* $Id$
*/
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
#include "irnode.h"
#include "ircons.h"
#include "irgmod.h"
#include "irgwalk.h"
+#include "iredges.h"
#include "pdeq.h"
#include "../bearch.h"
* THIS FUNCTIONS MUST BE CALLED AFTER REGISTER ALLOCATION.
*/
static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) {
- ia32_transform_env_t tenv;
+ ir_graph *irg;
ir_node *in1, *in2, *noreg, *nomem, *res;
+ ir_node *noreg_fp, *block;
+ ir_mode *mode = get_irn_mode(irn);
+ dbg_info *dbg = get_irn_dbg_info(irn);
const arch_register_t *in1_reg, *in2_reg, *out_reg, **slots;
+ int i, arity;
/* Return if AM node or not a Sub or xSub */
- if (get_ia32_op_type(irn) != ia32_Normal || !(is_ia32_Sub(irn) || is_ia32_xSub(irn)))
+ if (!(is_ia32_Sub(irn) || is_ia32_xSub(irn)) || get_ia32_op_type(irn) != ia32_Normal)
return;
noreg = ia32_new_NoReg_gp(cg);
+ noreg_fp = ia32_new_NoReg_fp(cg);
nomem = new_rd_NoMem(cg->irg);
in1 = get_irn_n(irn, 2);
in2 = get_irn_n(irn, 3);
in2_reg = arch_get_irn_register(cg->arch_env, in2);
out_reg = get_ia32_out_reg(irn, 0);
- tenv.block = get_nodes_block(irn);
- tenv.dbg = get_irn_dbg_info(irn);
- tenv.irg = cg->irg;
- tenv.irn = irn;
- tenv.mode = get_ia32_res_mode(irn);
- tenv.cg = cg;
- DEBUG_ONLY(tenv.mod = cg->mod;)
+ irg = cg->irg;
+ block = get_nodes_block(irn);
/* in case of sub and OUT == SRC2 we can transform the sequence into neg src2 -- add */
- if (REGS_ARE_EQUAL(out_reg, in2_reg)) {
- /* generate the neg src2 */
- res = gen_Minus_ex(&tenv, in2);
- arch_set_irn_register(cg->arch_env, res, in2_reg);
-
- /* add to schedule */
- sched_add_before(irn, res);
-
- /* generate the add */
- if (mode_is_float(tenv.mode)) {
- res = new_rd_ia32_xAdd(tenv.dbg, tenv.irg, tenv.block, noreg, noreg, res, in1, nomem);
- set_ia32_am_support(res, ia32_am_Source);
- }
- else {
- res = new_rd_ia32_Add(tenv.dbg, tenv.irg, tenv.block, noreg, noreg, res, in1, nomem);
- set_ia32_am_support(res, ia32_am_Full);
- set_ia32_commutative(res);
- }
- set_ia32_res_mode(res, tenv.mode);
+ if (!REGS_ARE_EQUAL(out_reg, in2_reg))
+ return;
- SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(tenv.cg, irn));
- /* copy register */
- slots = get_ia32_slots(res);
- slots[0] = in2_reg;
+ /* generate the neg src2 */
+ if(mode_is_float(mode)) {
+ int size;
+ ident *name;
+
+ res = new_rd_ia32_xXor(dbg, irg, block, noreg, noreg, in2, noreg_fp, nomem);
+ size = get_mode_size_bits(mode);
+ name = ia32_gen_fp_known_const(size == 32 ? ia32_SSIGN : ia32_DSIGN);
+ set_ia32_am_sc(res, name);
+ set_ia32_op_type(res, ia32_AddrModeS);
+ set_ia32_ls_mode(res, mode);
+ } else {
+ res = new_rd_ia32_Neg(dbg, irg, block, noreg, noreg, in2, nomem);
+ }
+ arch_set_irn_register(cg->arch_env, res, in2_reg);
- /* add to schedule */
- sched_add_before(irn, res);
+ /* add to schedule */
+ sched_add_before(irn, res);
- /* remove the old sub */
- sched_remove(irn);
+ /* generate the add */
+ if (mode_is_float(mode)) {
+ res = new_rd_ia32_xAdd(dbg, irg, block, noreg, noreg, res, in1, nomem);
+ set_ia32_am_support(res, ia32_am_Source);
+ }
+ else {
+ res = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, res, in1, nomem);
+ set_ia32_am_support(res, ia32_am_Full);
+ set_ia32_commutative(res);
+ }
+
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn));
+ /* copy register */
+ slots = get_ia32_slots(res);
+ slots[0] = in2_reg;
- DBG_OPT_SUB2NEGADD(irn, res);
+ /* exchange the add and the sub */
+ edges_reroute(irn, res, irg);
- /* exchange the add and the sub */
- exchange(irn, res);
+ /* add to schedule */
+ sched_add_before(irn, res);
+
+ /* remove the old sub */
+ sched_remove(irn);
+ arity = get_irn_arity(irn);
+ for(i = 0; i < arity; ++i) {
+ set_irn_n(irn, i, new_Bad());
}
+
+ DBG_OPT_SUB2NEGADD(irn, res);
}
/**
static void ia32_transform_lea_to_add(ir_node *irn, ia32_code_gen_t *cg) {
ia32_am_flavour_t am_flav;
int imm = 0;
+ dbg_info *dbg = get_irn_dbg_info(irn);
+ ir_graph *irg;
ir_node *res = NULL;
ir_node *nomem, *noreg, *base, *index, *op1, *op2;
- char *offs;
- ia32_transform_env_t tenv;
+ ir_node *block;
+ int offs = 0;
const arch_register_t *out_reg, *base_reg, *index_reg;
/* must be a LEA */
am_flav = get_ia32_am_flavour(irn);
+ /* mustn't have a symconst */
if (get_ia32_am_sc(irn))
return;
base = get_irn_n(irn, 0);
index = get_irn_n(irn,1);
- offs = get_ia32_am_offs(irn);
-
- /* offset has a explicit sign -> we need to skip + */
- if (offs && offs[0] == '+')
- offs++;
+ if (am_flav & ia32_O) {
+ offs = get_ia32_am_offs_int(irn);
+ }
out_reg = arch_get_irn_register(cg->arch_env, irn);
base_reg = arch_get_irn_register(cg->arch_env, base);
index_reg = arch_get_irn_register(cg->arch_env, index);
- tenv.block = get_nodes_block(irn);
- tenv.dbg = get_irn_dbg_info(irn);
- tenv.irg = cg->irg;
- tenv.irn = irn;
- DEBUG_ONLY(tenv.mod = cg->mod;)
- tenv.mode = get_irn_mode(irn);
- tenv.cg = cg;
+ irg = cg->irg;
+ block = get_nodes_block(irn);
switch(get_ia32_am_flavour(irn)) {
case ia32_am_B:
break;
}
- res = new_rd_ia32_Add(tenv.dbg, tenv.irg, tenv.block, noreg, noreg, op1, op2, nomem);
+ res = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, op1, op2, nomem);
arch_set_irn_register(cg->arch_env, res, out_reg);
set_ia32_op_type(res, ia32_Normal);
set_ia32_commutative(res);
- set_ia32_res_mode(res, tenv.mode);
if (imm) {
- set_ia32_cnst(res, offs);
- set_ia32_immop_type(res, ia32_ImmConst);
+ tarval *tv = new_tarval_from_long(offs, mode_Iu);
+ set_ia32_Immop_tarval(res, tv);
}
SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn));
DBG_OPT_LEA2ADD(irn, res);
- res = new_rd_Proj(tenv.dbg, tenv.irg, tenv.block, res, tenv.mode, pn_ia32_Add_res);
-
- /* add result Proj to schedule */
- sched_add_before(irn, res);
-
/* remove the old LEA */
sched_remove(irn);
! is_ia32_Conv_I2I(irn) && \
! is_ia32_Conv_I2I8Bit(irn) && \
! is_ia32_CmpCMov(irn) && \
+ ! is_ia32_PsiCondCMov(irn) && \
! is_ia32_CmpSet(irn);
}
}
}
- /* 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. */
+ /* check xCmp: try to avoid unordered cmp */
+ if ((is_ia32_xCmp(irn) || is_ia32_xCmpCMov(irn) || is_ia32_xCmpSet(irn)) &&
+ op_tp == ia32_Normal &&
+ ! is_ia32_ImmConst(irn) && ! is_ia32_ImmSymConst(irn))
+ {
+ long pnc = get_ia32_pncode(irn);
+
+ if (pnc & pn_Cmp_Uo) {
+ ir_node *tmp;
+ int idx1 = 2, idx2 = 3;
+
+ if (is_ia32_xCmpCMov(irn)) {
+ idx1 = 0;
+ idx2 = 1;
+ }
+
+ tmp = get_irn_n(irn, idx1);
+ set_irn_n(irn, idx1, get_irn_n(irn, idx2));
+ set_irn_n(irn, idx2, tmp);
+
+ set_ia32_pncode(irn, get_negated_pnc(pnc, mode_E));
+ }
+ }
+
+ /*
+ 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 0
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)));
}
-
- /* check if there is a sub which need to be transformed */
- ia32_transform_sub_to_neg_add(irn, cg);
-
- /* transform a LEA into an Add if possible */
- ia32_transform_lea_to_add(irn, cg);
+#endif
}
-end:
-
- /* check for peephole optimization */
- ia32_peephole_optimization(irn, cg);
+end: ;
}
/**
*/
static void fix_am_source(ir_node *irn, void *env) {
ia32_code_gen_t *cg = env;
- ir_node *base, *index;
+ ir_node *base, *index, *noreg;
const arch_register_t *reg_base, *reg_index;
const ia32_register_req_t **reqs;
int n_res, i;
/* check only ia32 nodes with source address mode */
if (! is_ia32_irn(irn) || get_ia32_op_type(irn) != ia32_AddrModeS)
return;
+ /* no need to fix unary operations */
+ if (get_irn_arity(irn) == 4)
+ return;
base = get_irn_n(irn, 0);
index = get_irn_n(irn, 1);
reg_index = arch_get_irn_register(cg->arch_env, index);
reqs = get_ia32_out_req_all(irn);
+ noreg = ia32_new_NoReg_gp(cg);
+
n_res = get_ia32_n_res(irn);
for (i = 0; i < n_res; i++) {
set_ia32_am_support(load, ia32_am_Source);
set_ia32_am_scale(load, get_ia32_am_scale(irn));
set_ia32_am_sc(load, get_ia32_am_sc(irn));
- add_ia32_am_offs(load, get_ia32_am_offs(irn));
+ add_ia32_am_offs_int(load, get_ia32_am_offs_int(irn));
set_ia32_frame_ent(load, get_ia32_frame_ent(irn));
if (is_ia32_use_frame(irn))
set_irn_n(irn, 3, load);
/* this is a normal node now */
+ set_irn_n(irn, 0, noreg);
+ set_irn_n(irn, 1, noreg);
set_ia32_op_type(irn, ia32_Normal);
break;
fix_am_source(irn, env);
}
+ for (irn = sched_first(block); ! sched_is_end(irn); irn = next) {
+ ia32_code_gen_t *cg = env;
+
+ next = sched_next(irn);
+
+ /* check if there is a sub which need to be transformed */
+ ia32_transform_sub_to_neg_add(irn, cg);
+
+ /* transform a LEA into an Add if possible */
+ ia32_transform_lea_to_add(irn, cg);
+ }
+
/* second: insert copies and finish irg */
for (irn = sched_first(block); ! sched_is_end(irn); irn = next) {
next = sched_next(irn);