_n_ress(method_type);
1899 assert(n <= 2 && "more than two results not supported");
1901 /* In case of 64bit returns, we will have two 32bit values */
1903 tp = get_method_res_type(method_type, 0);
1904 mode = get_type_mode(tp);
1906 assert(!mode_is_float(mode) && "two FP results not supported");
1908 tp = get_method_res_type(method_type, 1);
1909 mode = get_type_mode(tp);
1911 assert(!mode_is_float(mode) && "mixed INT, FP results not supported");
1913 be_abi_call_res_reg(abi, 0, &ia32_gp_regs[REG_EAX], ABI_CONTEXT_BOTH);
1914 be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX], ABI_CONTEXT_BOTH);
1917 const arch_register_t *reg;
1919 tp = get_method_res_type(method_type, 0);
1920 assert(is_atomic_type(tp));
1921 mode = get_type_mode(tp);
1923 reg = mode_is_float(mode) ? &ia32_vfp_regs[REG_VF0] : &ia32_gp_regs[REG_EAX];
1925 be_abi_call_res_reg(abi, 0, reg, ABI_CONTEXT_BOTH);
1929 static int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn)
1933 if (!is_ia32_irn(irn)) {
1937 if (is_ia32_NoReg_GP(irn) || is_ia32_NoReg_VFP(irn) || is_ia32_NoReg_XMM(irn)
1938 || is_ia32_ChangeCW(irn) || is_ia32_Immediate(irn))
1945 * Initializes the code generator interface.
1947 static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self)
1950 return &ia32_code_gen_if;
1954 * Returns the estimated execution time of an ia32 irn.
1956 static sched_timestep_t ia32_sched_exectime(void *env, const ir_node *irn)
1959 return is_ia32_irn(irn) ? ia32_get_op_estimated_cost(irn) : 1;
1962 list_sched_selector_t ia32_sched_selector;
1965 * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
1967 static const list_sched_selector_t *ia32_get_list_sched_selector(
1968 const void *self, list_sched_selector_t *selector)
1971 memcpy(&ia32_sched_selector, selector, sizeof(ia32_sched_selector));
1972 ia32_sched_selector.exectime = ia32_sched_exectime;
1973 ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
1974 return &ia32_sched_selector;
1977 static const ilp_sched_selector_t *ia32_get_ilp_sched_selector(const void *self)
1984 * Returns the necessary byte alignment for storing a register of given class.
1986 static int ia32_get_reg_class_alignment(const arch_register_class_t *cls)
1988 ir_mode *mode = arch_register_class_mode(cls);
1989 int bytes = get_mode_size_bytes(mode);
1991 if (mode_is_float(mode) && bytes > 8)
1996 static const be_execution_unit_t ***ia32_get_allowed_execution_units(
1999 static const be_execution_unit_t *_allowed_units_BRANCH[] = {
2000 &ia32_execution_units_BRANCH[IA32_EXECUNIT_TP_BRANCH_BRANCH1],
2001 &ia32_execution_units_BRANCH[IA32_EXECUNIT_TP_BRANCH_BRANCH2],
2004 static const be_execution_unit_t *_allowed_units_GP[] = {
2005 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EAX],
2006 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EBX],
2007 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_ECX],
2008 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EDX],
2009 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_ESI],
2010 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EDI],
2011 &ia32_execution_units_GP[IA32_EXECUNIT_TP_GP_GP_EBP],
2014 static const be_execution_unit_t *_allowed_units_DUMMY[] = {
2015 &be_machine_execution_units_DUMMY[0],
2018 static const be_execution_unit_t **_units_callret[] = {
2019 _allowed_units_BRANCH,
2022 static const be_execution_unit_t **_units_other[] = {
2026 static const be_execution_unit_t **_units_dummy[] = {
2027 _allowed_units_DUMMY,
2030 const be_execution_unit_t ***ret;
2032 if (is_ia32_irn(irn)) {
2033 ret = get_ia32_exec_units(irn);
2034 } else if (is_be_node(irn)) {
2035 if (be_is_Return(irn)) {
2036 ret = _units_callret;
2037 } else if (be_is_Barrier(irn)) {
2051 * Return the abstract ia32 machine.
2053 static const be_machine_t *ia32_get_machine(const void *self)
2055 const ia32_isa_t *isa = self;
2060 * Return irp irgs in the desired order.
2062 static ir_graph **ia32_get_irg_list(const void *self, ir_graph ***irg_list)
2069 static void ia32_mark_remat(ir_node *node)
2071 if (is_ia32_irn(node)) {
2072 set_ia32_is_remat(node);
2077 * Check if Mux(sel, t, f) would represent an Abs (or -Abs).
2079 static bool mux_is_abs(ir_node *sel, ir_node *mux_true, ir_node *mux_false)
2088 cmp = get_Proj_pred(sel);
2092 /* must be <, <=, >=, > */
2093 pnc = get_Proj_proj(sel);
2108 if (!is_negated_value(mux_true, mux_false))
2111 /* must be x cmp 0 */
2112 cmp_right = get_Cmp_right(cmp);
2113 if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
2116 cmp_left = get_Cmp_left(cmp);
2117 if (cmp_left != mux_true && cmp_left != mux_false)
2124 * Check if Mux(sel, mux_true, mux_false) would represent a Max or Min operation
2126 static bool mux_is_float_min_max(ir_node *sel, ir_node *mux_true,
2136 cmp = get_Proj_pred(sel);
2140 cmp_l = get_Cmp_left(cmp);
2141 cmp_r = get_Cmp_right(cmp);
2142 if (!mode_is_float(get_irn_mode(cmp_l)))
2145 /* check for min/max. They're defined as (C-Semantik):
2146 * min(a, b) = a < b ? a : b
2147 * or min(a, b) = a <= b ? a : b
2148 * max(a, b) = a > b ? a : b
2149 * or max(a, b) = a >= b ? a : b
2150 * (Note we only handle float min/max here)
2152 pnc = get_Proj_proj(sel);
2157 if (cmp_l == mux_true && cmp_r == mux_false)
2163 if (cmp_l == mux_true && cmp_r == mux_false)
2169 if (cmp_l == mux_false && cmp_r == mux_true)
2175 if (cmp_l == mux_false && cmp_r == mux_true)
2186 static bool mux_is_set(ir_node *sel, ir_node *mux_true, ir_node *mux_false)
2188 ir_mode *mode = get_irn_mode(mux_true);
2191 if (!mode_is_int(mode) && !mode_is_reference(mode)
2195 if (is_Const(mux_true) && is_Const(mux_false)) {
2196 /* we can create a set plus up two 3 instructions for any combination of constants */
2203 static bool mux_is_float_const_const(ir_node *sel, ir_node *mux_true,
2208 if (!mode_is_float(get_irn_mode(mux_true)))
2211 return is_Const(mux_true) && is_Const(mux_false);
2214 static bool mux_is_doz(ir_node *sel, ir_node *mux_true, ir_node *mux_false)
2225 cmp = get_Proj_pred(sel);
2229 mode = get_irn_mode(mux_true);
2230 if (mode_is_signed(mode) || mode_is_float(mode))
2233 pn = get_Proj_proj(sel);
2234 cmp_left = get_Cmp_left(cmp);
2235 cmp_right = get_Cmp_right(cmp);
2236 if ((pn & pn_Cmp_Gt) &&
2237 is_Const(mux_false) && is_Const_null(mux_false) && is_Sub(mux_true) &&
2238 get_Sub_left(mux_true) == cmp_left &&
2239 get_Sub_right(mux_true) == cmp_right) {
2240 /* Mux(a >=u b, a - b, 0) unsigned Doz */
2243 if ((pn & pn_Cmp_Lt) &&
2244 is_Const(mux_true) && is_Const_null(mux_true) && is_Sub(mux_false) &&
2245 get_Sub_left(mux_false) == cmp_left &&
2246 get_Sub_right(mux_false) == cmp_right) {
2247 /* Mux(a <=u b, 0, a - b) unsigned Doz */
2254 static int ia32_is_mux_allowed(ir_node *sel, ir_node *mux_false,
2259 /* we can handle Abs for all modes and compares */
2260 if (mux_is_abs(sel, mux_true, mux_false))
2262 /* we can handle Set for all modes and compares */
2263 if (mux_is_set(sel, mux_true, mux_false))
2265 /* SSE has own min/max operations */
2266 if (ia32_cg_config.use_sse2
2267 && mux_is_float_min_max(sel, mux_true, mux_false))
2269 /* we can handle Mux(?, Const[f], Const[f]) */
2270 if (mux_is_float_const_const(sel, mux_true, mux_false)) {
2271 #ifdef FIRM_GRGEN_BE
2272 /* well, some code selectors can't handle it */
2273 if (be_transformer != TRANSFORMER_PBQP
2274 || be_transformer != TRANSFORMER_RAND)
2281 /* no support for 64bit inputs to cmov */
2282 mode = get_irn_mode(mux_true);
2283 if (get_mode_size_bits(mode) > 32)
2285 /* we can't handle MuxF yet */
2286 if (mode_is_float(mode))
2289 if (mux_is_doz(sel, mux_true, mux_false))
2292 /* Check Cmp before the node */
2294 ir_node *cmp = get_Proj_pred(sel);
2296 ir_mode *cmp_mode = get_irn_mode(get_Cmp_left(cmp));
2298 /* we can't handle 64bit compares */
2299 if (get_mode_size_bits(cmp_mode) > 32)
2302 /* we can't handle float compares */
2303 if (mode_is_float(cmp_mode))
2308 /* did we disable cmov generation? */
2309 if (!ia32_cg_config.use_cmov)
2312 /* we can use a cmov */
2316 static asm_constraint_flags_t ia32_parse_asm_constraint(const char **c)
2320 /* we already added all our simple flags to the flags modifier list in
2321 * init, so this flag we don't know. */
2322 return ASM_CONSTRAINT_FLAG_INVALID;
2325 static int ia32_is_valid_clobber(const char *clobber)
2327 return ia32_get_clobber_register(clobber) != NULL;
2331 * Create the trampoline code.
2333 static ir_node *ia32_create_trampoline_fkt(ir_node *block, ir_node *mem, ir_node *trampoline, ir_node *env, ir_node *callee)
2335 ir_node *st, *p = trampoline;
2336 ir_mode *mode = get_irn_mode(p);
2339 st = new_r_Store(block, mem, p, new_Const_long(mode_Bu, 0xb9), 0);
2340 mem = new_r_Proj(st, mode_M, pn_Store_M);
2341 p = new_r_Add(block, p, new_Const_long(mode_Iu, 1), mode);
2342 st = new_r_Store(block, mem, p, env, 0);
2343 mem = new_r_Proj(st, mode_M, pn_Store_M);
2344 p = new_r_Add(block, p, new_Const_long(mode_Iu, 4), mode);
2346 st = new_r_Store(block, mem, p, new_Const_long(mode_Bu, 0xe9), 0);
2347 mem = new_r_Proj(st, mode_M, pn_Store_M);
2348 p = new_r_Add(block, p, new_Const_long(mode_Iu, 1), mode);
2349 st = new_r_Store(block, mem, p, callee, 0);
2350 mem = new_r_Proj(st, mode_M, pn_Store_M);
2351 p = new_r_Add(block, p, new_Const_long(mode_Iu, 4), mode);
2357 * Returns the libFirm configuration parameter for this backend.
2359 static const backend_params *ia32_get_libfirm_params(void)
2361 static const ir_settings_if_conv_t ifconv = {
2362 4, /* maxdepth, doesn't matter for Mux-conversion */
2363 ia32_is_mux_allowed /* allows or disallows Mux creation for given selector */
2365 static const ir_settings_arch_dep_t ad = {
2366 1, /* also use subs */
2367 4, /* maximum shifts */
2368 31, /* maximum shift amount */
2369 ia32_evaluate_insn, /* evaluate the instruction sequence */
2371 1, /* allow Mulhs */
2372 1, /* allow Mulus */
2373 32, /* Mulh allowed up to 32 bit */
2375 static backend_params p = {
2376 1, /* need dword lowering */
2377 1, /* support inline assembly */
2378 NULL, /* will be set later */
2379 ia32_create_intrinsic_fkt,
2380 &intrinsic_env, /* context for ia32_create_intrinsic_fkt */
2381 NULL, /* ifconv info will be set below */
2382 NULL, /* float arithmetic mode, will be set below */
2383 12, /* size of trampoline code */
2384 4, /* alignment of trampoline code */
2385 ia32_create_trampoline_fkt,
2386 4 /* alignment of stack parameter */
2389 ia32_setup_cg_config();
2391 /* doesn't really belong here, but this is the earliest place the backend
2393 init_asm_constraints();
2396 p.if_conv_info = &ifconv;
2397 if (! ia32_cg_config.use_sse2)
2398 p.mode_float_arithmetic = mode_E;
2402 static const lc_opt_enum_int_items_t gas_items[] = {
2403 { "elf", OBJECT_FILE_FORMAT_ELF },
2404 { "mingw", OBJECT_FILE_FORMAT_COFF },
2405 { "macho", OBJECT_FILE_FORMAT_MACH_O },
2409 static lc_opt_enum_int_var_t gas_var = {
2410 (int*) &be_gas_object_file_format, gas_items
2413 #ifdef FIRM_GRGEN_BE
2414 static const lc_opt_enum_int_items_t transformer_items[] = {
2415 { "default", TRANSFORMER_DEFAULT },
2416 { "pbqp", TRANSFORMER_PBQP },
2417 { "random", TRANSFORMER_RAND },
2421 static lc_opt_enum_int_var_t transformer_var = {
2422 (int*)&be_transformer, transformer_items
2426 static const lc_opt_table_entry_t ia32_options[] = {
2427 LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
2428 #ifdef FIRM_GRGEN_BE
2429 LC_OPT_ENT_ENUM_INT("transformer", "the transformer used for code selection", &transformer_var),
2431 LC_OPT_ENT_INT("stackalign", "set power of two stack alignment for calls",
2432 &ia32_isa_template.arch_env.stack_alignment),
2436 const arch_isa_if_t ia32_isa_if = {
2439 ia32_handle_intrinsics,
2440 ia32_get_n_reg_class,
2442 ia32_get_reg_class_for_mode,
2444 ia32_get_code_generator_if,
2445 ia32_get_list_sched_selector,
2446 ia32_get_ilp_sched_selector,
2447 ia32_get_reg_class_alignment,
2448 ia32_get_libfirm_params,
2449 ia32_get_allowed_execution_units,
2453 ia32_parse_asm_constraint,
2454 ia32_is_valid_clobber
2457 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_ia32);
2458 void be_init_arch_ia32(void)
2460 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2461 lc_opt_entry_t *ia32_grp = lc_opt_get_grp(be_grp, "ia32");
2463 lc_opt_add_table(ia32_grp, ia32_options);
2464 be_register_isa_if("ia32", &ia32_isa_if);
2466 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.cg");
2468 ia32_init_emitter();
2470 ia32_init_optimize();
2471 ia32_init_transform();
2473 ia32_init_architecture();