#include "ia32_optimize.h"
#include "bearch_ia32_t.h"
#include "gen_ia32_regalloc_if.h"
+#include "ia32_common_transform.h"
#include "ia32_transform.h"
#include "ia32_dbg_stat.h"
#include "ia32_util.h"
return new_node;
}
+/**
+ * Replace Cmp(x, 0) by a Test(x, x)
+ */
+static void peephole_ia32_Cmp(ir_node *const node)
+{
+ ir_node *right;
+ ia32_immediate_attr_t const *imm;
+ dbg_info *dbgi;
+ ir_graph *irg;
+ ir_node *block;
+ ir_node *noreg;
+ ir_node *nomem;
+ ir_node *op;
+ ia32_attr_t const *attr;
+ int ins_permuted;
+ int cmp_unsigned;
+ ir_node *test;
+ arch_register_t const *reg;
+ ir_edge_t const *edge;
+ ir_edge_t const *tmp;
+
+ if (get_ia32_op_type(node) != ia32_Normal)
+ return;
+
+ right = get_irn_n(node, n_ia32_Cmp_right);
+ if (!is_ia32_Immediate(right))
+ return;
+
+ imm = get_ia32_immediate_attr_const(right);
+ if (imm->symconst != NULL || imm->offset != 0)
+ return;
+
+ dbgi = get_irn_dbg_info(node);
+ irg = current_ir_graph;
+ block = get_nodes_block(node);
+ noreg = ia32_new_NoReg_gp(cg);
+ nomem = get_irg_no_mem(irg);
+ op = get_irn_n(node, n_ia32_Cmp_left);
+ attr = get_irn_generic_attr(node);
+ ins_permuted = attr->data.ins_permuted;
+ cmp_unsigned = attr->data.cmp_unsigned;
+
+ if (is_ia32_Cmp(node)) {
+ test = new_rd_ia32_Test(dbgi, irg, block, noreg, noreg, nomem,
+ op, op, ins_permuted, cmp_unsigned);
+ } else {
+ test = new_rd_ia32_Test8Bit(dbgi, irg, block, noreg, noreg, nomem,
+ op, op, ins_permuted, cmp_unsigned);
+ }
+ set_ia32_ls_mode(test, get_ia32_ls_mode(node));
+
+ reg = arch_get_irn_register(arch_env, node);
+ arch_set_irn_register(arch_env, test, reg);
+
+ foreach_out_edge_safe(node, edge, tmp) {
+ ir_node *const user = get_edge_src_irn(edge);
+
+ if (is_Proj(user))
+ exchange(user, test);
+ }
+
+ sched_add_before(node, test);
+ be_peephole_exchange(node, test);
+}
+
/**
* Peephole optimization for Test instructions.
* We can remove the Test, if a zero flags was produced which is still
}
}
- /* ensure, that the 3 byte return is generated
- * actually the emitter tests again if the block beginning has a label and
- * isn't just a fallthrough */
+ /* ensure, that the 3 byte return is generated */
be_Return_set_emit_pop(node, 1);
}
*/
static void peephole_IncSP_Store_to_push(ir_node *irn)
{
- int i, maxslot, inc_ofs;
- ir_node *node;
- ir_node *stores[MAXPUSH_OPTIMIZE];
- ir_node *block;
- ir_graph *irg;
- ir_node *curr_sp;
- ir_mode *spmode;
+ int i;
+ int maxslot;
+ int inc_ofs;
+ ir_node *node;
+ ir_node *stores[MAXPUSH_OPTIMIZE];
+ ir_node *block;
+ ir_graph *irg;
+ ir_node *curr_sp;
+ ir_mode *spmode;
+ ir_node *first_push = NULL;
+ ir_edge_t const *edge;
+ ir_edge_t const *next;
memset(stores, 0, sizeof(stores));
/* unfortunately we can't support the full AMs possible for push at the
* moment. TODO: fix this */
- if (get_ia32_am_scale(node) > 0 || !is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
+ if (!is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
break;
offset = get_ia32_am_offs_int(node);
/* we should NEVER access uninitialized stack BELOW the current SP */
assert(offset >= 0);
- offset = inc_ofs - 4 - offset;
-
/* storing at half-slots is bad */
if ((offset & 3) != 0)
break;
- if (offset < 0 || offset >= MAXPUSH_OPTIMIZE * 4)
+ if (inc_ofs - 4 < offset || offset >= MAXPUSH_OPTIMIZE * 4)
continue;
storeslot = offset >> 2;
maxslot = storeslot;
}
- curr_sp = be_get_IncSP_pred(irn);
+ curr_sp = irn;
+
+ for (i = -1; i < maxslot; ++i) {
+ if (stores[i + 1] == NULL)
+ break;
+ }
/* walk through the Stores and create Pushs for them */
block = get_nodes_block(irn);
spmode = get_irn_mode(irn);
irg = cg->irg;
- for (i = 0; i <= maxslot; ++i) {
+ for (; i >= 0; --i) {
const arch_register_t *spreg;
ir_node *push;
ir_node *val, *mem, *mem_proj;
ir_node *store = stores[i];
ir_node *noreg = ia32_new_NoReg_gp(cg);
- if (store == NULL)
- break;
-
val = get_irn_n(store, n_ia32_unary_op);
mem = get_irn_n(store, n_ia32_mem);
spreg = arch_get_irn_register(cg->arch_env, curr_sp);
push = new_rd_ia32_Push(get_irn_dbg_info(store), irg, block, noreg, noreg, mem, val, curr_sp);
- sched_add_before(irn, push);
+ if (first_push == NULL)
+ first_push = push;
+
+ sched_add_after(curr_sp, push);
/* create stackpointer Proj */
curr_sp = new_r_Proj(irg, block, push, spmode, pn_ia32_Push_stack);
mem_proj = new_r_Proj(irg, block, push, mode_M, pn_ia32_Push_M);
/* use the memproj now */
- exchange(store, mem_proj);
-
- /* we can remove the Store now */
- sched_remove(store);
+ be_peephole_exchange(store, mem_proj);
inc_ofs -= 4;
}
+ foreach_out_edge_safe(irn, edge, next) {
+ ir_node *const src = get_edge_src_irn(edge);
+ int const pos = get_edge_src_pos(edge);
+
+ if (src == first_push)
+ continue;
+
+ set_irn_n(src, pos, curr_sp);
+ }
+
be_set_IncSP_offset(irn, inc_ofs);
- be_set_IncSP_pred(irn, curr_sp);
}
/**
maxslot = -1;
pred_sp = be_get_IncSP_pred(irn);
for (node = sched_prev(irn); !sched_is_end(node); node = sched_prev(node)) {
- ir_node *mem;
int offset;
int loadslot;
const arch_register_t *sreg, *dreg;
* but we do not check this */
break;
}
- /* Load has to be attached to Spill-Mem */
- mem = skip_Proj(get_irn_n(node, n_ia32_mem));
- if (!is_Phi(mem) && !is_ia32_Store(mem) && !is_ia32_Push(mem))
- break;
/* should have NO index */
- if (get_ia32_am_scale(node) > 0 || !is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
+ if (!is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
break;
offset = get_ia32_am_offs_int(node);
set_Proj_pred(proj, pop);
}
-
/* we can remove the Load now */
sched_remove(load);
kill_node(load);
}
+
be_set_IncSP_offset(irn, -ofs);
be_set_IncSP_pred(irn, pred_sp);
-
}
* @return the new stack value
*/
static ir_node *create_push(dbg_info *dbgi, ir_graph *irg, ir_node *block,
- ir_node *stack, ir_node *schedpoint,
- const arch_register_t *reg)
+ ir_node *stack, ir_node *schedpoint)
{
const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
- ir_node *noreg, *nomem, *push, *val;
-
- val = new_rd_ia32_ProduceVal(NULL, irg, block);
- arch_set_irn_register(arch_env, val, reg);
- sched_add_before(schedpoint, val);
- noreg = ia32_new_NoReg_gp(cg);
- nomem = get_irg_no_mem(irg);
- push = new_rd_ia32_Push(dbgi, irg, block, noreg, noreg, nomem, val, stack);
+ ir_node *val = ia32_new_Unknown_gp(cg);
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
+ ir_node *nomem = get_irg_no_mem(irg);
+ ir_node *push = new_rd_ia32_Push(dbgi, irg, block, noreg, noreg, nomem, val, stack);
sched_add_before(schedpoint, push);
stack = new_r_Proj(irg, block, push, mode_Iu, pn_ia32_Push_stack);
dbgi = get_irn_dbg_info(node);
block = get_nodes_block(node);
stack = be_get_IncSP_pred(node);
- reg = &ia32_gp_regs[REG_EAX];
-
- stack = create_push(dbgi, irg, block, stack, node, reg);
+ stack = create_push(dbgi, irg, block, stack, node);
if (offset == +8) {
- stack = create_push(dbgi, irg, block, stack, node, reg);
+ stack = create_push(dbgi, irg, block, stack, node);
}
}
/* register peephole optimisations */
clear_irp_opcodes_generic_func();
- register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
- register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
- register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
- register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test);
+ register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
+ register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
+ register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
+ register_peephole_optimisation(op_ia32_Cmp, peephole_ia32_Cmp);
+ register_peephole_optimisation(op_ia32_Cmp8Bit, peephole_ia32_Cmp);
+ register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test);
register_peephole_optimisation(op_ia32_Test8Bit, peephole_ia32_Test);
- register_peephole_optimisation(op_be_Return, peephole_ia32_Return);
+ register_peephole_optimisation(op_be_Return, peephole_ia32_Return);
if (! ia32_cg_config.use_imul_mem_imm32)
register_peephole_optimisation(op_ia32_IMul, peephole_ia32_Imul_split);
if (ia32_cg_config.use_pxor)
sched_remove(node);
}
- be_kill_node(node);
+ kill_node(node);
}
static void optimize_conv_store(ir_node *node)
set_irn_n(node, n_ia32_Store_val, get_irn_n(pred, n_ia32_Conv_I2I_val));
if(get_irn_n_edges(pred_proj) == 0) {
- be_kill_node(pred_proj);
+ kill_node(pred_proj);
if(pred != pred_proj)
- be_kill_node(pred);
+ kill_node(pred);
}
}
exchange(node, result_conv);
if(get_irn_n_edges(pred_proj) == 0) {
- be_kill_node(pred_proj);
+ kill_node(pred_proj);
if(pred != pred_proj)
- be_kill_node(pred);
+ kill_node(pred);
}
optimize_conv_conv(result_conv);
}