#include "ircons.h"
#include "irgmod.h"
#include "irgopt.h"
-
-#include "bitset.h"
+#include "irbitset.h"
+#include "pdeq.h"
#include "debug.h"
#include "../beabi.h" /* the general register allocator interface */
#include "ia32_optimize.h"
#include "ia32_x87.h"
#include "ia32_dbg_stat.h"
+#include "ia32_finish.h"
+#include "ia32_util.h"
#define DEBUG_MODULE "firm.be.ia32.isa"
ia32_abi_env_t *env = self;
if (!env->flags.try_omit_fp) {
- int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode);
ir_node *bl = get_irg_start_block(env->irg);
ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
else {
const ia32_isa_t *isa = (ia32_isa_t *)env->isa;
ir_mode *mode_bp = env->isa->bp->reg_class->mode;
- int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode);
/* gcc always emits a leave at the end of a routine */
if (1 || ARCH_AMD(isa->opt_arch)) {
*/
static ir_type *ia32_abi_get_between_type(void *self)
{
+#define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
static ir_type *omit_fp_between_type = NULL;
static ir_type *between_type = NULL;
ia32_abi_env_t *env = self;
- if(!between_type) {
+ if ( !between_type) {
entity *old_bp_ent;
entity *ret_addr_ent;
entity *omit_fp_ret_addr_ent;
- ir_type *old_bp_type = new_type_primitive(new_id_from_str("bp"), mode_P);
- ir_type *ret_addr_type = new_type_primitive(new_id_from_str("return_addr"), mode_P);
+ ir_type *old_bp_type = new_type_primitive(IDENT("bp"), mode_P);
+ ir_type *ret_addr_type = new_type_primitive(IDENT("return_addr"), mode_P);
- between_type = new_type_class(new_id_from_str("ia32_between_type"));
- old_bp_ent = new_entity(between_type, new_id_from_str("old_bp"), old_bp_type);
- ret_addr_ent = new_entity(between_type, new_id_from_str("ret_addr"), ret_addr_type);
+ between_type = new_type_struct(IDENT("ia32_between_type"));
+ old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
+ ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type);
set_entity_offset_bytes(old_bp_ent, 0);
set_entity_offset_bytes(ret_addr_ent, get_type_size_bytes(old_bp_type));
set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
+ set_type_state(between_type, layout_fixed);
- omit_fp_between_type = new_type_class(new_id_from_str("ia32_between_type_omit_fp"));
- omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, new_id_from_str("ret_addr"), ret_addr_type);
+ omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
+ omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type);
set_entity_offset_bytes(omit_fp_ret_addr_ent, 0);
set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
+ set_type_state(omit_fp_between_type, layout_fixed);
}
return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
+#undef IDENT
}
/**
*/
static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn)
{
- int cost;
- switch (get_ia32_irn_opcode(irn)) {
- case iro_ia32_xDiv:
- case iro_ia32_DivMod:
- cost = 8;
- break;
-
- case iro_ia32_xLoad:
- case iro_ia32_l_Load:
- case iro_ia32_Load:
- case iro_ia32_Push:
- case iro_ia32_Pop:
- cost = 10;
- break;
-
- case iro_ia32_xStore:
- case iro_ia32_l_Store:
- case iro_ia32_Store:
- case iro_ia32_Store8Bit:
- cost = 50;
- break;
-
- case iro_ia32_MulS:
- case iro_ia32_Mul:
- case iro_ia32_Mulh:
- case iro_ia32_xMul:
- case iro_ia32_l_MulS:
- case iro_ia32_l_Mul:
- cost = 2;
- break;
-
- default:
- cost = 1;
- }
-
- return cost;
+ int cost;
+
+ if(is_Proj(irn))
+ return 0;
+
+ switch (get_ia32_irn_opcode(irn)) {
+ case iro_ia32_xDiv:
+ case iro_ia32_DivMod:
+ cost = 8;
+ break;
+
+ case iro_ia32_xLoad:
+ case iro_ia32_l_Load:
+ case iro_ia32_Load:
+ cost = 25;
+ break;
+
+ case iro_ia32_Push:
+ case iro_ia32_Pop:
+ cost = 5;
+ break;
+
+ case iro_ia32_xStore:
+ case iro_ia32_l_Store:
+ case iro_ia32_Store:
+ case iro_ia32_Store8Bit:
+ cost = 50;
+ break;
+
+ case iro_ia32_MulS:
+ case iro_ia32_Mul:
+ case iro_ia32_Mulh:
+ case iro_ia32_xMul:
+ case iro_ia32_l_MulS:
+ case iro_ia32_l_Mul:
+ cost = 2;
+ break;
+
+ default:
+ cost = 1;
+ }
+
+ return cost;
}
/**
}
else {
/* normal add: inverse == sub */
- ir_node *proj = get_irn_out_edge_first(irn)->src;
- assert(proj && is_Proj(proj));
+ ir_node *proj = ia32_get_res_proj(irn);
+ assert(proj);
inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, i ^ 1), nomem);
pnc = pn_ia32_Sub_res;
}
else {
/* normal sub */
- ir_node *proj = get_irn_out_edge_first(irn)->src;
- assert(proj && is_Proj(proj));
+ ir_node *proj = ia32_get_res_proj(irn);
+ assert(proj);
if (i == 2) {
inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, 3), nomem);
inverse->costs += 1;
}
break;
- case iro_ia32_Not:
- inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), nomem);
+ case iro_ia32_Not: {
+ ir_node *proj = ia32_get_res_proj(irn);
+ assert(proj);
+
+ inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, proj, nomem);
pnc = pn_ia32_Not_res;
inverse->costs += 1;
break;
- case iro_ia32_Minus:
- inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), nomem);
- pnc = pn_ia32_Minus_res;
+ }
+ case iro_ia32_Minus: {
+ ir_node *proj = ia32_get_res_proj(irn);
+ assert(proj);
+
+ inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, proj, nomem);
+ pnc = pn_ia32_Minus_res;
inverse->costs += 1;
break;
+ }
default:
/* inverse operation not supported */
return NULL;
return inverse;
}
+/**
+ * Check if irn can load it's operand at position i from memory (source addressmode).
+ * @param self Pointer to irn ops itself
+ * @param irn The irn to be checked
+ * @param i The operands position
+ * @return Non-Zero if operand can be loaded
+ */
+static int ia32_possible_memory_operand(const void *self, const ir_node *irn, unsigned int i) {
+ if (! is_ia32_irn(irn) || /* must be an ia32 irn */
+ get_irn_arity(irn) != 5 || /* must be a binary operation */
+ get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
+ ! (get_ia32_am_support(irn) & ia32_am_Source) || /* must be capable of source addressmode */
+ (i != 2 && i != 3) || /* a "real" operand position must be requested */
+ (i == 2 && ! is_ia32_commutative(irn)) || /* if first operand requested irn must be commutative */
+ is_ia32_use_frame(irn)) /* must not already use frame */
+ return 0;
+
+ return 1;
+}
+
+static void ia32_perform_memory_operand(const void *self, ir_node *irn, ir_node *reload, unsigned int i) {
+ assert(ia32_possible_memory_operand(self, irn, i) && "Cannot perform memory operand change");
+ assert(get_nodes_block(reload) == get_nodes_block(irn) && "Reload must be in same block as irn.");
+
+ if (get_irn_n_edges(reload) > 1)
+ return;
+
+ if (i == 2) {
+ ir_node *tmp = get_irn_n(irn, 3);
+ set_irn_n(irn, 3, get_irn_n(irn, 2));
+ set_irn_n(irn, 2, tmp);
+ }
+
+ set_ia32_am_support(irn, ia32_am_Source);
+ set_ia32_op_type(irn, ia32_AddrModeS);
+ set_ia32_am_flavour(irn, ia32_B);
+ set_ia32_ls_mode(irn, get_irn_mode(reload));
+ set_ia32_frame_ent(irn, be_get_frame_entity(reload));
+ set_ia32_use_frame(irn);
+
+ set_irn_n(irn, 0, be_get_Reload_frame(reload));
+ set_irn_n(irn, 4, be_get_Reload_mem(reload));
+
+ /*
+ Input at position one is index register, which is NoReg.
+ We would need cg object to get a real noreg, but we cannot
+ access it from here.
+ */
+ set_irn_n(irn, 3, get_irn_n(irn, 1));
+
+ DBG_OPT_AM_S(reload, irn);
+}
+
static const be_abi_callbacks_t ia32_abi_callbacks = {
ia32_abi_init,
free,
ia32_get_frame_entity,
ia32_set_stack_bias,
ia32_get_inverse,
- ia32_get_op_estimated_cost
+ ia32_get_op_estimated_cost,
+ ia32_possible_memory_operand,
+ ia32_perform_memory_operand,
};
ia32_irn_ops_t ia32_irn_ops = {
FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
/* 1st: transform constants and psi condition trees */
- irg_walk_blkwise_graph(cg->irg, ia32_place_consts_set_modes, ia32_transform_psi_cond_tree, cg);
+ ia32_pre_transform_phase(cg);
/* 2nd: transform all remaining nodes */
ia32_register_transformers();
dom = be_compute_dominance_frontiers(cg->irg);
irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_node, cg);
be_free_dominance_frontiers(dom);
- be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
+
+ if (cg->dump)
+ be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
/* 3rd: optimize address mode */
FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
ia32_optimize_addressmode(cg);
- be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
- DEBUG_ONLY(cg->mod = old_mod;)
-}
-static INLINE int need_constraint_copy(ir_node *irn) {
- return \
- ! is_ia32_Lea(irn) && \
- ! is_ia32_Conv_I2I(irn) && \
- ! is_ia32_Conv_I2I8Bit(irn) && \
- ! is_ia32_CmpCMov(irn) && \
- ! is_ia32_CmpSet(irn);
+ if (cg->dump)
+ be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
+
+ DEBUG_ONLY(cg->mod = old_mod;)
}
/**
- * Insert copies for all ia32 nodes where the should_be_same requirement
- * is not fulfilled.
- * Transform Sub into Neg -- Add if IN2 == OUT
+ * Dummy functions for hooks we don't need but which must be filled.
*/
-static void ia32_finish_node(ir_node *irn, void *env) {
- ia32_code_gen_t *cg = env;
- const ia32_register_req_t **reqs;
- const arch_register_t *out_reg, *in_reg, *in2_reg;
- int n_res, i;
- ir_node *copy, *in_node, *block, *in2_node;
- ia32_op_type_t op_tp;
+static void ia32_before_sched(void *self) {
+}
- if (is_ia32_irn(irn)) {
- /* AM Dest nodes don't produce any values */
- op_tp = get_ia32_op_type(irn);
- if (op_tp == ia32_AddrModeD)
- goto end;
-
- reqs = get_ia32_out_req_all(irn);
- n_res = get_ia32_n_res(irn);
- block = get_nodes_block(irn);
-
- /* check all OUT requirements, if there is a should_be_same */
- if ((op_tp == ia32_Normal || op_tp == ia32_AddrModeS) && need_constraint_copy(irn))
- {
- for (i = 0; i < n_res; i++) {
- if (arch_register_req_is(&(reqs[i]->req), should_be_same)) {
- /* get in and out register */
- out_reg = get_ia32_out_reg(irn, i);
- in_node = get_irn_n(irn, reqs[i]->same_pos);
- in_reg = arch_get_irn_register(cg->arch_env, in_node);
-
- /* don't copy ignore nodes */
- if (arch_irn_is(cg->arch_env, in_node, ignore) && is_Proj(in_node))
- continue;
-
- /* check if in and out register are equal */
- if (! REGS_ARE_EQUAL(out_reg, in_reg)) {
- /* in case of a commutative op: just exchange the in's */
- /* beware: the current op could be everything, so test for ia32 */
- /* commutativity first before getting the second in */
- if (is_ia32_commutative(irn)) {
- in2_node = get_irn_n(irn, reqs[i]->same_pos ^ 1);
- in2_reg = arch_get_irn_register(cg->arch_env, in2_node);
-
- if (REGS_ARE_EQUAL(out_reg, in2_reg)) {
- set_irn_n(irn, reqs[i]->same_pos, in2_node);
- set_irn_n(irn, reqs[i]->same_pos ^ 1, in_node);
- }
- else
- goto insert_copy;
- }
- else {
-insert_copy:
- DBG((cg->mod, LEVEL_1, "inserting copy for %+F in_pos %d\n", irn, reqs[i]->same_pos));
- /* create copy from in register */
- copy = be_new_Copy(arch_register_get_class(in_reg), cg->irg, block, in_node);
-
- DBG_OPT_2ADDRCPY(copy);
-
- /* destination is the out register */
- arch_set_irn_register(cg->arch_env, copy, out_reg);
-
- /* insert copy before the node into the schedule */
- sched_add_before(irn, copy);
-
- /* set copy as in */
- set_irn_n(irn, reqs[i]->same_pos, copy);
- }
- }
- }
- }
- }
+static void remove_unused_nodes(ir_node *irn, bitset_t *already_visited) {
+ int i;
+ ir_mode *mode;
+ ir_node *mem_proj;
- /* If we have a CondJmp/CmpSet/xCmpSet with immediate, we need to */
- /* check if it's the right operand, otherwise we have */
- /* to change it, as CMP doesn't support immediate as */
- /* left operands. */
- if ((is_ia32_CondJmp(irn) || is_ia32_CmpSet(irn) || is_ia32_xCmpSet(irn)) &&
- (is_ia32_ImmConst(irn) || is_ia32_ImmSymConst(irn)) &&
- op_tp == ia32_AddrModeS)
- {
- set_ia32_op_type(irn, ia32_AddrModeD);
- set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
- }
+ if (is_Block(irn))
+ return;
- /* check if there is a sub which need to be transformed */
- ia32_transform_sub_to_neg_add(irn, cg);
+ mode = get_irn_mode(irn);
- /* transform a LEA into an Add if possible */
- ia32_transform_lea_to_add(irn, cg);
- }
-end:
+ /* check if we already saw this node or the node has more than one user */
+ if (bitset_contains_irn(already_visited, irn) || get_irn_n_edges(irn) > 1)
+ return;
- /* check for peephole optimization */
- ia32_peephole_optimization(irn, cg);
-}
+ /* mark irn visited */
+ bitset_add_irn(already_visited, irn);
-static void ia32_finish_irg_walker(ir_node *block, void *env) {
- ir_node *irn, *next;
+ /* non-Tuple nodes with one user: ok, return */
+ if (get_irn_n_edges(irn) >= 1 && mode != mode_T)
+ return;
- for (irn = sched_first(block); !sched_is_end(irn); irn = next) {
- next = sched_next(irn);
- ia32_finish_node(irn, env);
+ /* tuple node has one user which is not the mem proj-> ok */
+ if (mode == mode_T && get_irn_n_edges(irn) == 1) {
+ mem_proj = ia32_get_proj_for_mode(irn, mode_M);
+ if (! mem_proj)
+ return;
}
-}
-/**
- * Add Copy nodes for not fulfilled should_be_equal constraints
- */
-static void ia32_finish_irg(ir_graph *irg, ia32_code_gen_t *cg) {
- irg_block_walk_graph(irg, NULL, ia32_finish_irg_walker, cg);
-}
+ for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
+ ir_node *pred = get_irn_n(irn, i);
+ /* do not follow memory edges or we will accidentally remove stores */
+ if (is_Proj(pred) && get_irn_mode(pred) == mode_M)
+ continue;
+ set_irn_n(irn, i, new_Bad());
-/**
- * Dummy functions for hooks we don't need but which must be filled.
- */
-static void ia32_before_sched(void *self) {
+ /*
+ The current node is about to be removed: if the predecessor
+ has only this node as user, it need to be removed as well.
+ */
+ if (get_irn_n_edges(pred) <= 1)
+ remove_unused_nodes(pred, already_visited);
+ }
+
+ if (sched_is_scheduled(irn))
+ sched_remove(irn);
+}
+
+static void remove_unused_loads_walker(ir_node *irn, void *env) {
+ bitset_t *already_visited = env;
+ if (is_ia32_Ld(irn) && ! bitset_contains_irn(already_visited, irn))
+ remove_unused_nodes(irn, env);
}
/**
* simulator and the emitter.
*/
static void ia32_before_ra(void *self) {
- ia32_code_gen_t *cg = self;
+ ia32_code_gen_t *cg = self;
+ bitset_t *already_visited = bitset_irg_malloc(cg->irg);
cg->blk_sched = sched_create_block_schedule(cg->irg);
+
+ /*
+ Handle special case:
+ There are sometimes unused loads, only pinned by memory.
+ We need to remove those Loads and all other nodes which won't be used
+ after removing the Load from schedule.
+ */
+ irg_walk_graph(cg->irg, remove_unused_loads_walker, NULL, already_visited);
+ bitset_free(already_visited);
}
transform_to_Load(&tenv);
}
else if (be_is_Spill(node)) {
+ ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
/* we always spill the whole register */
tenv.dbg = get_irn_dbg_info(node);
tenv.irn = node;
- tenv.mode = fix_spill_mode(cg, get_irn_mode(be_get_Spill_context(node)));
+ tenv.mode = fix_spill_mode(cg, get_irn_mode(spillval));
transform_to_Store(&tenv);
}
}
*/
static void ia32_after_ra(void *self) {
ia32_code_gen_t *cg = self;
+
irg_block_walk_graph(cg->irg, NULL, ia32_after_ra_walker, self);
/* if we do x87 code generation, rewrite all the virtual instructions and registers */
ir_graph *irg = cg->irg;
ia32_finish_irg(irg, cg);
- be_dump(irg, "-finished", dump_ir_block_graph_sched);
+ if (cg->dump)
+ be_dump(irg, "-finished", dump_ir_block_graph_sched);
ia32_gen_routine(cg->isa->out, irg, cg);
cur_reg_set = NULL;
cg->gp_to_fp = NULL;
cg->fp_kind = isa->fp_kind;
cg->used_fp = fp_none;
+ cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
/* copy optimizations from isa for easier access */
- cg->opt = isa->opt;
+ cg->opt = isa->opt;
+ cg->arch = isa->arch;
+ cg->opt_arch = isa->opt_arch;
/* enter it */
isa->cg = cg;
*
*****************************************************************/
+/**
+ * Set output modes for GCC
+ */
+static const tarval_mode_info mo_integer = {
+ TVO_DECIMAL,
+ NULL,
+ NULL,
+};
+
+/*
+* set the tarval output mode to C-semantics
+*/
+static void set_tarval_output_modes(void)
+{
+ set_tarval_mode_output_option(get_modeLs(), &mo_integer);
+ set_tarval_mode_output_option(get_modeLu(), &mo_integer);
+ set_tarval_mode_output_option(get_modeIs(), &mo_integer);
+ set_tarval_mode_output_option(get_modeIu(), &mo_integer);
+ set_tarval_mode_output_option(get_modeHs(), &mo_integer);
+ set_tarval_mode_output_option(get_modeHu(), &mo_integer);
+ set_tarval_mode_output_option(get_modeBs(), &mo_integer);
+ set_tarval_mode_output_option(get_modeBu(), &mo_integer);
+ set_tarval_mode_output_option(get_modeC(), &mo_integer);
+ set_tarval_mode_output_option(get_modeU(), &mo_integer);
+ set_tarval_mode_output_option(get_modeIu(), &mo_integer);
+}
+
+
/**
* The template that generates a new ISA object.
* Note that this template can be changed by command line
if (inited)
return NULL;
+ set_tarval_output_modes();
+
isa = xmalloc(sizeof(*isa));
memcpy(isa, &ia32_isa_template, sizeof(*isa));
/* patch register names of x87 registers */
if (USE_x87(isa)) {
- ia32_st_regs[0].name = "st";
- ia32_st_regs[1].name = "st(1)";
- ia32_st_regs[2].name = "st(2)";
- ia32_st_regs[3].name = "st(3)";
- ia32_st_regs[4].name = "st(4)";
- ia32_st_regs[5].name = "st(5)";
- ia32_st_regs[6].name = "st(6)";
- ia32_st_regs[7].name = "st(7)";
+ ia32_st_regs[0].name = "st";
+ ia32_st_regs[1].name = "st(1)";
+ ia32_st_regs[2].name = "st(2)";
+ ia32_st_regs[3].name = "st(3)";
+ ia32_st_regs[4].name = "st(4)";
+ ia32_st_regs[5].name = "st(5)";
+ ia32_st_regs[6].name = "st(6)";
+ ia32_st_regs[7].name = "st(7)";
}
#ifndef NDEBUG
* Return the register class for index i.
*/
static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
- const ia32_isa_t *isa = self;
assert(i >= 0 && i < 3 && "Invalid ia32 register class requested.");
if (i == 0)
return &ia32_reg_classes[CLASS_ia32_gp];
}
int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
- return is_ia32_irn(irn);
+ return is_ia32_irn(irn) ? 1 : -1;
}
/**
projects / libfirm / blobdiff