#include "irprog_t.h"
#include "ircons.h"
#include "firm_types.h"
+#include "iredges.h"
+#include "tv.h"
+#include "irgmod.h"
+
+#include "../benode_t.h"
#include "ia32_new_nodes.h"
#include "bearch_ia32_t.h"
+#undef is_NoMem
+#define is_NoMem(irn) (get_irn_op(irn) == op_NoMem)
+
/**
* creates a unique ident by adding a number to a tag
*
tenv.block = get_nodes_block(irn);
tenv.cg = cg;
tenv.irg = cg->irg;
- tenv.mod = firm_dbg_register("ir.be.ia32.optimize");
+ tenv.mod = cg->mod;
/* Loop over all predecessors and check for Sym/Const nodes */
for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
}
}
}
+
+
+/******************************************************************
+ * _ _ __ __ _
+ * /\ | | | | | \/ | | |
+ * / \ __| | __| |_ __ ___ ___ ___| \ / | ___ __| | ___
+ * / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \
+ * / ____ \ (_| | (_| | | | __/\__ \__ \ | | | (_) | (_| | __/
+ * /_/ \_\__,_|\__,_|_| \___||___/___/_| |_|\___/ \__,_|\___|
+ *
+ ******************************************************************/
+
+static int node_is_comm(const ir_node *irn) {
+ if (is_ia32_Add(irn) ||
+ is_ia32_fAdd(irn) ||
+ is_ia32_Mul(irn) ||
+ is_ia32_Mulh(irn) ||
+ is_ia32_fMul(irn) ||
+ is_ia32_And(irn) ||
+ is_ia32_fAnd(irn) ||
+ is_ia32_Or(irn) ||
+ is_ia32_fOr(irn) ||
+ is_ia32_Eor(irn) ||
+ is_ia32_fEor(irn) ||
+ is_ia32_Min(irn) ||
+ is_ia32_fMin(irn) ||
+ is_ia32_Max(irn) ||
+ is_ia32_fMax(irn))
+ {
+ return 1;
+ }
+
+ return 0;
+}
+
+static int ia32_get_irn_n_edges(const ir_node *irn) {
+ const ir_edge_t *edge;
+ int cnt = 0;
+
+ foreach_out_edge(irn, edge) {
+ cnt++;
+ }
+
+ return cnt;
+}
+
+/**
+ * Returns the first mode_M Proj connected to irn.
+ */
+static ir_node *get_mem_proj(const ir_node *irn) {
+ const ir_edge_t *edge;
+ ir_node *src;
+
+ assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
+
+ foreach_out_edge(irn, edge) {
+ src = get_edge_src_irn(edge);
+
+ assert(is_Proj(src) && "Proj expected");
+
+ if (get_irn_mode(src) == mode_M)
+ return src;
+ }
+
+ return NULL;
+}
+
+/**
+ * Returns the Proj with number 0 connected to irn.
+ */
+static ir_node *get_res_proj(const ir_node *irn) {
+ const ir_edge_t *edge;
+ ir_node *src;
+
+ assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
+
+ foreach_out_edge(irn, edge) {
+ src = get_edge_src_irn(edge);
+
+ assert(is_Proj(src) && "Proj expected");
+
+ if (get_Proj_proj(src) == 0)
+ return src;
+ }
+
+ return NULL;
+}
+
+/**
+ * Determines if pred is a Proj and if is_op_func returns true for it's predecessor.
+ *
+ * @param pred The node to be checked
+ * @param is_op_func The check-function
+ * @return 1 if conditions are fulfilled, 0 otherwise
+ */
+static int pred_is_specific_node(const ir_node *pred, int (*is_op_func)(const ir_node *n)) {
+ if (is_Proj(pred) && is_op_func(get_Proj_pred(pred))) {
+ return 1;
+ }
+
+ return 0;
+}
+
+/**
+ * Determines if pred is a Proj and if is_op_func returns true for it's predecessor
+ * and if the predecessor is in block bl.
+ *
+ * @param bl The block
+ * @param pred The node to be checked
+ * @param is_op_func The check-function
+ * @return 1 if conditions are fulfilled, 0 otherwise
+ */
+static int pred_is_specific_nodeblock(const ir_node *bl, const ir_node *pred,
+ int (*is_op_func)(const ir_node *n))
+{
+ if (is_Proj(pred)) {
+ pred = get_Proj_pred(pred);
+ if ((bl == get_nodes_block(pred)) && is_op_func(pred)) {
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
+/**
+ * Folds Add or Sub to LEA if possible
+ */
+static ir_node *fold_addr(ir_node *irn, firm_dbg_module_t *mod, ir_node *noreg) {
+ ir_graph *irg = get_irn_irg(irn);
+ ir_mode *mode = get_irn_mode(irn);
+ dbg_info *dbg = get_irn_dbg_info(irn);
+ ir_node *block = get_nodes_block(irn);
+ ir_node *res = irn;
+ char *offs = NULL;
+ char *offs_cnst = NULL;
+ char *offs_lea = NULL;
+ int scale = 0;
+ int isadd = 0;
+ int dolea = 0;
+ ir_node *left, *right, *temp;
+ ir_node *base, *index;
+ ia32_am_flavour_t am_flav;
+
+ if (is_ia32_Add(irn))
+ isadd = 1;
+
+ left = get_irn_n(irn, 2);
+ right = get_irn_n(irn, 3);
+
+ base = left;
+ index = noreg;
+ offs = NULL;
+ scale = 0;
+ am_flav = 0;
+
+ /* "normalize" arguments in case of add with two operands */
+ if (isadd && ! be_is_NoReg(right)) {
+ /* put LEA == ia32_am_O as right operand */
+ if (is_ia32_Lea(left) && get_ia32_am_flavour(left) == ia32_am_O) {
+ set_irn_n(irn, 2, right);
+ set_irn_n(irn, 3, left);
+ temp = left;
+ left = right;
+ right = temp;
+ }
+
+ /* put LEA != ia32_am_O as left operand */
+ if (is_ia32_Lea(right) && get_ia32_am_flavour(right) != ia32_am_O) {
+ set_irn_n(irn, 2, right);
+ set_irn_n(irn, 3, left);
+ temp = left;
+ left = right;
+ right = temp;
+ }
+
+ /* put SHL as left operand iff left is NOT a LEA */
+ if (! is_ia32_Lea(left) && pred_is_specific_node(right, is_ia32_Shl)) {
+ set_irn_n(irn, 2, right);
+ set_irn_n(irn, 3, left);
+ temp = left;
+ left = right;
+ right = temp;
+ }
+ }
+
+ /* check if operand is either const */
+ if (get_ia32_cnst(irn)) {
+ DBG((mod, LEVEL_1, "\tfound op with imm"));
+
+ offs_cnst = get_ia32_cnst(irn);
+ dolea = 1;
+ }
+
+ /* determine the operand which needs to be checked */
+ if (be_is_NoReg(right)) {
+ temp = left;
+ }
+ else {
+ temp = right;
+ }
+
+ /* check if right operand is AMConst (LEA with ia32_am_O) */
+ if (is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) {
+ DBG((mod, LEVEL_1, "\tgot op with LEA am_O"));
+
+ offs_lea = get_ia32_am_offs(temp);
+ dolea = 1;
+ }
+
+ if (isadd) {
+ /* default for add -> make right operand to index */
+ index = right;
+ dolea = 1;
+
+ DBG((mod, LEVEL_1, "\tgot LEA candidate with index %+F\n", index));
+
+ /* determine the operand which needs to be checked */
+ temp = left;
+ if (is_ia32_Lea(left)) {
+ temp = right;
+ }
+
+ /* check for SHL 1,2,3 */
+ if (pred_is_specific_node(temp, is_ia32_Shl)) {
+ temp = get_Proj_pred(temp);
+
+ if (get_ia32_Immop_tarval(temp)) {
+ scale = get_tarval_long(get_ia32_Immop_tarval(temp));
+
+ if (scale <= 3) {
+ scale = 1 << scale;
+ index = get_irn_n(temp, 2);
+
+ DBG((mod, LEVEL_1, "\tgot scaled index %+F\n", index));
+ }
+ }
+ }
+
+ /* fix base */
+ if (! be_is_NoReg(index)) {
+ /* if we have index, but left == right -> no base */
+ if (left == right) {
+ base = noreg;
+ }
+ else if (! is_ia32_Lea(left) && (index != right)) {
+ /* index != right -> we found a good Shl */
+ /* left != LEA -> this Shl was the left operand */
+ /* -> base is right operand */
+ base = right;
+ }
+ }
+ }
+
+ /* Try to assimilate a LEA as left operand */
+ if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) {
+ am_flav = get_ia32_am_flavour(left);
+
+ /* If we have an Add with a real right operand (not NoReg) and */
+ /* the LEA contains already an index calculation then we create */
+ /* a new LEA. */
+ if (isadd && !be_is_NoReg(index) && (am_flav & ia32_am_I)) {
+ DBG((mod, LEVEL_1, "\tleave old LEA, creating new one\n"));
+ }
+ else {
+ DBG((mod, LEVEL_1, "\tgot LEA as left operand ... assimilating\n"));
+ offs = get_ia32_am_offs(left);
+ base = get_irn_n(left, 0);
+ index = get_irn_n(left, 1);
+ scale = get_ia32_am_scale(left);
+ }
+ }
+
+ /* ok, we can create a new LEA */
+ if (dolea) {
+ res = new_rd_ia32_Lea(dbg, irg, block, base, index, mode_Is);
+
+ /* add the old offset of a previous LEA */
+ if (offs) {
+ add_ia32_am_offs(res, offs);
+ }
+
+ /* add the new offset */
+ if (isadd) {
+ if (offs_cnst) {
+ add_ia32_am_offs(res, offs_cnst);
+ }
+ if (offs_lea) {
+ add_ia32_am_offs(res, offs_lea);
+ }
+ }
+ else {
+ /* either lea_O-cnst, -cnst or -lea_O */
+ if (offs_cnst) {
+ if (offs_lea) {
+ add_ia32_am_offs(res, offs_lea);
+ }
+
+ sub_ia32_am_offs(res, offs_cnst);
+ }
+ else {
+ sub_ia32_am_offs(res, offs_lea);
+ }
+ }
+
+ /* set scale */
+ set_ia32_am_scale(res, scale);
+
+ am_flav = ia32_am_N;
+ /* determine new am flavour */
+ if (offs || offs_cnst || offs_lea) {
+ am_flav |= ia32_O;
+ }
+ if (! be_is_NoReg(base)) {
+ am_flav |= ia32_B;
+ }
+ if (! be_is_NoReg(index)) {
+ am_flav |= ia32_I;
+ }
+ if (scale > 0) {
+ am_flav |= ia32_S;
+ }
+ set_ia32_am_flavour(res, am_flav);
+
+ set_ia32_op_type(res, ia32_AddrModeS);
+
+ DBG((mod, LEVEL_1, "\tLEA [%+F + %+F * %d + %s]\n", base, index, scale, get_ia32_am_offs(res)));
+
+ /* get the result Proj of the Add/Sub */
+ irn = get_res_proj(irn);
+
+ assert(irn && "Couldn't find result proj");
+
+ /* exchange the old op with the new LEA */
+ exchange(irn, res);
+ }
+
+ return res;
+}
+
+/**
+ * Optimizes a pattern around irn to address mode if possible.
+ */
+void ia32_optimize_am(ir_node *irn, void *env) {
+ ia32_code_gen_t *cg = env;
+ ir_graph *irg = cg->irg;
+ firm_dbg_module_t *mod = cg->mod;
+ 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;
+
+ 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. */
+ if (! pred_is_specific_nodeblock(block, left, is_ia32_Load) &&
+ ! pred_is_specific_nodeblock(block, right, is_ia32_Load))
+ {
+ res = fold_addr(irn, mod, noreg_gp);
+ }
+ }
+
+ /* 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)) {
+ left = get_irn_n(irn, 0);
+
+ if (is_ia32_Lea(left)) {
+ /* 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_op_type(irn, get_ia32_op_type(left));
+
+ /* set base and index */
+ set_irn_n(irn, 0, get_irn_n(left, 0));
+ set_irn_n(irn, 1, get_irn_n(left, 1));
+ }
+ }
+ /* check if at least one operand is a Load */
+ else if (pred_is_specific_nodeblock(block, get_irn_n(irn, 2), is_ia32_Ld) ||
+ pred_is_specific_nodeblock(block, get_irn_n(irn, 3), is_ia32_Ld))
+ {
+ 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_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_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 ((addr_b == get_irn_n(load, 0)) && (addr_i == get_irn_n(load, 1)))
+ {
+ /* 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);
+
+ /* 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);
+ }
+ } /* 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;
+ }
+
+ /* normalize commutative ops */
+ if (node_is_comm(irn)) {
+ /* Assure that left operand is always a Load if there is one */
+ /* because non-commutative ops can only use Source AM if the */
+ /* left operand is a Load, so we only need to check the left */
+ /* operand afterwards. */
+ if (pred_is_specific_nodeblock(block, right, is_ia32_Ld)) {
+ set_irn_n(irn, 2, right);
+ set_irn_n(irn, 3, left);
+
+ temp = left;
+ left = right;
+ right = temp;
+ }
+ }
+
+ /* optimize op -> Load iff Load is only used by this op */
+ /* and left operand is a Load which only used by this irn */
+ if (check_am_src &&
+ pred_is_specific_nodeblock(block, left, is_ia32_Ld) &&
+ (ia32_get_irn_n_edges(left) == 1))
+ {
+ left = get_Proj_pred(left);
+
+ addr_b = get_irn_n(left, 0);
+ addr_i = get_irn_n(left, 1);
+
+ /* 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(left));
+ set_ia32_am_scale(irn, get_ia32_am_scale(left));
+ set_ia32_am_flavour(irn, get_ia32_am_flavour(left));
+ set_ia32_op_type(irn, ia32_AddrModeS);
+
+ /* connect to Load memory */
+ if (get_irn_arity(irn) == 5) {
+ /* binary AMop */
+ set_irn_n(irn, 4, get_irn_n(left, 2));
+ }
+ else {
+ /* unary AMop */
+ set_irn_n(irn, 3, get_irn_n(left, 2));
+ }
+
+ /* disconnect from Load */
+ set_irn_n(irn, 2, noreg_gp);
+
+ /* If Load has a memory Proj, connect it to the op */
+ mem_proj = get_mem_proj(left);
+ if (mem_proj) {
+ set_Proj_pred(mem_proj, irn);
+ set_Proj_proj(mem_proj, 1);
+ }
+ }
+ }
+ }
+}