static x87_state *empty = (x87_state *)&_empty;
/** The type of an instruction simulator function. */
-typedef int (*sim_func)(x87_state *state, ir_node *n, const arch_env_t *env);
+typedef int (*sim_func)(x87_state *state, ir_node *n);
/**
* A block state: Every block has a x87 state at the beginning and at the end.
* The x87 simulator.
*/
struct _x87_simulator {
- struct obstack obst; /**< an obstack for fast allocating */
- pmap *blk_states; /**< map blocks to states */
- const arch_env_t *env; /**< architecture environment */
+ struct obstack obst; /**< An obstack for fast allocating. */
+ pmap *blk_states; /**< Map blocks to states. */
+ const arch_env_t *env; /**< The architecture environment. */
unsigned char *live; /**< Liveness information. */
- unsigned n_idx; /**< cached get_irg_last_idx() result */
+ unsigned n_idx; /**< The cached get_irg_last_idx() result. */
};
/**
x87_set_st(state, reg_idx, node, 0);
} /* x87_set_tos */
+#if 0
/**
* Flush the x87 stack.
*
state->depth = 0;
state->tos = 0;
} /* x87_flush */
+#endif
/**
* Swap st(0) with st(pos).
return res;
} /* x87_alloc_state */
+#if 0
/**
* Create a new empty x87 state.
*
x87_flush(res);
return res;
} /* x87_alloc_empty_state */
+#endif
/**
* Clone a x87 state.
return NULL;
} /* get_irn_Proj_for_mode */
+/**
+ * Wrap the arch_* function here so we can check for errors.
+ */
+static const arch_register_t *x87_get_irn_register(x87_simulator *sim, const ir_node *irn) {
+ const arch_register_t *res;
+
+ res = arch_get_irn_register(sim->env, irn);
+ assert(res->reg_class->regs == ia32_vfp_regs);
+ return res;
+}
+
/* -------------- x87 perm --------------- */
/**
*
* @param state the x87 state
* @param pos parameter for fxch
- * @param dst_block the block of the user
+ * @param block the block were fxch is inserted
*
* Creates a new fxch node and reroute the user of the old node
* to the fxch.
*
* @return the fxch node
*/
-static ir_node *x87_fxch_shuffle(x87_state *state, int pos, ir_node *block, ir_node *dst_block)
+static ir_node *x87_fxch_shuffle(x87_state *state, int pos, ir_node *block)
{
- const ir_edge_t *edge;
- ir_node *n = x87_get_st_node(state, pos);
- ir_node *user = NULL;
- ir_node *fxch;
- int node_idx;
- ia32_attr_t *attr;
-
- if (block == get_nodes_block(n)) {
- /* this is a node from out block: change it's user */
- foreach_out_edge(n, edge) {
- ir_node *succ = get_edge_src_irn(edge);
-
- if (is_Phi(succ) && get_nodes_block(succ) == dst_block) {
- user = succ;
- node_idx = get_edge_src_pos(edge);
- break;
- }
- }
- assert(user);
- }
+ ir_node *fxch;
+ ia32_attr_t *attr;
- fxch = new_rd_ia32_fxch(NULL, get_irn_irg(block), block, n, get_irn_mode(n));
+ fxch = new_rd_ia32_fxch(NULL, get_irn_irg(block), block, mode_E);
attr = get_ia32_attr(fxch);
attr->x87[0] = &ia32_st_regs[pos];
attr->x87[2] = &ia32_st_regs[0];
- if (user) {
- DB((dbg, LEVEL_2, "%+F replaced input %d of %+F\n", fxch, node_idx, user));
- set_irn_n(user, node_idx, fxch);
- }
- else {
- /*
- * This is a node from a dominator block. Changing it's user might be wrong,
- * so just keep it alive.
- * The "right" solution would require a new Phi, but we don't care here.
- */
- keep_alive(fxch);
- }
+ keep_alive(fxch);
x87_fxch(state, pos);
return fxch;
* @return state
*/
static x87_state *x87_shuffle(x87_simulator *sim, ir_node *block, x87_state *state, ir_node *dst_block, const x87_state *dst_state) {
- int i, n_cycles, k, ri;
+ int i, n_cycles, k, ri;
unsigned cycles[4], all_mask;
- char cycle_idx[4][8];
- ir_node *fxch;
- ir_node *before, *after;
+ char cycle_idx[4][8];
+ ir_node *fxch, *before, *after;
assert(state->depth == dst_state->depth);
for (ri = 0; ri < n_cycles; ++ri) {
if ((cycles[ri] & 1) == 0) {
/* this cycle does not include the tos */
- fxch = x87_fxch_shuffle(state, cycle_idx[ri][0], block, dst_block);
+ fxch = x87_fxch_shuffle(state, cycle_idx[ri][0], block);
if (after)
sched_add_after(after, fxch);
else
after = fxch;
}
for (k = 1; cycle_idx[ri][k] != -1; ++k) {
- fxch = x87_fxch_shuffle(state, cycle_idx[ri][k], block, dst_block);
+ fxch = x87_fxch_shuffle(state, cycle_idx[ri][k], block);
if (after)
sched_add_after(after, fxch);
else
}
if ((cycles[ri] & 1) == 0) {
/* this cycle does not include the tos */
- fxch = x87_fxch_shuffle(state, cycle_idx[ri][0], block, dst_block);
+ fxch = x87_fxch_shuffle(state, cycle_idx[ri][0], block);
sched_add_after(after, fxch);
}
}
* Create a fxch node before another node.
*
* @param state the x87 state
- * @param n the node before the fxch
+ * @param n the node after the fxch
* @param pos exchange st(pos) with st(0)
* @param op_idx if >= 0, replace input op_idx of n with the fxch result
*
* @return the fxch
*/
static ir_node *x87_create_fxch(x87_state *state, ir_node *n, int pos, int op_idx) {
- ir_node *fxch, *pred;
+ ir_node *fxch;
ia32_attr_t *attr;
x87_fxch(state, pos);
- if (op_idx >= 0)
- pred = get_irn_n(n, op_idx);
- else
- pred = x87_get_st_node(state, pos);
-
- fxch = new_rd_ia32_fxch(NULL, get_irn_irg(n), get_nodes_block(n), pred, get_irn_mode(pred));
+ fxch = new_rd_ia32_fxch(NULL, get_irn_irg(n), get_nodes_block(n), mode_E);
attr = get_ia32_attr(fxch);
attr->x87[0] = &ia32_st_regs[pos];
attr->x87[2] = &ia32_st_regs[0];
- if (op_idx >= 0)
- set_irn_n(n, op_idx, fxch);
+ keep_alive(fxch);
sched_add_before(n, fxch);
DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(fxch), attr->x87[0]->name, attr->x87[2]->name));
* Create a fpush before node n.
*
* @param state the x87 state
- * @param n the node before the fpush
+ * @param n the node after the fpush
* @param pos push st(pos) on stack
- * @param op_idx if >= 0, replace input op_idx of n with the fpush result
+ * @param op_idx replace input op_idx of n with the fpush result
*/
-static void x87_create_fpush(const arch_env_t *env, x87_state *state, ir_node *n, int pos, int op_idx) {
- ir_node *fpush, *pred = get_irn_n(n, op_idx);
- ia32_attr_t *attr;
- const arch_register_t *out = arch_get_irn_register(env, pred);
+static void x87_create_fpush(x87_state *state, ir_node *n, int pos, int op_idx) {
+ ir_node *fpush, *pred = get_irn_n(n, op_idx);
+ ia32_attr_t *attr;
+ const arch_register_t *out = x87_get_irn_register(state->sim, pred);
x87_push_dbl(state, arch_register_get_index(out), pred);
- fpush = new_rd_ia32_fpush(NULL, get_irn_irg(n), get_nodes_block(n), pred, get_irn_mode(pred));
+ fpush = new_rd_ia32_fpush(NULL, get_irn_irg(n), get_nodes_block(n), mode_E);
attr = get_ia32_attr(fpush);
attr->x87[0] = &ia32_st_regs[pos];
attr->x87[2] = &ia32_st_regs[0];
- if (op_idx >= 0)
- set_irn_n(n, op_idx, fpush);
+ keep_alive(fpush);
sched_add_before(n, fpush);
+
DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(fpush), attr->x87[0]->name, attr->x87[2]->name));
} /* x87_create_fpush */
* Create a fpop before node n.
*
* @param state the x87 state
- * @param n the node before the fpop
+ * @param n the node after the fpop
* @param num pop 1 or 2 values
* @param pred node to use as predecessor of the fpop
*
* @return the fpop node
*/
-static ir_node *x87_create_fpop(const arch_env_t *env, x87_state *state, ir_node *n, int num, ir_node *pred) {
- ir_node *fpop;
+static ir_node *x87_create_fpop(x87_state *state, ir_node *n, int num, ir_node *pred) {
+ ir_node *fpop = pred;
ia32_attr_t *attr;
while (num > 0) {
+ keep_alive(pred);
+
x87_pop(state);
- fpop = new_rd_ia32_fpop(NULL, get_irn_irg(n), get_nodes_block(n), pred, mode_E);
+ fpop = new_rd_ia32_fpop(NULL, get_irn_irg(n), get_nodes_block(n), mode_E);
attr = get_ia32_attr(fpop);
attr->x87[0] = &ia32_st_regs[0];
attr->x87[1] = &ia32_st_regs[0];
* Updates a live set over a single step from a given node to its predecessor.
* Everything defined at the node is removed from the set, the uses of the node get inserted.
*
- * @param arch_env The architecture environment.
+ * @param sim The simulator handle.
* @param irn The node at which liveness should be computed.
* @param live The bitset of registers live before @p irn. This set gets modified by updating it to
* the registers live after irn.
*
* @return The live bitset.
*/
-static unsigned vfp_liveness_transfer(const arch_env_t *arch_env, ir_node *irn, unsigned live)
+static unsigned vfp_liveness_transfer(x87_simulator *sim, ir_node *irn, unsigned live)
{
int i, n;
const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
+ const arch_env_t *arch_env = sim->env;
if (arch_irn_consider_in_reg_alloc(arch_env, cls, irn)) {
- const arch_register_t *reg = arch_get_irn_register(arch_env, irn);
- live &= ~(1 << reg->index);
+ const arch_register_t *reg = x87_get_irn_register(sim, irn);
+ live &= ~(1 << arch_register_get_index(reg));
}
for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
ir_node *op = get_irn_n(irn, i);
if (mode_is_float(get_irn_mode(op)) && arch_irn_consider_in_reg_alloc(arch_env, cls, op)) {
- const arch_register_t *reg = arch_get_irn_register(arch_env, op);
- live |= 1 << reg->index;
+ const arch_register_t *reg = x87_get_irn_register(sim, op);
+ live |= 1 << arch_register_get_index(reg);
}
}
return live;
/**
* Put all live virtual registers at the end of a block into a bitset.
*
- * @param env the architecture environment
+ * @param sim the simulator handle
* @param lv the liveness information
* @param bl the block
*
* @return The live bitset at the end of this block
*/
-static unsigned vfp_liveness_end_of_block(const arch_env_t *env, be_lv_t *lv, const ir_node *bl)
+static unsigned vfp_liveness_end_of_block(x87_simulator *sim, be_lv_t *lv, const ir_node *bl)
{
int i;
unsigned live = 0;
const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
+ const arch_env_t *arch_env = sim->env;
be_lv_foreach(lv, bl, be_lv_state_end, i) {
ir_node *irn = be_lv_get_irn(lv, bl, i);
- if (arch_irn_consider_in_reg_alloc(env, cls, irn)) {
- const arch_register_t *reg = arch_get_irn_register(env, irn);
- live |= 1 << reg->index;
+ if (arch_irn_consider_in_reg_alloc(arch_env, cls, irn)) {
+ const arch_register_t *reg = x87_get_irn_register(sim, irn);
+ live |= 1 << arch_register_get_index(reg);
}
}
return live;
} /* vfp_liveness_end_of_block */
+/** get the register mask from an arch_register */
+#define REGMASK(reg) (1 << (arch_register_get_index(reg)))
+
/**
- * Return a bitset of registers which are live at a node.
+ * Return a bitset of argument registers which are live at the end of a node.
*
* @param sim the simulator handle
* @param pos the node
+ * @param kill kill mask for the output registers
*
* @return The live bitset.
*/
-static unsigned vfp_liveness_nodes_live_at(x87_simulator *sim, const ir_node *pos)
+static unsigned vfp_live_args_after(x87_simulator *sim, const ir_node *pos, unsigned kill)
{
unsigned idx = get_irn_idx(pos);
assert(idx < sim->n_idx);
- return sim->live[idx];
-} /* vfp_liveness_nodes_live_at */
+ return sim->live[idx] & ~kill;
+} /* vfp_live_args_after */
/**
* Calculate the liveness for a whole block and cache it.
* @param blk the block
*/
static void update_liveness(x87_simulator *sim, be_lv_t *lv, ir_node *blk) {
- unsigned live = vfp_liveness_end_of_block(sim->env, lv, blk);
+ unsigned live = vfp_liveness_end_of_block(sim, lv, blk);
unsigned idx;
ir_node *irn;
/* now iterate through the block backward and cache the results */
sched_foreach_reverse(blk, irn) {
+ /* stop at the first Phi: this produces the live-in */
+ if (is_Phi(irn))
+ break;
+
idx = get_irn_idx(irn);
sim->live[idx] = live;
- live = vfp_liveness_transfer(sim->env, irn, live);
+ live = vfp_liveness_transfer(sim, irn, live);
}
idx = get_irn_idx(blk);
sim->live[idx] = live;
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
* @param tmpl the template containing the 4 possible x87 opcodes
*/
-static int sim_binop(x87_state *state, ir_node *n, const arch_env_t *env, const exchange_tmpl *tmpl) {
+static int sim_binop(x87_state *state, ir_node *n, const exchange_tmpl *tmpl) {
int op2_idx, op1_idx = -1;
int out_idx, do_pop =0;
ia32_attr_t *attr;
ir_op *dst;
- const arch_register_t *op1 = arch_get_irn_register(env, get_irn_n(n, BINOP_IDX_1));
- const arch_register_t *op2 = arch_get_irn_register(env, get_irn_n(n, BINOP_IDX_2));
- const arch_register_t *out = arch_get_irn_register(env, n);
- unsigned live = vfp_liveness_nodes_live_at(state->sim, n);
+ x87_simulator *sim = state->sim;
+ const arch_register_t *op1 = x87_get_irn_register(sim, get_irn_n(n, BINOP_IDX_1));
+ const arch_register_t *op2 = x87_get_irn_register(sim, get_irn_n(n, BINOP_IDX_2));
+ const arch_register_t *out = x87_get_irn_register(sim, n);
+ unsigned live = vfp_live_args_after(sim, n, REGMASK(out));
- DB((dbg, LEVEL_1, ">>> %s %s, %s -> %s\n", get_irn_opname(n),
+ DB((dbg, LEVEL_1, ">>> %+F %s, %s -> %s\n", n,
arch_register_get_name(op1), arch_register_get_name(op2),
arch_register_get_name(out)));
DEBUG_ONLY(vfp_dump_live(live));
op1_idx = x87_on_stack(state, arch_register_get_index(op1));
op2_idx = x87_on_stack(state, arch_register_get_index(op2));
- if (op2->index != REG_VFP_NOREG) {
+ if (arch_register_get_index(op2) != REG_VFP_NOREG) {
/* second operand is a vfp register */
- if (is_vfp_live(op2->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op2), live)) {
/* Second operand is live. */
- if (is_vfp_live(op1->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
/* Both operands are live: push the first one.
This works even for op1 == op2. */
- x87_create_fpush(env, state, n, op2_idx, BINOP_IDX_2);
+ x87_create_fpush(state, n, op2_idx, BINOP_IDX_2);
out_idx = op2_idx = 0;
++op1_idx;
dst = tmpl->normal_op;
}
else {
/* Second operand is dead. */
- if (is_vfp_live(op1->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
/* First operand is live: bring second to tos. */
if (op2_idx != 0) {
x87_create_fxch(state, n, op2_idx, BINOP_IDX_2);
}
else {
/* second operand is an address mode */
- if (is_vfp_live(op1->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
/* first operand is live: push it here */
- x87_create_fpush(env, state, n, op1_idx, BINOP_IDX_1);
+ x87_create_fpush(state, n, op1_idx, BINOP_IDX_1);
}
else {
/* first operand is dead: bring it to tos */
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
* @param op the x87 opcode that will replace n's opcode
*/
-static int sim_unop(x87_state *state, ir_node *n, const arch_env_t *env, ir_op *op) {
+static int sim_unop(x87_state *state, ir_node *n, ir_op *op) {
int op1_idx, out_idx;
- const arch_register_t *op1 = arch_get_irn_register(env, get_irn_n(n, UNOP_IDX));
- const arch_register_t *out = arch_get_irn_register(env, n);
+ x87_simulator *sim = state->sim;
+ const arch_register_t *op1 = x87_get_irn_register(sim, get_irn_n(n, UNOP_IDX));
+ const arch_register_t *out = x87_get_irn_register(sim, n);
ia32_attr_t *attr;
- unsigned live = vfp_liveness_nodes_live_at(state->sim, n);
+ unsigned live = vfp_live_args_after(sim, n, REGMASK(out));
- DB((dbg, LEVEL_1, ">>> %s -> %s\n", get_irn_opname(n), out->name));
+ DB((dbg, LEVEL_1, ">>> %+F -> %s\n", n, out->name));
DEBUG_ONLY(vfp_dump_live(live));
op1_idx = x87_on_stack(state, arch_register_get_index(op1));
- if (is_vfp_live(op1->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
/* push the operand here */
- x87_create_fpush(env, state, n, op1_idx, UNOP_IDX);
+ x87_create_fpush(state, n, op1_idx, UNOP_IDX);
}
else {
/* operand is dead, bring it to tos */
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
* @param op the x87 opcode that will replace n's opcode
*/
-static int sim_load(x87_state *state, ir_node *n, const arch_env_t *env, ir_op *op) {
- const arch_register_t *out = arch_get_irn_register(env, n);
+static int sim_load(x87_state *state, ir_node *n, ir_op *op) {
+ const arch_register_t *out = x87_get_irn_register(state->sim, n);
ia32_attr_t *attr;
- DB((dbg, LEVEL_1, ">>> %s -> %s\n", get_irn_opname(n), arch_register_get_name(out)));
+ DB((dbg, LEVEL_1, ">>> %+F -> %s\n", n, arch_register_get_name(out)));
x87_push(state, arch_register_get_index(out), x87_patch_insn(n, op));
+ assert(out == x87_get_irn_register(state->sim, n));
attr = get_ia32_attr(n);
attr->x87[2] = out = &ia32_st_regs[0];
DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), arch_register_get_name(out)));
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
* @param op the x87 store opcode
* @param op_p the x87 store and pop opcode
*/
-static int sim_store(x87_state *state, ir_node *n, const arch_env_t *env, ir_op *op, ir_op *op_p) {
+static int sim_store(x87_state *state, ir_node *n, ir_op *op, ir_op *op_p) {
+ x87_simulator *sim = state->sim;
ir_node *val = get_irn_n(n, STORE_VAL_IDX);
- const arch_register_t *op2 = arch_get_irn_register(env, val);
- unsigned live = vfp_liveness_nodes_live_at(state->sim, n);
+ const arch_register_t *op2 = x87_get_irn_register(sim, val);
+ unsigned live = vfp_live_args_after(sim, n, 0);
int insn = 0;
ia32_attr_t *attr;
int op2_idx, depth;
op2_idx = x87_on_stack(state, arch_register_get_index(op2));
assert(op2_idx >= 0);
- DB((dbg, LEVEL_1, ">>> %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2)));
+ DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
mode = get_ia32_ls_mode(n);
depth = x87_get_depth(state);
if (! (mode == mode_E && depth < N_x87_REGS) && op2_idx != 0)
x87_create_fxch(state, n, op2_idx, STORE_VAL_IDX);
- if (is_vfp_live(op2->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op2), live)) {
/*
Problem: fst doesn't support mode_E (spills), only fstp does
Solution:
if (mode == mode_E) {
if (depth < N_x87_REGS) {
/* ok, we have a free register: push + fstp */
- x87_create_fpush(env, state, n, op2_idx, STORE_VAL_IDX);
+ x87_create_fpush(state, n, op2_idx, STORE_VAL_IDX);
x87_pop(state);
x87_patch_insn(n, op_p);
}
assert(mem && "Store memory not found");
- arch_set_irn_register(env, rproj, op2);
+ arch_set_irn_register(sim->env, rproj, op2);
/* reroute all former users of the store memory to the load memory */
edges_reroute(mem, mproj, irg);
#define _GEN_BINOP(op, rev) \
-static int sim_##op(x87_state *state, ir_node *n, const arch_env_t *env) { \
+static int sim_##op(x87_state *state, ir_node *n) { \
exchange_tmpl tmpl = { op_ia32_##op, op_ia32_##rev, op_ia32_##op##p, op_ia32_##rev##p }; \
- return sim_binop(state, n, env, &tmpl); \
+ return sim_binop(state, n, &tmpl); \
}
#define GEN_BINOP(op) _GEN_BINOP(op, op)
#define GEN_BINOPR(op) _GEN_BINOP(op, op##r)
#define GEN_LOAD2(op, nop) \
-static int sim_##op(x87_state *state, ir_node *n, const arch_env_t *env) { \
- return sim_load(state, n, env, op_ia32_##nop); \
+static int sim_##op(x87_state *state, ir_node *n) { \
+ return sim_load(state, n, op_ia32_##nop); \
}
#define GEN_LOAD(op) GEN_LOAD2(op, op)
#define GEN_UNOP(op) \
-static int sim_##op(x87_state *state, ir_node *n, const arch_env_t *env) { \
- return sim_unop(state, n, env, op_ia32_##op); \
+static int sim_##op(x87_state *state, ir_node *n) { \
+ return sim_unop(state, n, op_ia32_##op); \
}
#define GEN_STORE(op) \
-static int sim_##op(x87_state *state, ir_node *n, const arch_env_t *env) { \
- return sim_store(state, n, env, op_ia32_##op, op_ia32_##op##p); \
+static int sim_##op(x87_state *state, ir_node *n) { \
+ return sim_store(state, n, op_ia32_##op, op_ia32_##op##p); \
}
/* all stubs */
GEN_BINOPR(fsub)
GEN_BINOP(fmul)
GEN_BINOPR(fdiv)
+GEN_BINOP(fprem)
GEN_UNOP(fabs)
GEN_UNOP(fchs)
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
*/
-static int sim_fCondJmp(x87_state *state, ir_node *n, const arch_env_t *env) {
+static int sim_fCondJmp(x87_state *state, ir_node *n) {
int op2_idx, op1_idx = -1, pop_cnt = 0;
ia32_attr_t *attr;
ir_op *dst;
- const arch_register_t *op1 = arch_get_irn_register(env, get_irn_n(n, BINOP_IDX_1));
- const arch_register_t *op2 = arch_get_irn_register(env, get_irn_n(n, BINOP_IDX_2));
- unsigned live = vfp_liveness_nodes_live_at(state->sim, n);
+ x87_simulator *sim = state->sim;
+ const arch_register_t *op1 = x87_get_irn_register(sim, get_irn_n(n, BINOP_IDX_1));
+ const arch_register_t *op2 = x87_get_irn_register(sim, get_irn_n(n, BINOP_IDX_2));
+ unsigned live = vfp_live_args_after(sim, n, 0);
- DB((dbg, LEVEL_1, ">>> %s %s, %s\n", get_irn_opname(n),
+ DB((dbg, LEVEL_1, ">>> %+F %s, %s\n", n,
arch_register_get_name(op1), arch_register_get_name(op2)));
DEBUG_ONLY(vfp_dump_live(live));
op2_idx = x87_on_stack(state, arch_register_get_index(op2));
/* BEWARE: check for comp a,a cases, they might happen */
- if (op2->index != REG_VFP_NOREG) {
+ if (arch_register_get_index(op2) != REG_VFP_NOREG) {
/* second operand is a vfp register */
- if (is_vfp_live(op2->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op2), live)) {
/* second operand is live */
- if (is_vfp_live(op1->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
/* both operands are live: move one of them to tos */
if (op2_idx == 0) {
XCHG(op2_idx, op1_idx);
}
else {
/* second operand is dead */
- if (is_vfp_live(op1->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
/* first operand is live: bring second to tos.
This means further, op1_idx != op2_idx. */
if (op2_idx != 0) {
}
else {
/* second operand is an address mode */
- if (is_vfp_live(op1->index, live)) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
/* first operand is live: bring it to TOS */
if (op1_idx != 0) {
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
/* patch the operation */
attr = get_ia32_attr(n);
- attr->x87[0] = op1 = &ia32_st_regs[op1_idx];
+ attr->x87[1] = op1 = &ia32_st_regs[op1_idx];
if (op2_idx >= 0)
- attr->x87[1] = op2 = &ia32_st_regs[op2_idx];
+ attr->x87[2] = op2 = &ia32_st_regs[op2_idx];
if (op2_idx >= 0)
DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(n),
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
*/
-static int sim_Copy(x87_state *state, ir_node *n, const arch_env_t *env) {
+static int sim_Copy(x87_state *state, ir_node *n) {
ir_mode *mode = get_irn_mode(n);
if (mode_is_float(mode)) {
- const arch_register_t *op1 = arch_get_irn_register(env, get_irn_n(n, 0));
- const arch_register_t *out = arch_get_irn_register(env, n);
- ir_node *node, *next;
- ia32_attr_t *attr;
- int op1_idx, out_idx;
- unsigned live = vfp_liveness_nodes_live_at(state->sim, n);
-
- op1_idx = x87_on_stack(state, arch_register_get_index(op1));
-
- DB((dbg, LEVEL_1, ">>> %s %s -> %s\n", get_irn_opname(n),
+ x87_simulator *sim = state->sim;
+ ir_node *pred = get_irn_n(n, 0);
+ const arch_register_t *out = x87_get_irn_register(sim, n);
+ const arch_register_t *op1 = x87_get_irn_register(sim, pred);
+ ir_node *node, *next;
+ ia32_attr_t *attr;
+ int op1_idx, out_idx;
+ unsigned live = vfp_live_args_after(sim, n, REGMASK(out));
+ ir_node *(*cnstr)(dbg_info *, ir_graph *, ir_node *, ir_mode *);
+
+ DB((dbg, LEVEL_1, ">>> %+F %s -> %s\n", n,
arch_register_get_name(op1), arch_register_get_name(out)));
DEBUG_ONLY(vfp_dump_live(live));
- if (is_vfp_live(op1->index, live)) {
- /* operand is still live,a real copy */
- node = new_rd_ia32_fpush(get_irn_dbg_info(n), get_irn_irg(n), get_nodes_block(n), get_irn_n(n, 0), mode);
- arch_set_irn_register(env, node, out);
+ /* Do not copy constants, recreate them. */
+ switch (get_ia32_irn_opcode(pred)) {
+ case iro_ia32_fldz:
+ cnstr = new_rd_ia32_fldz;
+ break;
+ case iro_ia32_fld1:
+ cnstr = new_rd_ia32_fld1;
+ break;
+ case iro_ia32_fldpi:
+ cnstr = new_rd_ia32_fldpi;
+ break;
+ case iro_ia32_fldl2e:
+ cnstr = new_rd_ia32_fldl2e;
+ break;
+ case iro_ia32_fldl2t:
+ cnstr = new_rd_ia32_fldl2t;
+ break;
+ case iro_ia32_fldlg2:
+ cnstr = new_rd_ia32_fldlg2;
+ break;
+ case iro_ia32_fldln2:
+ cnstr = new_rd_ia32_fldln2;
+ break;
+ default:
+ goto no_constant;
+ }
+
+ /* copy a constant */
+ node = (*cnstr)(get_irn_dbg_info(n), get_irn_irg(n), get_nodes_block(n), mode);
+ arch_set_irn_register(sim->env, node, out);
+
+ x87_push(state, arch_register_get_index(out), node);
+
+ attr = get_ia32_attr(node);
+ attr->x87[2] = out = &ia32_st_regs[0];
+
+ next = sched_next(n);
+ sched_remove(n);
+ exchange(n, node);
+ sched_add_before(next, node);
+ DB((dbg, LEVEL_1, ">>> %+F -> %s\n", node, arch_register_get_name(out)));
+ return 0;
+
+no_constant:
+ /* handle the infamous unknown value */
+ if (arch_register_get_index(op1) == REG_VFP_UKNWN) {
+ /* This happens before Phi nodes */
+ if (x87_state_is_empty(state)) {
+ /* create some value */
+ x87_patch_insn(n, op_ia32_fldz);
+ attr = get_ia32_attr(n);
+ attr->x87[2] = out = &ia32_st_regs[0];
+ DB((dbg, LEVEL_1, "<<< %+F -> %s\n", n,
+ arch_register_get_name(out)));
+ } else {
+ /* Just copy one. We need here an fpush that can hold a
+ a register, so use the fpushCopy. */
+ node = new_rd_ia32_fpushCopy(get_irn_dbg_info(n), get_irn_irg(n), get_nodes_block(n), get_irn_n(n, 0), mode);
+ arch_set_irn_register(sim->env, node, out);
+
+ x87_push(state, arch_register_get_index(out), node);
+
+ attr = get_ia32_attr(node);
+ attr->x87[0] = op1 =
+ attr->x87[2] = out = &ia32_st_regs[0];
+
+ next = sched_next(n);
+ sched_remove(n);
+ exchange(n, node);
+ sched_add_before(next, node);
+ DB((dbg, LEVEL_1, "<<< %+F %s -> %s\n", node,
+ arch_register_get_name(op1),
+ arch_register_get_name(out)));
+ }
+ return 0;
+ }
+
+ op1_idx = x87_on_stack(state, arch_register_get_index(op1));
+
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
+ /* Operand is still live,a real copy. We need here an fpush that can hold a
+ a register, so use the fpushCopy. */
+ node = new_rd_ia32_fpushCopy(get_irn_dbg_info(n), get_irn_irg(n), get_nodes_block(n), get_irn_n(n, 0), mode);
+ arch_set_irn_register(sim->env, node, out);
x87_push(state, arch_register_get_index(out), node);
sched_remove(n);
exchange(n, node);
sched_add_before(next, node);
- DB((dbg, LEVEL_1, ">>> %s %s -> %s\n", get_irn_opname(node), op1->name, out->name));
+ DB((dbg, LEVEL_1, ">>> %+F %s -> %s\n", node, op1->name, out->name));
}
else {
out_idx = x87_on_stack(state, arch_register_get_index(out));
x87_pop(state);
x87_set_st(state, arch_register_get_index(out), n, out_idx - 1);
}
- DB((dbg, LEVEL_1, ">>> %s %s\n", get_irn_opname(n), op1->name));
+ DB((dbg, LEVEL_1, ">>> %+F %s\n", n, op1->name));
}
else {
/* just a virtual copy */
}
}
}
-
return 0;
} /* sim_Copy */
*
* Should not happen, spills are lowered before x87 simulator see them.
*/
-static int sim_Spill(x87_state *state, ir_node *n, const arch_env_t *env) {
+static int sim_Spill(x87_state *state, ir_node *n) {
assert(0 && "Spill not lowered");
- return sim_fst(state, n, env);
+ return sim_fst(state, n);
} /* sim_Spill */
/**
*
* Should not happen, reloads are lowered before x87 simulator see them.
*/
-static int sim_Reload(x87_state *state, ir_node *n, const arch_env_t *env) {
+static int sim_Reload(x87_state *state, ir_node *n) {
assert(0 && "Reload not lowered");
- return sim_fld(state, n, env);
+ return sim_fld(state, n);
} /* sim_Reload */
/**
* @param n the node that should be simulated (and patched)
* @param env the architecture environment
*/
-static int sim_Return(x87_state *state, ir_node *n, const arch_env_t *env) {
+static int sim_Return(x87_state *state, ir_node *n) {
int n_res = be_Return_get_n_rets(n);
int i, n_float_res = 0;
return 0;
} /* sim_Return */
+typedef struct _perm_data_t {
+ const arch_register_t *in;
+ const arch_register_t *out;
+} perm_data_t;
+
+/**
+ * Simulate a be_Perm.
+ *
+ * @param state the x87 state
+ * @param irn the node that should be simulated (and patched)
+ */
+static int sim_Perm(x87_state *state, ir_node *irn) {
+ int i, n;
+ x87_simulator *sim = state->sim;
+ ir_node *pred = get_irn_n(irn, 0);
+ int *stack_pos;
+ const ir_edge_t *edge;
+
+ /* handle only floating point Perms */
+ if (! mode_is_float(get_irn_mode(pred)))
+ return 0;
+
+ DB((dbg, LEVEL_1, ">>> %+F\n", irn));
+
+ /* Perm is a pure virtual instruction on x87.
+ All inputs must be on the FPU stack and are pairwise
+ different from each other.
+ So, all we need to do is to permutate the stack state. */
+ n = get_irn_arity(irn);
+ NEW_ARR_A(int, stack_pos, n);
+
+ /* collect old stack positions */
+ for (i = 0; i < n; ++i) {
+ const arch_register_t *inreg = x87_get_irn_register(sim, get_irn_n(irn, i));
+ int idx = x87_on_stack(state, arch_register_get_index(inreg));
+
+ assert(idx >= 0 && "Perm argument not on x87 stack");
+
+ stack_pos[i] = idx;
+ }
+ /* now do the permutation */
+ foreach_out_edge(irn, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ const arch_register_t *out = x87_get_irn_register(sim, proj);
+ long num = get_Proj_proj(proj);
+
+ assert(0 <= num && num < n && "More Proj's than Perm inputs");
+ x87_set_st(state, arch_register_get_index(out), proj, stack_pos[(unsigned)num]);
+ }
+ DB((dbg, LEVEL_1, "<<< %+F\n", irn));
+
+ return 0;
+} /* be_Perm */
+
/**
* Kill any dead registers at block start by popping them from the stack.
*
x87_state *state = start_state;
ir_node *first_insn = sched_first(block);
ir_node *keep = NULL;
- unsigned live = vfp_liveness_nodes_live_at(sim, block);
+ unsigned live = vfp_live_args_after(sim, block, 0);
unsigned kill_mask;
int i, depth, num_pop;
depth -= num_pop;
kill_mask >>= num_pop;
- keep = x87_create_fpop(sim->env, state, first_insn, num_pop, keep);
+ keep = x87_create_fpop(state, first_insn, num_pop, keep);
}
- add_End_keepalive(get_irg_end(get_irn_irg(block)), keep);
+ keep_alive(keep);
}
return state;
} /* x87_kill_deads */
}
/* simulate it */
- node_inserted = (*func)(state, n, sim->env);
+ node_inserted = (*func)(state, n);
/*
sim_func might have added additional nodes after n,
DB((dbg, LEVEL_1, "--------------------------------\n"
"x87 Simulator started for %+F\n", irg));
- /* set the generic function pointer of instruction we must simulate */
+ /* set the generic function pointer of instruction we must simulate */
clear_irp_opcodes_generic_func();
#define ASSOC(op) (op_ ## op)->ops.generic = (op_func)(sim_##op)
ASSOC_IA32(fsub);
ASSOC_IA32(fmul);
ASSOC_IA32(fdiv);
- ASSOC_IA32(fldz);
+ ASSOC_IA32(fprem);
ASSOC_IA32(fabs);
ASSOC_IA32(fchs);
ASSOC_IA32(fsin);
ASSOC_BE(Spill);
ASSOC_BE(Reload);
ASSOC_BE(Return);
+ ASSOC_BE(Perm);
ASSOC(Phi);
#undef ASSOC_BE
#undef ASSOC_IA32
* Needs a block-schedule.
*/
void x87_simulate_graph(const arch_env_t *env, ir_graph *irg, ir_node **blk_list) {
- ir_node *block, *start_block;
- waitq *worklist;
- blk_state *bl_state;
+ ir_node *block, *start_block;
+ waitq *worklist;
+ blk_state *bl_state;
x87_simulator sim;
- int i, n;
- be_lv_t *lv;
+ int i, n;
+ be_lv_t *lv;
+ ir_graph *rem = current_ir_graph;
+
+ current_ir_graph = irg;
/* create the simulator */
x87_init_simulator(&sim, irg, env);
/* kill it */
del_waitq(worklist);
x87_destroy_simulator(&sim);
+ current_ir_graph = rem;
} /* x87_simulate_graph */