-void ia32_optimize_am(ir_node *irn, void *env) {
- ia32_code_gen_t *cg = env;
- ir_node *res = irn;
- dbg_info *dbg;
- ir_mode *mode;
- ir_node *block, *noreg_gp, *noreg_fp;
- ir_node *left, *right, *temp;
- ir_node *store, *load, *mem_proj;
- ir_node *succ, *addr_b, *addr_i;
- int check_am_src = 0;
- DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;)
-
- if (! is_ia32_irn(irn))
- return;
-
- dbg = get_irn_dbg_info(irn);
- mode = get_irn_mode(irn);
- block = get_nodes_block(irn);
- noreg_gp = ia32_new_NoReg_gp(cg);
- noreg_fp = ia32_new_NoReg_fp(cg);
-
- DBG((mod, LEVEL_1, "checking for AM\n"));
-
- /* 1st part: check for address calculations and transform the into Lea */
-
- /* Following cases can occur: */
- /* - Sub (l, imm) -> LEA [base - offset] */
- /* - Sub (l, r == LEA with ia32_am_O) -> LEA [base - offset] */
- /* - Add (l, imm) -> LEA [base + offset] */
- /* - Add (l, r == LEA with ia32_am_O) -> LEA [base + offset] */
- /* - Add (l == LEA with ia32_am_O, r) -> LEA [base + offset] */
- /* - Add (l, r) -> LEA [base + index * scale] */
- /* with scale > 1 iff l/r == shl (1,2,3) */
-
- if (is_ia32_Sub(irn) || is_ia32_Add(irn)) {
- left = get_irn_n(irn, 2);
- right = get_irn_n(irn, 3);
-
- /* Do not try to create a LEA if one of the operands is a Load. */
- /* check is irn is a candidate for address calculation */
- if (is_candidate(block, irn, 1)) {
- DBG((mod, LEVEL_1, "\tfound address calculation candidate %+F ... ", irn));
- res = fold_addr(cg, irn, noreg_gp);
-
- if (res == irn)
- DB((mod, LEVEL_1, "transformed into %+F\n", res));
- else
- DB((mod, LEVEL_1, "not transformed\n"));
- }
- }
-
- /* 2nd part: fold following patterns: */
- /* - Load -> LEA into Load } TODO: If the LEA is used by more than one Load/Store */
- /* - Store -> LEA into Store } it might be better to keep the LEA */
- /* - op -> Load into AMop with am_Source */
- /* conditions: */
- /* - op is am_Source capable AND */
- /* - the Load is only used by this op AND */
- /* - the Load is in the same block */
- /* - Store -> op -> Load into AMop with am_Dest */
- /* conditions: */
- /* - op is am_Dest capable AND */
- /* - the Store uses the same address as the Load AND */
- /* - the Load is only used by this op AND */
- /* - the Load and Store are in the same block AND */
- /* - nobody else uses the result of the op */
-
- if ((res == irn) && (get_ia32_am_support(irn) != ia32_am_None) && !is_ia32_Lea(irn)) {
- /* 1st: check for Load/Store -> LEA */
- if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn)) {
- left = get_irn_n(irn, 0);
-
- if (is_ia32_Lea(left)) {
- DBG((mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
-
- /* get the AM attributes from the LEA */
- add_ia32_am_offs(irn, get_ia32_am_offs(left));
- set_ia32_am_scale(irn, get_ia32_am_scale(left));
- set_ia32_am_flavour(irn, get_ia32_am_flavour(left));
-
- set_ia32_am_sc(irn, get_ia32_am_sc(left));
- if (is_ia32_am_sc_sign(left))
- set_ia32_am_sc_sign(irn);
-
- set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD);
-
- /* set base and index */
- set_irn_n(irn, 0, get_irn_n(left, 0));
- set_irn_n(irn, 1, get_irn_n(left, 1));
-
- /* clear remat flag */
- set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
- }
- }
- /* check if the node is an address mode candidate */
- else if (is_candidate(block, irn, 0)) {
- DBG((mod, LEVEL_1, "\tfound address mode candidate %+F ... ", irn));
-
- left = get_irn_n(irn, 2);
- if (get_irn_arity(irn) == 4) {
- /* it's an "unary" operation */
- right = left;
- }
- else {
- right = get_irn_n(irn, 3);
- }
-
- /* normalize commutative ops */
- if (node_is_ia32_comm(irn)) {
- /* Assure that right operand is always a Load if there is one */
- /* because non-commutative ops can only use Dest AM if the right */
- /* operand is a load, so we only need to check right operand. */
- if (pred_is_specific_nodeblock(block, left, is_ia32_Ld))
- {
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
-
- temp = left;
- left = right;
- right = temp;
- }
- }
-
- /* check for Store -> op -> Load */
-
- /* Store -> op -> Load optimization is only possible if supported by op */
- /* and if right operand is a Load */
- if ((get_ia32_am_support(irn) & ia32_am_Dest) &&
- pred_is_specific_nodeblock(block, right, is_ia32_Ld))
- {
-
- /* An address mode capable op always has a result Proj. */
- /* If this Proj is used by more than one other node, we don't need to */
- /* check further, otherwise we check for Store and remember the address, */
- /* the Store points to. */
-
- succ = get_res_proj(irn);
- assert(succ && "Couldn't find result proj");
-
- addr_b = NULL;
- addr_i = NULL;
- store = NULL;
-
- /* now check for users and Store */
- if (ia32_get_irn_n_edges(succ) == 1) {
- succ = get_edge_src_irn(get_irn_out_edge_first(succ));
-
- if (is_ia32_fStore(succ) || is_ia32_Store(succ)) {
- store = succ;
- addr_b = get_irn_n(store, 0);
-
- /* Could be that the Store is connected to the address */
- /* calculating LEA while the Load is already transformed. */
- if (is_ia32_Lea(addr_b)) {
- succ = addr_b;
- addr_b = get_irn_n(succ, 0);
- addr_i = get_irn_n(succ, 1);
- }
- else {
- addr_i = noreg_gp;
- }
- }
- }
-
- if (store) {
- /* we found a Store as single user: Now check for Load */
-
- /* Extra check for commutative ops with two Loads */
- /* -> put the interesting Load right */
- if (node_is_ia32_comm(irn) &&
- pred_is_specific_nodeblock(block, left, is_ia32_Ld))
- {
- if ((addr_b == get_irn_n(get_Proj_pred(left), 0)) &&
- (addr_i == get_irn_n(get_Proj_pred(left), 1)))
- {
- /* We exchange left and right, so it's easier to kill */
- /* the correct Load later and to handle unary operations. */
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
-
- temp = left;
- left = right;
- right = temp;
- }
- }
-
- /* skip the Proj for easier access */
- load = get_Proj_pred(right);
-
- /* Compare Load and Store address */
- if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
- /* Right Load is from same address, so we can */
- /* disconnect the Load and Store here */
-
- /* set new base, index and attributes */
- set_irn_n(irn, 0, addr_b);
- set_irn_n(irn, 1, addr_i);
- add_ia32_am_offs(irn, get_ia32_am_offs(load));
- set_ia32_am_scale(irn, get_ia32_am_scale(load));
- set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
- set_ia32_op_type(irn, ia32_AddrModeD);
- set_ia32_frame_ent(irn, get_ia32_frame_ent(load));
- set_ia32_ls_mode(irn, get_ia32_ls_mode(load));
-
- set_ia32_am_sc(irn, get_ia32_am_sc(load));
- if (is_ia32_am_sc_sign(load))
- set_ia32_am_sc_sign(irn);
-
- if (is_ia32_use_frame(load))
- set_ia32_use_frame(irn);
-
- /* connect to Load memory and disconnect Load */
- if (get_irn_arity(irn) == 5) {
- /* binary AMop */
- set_irn_n(irn, 4, get_irn_n(load, 2));
- set_irn_n(irn, 3, noreg_gp);
- }
- else {
- /* unary AMop */
- set_irn_n(irn, 3, get_irn_n(load, 2));
- set_irn_n(irn, 2, noreg_gp);
- }
-
- /* connect the memory Proj of the Store to the op */
- mem_proj = get_mem_proj(store);
- set_Proj_pred(mem_proj, irn);
- set_Proj_proj(mem_proj, 1);
-
- /* clear remat flag */
- set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
-
- DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store));
- }
- } /* if (store) */
- else if (get_ia32_am_support(irn) & ia32_am_Source) {
- /* There was no store, check if we still can optimize for source address mode */
- check_am_src = 1;
- }
- } /* if (support AM Dest) */
- else if (get_ia32_am_support(irn) & ia32_am_Source) {
- /* op doesn't support am AM Dest -> check for AM Source */
- check_am_src = 1;
- }