*
* $Id$
*/
-
#ifdef HAVE_CONFIG_H
-#include "config.h"
-#endif /* HAVE_CONFIG_H */
+#include <config.h>
+#endif
#include <assert.h>
#include "iredges_t.h"
#include "irgmod.h"
#include "ircons.h"
+#include "irgwalk.h"
#include "obst.h"
#include "pmap.h"
#include "pdeq.h"
#define N_x87_REGS 8
/* first and second binop index */
-#define BINOP_IDX_1 2
+#define BINOP_IDX_1 2
#define BINOP_IDX_2 3
/* the unop index */
-#define UNOP_IDX 0
+#define UNOP_IDX 0
/* the store val index */
#define STORE_VAL_IDX 2
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.
#define PTR_TO_BLKSTATE(p) ((blk_state *)(p))
+/** liveness bitset for vfp registers. */
+typedef unsigned char vfp_liveness;
+
/**
* 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 */
- be_lv_t *lv; /**< Liveness information. */
+ struct obstack obst; /**< An obstack for fast allocating. */
+ pmap *blk_states; /**< Map blocks to states. */
+ const arch_env_t *arch_env; /**< The architecture environment. */
+ be_lv_t *lv; /**< intrablock liveness. */
+ vfp_liveness *live; /**< Liveness information. */
+ unsigned n_idx; /**< The cached get_irg_last_idx() result. */
+ waitq *worklist; /**< list of blocks to process. */
};
/**
return state->depth;
}
+#if 0
/**
* Check if the state is empty.
*
static int x87_state_is_empty(const x87_state *state) {
return state->depth == 0;
}
+#endif
/**
* Return the virtual register index at st(pos).
static ir_node *x87_get_st_node(const x87_state *state, int pos) {
assert(pos < state->depth);
return state->st[MASK_TOS(state->tos + pos)].node;
-}
+} /* x87_get_st_node */
#ifdef DEBUG_libfirm
/**
int i;
for (i = state->depth - 1; i >= 0; --i) {
- DB((dbg, LEVEL_2, "vf%d ", x87_get_st_reg(state, i)));
+ DB((dbg, LEVEL_2, "vf%d(%+F) ", x87_get_st_reg(state, i),
+ x87_get_st_node(state, i)));
}
DB((dbg, LEVEL_2, "<-- TOS\n"));
-}
+} /* x87_dump_stack */
#endif /* DEBUG_libfirm */
/**
state->st[MASK_TOS(state->tos + pos)].reg_idx = reg_idx;
state->st[MASK_TOS(state->tos + pos)].node = node;
- DB((dbg, LEVEL_2, "After SET_REG:\n ")); DEBUG_ONLY(x87_dump_stack(state));
-}
+ DB((dbg, LEVEL_2, "After SET_REG: "));
+ DEBUG_ONLY(x87_dump_stack(state));
+} /* x87_set_st */
/**
* Set the tos virtual register.
*/
static void x87_set_tos(x87_state *state, int reg_idx, ir_node *node) {
x87_set_st(state, reg_idx, node, 0);
-}
+} /* x87_set_tos */
+#if 0
/**
* Flush the x87 stack.
*
static void x87_flush(x87_state *state) {
state->depth = 0;
state->tos = 0;
-}
+} /* x87_flush */
+#endif
/**
* Swap st(0) with st(pos).
state->st[MASK_TOS(state->tos + pos)] = state->st[MASK_TOS(state->tos)];
state->st[MASK_TOS(state->tos)] = entry;
- DB((dbg, LEVEL_2, "After FXCH:\n ")); DEBUG_ONLY(x87_dump_stack(state));
-}
+ DB((dbg, LEVEL_2, "After FXCH: ")); DEBUG_ONLY(x87_dump_stack(state));
+} /* x87_fxch */
/**
* Convert a virtual register to the stack index.
if (state->st[MASK_TOS(tos + i)].reg_idx == reg_idx)
return i;
return -1;
-}
+} /* x87_on_stack */
/**
* Push a virtual Register onto the stack, double pushed allowed.
state->st[state->tos].reg_idx = reg_idx;
state->st[state->tos].node = node;
- DB((dbg, LEVEL_2, "After PUSH:\n ")); DEBUG_ONLY(x87_dump_stack(state));
-}
+ DB((dbg, LEVEL_2, "After PUSH: ")); DEBUG_ONLY(x87_dump_stack(state));
+} /* x87_push_dbl */
/**
* Push a virtual Register onto the stack, double pushes are NOT allowed..
assert(x87_on_stack(state, reg_idx) == -1 && "double push");
x87_push_dbl(state, reg_idx, node);
-}
+} /* x87_push */
/**
* Pop a virtual Register from the stack.
--state->depth;
state->tos = MASK_TOS(state->tos + 1);
- DB((dbg, LEVEL_2, "After POP:\n ")); DEBUG_ONLY(x87_dump_stack(state));
-}
+ DB((dbg, LEVEL_2, "After POP: ")); DEBUG_ONLY(x87_dump_stack(state));
+} /* x87_pop */
/**
* Returns the block state of a block.
}
return PTR_TO_BLKSTATE(entry->value);
-}
+} /* x87_get_bl_state */
/**
* Creates a new x87 state.
*/
static x87_state *x87_alloc_state(x87_simulator *sim) {
x87_state *res = obstack_alloc(&sim->obst, sizeof(*res));
+
res->sim = sim;
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.
memcpy(res, src, sizeof(*res));
return res;
-}
+} /* x87_clone_state */
/**
* Patch a virtual instruction into a x87 one and return
else if (mode_is_float(mode))
set_irn_mode(n, mode_E);
return res;
-}
+} /* x87_patch_insn */
/**
* Returns the first Proj of a mode_T node having a given mode.
}
return NULL;
+} /* get_irn_Proj_for_mode */
+
+/**
+ * Wrap the arch_* function here so we can check for errors.
+ */
+static INLINE 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->arch_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;
-}
+} /* x87_fxch_shuffle */
/**
* Calculate the necessary permutations to reach dst_state.
* @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);
}
}
return state;
-}
+} /* x87_shuffle */
/**
* 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;
+ ir_graph *irg = get_irn_irg(n);
+ ir_node *block = get_nodes_block(n);
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, irg, block, 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));
return fxch;
-}
+} /* x87_create_fxch */
/**
* 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];
attr->x87[2] = &ia32_st_regs[0];
+ keep_alive(fpop);
sched_add_before(n, fpop);
DB((dbg, LEVEL_1, "<<< %s %s\n", get_irn_opname(fpop), attr->x87[0]->name));
--num;
}
return fpop;
+} /* x87_create_fpop */
+
+/**
+ * Creates an fldz before node n
+ *
+ * @param state the x87 state
+ * @param n the node after the fldz
+ *
+ * @return the fldz node
+ */
+static ir_node *x87_create_fldz(x87_state *state, ir_node *n, int regidx) {
+ ir_graph *irg = get_irn_irg(n);
+ ir_node *block = get_nodes_block(n);
+ ir_node *fldz;
+
+ fldz = new_rd_ia32_fldz(NULL, irg, block, mode_E);
+
+ sched_add_before(n, fldz);
+ DB((dbg, LEVEL_1, "<<< %s\n", get_irn_opname(fldz)));
+ keep_alive(fldz);
+
+ x87_push(state, regidx, fldz);
+
+ return fldz;
}
/* --------------------------------- liveness ------------------------------------------ */
* The liveness transfer function.
* 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 live.
+ *
+ * @return The live bitset.
*/
-static unsigned vfp_liveness_transfer(const arch_env_t *arch_env, ir_node *irn, unsigned live)
+static vfp_liveness vfp_liveness_transfer(x87_simulator *sim, ir_node *irn, vfp_liveness live)
{
int i, n;
const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
+ const arch_env_t *arch_env = sim->arch_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;
-}
+} /* vfp_liveness_transfer */
/**
* Put all live virtual registers at the end of a block into a bitset.
- * @param arch_env The architecture environment.
- * @param bl The block.
- * @return The live bitset.
+ *
+ * @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(x87_simulator *sim, const ir_node *bl)
+static vfp_liveness vfp_liveness_end_of_block(x87_simulator *sim, const ir_node *block)
{
int i;
- unsigned live = 0;
+ vfp_liveness live = 0;
const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
+ const arch_env_t *arch_env = sim->arch_env;
+ const be_lv_t *lv = sim->lv;
- be_lv_foreach(sim->lv, bl, be_lv_state_end, i) {
- ir_node *irn = be_lv_get_irn(sim->lv, bl, i);
- if (arch_irn_consider_in_reg_alloc(sim->env, cls, irn)) {
- const arch_register_t *reg = arch_get_irn_register(sim->env, irn);
- live |= 1 << reg->index;
- }
+ be_lv_foreach(lv, block, be_lv_state_end, i) {
+ const arch_register_t *reg;
+ const ir_node *node = be_lv_get_irn(lv, block, i);
+ if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
+ continue;
+
+ reg = x87_get_irn_register(sim, node);
+ 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)))
/**
- * Compute a bitset of registers which are live at another node.
- * @param arch_env The architecture environment.
- * @param pos The 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)
{
- const ir_node *bl = is_Block(pos) ? pos : get_nodes_block(pos);
+ unsigned idx = get_irn_idx(pos);
+
+ assert(idx < sim->n_idx);
+ return sim->live[idx] & ~kill;
+} /* vfp_live_args_after */
+
+/**
+ * Calculate the liveness for a whole block and cache it.
+ *
+ * @param sim the simulator handle
+ * @param lv the liveness handle
+ * @param block the block
+ */
+static void update_liveness(x87_simulator *sim, ir_node *block) {
+ vfp_liveness live = vfp_liveness_end_of_block(sim, block);
+ unsigned idx;
ir_node *irn;
- unsigned live;
- live = vfp_liveness_end_of_block(sim, bl);
+ /* now iterate through the block backward and cache the results */
+ sched_foreach_reverse(block, irn) {
+ /* stop at the first Phi: this produces the live-in */
+ if (is_Phi(irn))
+ break;
- sched_foreach_reverse(bl, irn) {
- /*
- * If we encounter the node we want to insert the Perm after,
- * exit immediately, so that this node is still live
- */
- if (irn == pos)
- return live;
+ idx = get_irn_idx(irn);
+ sim->live[idx] = live;
- live = vfp_liveness_transfer(sim->env, irn, live);
+ live = vfp_liveness_transfer(sim, irn, live);
}
-
- return live;
-}
+ idx = get_irn_idx(block);
+ sim->live[idx] = live;
+} /* update_liveness */
/**
* Returns true if a register is live in a set.
* @param reg_idx the vfp register index
* @param live a live bitset
*/
-static unsigned is_vfp_live(int reg_idx, unsigned live) {
- return live & (1 << reg_idx);
-}
+#define is_vfp_live(reg_idx, live) ((live) & (1 << (reg_idx)))
#ifdef DEBUG_libfirm
/**
*
* @param live the live bitset
*/
-static void vfp_dump_live(unsigned live) {
+static void vfp_dump_live(vfp_liveness live) {
int i;
- DB((dbg, LEVEL_2, "Live registers here: \n"));
+ DB((dbg, LEVEL_2, "Live after: "));
for (i = 0; i < 8; ++i) {
if (live & (1 << i)) {
- DB((dbg, LEVEL_2, " vf%d", i));
+ DB((dbg, LEVEL_2, "vf%d ", i));
}
}
DB((dbg, LEVEL_2, "\n"));
-}
+} /* vfp_dump_live */
#endif /* DEBUG_libfirm */
/* --------------------------------- simulators ---------------------------------------- */
*
* @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) {
- int op2_idx, op1_idx = -1;
- int out_idx, do_pop =0;
+static int sim_binop(x87_state *state, ir_node *n, const exchange_tmpl *tmpl) {
+ int op2_idx, op1_idx;
+ int out_idx, do_pop = 0;
ia32_attr_t *attr;
+ ir_node *patched_insn;
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);
-
- DB((dbg, LEVEL_1, ">>> %s %s, %s -> %s\n", get_irn_opname(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);
+ int reg_index_1 = arch_register_get_index(op1);
+ int reg_index_2 = arch_register_get_index(op2);
+ vfp_liveness live = vfp_live_args_after(sim, n, REGMASK(out));
+
+ 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));
+ DB((dbg, LEVEL_1, "Stack before: "));
+ DEBUG_ONLY(x87_dump_stack(state));
- op1_idx = x87_on_stack(state, arch_register_get_index(op1));
- op2_idx = x87_on_stack(state, arch_register_get_index(op2));
+ op1_idx = x87_on_stack(state, reg_index_1);
+ assert(op1_idx >= 0);
- if (op2->index != REG_VFP_NOREG) {
+ if (reg_index_2 != REG_VFP_NOREG) {
/* second operand is a vfp register */
+ op2_idx = x87_on_stack(state, reg_index_2);
+ assert(op2_idx >= 0);
- 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);
- out_idx = op2_idx = 0;
- ++op1_idx;
+ x87_create_fpush(state, n, op1_idx, BINOP_IDX_2);
+ /* now do fxxx (tos=tos X op) */
+ op1_idx = 0;
+ op2_idx += 1;
+ out_idx = 0;
dst = tmpl->normal_op;
- do_pop = 0;
- }
- else {
+ } else {
/* Second live, first operand is dead here, bring it to tos. */
if (op1_idx != 0) {
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
if (op2_idx == 0)
op2_idx = op1_idx;
+ op1_idx = 0;
}
- op1_idx = out_idx = 0;
+ /* now do fxxx (tos=tos X op) */
+ out_idx = 0;
dst = tmpl->normal_op;
- do_pop = 0;
}
- }
- else {
+ } 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);
if (op1_idx == 0)
op1_idx = op2_idx;
+ op2_idx = 0;
}
- op2_idx = out_idx = 0;
- dst = tmpl->normal_op;
- do_pop = 0;
- }
- else {
+ /* now do fxxxr (tos = op X tos) */
+ out_idx = 0;
+ dst = tmpl->reverse_op;
+ } else {
/* Both operands are dead here, pop them from the stack. */
if (op2_idx == 0) {
- out_idx = op1_idx;
- XCHG(op2_idx, op1_idx);
- if (op1_idx == op2_idx) {
- /* Both are identically, no pop needed. */
- dst = tmpl->reverse_op;
- do_pop = 0;
- }
- else {
- dst = tmpl->reverse_pop_op;
- do_pop = 1;
- }
- }
- else if (op1_idx == 0) {
- out_idx = op2_idx;
- if (op1_idx == op2_idx) {
- /* Both are identically, no pop needed. */
+ if (op1_idx == 0) {
+ /* Both are identically and on tos, no pop needed. */
+ /* here fxxx (tos = tos X tos) */
dst = tmpl->normal_op;
- do_pop = 0;
- }
- else {
+ out_idx = 0;
+ } else {
+ /* now do fxxxp (op = op X tos, pop) */
dst = tmpl->normal_pop_op;
do_pop = 1;
+ out_idx = op1_idx;
}
- }
- else {
- /* Bring the first on top. */
- x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
+ } else if (op1_idx == 0) {
+ assert(op1_idx != op2_idx);
+ /* now do fxxxrp (op = tos X op, pop) */
+ dst = tmpl->reverse_pop_op;
+ do_pop = 1;
+ out_idx = op2_idx;
+ } else {
+ /* Bring the second on top. */
+ x87_create_fxch(state, n, op2_idx, BINOP_IDX_2);
if (op1_idx == op2_idx) {
- /* Both are identically, no pop needed. */
- out_idx = op1_idx = op2_idx = 0;
- dst = tmpl->normal_op;
- do_pop = 0;
- }
- else {
+ /* Both are identically and on tos now, no pop needed. */
op1_idx = 0;
- out_idx = op2_idx;
+ op2_idx = 0;
+ /* use fxxx (tos = tos X tos) */
+ dst = tmpl->normal_op;
+ out_idx = 0;
+ } else {
+ /* op2 is on tos now */
+ op2_idx = 0;
+ /* use fxxxp (op = op X tos, pop) */
dst = tmpl->normal_pop_op;
+ out_idx = op1_idx;
do_pop = 1;
}
}
}
}
- }
- else {
+ } 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);
- }
- else {
+ x87_create_fpush(state, n, op1_idx, BINOP_IDX_1);
+ op1_idx = 0;
+ /* use fxxx (tos = tos X mem) */
+ dst = tmpl->normal_op;
+ out_idx = 0;
+ } else {
/* first operand is dead: bring it to tos */
- if (op1_idx != 0)
+ if (op1_idx != 0) {
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
+ op1_idx = 0;
+ }
+
+ /* use fxxxp (tos = tos X mem) */
+ dst = tmpl->normal_op;
+ out_idx = 0;
}
- op1_idx = out_idx = 0;
- dst = tmpl->normal_op;
- do_pop = 0;
}
- x87_set_st(state, arch_register_get_index(out), x87_patch_insn(n, dst), out_idx);
- if (do_pop)
+ patched_insn = x87_patch_insn(n, dst);
+ x87_set_st(state, arch_register_get_index(out), patched_insn, out_idx);
+ if (do_pop) {
x87_pop(state);
+ }
/* patch the operation */
attr = get_ia32_attr(n);
attr->x87[0] = op1 = &ia32_st_regs[op1_idx];
- if (op2_idx >= 0)
+ if (reg_index_2 != REG_VFP_NOREG) {
attr->x87[1] = op2 = &ia32_st_regs[op2_idx];
+ }
attr->x87[2] = out = &ia32_st_regs[out_idx];
- if (op2_idx >= 0)
+ if (reg_index_2 != REG_VFP_NOREG) {
DB((dbg, LEVEL_1, "<<< %s %s, %s -> %s\n", get_irn_opname(n),
arch_register_get_name(op1), arch_register_get_name(op2),
arch_register_get_name(out)));
- else
+ } else {
DB((dbg, LEVEL_1, "<<< %s %s, [AM] -> %s\n", get_irn_opname(n),
arch_register_get_name(op1),
arch_register_get_name(out)));
+ }
return 0;
-}
+} /* sim_binop */
/**
* Simulate a virtual Unop.
*
* @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);
+ op1_idx = 0;
}
else {
/* operand is dead, bring it to tos */
- if (op1_idx != 0)
+ if (op1_idx != 0) {
x87_create_fxch(state, n, op1_idx, UNOP_IDX);
+ op1_idx = 0;
+ }
}
x87_set_tos(state, arch_register_get_index(out), x87_patch_insn(n, op));
- op1_idx = out_idx = 0;
+ out_idx = 0;
attr = get_ia32_attr(n);
attr->x87[0] = op1 = &ia32_st_regs[0];
attr->x87[2] = out = &ia32_st_regs[0];
DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), out->name));
return 0;
-}
+} /* sim_unop */
/**
* Simulate a virtual Load instruction.
*
* @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)));
return 0;
-}
+} /* sim_load */
/**
* Rewire all users of @p old_val to @new_val iff they are scheduled after @p store.
}
}
}
-}
+} /* collect_and_rewire_users */
/**
* Simulate a virtual Store.
*
* @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;
+ int op2_reg_idx, op2_idx, depth;
+ int live_after_node;
ir_mode *mode;
- 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)));
+ op2_reg_idx = arch_register_get_index(op2);
+ if (op2_reg_idx == REG_VFP_UKNWN) {
+ // just take any value from stack
+ if(state->depth > 0) {
+ op2_idx = 0;
+ DEBUG_ONLY(op2 = NULL);
+ live_after_node = 1;
+ } else {
+ // produce a new value which we will consume imediately
+ x87_create_fldz(state, n, op2_reg_idx);
+ live_after_node = 0;
+ op2_idx = x87_on_stack(state, op2_reg_idx);
+ assert(op2_idx >= 0);
+ }
+ } else {
+ op2_idx = x87_on_stack(state, op2_reg_idx);
+ live_after_node = is_vfp_live(arch_register_get_index(op2), live);
+ assert(op2_idx >= 0);
+ DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
+ }
mode = get_ia32_ls_mode(n);
depth = x87_get_depth(state);
- /*
- We can only store the tos to memory.
- A store of mode_E with free registers
- pushes value to tos, so skip it here.
- */
- 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 (live_after_node) {
/*
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);
}
set_ia32_use_frame(vfld);
set_ia32_am_flavour(vfld, get_ia32_am_flavour(n));
set_ia32_op_type(vfld, ia32_am_Source);
- add_ia32_am_offs(vfld, get_ia32_am_offs(n));
+ add_ia32_am_offs_int(vfld, get_ia32_am_offs_int(n));
set_ia32_am_sc(vfld, get_ia32_am_sc(n));
set_ia32_ls_mode(vfld, get_ia32_ls_mode(n));
assert(mem && "Store memory not found");
- arch_set_irn_register(env, rproj, op2);
+ arch_set_irn_register(sim->arch_env, rproj, op2);
/* reroute all former users of the store memory to the load memory */
edges_reroute(mem, mproj, irg);
}
}
else {
+ /* we can only store the tos to memory */
+ if(op2_idx != 0)
+ x87_create_fxch(state, n, op2_idx, STORE_VAL_IDX);
+
/* mode != mode_E -> use normal fst */
x87_patch_insn(n, op);
}
}
else {
+ /* we can only store the tos to memory */
+ if(op2_idx != 0)
+ x87_create_fxch(state, n, op2_idx, STORE_VAL_IDX);
+
x87_pop(state);
x87_patch_insn(n, op_p);
}
DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2)));
return insn;
-}
+} /* sim_store */
/**
* Simulate a virtual Phi.
* Just for cosmetic reasons change the mode of Phi nodes to mode_E.
*
- * @param state the x87 state
- * @param n the node that should be simulated (and patched)
- * @param env the architecture environment
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ * @param arch_env the architecture environment
*/
-static int sim_Phi(x87_state *state, ir_node *n, const arch_env_t *env) {
+static int sim_Phi(x87_state *state, ir_node *n, const arch_env_t *arch_env) {
ir_mode *mode = get_irn_mode(n);
if (mode_is_float(mode))
set_irn_mode(n, mode_E);
return 0;
-}
-
+} /* sim_Phi */
#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)
GEN_LOAD(fild)
GEN_LOAD(fldz)
GEN_LOAD(fld1)
-GEN_LOAD2(fConst, fldConst)
GEN_STORE(fst)
GEN_STORE(fist)
*
* @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) {
- int op2_idx, op1_idx = -1, pop_cnt = 0;
+static int sim_fCondJmp(x87_state *state, ir_node *n) {
+ int op1_idx;
+ int op2_idx = -1;
+ int 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);
-
- DB((dbg, LEVEL_1, ">>> %s %s, %s\n", get_irn_opname(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));
+ int reg_index_1 = arch_register_get_index(op1);
+ int reg_index_2 = arch_register_get_index(op2);
+ unsigned live = vfp_live_args_after(sim, n, 0);
+
+ 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));
+ DB((dbg, LEVEL_1, "Stack before: "));
+ DEBUG_ONLY(x87_dump_stack(state));
- op1_idx = x87_on_stack(state, arch_register_get_index(op1));
- op2_idx = x87_on_stack(state, arch_register_get_index(op2));
+ op1_idx = x87_on_stack(state, reg_index_1);
+ assert(op1_idx >= 0);
/* BEWARE: check for comp a,a cases, they might happen */
- if (op2->index != REG_VFP_NOREG) {
+ if (reg_index_2 != REG_VFP_NOREG) {
/* second operand is a vfp register */
+ op2_idx = x87_on_stack(state, reg_index_2);
+ assert(op2_idx >= 0);
- 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)) {
- /* both operands are live: move one of them to tos */
- if (op2_idx == 0) {
- XCHG(op2_idx, op1_idx);
- dst = op_ia32_fcomrJmp;
- }
- else if (op1_idx == 0) {
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
+ /* both operands are live */
+
+ if (op1_idx == 0) {
+ /* res = tos X op */
dst = op_ia32_fcomJmp;
- }
- else {
- /* bring the first on top */
+ } else if (op2_idx == 0) {
+ /* res = op X tos */
+ dst = op_ia32_fcomrJmp;
+ } else {
+ /* bring the first one to tos */
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
- if (op1_idx == op2_idx)
- op2_idx = 0;
+ if (op2_idx == 0)
+ op2_idx = op1_idx;
op1_idx = 0;
+ /* res = tos X op */
dst = op_ia32_fcomJmp;
}
- }
- else {
+ } else {
/* second live, first operand is dead here, bring it to tos.
This means further, op1_idx != op2_idx. */
+ assert(op1_idx != op2_idx);
if (op1_idx != 0) {
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
if (op2_idx == 0)
op2_idx = op1_idx;
+ op1_idx = 0;
}
- op1_idx = 0;
+ /* res = tos X op, pop */
dst = op_ia32_fcompJmp;
pop_cnt = 1;
}
- }
- else {
+ } 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. */
+ assert(op1_idx != op2_idx);
if (op2_idx != 0) {
x87_create_fxch(state, n, op2_idx, BINOP_IDX_2);
if (op1_idx == 0)
op1_idx = op2_idx;
+ op2_idx = 0;
}
- op2_idx = 0;
+ /* res = op X tos, pop */
dst = op_ia32_fcomrpJmp;
pop_cnt = 1;
- }
- else {
+ } else {
/* both operands are dead here, check first for identity. */
if (op1_idx == op2_idx) {
- /* identically, one one needed */
+ /* identically, one pop needed */
if (op1_idx != 0) {
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
- op1_idx = op2_idx = 0;
+ op1_idx = 0;
+ op2_idx = 0;
}
+ /* res = tos X op, pop */
dst = op_ia32_fcompJmp;
pop_cnt = 1;
}
if (op1_idx != 0) {
/* bring the first on top */
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
+ assert(op2_idx != 0);
op1_idx = 0;
}
+ /* res = tos X op, pop, pop */
dst = op_ia32_fcomppJmp;
pop_cnt = 2;
- }
- else if (op1_idx == 1) {
+ } else if (op1_idx == 1) {
/* good, first operand is already in the right place, move the second */
if (op2_idx != 0) {
/* bring the first on top */
x87_create_fxch(state, n, op2_idx, BINOP_IDX_2);
+ assert(op1_idx != 0);
op2_idx = 0;
}
dst = op_ia32_fcomrppJmp;
pop_cnt = 2;
- }
- else {
+ } else {
/* if one is already the TOS, we need two fxch */
if (op1_idx == 0) {
/* first one is TOS, move to st(1) */
x87_create_fxch(state, n, 1, BINOP_IDX_1);
+ assert(op2_idx != 1);
op1_idx = 1;
x87_create_fxch(state, n, op2_idx, BINOP_IDX_2);
op2_idx = 0;
+ /* res = op X tos, pop, pop */
dst = op_ia32_fcomrppJmp;
pop_cnt = 2;
- }
- else if (op2_idx == 0) {
+ } else if (op2_idx == 0) {
/* second one is TOS, move to st(1) */
x87_create_fxch(state, n, 1, BINOP_IDX_2);
+ assert(op1_idx != 1);
op2_idx = 1;
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
op1_idx = 0;
- dst = op_ia32_fcomrppJmp;
+ /* res = tos X op, pop, pop */
+ dst = op_ia32_fcomppJmp;
pop_cnt = 2;
- }
- else {
+ } else {
/* none of them is either TOS or st(1), 3 fxch needed */
x87_create_fxch(state, n, op2_idx, BINOP_IDX_2);
+ assert(op1_idx != 0);
x87_create_fxch(state, n, 1, BINOP_IDX_2);
+ op2_idx = 1;
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
op1_idx = 0;
- op2_idx = 1;
+ /* res = tos X op, pop, pop */
dst = op_ia32_fcomppJmp;
pop_cnt = 2;
}
}
}
}
- }
- else {
+ } 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);
op1_idx = 0;
}
dst = op_ia32_fcomJmp;
- }
- else {
+ } else {
/* first operand is dead: bring it to tos */
if (op1_idx != 0) {
x87_create_fxch(state, n, op1_idx, BINOP_IDX_1);
op1_idx = 0;
}
+ dst = op_ia32_fcompJmp;
+ pop_cnt = 1;
}
- dst = op_ia32_fcompJmp;
- pop_cnt = 1;
}
x87_patch_insn(n, dst);
- if (pop_cnt > 1)
+ assert(pop_cnt < 3);
+ if (pop_cnt >= 2)
x87_pop(state);
- if (pop_cnt > 0)
+ if (pop_cnt >= 1)
x87_pop(state);
/* patch the operation */
attr = get_ia32_attr(n);
- attr->x87[0] = op1 = &ia32_st_regs[op1_idx];
- if (op2_idx >= 0)
- attr->x87[1] = op2 = &ia32_st_regs[op2_idx];
+ op1 = &ia32_st_regs[op1_idx];
+ attr->x87[0] = op1;
+ if (op2_idx >= 0) {
+ op2 = &ia32_st_regs[op2_idx];
+ attr->x87[1] = op2;
+ }
+ attr->x87[2] = NULL;
if (op2_idx >= 0)
DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(n),
arch_register_get_name(op1)));
return 0;
+} /* sim_fCondJmp */
+
+static ir_node *create_Copy(x87_state *state, ir_node *n) {
+ x87_simulator *sim = state->sim;
+ ir_graph *irg = get_irn_irg(n);
+ dbg_info *n_dbg = get_irn_dbg_info(n);
+ ir_mode *mode = get_irn_mode(n);
+ ir_node *block = get_nodes_block(n);
+ ir_node *pred = get_irn_n(n, 0);
+ ir_node *(*cnstr)(dbg_info *, ir_graph *, ir_node *, ir_mode *) = NULL;
+ ir_node *res;
+ const arch_register_t *out;
+ const arch_register_t *op1;
+ ia32_attr_t *attr;
+
+ /* Do not copy constants, recreate them. */
+ switch (get_ia32_irn_opcode(pred)) {
+ case iro_ia32_Unknown_VFP:
+ 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;
+ }
+
+ out = x87_get_irn_register(sim, n);
+ op1 = x87_get_irn_register(sim, pred);
+
+ if(cnstr != NULL) {
+ /* copy a constant */
+ res = (*cnstr)(n_dbg, irg, block, mode);
+
+ x87_push(state, arch_register_get_index(out), res);
+
+ attr = get_ia32_attr(res);
+ attr->x87[2] = out = &ia32_st_regs[0];
+ } else {
+ int op1_idx = x87_on_stack(state, arch_register_get_index(op1));
+
+ res = new_rd_ia32_fpushCopy(n_dbg, irg, block, pred, mode);
+
+ x87_push(state, arch_register_get_index(out), res);
+
+ attr = get_ia32_attr(res);
+ attr->x87[0] = op1 = &ia32_st_regs[op1_idx];
+ attr->x87[2] = out = &ia32_st_regs[0];
+ }
+ arch_set_irn_register(sim->arch_env, res, out);
+
+ return res;
}
/**
*
* @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) {
+ x87_simulator *sim;
+ ir_node *pred;
+ const arch_register_t *out;
+ const arch_register_t *op1;
+ ir_node *node, *next;
+ ia32_attr_t *attr;
+ int op1_idx, out_idx;
+ unsigned live;
+
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);
+ if (!mode_is_float(mode))
+ return 0;
- op1_idx = x87_on_stack(state, arch_register_get_index(op1));
+ sim = state->sim;
+ pred = get_irn_n(n, 0);
+ out = x87_get_irn_register(sim, n);
+ op1 = x87_get_irn_register(sim, pred);
+ live = vfp_live_args_after(sim, n, REGMASK(out));
- DB((dbg, LEVEL_1, ">>> %s %s -> %s\n", get_irn_opname(n),
- arch_register_get_name(op1), arch_register_get_name(out)));
- DEBUG_ONLY(vfp_dump_live(live));
+ 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);
+ /* handle the infamous unknown value */
+ if (arch_register_get_index(op1) == REG_VFP_UKNWN) {
+ /* Operand is still live, a real copy. We need here an fpush that can
+ hold a a register, so use the fpushCopy or recreate constants */
+ node = create_Copy(state, n);
- x87_push(state, arch_register_get_index(out), node);
+ assert(is_ia32_fldz(node));
+ next = sched_next(n);
+ sched_remove(n);
+ exchange(n, node);
+ sched_add_before(next, node);
- attr = get_ia32_attr(node);
- attr->x87[0] = op1 = &ia32_st_regs[op1_idx];
- attr->x87[2] = out = &ia32_st_regs[0];
+ DB((dbg, LEVEL_1, "<<< %+F %s -> %s\n", node, op1->name,
+ arch_get_irn_register(sim->arch_env, node)->name));
+ return 0;
+ }
- next = sched_next(n);
- 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));
- }
- else {
- out_idx = x87_on_stack(state, arch_register_get_index(out));
-
- if (out_idx >= 0 && out_idx != op1_idx) {
- /* op1 must be killed and placed where out is */
- if (out_idx == 0) {
- /* best case, simple remove and rename */
- x87_patch_insn(n, op_ia32_Pop);
- attr = get_ia32_attr(n);
- attr->x87[0] = op1 = &ia32_st_regs[0];
-
- x87_pop(state);
- x87_set_st(state, arch_register_get_index(out), n, op1_idx - 1);
- }
- else {
- /* move op1 to tos, store and pop it */
- if (op1_idx != 0) {
- x87_create_fxch(state, n, op1_idx, 0);
- op1_idx = 0;
- }
- x87_patch_insn(n, op_ia32_Pop);
- attr = get_ia32_attr(n);
- attr->x87[0] = op1 = &ia32_st_regs[out_idx];
+ 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 or recreate constants */
+ node = create_Copy(state, n);
- x87_pop(state);
- x87_set_st(state, arch_register_get_index(out), n, out_idx - 1);
+ next = sched_next(n);
+ sched_remove(n);
+ exchange(n, node);
+ sched_add_before(next, node);
+ DB((dbg, LEVEL_1, "<<< %+F %s -> %s\n", node, op1->name,
+ arch_get_irn_register(sim->arch_env, node)->name));
+ } else {
+ out_idx = x87_on_stack(state, arch_register_get_index(out));
+
+ if (out_idx >= 0 && out_idx != op1_idx) {
+ /* Matze: out already on stack? how can this happen? */
+ assert(0);
+
+ /* op1 must be killed and placed where out is */
+ if (out_idx == 0) {
+ /* best case, simple remove and rename */
+ x87_patch_insn(n, op_ia32_Pop);
+ attr = get_ia32_attr(n);
+ attr->x87[0] = op1 = &ia32_st_regs[0];
+
+ x87_pop(state);
+ x87_set_st(state, arch_register_get_index(out), n, op1_idx - 1);
+ } else {
+ /* move op1 to tos, store and pop it */
+ if (op1_idx != 0) {
+ x87_create_fxch(state, n, op1_idx, 0);
+ op1_idx = 0;
}
- DB((dbg, LEVEL_1, ">>> %s %s\n", get_irn_opname(n), op1->name));
- }
- else {
- /* just a virtual copy */
- x87_set_st(state, arch_register_get_index(out), get_unop_op(n), op1_idx);
- sched_remove(n);
- DB((dbg, LEVEL_1, ">>> KILLED %s\n", get_irn_opname(n)));
- exchange(n, get_unop_op(n));
+ x87_patch_insn(n, op_ia32_Pop);
+ attr = get_ia32_attr(n);
+ attr->x87[0] = op1 = &ia32_st_regs[out_idx];
+
+ x87_pop(state);
+ x87_set_st(state, arch_register_get_index(out), n, out_idx - 1);
}
+ DB((dbg, LEVEL_1, "<<< %+F %s\n", n, op1->name));
+ } else {
+ /* just a virtual copy */
+ x87_set_st(state, arch_register_get_index(out), get_unop_op(n), op1_idx);
+ sched_remove(n);
+ DB((dbg, LEVEL_1, "<<< KILLED %s\n", get_irn_opname(n)));
+ exchange(n, get_unop_op(n));
}
}
return 0;
-}
+} /* sim_Copy */
/**
* Simulate a be_Call.
*
- * @param state the x87 state
- * @param n the node that should be simulated
- * @param env the architecture environment
+ * @param state the x87 state
+ * @param n the node that should be simulated
+ * @param arch_env the architecture environment
*/
-static int sim_Call(x87_state *state, ir_node *n, const arch_env_t *env) {
+static int sim_Call(x87_state *state, ir_node *n, const arch_env_t *arch_env) {
ir_type *call_tp = be_Call_get_type(n);
/* at the begin of a call the x87 state should be empty */
}
return 0;
-}
+} /* sim_Call */
/**
* Simulate a be_Spill.
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
*
* 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 */
/**
* Simulate a be_Reload.
*
* @param state the x87 state
* @param n the node that should be simulated (and patched)
- * @param env the architecture environment
*
* 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 */
/**
* Simulate a be_Return.
*
* @param state the x87 state
* @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;
x87_pop(state);
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 */
/**
* Run a simulation and fix all virtual instructions for a block.
* @return non-zero if simulation is complete,
* zero if the simulation must be rerun
*/
-static int x87_simulate_block(x87_simulator *sim, ir_node *block) {
+static void x87_simulate_block(x87_simulator *sim, ir_node *block) {
ir_node *n, *next;
blk_state *bl_state = x87_get_bl_state(sim, block);
x87_state *state = bl_state->begin;
const ir_edge_t *edge;
ir_node *start_block;
- /* if we have no assigned start state, we must wait ... */
- if (! state)
- return 0;
+ assert(state != NULL);
+ // already processed?
+ if(bl_state->end != NULL)
+ return;
- assert(bl_state->end == NULL);
+ //update_liveness(sim, block);
DB((dbg, LEVEL_1, "Simulate %+F\n", block));
+ DB((dbg, LEVEL_2, "State at Block begin:\n "));
+ DEBUG_ONLY(x87_dump_stack(state));
/* at block begin, kill all dead registers */
state = x87_kill_deads(sim, block, state);
- /* beware, n might changed */
+ /* beware, n might change */
for (n = sched_first(block); !sched_is_end(n); n = next) {
+ int node_inserted;
+ sim_func func;
ir_op *op = get_irn_op(n);
next = sched_next(n);
- if (op->ops.generic) {
- int node_inserted;
- sim_func func = (sim_func)op->ops.generic;
-
- /* have work to do */
- if (state == bl_state->begin) {
- /* create a new state, will be changed */
- state = x87_clone_state(sim, state);
- }
+ if (op->ops.generic == NULL)
+ continue;
- /* simulate it */
- node_inserted = (*func)(state, n, sim->env);
+ func = (sim_func)op->ops.generic;
- /*
- sim_func might have added additional nodes after n,
- so update next node
- beware: n must not be changed by sim_func
- (i.e. removed from schedule) in this case
- */
- if (node_inserted)
- next = sched_next(n);
+ /* have work to do */
+ if (state == bl_state->begin) {
+ /* create a new state, will be changed */
+ state = x87_clone_state(sim, state);
}
+
+ /* simulate it */
+ node_inserted = (*func)(state, n);
+
+ /*
+ sim_func might have added additional nodes after n,
+ so update next node
+ beware: n must not be changed by sim_func
+ (i.e. removed from schedule) in this case
+ */
+ if (node_inserted)
+ next = sched_next(n);
}
start_block = get_irg_start_block(get_irn_irg(block));
/* check if the state must be shuffled */
foreach_block_succ(block, edge) {
ir_node *succ = get_edge_src_irn(edge);
- blk_state *succ_state = x87_get_bl_state(sim, succ);
+ blk_state *succ_state;
+
+ if(succ == start_block)
+ continue;
+
+ succ_state = x87_get_bl_state(sim, succ);
- if (succ_state->begin && succ != start_block) {
- /* There is already a begin state for this block, bad.
+ if (succ_state->begin == NULL) {
+ succ_state->begin = state;
+ waitq_put(sim->worklist, succ);
+ } else {
+ /* There is already a begin state for the successor, bad.
Do the necessary permutations.
Note that critical edges are removed, so this is always possible. */
x87_shuffle(sim, block, state, succ, succ_state->begin);
-
- /* Note further, that there can be only one such situation,
- so we can break here. */
- break;
}
}
bl_state->end = state;
- /* now propagate the state to all successor blocks */
- foreach_block_succ(block, edge) {
- ir_node *succ = get_edge_src_irn(edge);
- blk_state *succ_state = x87_get_bl_state(sim, succ);
-
- if (! succ_state->begin)
- succ_state->begin = state;
- }
-
- return 1;
-}
+ DB((dbg, LEVEL_2, "State at Block end:\n ")); DEBUG_ONLY(x87_dump_stack(state));
+} /* x87_simulate_block */
/**
* Create a new x87 simulator.
*
- * @param sim a simulator handle, will be initialized
- * @param irg the current graph
- * @param env the architecture environment
+ * @param sim a simulator handle, will be initialized
+ * @param irg the current graph
+ * @param arch_env the architecture environment
*/
-static void x87_init_simulator(x87_simulator *sim, ir_graph *irg, const arch_env_t *env) {
+static void x87_init_simulator(x87_simulator *sim, ir_graph *irg,
+ const arch_env_t *arch_env)
+{
obstack_init(&sim->obst);
sim->blk_states = pmap_create();
- sim->env = env;
- sim->lv = be_liveness(irg);
+ sim->arch_env = arch_env;
+ sim->n_idx = get_irg_last_idx(irg);
+ sim->live = obstack_alloc(&sim->obst, sizeof(*sim->live) * sim->n_idx);
FIRM_DBG_REGISTER(dbg, "firm.be.ia32.x87");
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)
#define ASSOC_IA32(op) (op_ia32_v ## op)->ops.generic = (op_func)(sim_##op)
#define ASSOC_BE(op) (op_be_ ## op)->ops.generic = (op_func)(sim_##op)
- ASSOC_IA32(fConst);
ASSOC_IA32(fld);
ASSOC_IA32(fild);
ASSOC_IA32(fld1);
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
#undef ASSOC
-}
+} /* x87_init_simulator */
/**
* Destroy a x87 simulator.
static void x87_destroy_simulator(x87_simulator *sim) {
pmap_destroy(sim->blk_states);
obstack_free(&sim->obst, NULL);
- be_liveness_free(sim->lv);
DB((dbg, LEVEL_1, "x87 Simulator stopped\n\n"));
+} /* x87_destroy_simulator */
+
+static void update_liveness_walker(ir_node *block, void *data)
+{
+ x87_simulator *sim = data;
+ update_liveness(sim, block);
}
/**
*
* @param env the architecture environment
* @param irg the current graph
- * @param blk_list the block schedule list
*
* 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;
- pdeq *worklist;
- blk_state *bl_state;
+void x87_simulate_graph(const arch_env_t *arch_env, be_irg_t *birg) {
+ ir_node *block, *start_block;
+ blk_state *bl_state;
x87_simulator sim;
- int i;
+ ir_graph *irg = birg->irg;
/* create the simulator */
- x87_init_simulator(&sim, irg, env);
+ x87_init_simulator(&sim, irg, arch_env);
start_block = get_irg_start_block(irg);
bl_state = x87_get_bl_state(&sim, start_block);
bl_state->begin = empty;
empty->sim = ∼
- worklist = new_pdeq();
+ sim.worklist = new_waitq();
+ waitq_put(sim.worklist, start_block);
- /* create the worklist for the schedule. */
- for (i = 0; i < ARR_LEN(blk_list); ++i)
- pdeq_putr(worklist, blk_list[i]);
+ be_invalidate_liveness(birg);
+ be_assure_liveness(birg);
+ sim.lv = birg->lv;
+
+ /* Calculate the liveness for all nodes. We must precalculate this info,
+ * because the simulator adds new nodes (possible before Phi nodes) which
+ * would let a lazy calculation fail.
+ * On the other hand we reduce the computation amount due to
+ * precaching from O(n^2) to O(n) at the expense of O(n) cache memory.
+ */
+ irg_block_walk_graph(irg, update_liveness_walker, NULL, &sim);
/* iterate */
do {
- block = pdeq_getl(worklist);
- if (! x87_simulate_block(&sim, block)) {
- pdeq_putr(worklist, block);
- continue;
- }
- } while (! pdeq_empty(worklist));
+ block = waitq_get(sim.worklist);
+ x87_simulate_block(&sim, block);
+ } while (! pdeq_empty(sim.worklist));
/* kill it */
+ del_waitq(sim.worklist);
x87_destroy_simulator(&sim);
-}
+} /* x87_simulate_graph */