-/**
- * This file implements the x87 support and virtual to stack
- * register translation.
+/*
+ * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
*
- * $Id$
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
*/
+
+/**
+ * @file
+ * @brief This file implements the x87 support and virtual to stack
+ * register translation for the ia32 backend.
+ * @author Michael Beck
+ * @version $Id$
+ */
+#include "config.h"
+
#include <assert.h>
#include "irnode_t.h"
#include "irop_t.h"
#include "irprog.h"
#include "iredges_t.h"
+#include "irgmod.h"
+#include "ircons.h"
+#include "irgwalk.h"
#include "obst.h"
#include "pmap.h"
+#include "array_t.h"
#include "pdeq.h"
#include "irprintf.h"
+#include "debug.h"
+#include "error.h"
-#include "..\belive.h"
-#include "..\besched.h"
+#include "../belive_t.h"
+#include "../besched_t.h"
+#include "../benode_t.h"
+#include "bearch_ia32_t.h"
#include "ia32_new_nodes.h"
#include "gen_ia32_new_nodes.h"
#include "gen_ia32_regalloc_if.h"
#include "ia32_x87.h"
+#include "ia32_architecture.h"
#define N_x87_REGS 8
-/* first and second binop index */
-#define BINOP_IDX_1 2
-#define BINOP_IDX_2 3
-
/* the unop index */
-#define UNOP_IDX 0
+#define UNOP_IDX 0
#define MASK_TOS(x) ((x) & (N_x87_REGS - 1))
-/** the virtual floating point flag */
-#define irop_flag_vp (irop_flag_machine << 1)
+/** the debug handle */
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
+
+/* Forward declaration. */
+typedef struct _x87_simulator x87_simulator;
/**
* An exchange template.
* Note that our virtual functions have the same inputs
- * ant attributes as the real ones, so we can simple exchange
+ * and attributes as the real ones, so we can simple exchange
* their opcodes!
* Further, x87 supports inverse instructions, so we can handle them.
*/
typedef struct _exchange_tmpl {
- ir_op *normal_op; /**< the normal one */
- ir_op *reverse_op; /**< the reverse one if exists */
+ ir_op *normal_op; /**< the normal one */
+ ir_op *reverse_op; /**< the reverse one if exists */
+ ir_op *normal_pop_op; /**< the normal one with tos pop */
+ ir_op *reverse_pop_op; /**< the reverse one with tos pop */
} exchange_tmpl;
+/**
+ * An entry on the simulated x87 stack.
+ */
+typedef struct _st_entry {
+ int reg_idx; /**< the virtual register index of this stack value */
+ ir_node *node; /**< the node that produced this value */
+} st_entry;
+
/**
* The x87 state.
*/
typedef struct _x87_state {
- const arch_register_t *st[N_x87_REGS]; /**< the register stack */
- int depth; /**< the current stack depth */
- int tos; /**< position of the tos */
+ st_entry st[N_x87_REGS]; /**< the register stack */
+ int depth; /**< the current stack depth */
+ int tos; /**< position of the tos */
+ x87_simulator *sim; /**< The simulator. */
} x87_state;
/** An empty state, used for blocks without fp instructions. */
-static const x87_state _empty = { {0}, 0, 0 };
+static x87_state _empty = { { {0, NULL}, }, 0, 0, NULL };
static x87_state *empty = (x87_state *)&_empty;
-/** The type of an instruction simulator */
-typedef void (*sim_func)(x87_state *state, ir_node *n, const arch_env_t *env);
+enum {
+ NO_NODE_ADDED = 0, /**< No node was added. */
+ NODE_ADDED = 1 /**< A node was added by the simulator in the schedule. */
+};
+
+/**
+ * The type of an instruction simulator function.
+ *
+ * @param state the x87 state
+ * @param n the node to be simulated
+ *
+ * @return NODE_ADDED if a node was added AFTER n in schedule,
+ * NO_NODE_ADDED else
+ */
+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.
*/
-typedef 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 */
-} x87_simulator;
+struct _x87_simulator {
+ struct obstack obst; /**< An obstack for fast allocating. */
+ pmap *blk_states; /**< Map blocks to states. */
+ be_lv_t *lv; /**< intrablock liveness. */
+ vfp_liveness *live; /**< Liveness information. */
+ unsigned n_idx; /**< The cached get_irg_last_idx() result. */
+ waitq *worklist; /**< Worklist of blocks that must be processed. */
+ ia32_isa_t *isa; /**< the ISA object */
+};
/**
- * Check if the state is empty.
+ * Returns the current stack depth.
+ *
+ * @param state the x87 state
+ *
+ * @return the x87 stack depth
*/
-static int x87_state_is_empty(const x87_state *state) {
- return state->depth == 0;
-}
+static int x87_get_depth(const x87_state *state)
+{
+ return state->depth;
+} /* x87_get_depth */
+
+/**
+ * Return the virtual register index at st(pos).
+ *
+ * @param state the x87 state
+ * @param pos a stack position
+ *
+ * @return the vfp register index that produced the value at st(pos)
+ */
+static int x87_get_st_reg(const x87_state *state, int pos)
+{
+ assert(pos < state->depth);
+ return state->st[MASK_TOS(state->tos + pos)].reg_idx;
+} /* x87_get_st_reg */
+#ifdef DEBUG_libfirm
/**
- * Return the virtual register at st(pos).
+ * Return the node at st(pos).
+ *
+ * @param state the x87 state
+ * @param pos a stack position
+ *
+ * @return the IR node that produced the value at st(pos)
*/
-static const arch_register_t *x87_get_reg(const x87_state *state, int 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)];
-}
+ return state->st[MASK_TOS(state->tos + pos)].node;
+} /* x87_get_st_node */
/**
* Dump the stack for debugging.
+ *
+ * @param state the x87 state
*/
-static void x87_dump_stack(const x87_state *state) {
+static void x87_dump_stack(const x87_state *state)
+{
int i;
for (i = state->depth - 1; i >= 0; --i) {
- const arch_register_t *vreg = x87_get_reg(state, i);
- ir_printf("%s ", vreg->name);
+ DB((dbg, LEVEL_2, "vf%d(%+F) ", x87_get_st_reg(state, i),
+ x87_get_st_node(state, i)));
}
- ir_printf("<-- TOS\n");
-}
+ DB((dbg, LEVEL_2, "<-- TOS\n"));
+} /* x87_dump_stack */
+#endif /* DEBUG_libfirm */
/**
- * Set the tos virtual register
+ * Set a virtual register to st(pos).
+ *
+ * @param state the x87 state
+ * @param reg_idx the vfp register index that should be set
+ * @param node the IR node that produces the value of the vfp register
+ * @param pos the stack position where the new value should be entered
*/
-static void x87_set_tos_reg(x87_state *state, const arch_register_t *vreg) {
+static void x87_set_st(x87_state *state, int reg_idx, ir_node *node, int pos)
+{
assert(0 < state->depth);
- state->st[MASK_TOS(state->tos)] = vreg;
+ state->st[MASK_TOS(state->tos + pos)].reg_idx = reg_idx;
+ state->st[MASK_TOS(state->tos + pos)].node = node;
- printf("After INSTR:\n "); x87_dump_stack(state);
-}
+ DB((dbg, LEVEL_2, "After SET_REG: "));
+ DEBUG_ONLY(x87_dump_stack(state));
+} /* x87_set_st */
/**
- * Flush the x87 stack.
+ * Set the tos virtual register.
+ *
+ * @param state the x87 state
+ * @param reg_idx the vfp register index that should be set
+ * @param node the IR node that produces the value of the vfp register
*/
-static void x87_flush(x87_state *state) {
- state->depth = 0;
- state->tos = 0;
-}
+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 */
/**
* Swap st(0) with st(pos).
+ *
+ * @param state the x87 state
+ * @param pos the stack position to change the tos with
*/
-static void x87_fxch(x87_state *state, int pos) {
- const arch_register_t *vreg;
+static void x87_fxch(x87_state *state, int pos)
+{
+ st_entry entry;
assert(pos < state->depth);
- vreg = state->st[MASK_TOS(state->tos + pos)];
+ entry = state->st[MASK_TOS(state->tos + pos)];
state->st[MASK_TOS(state->tos + pos)] = state->st[MASK_TOS(state->tos)];
- state->st[MASK_TOS(state->tos)] = vreg;
+ state->st[MASK_TOS(state->tos)] = entry;
- printf("After FXCH:\n "); 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.
- * Return -1 if the virtual register was not found.
+ *
+ * @param state the x87 state
+ * @param reg_idx the register vfp index
+ *
+ * @return the stack position where the register is stacked
+ * or -1 if the virtual register was not found
*/
-static int x87_on_stack(const x87_state *state, const arch_register_t *vreg) {
+static int x87_on_stack(const x87_state *state, int reg_idx)
+{
int i, tos = state->tos;
for (i = 0; i < state->depth; ++i)
- if (state->st[MASK_TOS(tos + i)] == vreg)
+ 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.
+ * Push a virtual Register onto the stack, double pushed allowed.
+ *
+ * @param state the x87 state
+ * @param reg_idx the register vfp index
+ * @param node the node that produces the value of the vfp register
*/
-static void x87_push(x87_state *state, const arch_register_t *vreg) {
-// assert(x87_on_stack(state, vreg) == -1 && "double push");
+static void x87_push_dbl(x87_state *state, int reg_idx, ir_node *node)
+{
assert(state->depth < N_x87_REGS && "stack overrun");
++state->depth;
state->tos = MASK_TOS(state->tos - 1);
- state->st[state->tos] = vreg;
+ state->st[state->tos].reg_idx = reg_idx;
+ state->st[state->tos].node = node;
- printf("After PUSH:\n "); 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.
+ *
+ * @param state the x87 state
+ * @param reg_idx the register vfp index
+ * @param node the node that produces the value of the vfp register
+ * @param dbl_push if != 0 double pushes are allowed
+ */
+static void x87_push(x87_state *state, int reg_idx, ir_node *node)
+{
+ 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.
+ *
+ * @param state the x87 state
*/
-static void x87_pop(x87_state *state) {
+static void x87_pop(x87_state *state)
+{
assert(state->depth > 0 && "stack underrun");
--state->depth;
state->tos = MASK_TOS(state->tos + 1);
- printf("After POP:\n "); x87_dump_stack(state);
+ DB((dbg, LEVEL_2, "After POP: ")); DEBUG_ONLY(x87_dump_stack(state));
+} /* x87_pop */
+
+/**
+ * Empty the fpu stack
+ *
+ * @param state the x87 state
+ */
+static void x87_emms(x87_state *state)
+{
+ state->depth = 0;
+ state->tos = 0;
}
/**
* Returns the block state of a block.
+ *
+ * @param sim the x87 simulator handle
+ * @param block the current block
+ *
+ * @return the block state
*/
-static blk_state *x87_get_bl_state(x87_simulator *sim, ir_node *block) {
+static blk_state *x87_get_bl_state(x87_simulator *sim, ir_node *block)
+{
pmap_entry *entry = pmap_find(sim->blk_states, block);
if (! entry) {
return bl_state;
}
- return entry ? PTR_TO_BLKSTATE(entry->value) : NULL;
-}
+ return PTR_TO_BLKSTATE(entry->value);
+} /* x87_get_bl_state */
/**
- * Create a new x87 state.
+ * Creates a new x87 state.
+ *
+ * @param sim the x87 simulator handle
+ *
+ * @return a new x87 state
*/
-static x87_state *x87_alloc_state(x87_simulator *sim) {
+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 */
/**
- * Create a new empty x87 state.
+ * Clone a x87 state.
+ *
+ * @param sim the x87 simulator handle
+ * @param src the x87 state that will be cloned
+ *
+ * @return a cloned copy of the src state
*/
-static x87_state *x87_alloc_empty_state(x87_simulator *sim) {
+static x87_state *x87_clone_state(x87_simulator *sim, const x87_state *src)
+{
x87_state *res = x87_alloc_state(sim);
- x87_flush(res);
+ memcpy(res, src, sizeof(*res));
return res;
-}
+} /* x87_clone_state */
/**
- * Clone a x87 state.
+ * Patch a virtual instruction into a x87 one and return
+ * the node representing the result value.
+ *
+ * @param n the IR node to patch
+ * @param op the x87 opcode to patch in
*/
-static x87_state *x87_clone_state(x87_simulator *sim, const x87_state *src) {
- x87_state *res = x87_alloc_state(sim);
+static ir_node *x87_patch_insn(ir_node *n, ir_op *op)
+{
+ ir_mode *mode = get_irn_mode(n);
+ ir_node *res = n;
+
+ set_irn_op(n, op);
+
+ if (mode == mode_T) {
+ /* patch all Proj's */
+ const ir_edge_t *edge;
+
+ foreach_out_edge(n, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ if (is_Proj(proj)) {
+ mode = get_irn_mode(proj);
+ if (mode_is_float(mode)) {
+ res = proj;
+ set_irn_mode(proj, mode_E);
+ }
+ }
+ }
+ } else if (mode_is_float(mode))
+ set_irn_mode(n, mode_E);
+ return res;
+} /* x87_patch_insn */
- memcpy(res, src, sizeof(*res));
+/**
+ * Returns the first Proj of a mode_T node having a given mode.
+ *
+ * @param n the mode_T node
+ * @param m the desired mode of the Proj
+ * @return The first Proj of mode @p m found or NULL.
+ */
+static ir_node *get_irn_Proj_for_mode(ir_node *n, ir_mode *m)
+{
+ const ir_edge_t *edge;
+
+ assert(get_irn_mode(n) == mode_T && "Need mode_T node");
+
+ foreach_out_edge(n, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ if (get_irn_mode(proj) == m)
+ return proj;
+ }
+
+ 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(const ir_node *irn)
+{
+ const arch_register_t *res = arch_get_irn_register(irn);
+
+ assert(res->reg_class->regs == ia32_vfp_regs);
return res;
-}
+} /* x87_get_irn_register */
+
+/* -------------- x87 perm --------------- */
+
+/**
+ * Creates a fxch for shuffle.
+ *
+ * @param state the x87 state
+ * @param pos parameter for fxch
+ * @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 *fxch;
+ ia32_x87_attr_t *attr;
+
+ fxch = new_rd_ia32_fxch(NULL, get_irn_irg(block), block);
+ attr = get_ia32_x87_attr(fxch);
+ attr->x87[0] = &ia32_st_regs[pos];
+ attr->x87[2] = &ia32_st_regs[0];
+
+ keep_alive(fxch);
+
+ x87_fxch(state, pos);
+ return fxch;
+} /* x87_fxch_shuffle */
/**
* Calculate the necessary permutations to reach dst_state.
+ *
+ * These permutations are done with fxch instructions and placed
+ * at the end of the block.
+ *
+ * Note that critical edges are removed here, so we need only
+ * a shuffle if the current block has only one successor.
+ *
+ * @param sim the simulator handle
+ * @param block the current block
+ * @param state the current x87 stack state, might be modified
+ * @param dst_block the destination block
+ * @param dst_state destination state
+ *
+ * @return state
*/
-static x87_state *x87_shuffle(x87_simulator *sim, ir_node *block, x87_state *state, const x87_state *dst_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;
+ unsigned cycles[4], all_mask;
+ char cycle_idx[4][8];
+ ir_node *fxch, *before, *after;
+ (void) sim;
+ (void) dst_block;
+
assert(state->depth == dst_state->depth);
- if (state == empty)
- state = x87_clone_state(sim, state);
+ /* Some mathematics here:
+ If we have a cycle of length n that includes the tos,
+ we need n-1 exchange operations.
+ We can always add the tos and restore it, so we need
+ n+1 exchange operations for a cycle not containing the tos.
+ So, the maximum of needed operations is for a cycle of 7
+ not including the tos == 8.
+ This is the same number of ops we would need for using stores,
+ so exchange is cheaper (we save the loads).
+ On the other hand, we might need an additional exchange
+ in the next block to bring one operand on top, so the
+ number of ops in the first case is identical.
+ Further, no more than 4 cycles can exists (4 x 2).
+ */
+ all_mask = (1 << (state->depth)) - 1;
+
+ for (n_cycles = 0; all_mask; ++n_cycles) {
+ int src_idx, dst_idx;
+
+ /* find the first free slot */
+ for (i = 0; i < state->depth; ++i) {
+ if (all_mask & (1 << i)) {
+ all_mask &= ~(1 << i);
+
+ /* check if there are differences here */
+ if (x87_get_st_reg(state, i) != x87_get_st_reg(dst_state, i))
+ break;
+ }
+ }
+
+ if (! all_mask) {
+ /* no more cycles found */
+ break;
+ }
+
+ k = 0;
+ cycles[n_cycles] = (1 << i);
+ cycle_idx[n_cycles][k++] = i;
+ for (src_idx = i; ; src_idx = dst_idx) {
+ dst_idx = x87_on_stack(dst_state, x87_get_st_reg(state, src_idx));
+
+ if ((all_mask & (1 << dst_idx)) == 0)
+ break;
+
+ cycle_idx[n_cycles][k++] = dst_idx;
+ cycles[n_cycles] |= (1 << dst_idx);
+ all_mask &= ~(1 << dst_idx);
+ }
+ cycle_idx[n_cycles][k] = -1;
+ }
+
+ if (n_cycles <= 0) {
+ /* no permutation needed */
+ return state;
+ }
+
+ /* Hmm: permutation needed */
+ DB((dbg, LEVEL_2, "\n%+F needs permutation: from\n", block));
+ DEBUG_ONLY(x87_dump_stack(state));
+ DB((dbg, LEVEL_2, " to\n"));
+ DEBUG_ONLY(x87_dump_stack(dst_state));
+
+
+#ifdef DEBUG_libfirm
+ DB((dbg, LEVEL_2, "Need %d cycles\n", n_cycles));
+ for (ri = 0; ri < n_cycles; ++ri) {
+ DB((dbg, LEVEL_2, " Ring %d:\n ", ri));
+ for (k = 0; cycle_idx[ri][k] != -1; ++k)
+ DB((dbg, LEVEL_2, " st%d ->", cycle_idx[ri][k]));
+ DB((dbg, LEVEL_2, "\n"));
+ }
+#endif
+
+ after = NULL;
+
+ /*
+ * Find the place node must be insert.
+ * We have only one successor block, so the last instruction should
+ * be a jump.
+ */
+ before = sched_last(block);
+ assert(is_cfop(before));
+
+ /* now do the permutations */
+ 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);
+ if (after)
+ sched_add_after(after, fxch);
+ else
+ sched_add_before(before, fxch);
+ after = fxch;
+ }
+ for (k = 1; cycle_idx[ri][k] != -1; ++k) {
+ fxch = x87_fxch_shuffle(state, cycle_idx[ri][k], block);
+ if (after)
+ sched_add_after(after, fxch);
+ else
+ sched_add_before(before, fxch);
+ after = fxch;
+ }
+ if ((cycles[ri] & 1) == 0) {
+ /* this cycle does not include the tos */
+ fxch = x87_fxch_shuffle(state, cycle_idx[ri][0], block);
+ sched_add_after(after, fxch);
+ }
+ }
return state;
-}
+} /* x87_shuffle */
/**
- * Create a fxch before node n.
+ * Create a fxch node before another node.
+ *
+ * @param state the x87 state
+ * @param n the node after the fxch
+ * @param pos exchange st(pos) with st(0)
+ *
+ * @return the fxch
*/
-static void x87_create_fxch(x87_state *state, ir_node *n, int pos, int op_idx) {
- ir_node *fxch, *pred;
- ia32_attr_t *attr;
+static ir_node *x87_create_fxch(x87_state *state, ir_node *n, int pos)
+{
+ ir_node *fxch;
+ ia32_x87_attr_t *attr;
+ ir_graph *irg = get_irn_irg(n);
+ ir_node *block = get_nodes_block(n);
x87_fxch(state, pos);
- pred = get_irn_n(n, op_idx);
- fxch = new_rd_ia32_fxch(NULL, get_irn_irg(n), get_nodes_block(n), pred, get_irn_mode(pred));
- attr = get_ia32_attr(fxch);
+ fxch = new_rd_ia32_fxch(NULL, irg, block);
+ attr = get_ia32_x87_attr(fxch);
attr->x87[0] = &ia32_st_regs[pos];
attr->x87[2] = &ia32_st_regs[0];
- set_irn_n(n, op_idx, fxch);
+
+ keep_alive(fxch);
sched_add_before(n, fxch);
- printf("<<< %s %s, %s\n", get_irn_opname(fxch), attr->x87[0]->name, attr->x87[2]->name);
-}
+ 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 after the fpush
+ * @param pos push st(pos) on stack
+ * @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;
- ia32_attr_t *attr;
+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_x87_attr_t *attr;
+ const arch_register_t *out = x87_get_irn_register(pred);
- x87_push(state, arch_get_irn_register(env, n));
+ x87_push_dbl(state, arch_register_get_index(out), pred);
- pred = get_irn_n(n, op_idx);
- fpush = new_rd_ia32_fpush(NULL, get_irn_irg(n), get_nodes_block(n), pred, get_irn_mode(pred));
- attr = get_ia32_attr(fpush);
+ fpush = new_rd_ia32_fpush(NULL, get_irn_irg(n), get_nodes_block(n));
+ attr = get_ia32_x87_attr(fpush);
attr->x87[0] = &ia32_st_regs[pos];
attr->x87[2] = &ia32_st_regs[0];
- set_irn_n(n, op_idx, fpush);
+ keep_alive(fpush);
sched_add_before(n, fpush);
- printf("<<< %s %s, %s\n", get_irn_opname(fpush), attr->x87[0]->name, attr->x87[2]->name);
+
+ 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 after the fpop
+ * @param num pop 1 or 2 values
+ *
+ * @return the fpop node
+ */
+static ir_node *x87_create_fpop(x87_state *state, ir_node *n, int num)
+{
+ ir_node *fpop = NULL;
+ ia32_x87_attr_t *attr;
+
+ assert(num > 0);
+ while (num > 0) {
+ x87_pop(state);
+ if (ia32_cg_config.use_ffreep)
+ fpop = new_rd_ia32_ffreep(NULL, get_irn_irg(n), get_nodes_block(n));
+ else
+ fpop = new_rd_ia32_fpop(NULL, get_irn_irg(n), get_nodes_block(n));
+ attr = get_ia32_x87_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 ------------------------------------------ */
+
/**
- * Simulate a virtual binop
+ * 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 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 void sim_binop(x87_state *state, ir_node *n, const arch_env_t *env, const exchange_tmpl *tmpl) {
- int op1_idx, op2_idx = -1;
- int out_idx;
- 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);
- ia32_attr_t *attr = get_ia32_attr(n);
+static vfp_liveness vfp_liveness_transfer(ir_node *irn, vfp_liveness live)
+{
+ int i, n;
+ const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
- printf(">>> %s %s, %s -> %s\n", get_irn_opname(n), op1->name, op2->name, out->name);
+ if (get_irn_mode(irn) == mode_T) {
+ const ir_edge_t *edge;
- out_idx = x87_on_stack(state, out);
- op1_idx = x87_on_stack(state, op1);
+ foreach_out_edge(irn, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
- if (out_idx == op1_idx) {
- /* good, first operands match, bring it to top */
- if (out_idx != 0) {
- x87_create_fxch(state, n, out_idx, BINOP_IDX_1);
- op1_idx = 0;
- out_idx = 0;
+ if (arch_irn_consider_in_reg_alloc(cls, proj)) {
+ const arch_register_t *reg = x87_get_irn_register(proj);
+ live &= ~(1 << arch_register_get_index(reg));
+ }
}
+ }
- if (op2->index != REG_VFP_NOREG) {
- /* op2 is no address mode */
- op2_idx = x87_on_stack(state, op2);
+ if (arch_irn_consider_in_reg_alloc(cls, irn)) {
+ const arch_register_t *reg = x87_get_irn_register(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(cls, op)) {
+ const arch_register_t *reg = x87_get_irn_register(op);
+ live |= 1 << arch_register_get_index(reg);
}
- else
- op2_idx = REG_ST_NOREG;
+ }
+ return live;
+} /* vfp_liveness_transfer */
+
+/**
+ * Put all live virtual registers at the end of a block into a 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 vfp_liveness vfp_liveness_end_of_block(x87_simulator *sim, const ir_node *block)
+{
+ int i;
+ vfp_liveness live = 0;
+ const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
+ const be_lv_t *lv = sim->lv;
+
+ 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(cls, node))
+ continue;
- n->op = tmpl->normal_op;
+ reg = x87_get_irn_register(node);
+ live |= 1 << arch_register_get_index(reg);
}
- else {
- if (op2->index != REG_VFP_NOREG) {
- /* op2 is no address mode */
- op2_idx = x87_on_stack(state, op2);
+
+ 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 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_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] & ~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;
+
+ /* 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;
+
+ idx = get_irn_idx(irn);
+ sim->live[idx] = live;
+
+ live = vfp_liveness_transfer(irn, 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
+ */
+#define is_vfp_live(reg_idx, live) ((live) & (1 << (reg_idx)))
+
+#ifdef DEBUG_libfirm
+/**
+ * Dump liveness info.
+ *
+ * @param live the live bitset
+ */
+static void vfp_dump_live(vfp_liveness live)
+{
+ int i;
+
+ 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, "\n"));
+} /* vfp_dump_live */
+#endif /* DEBUG_libfirm */
- if (out_idx == op2_idx) {
- /* good, second operands match, bring it to top */
- if (out_idx != 0) {
- x87_create_fxch(state, n, out_idx, BINOP_IDX_2);
- op2_idx = 0;
- out_idx = 0;
+/* --------------------------------- simulators ---------------------------------------- */
+
+#define XCHG(a, b) do { int t = (a); (a) = (b); (b) = t; } while (0)
+
+/* Pseudocode:
+
+
+
+
+
+
+*/
+
+/**
+ * Simulate a virtual binop.
+ *
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ * @param tmpl the template containing the 4 possible x87 opcodes
+ *
+ * @return NO_NODE_ADDED
+ */
+static int sim_binop(x87_state *state, ir_node *n, const exchange_tmpl *tmpl)
+{
+ int op2_idx = 0, op1_idx;
+ int out_idx, do_pop = 0;
+ ia32_x87_attr_t *attr;
+ int permuted;
+ ir_node *patched_insn;
+ ir_op *dst;
+ x87_simulator *sim = state->sim;
+ ir_node *op1 = get_irn_n(n, n_ia32_binary_left);
+ ir_node *op2 = get_irn_n(n, n_ia32_binary_right);
+ const arch_register_t *op1_reg = x87_get_irn_register(op1);
+ const arch_register_t *op2_reg = x87_get_irn_register(op2);
+ const arch_register_t *out = x87_get_irn_register(n);
+ int reg_index_1 = arch_register_get_index(op1_reg);
+ int reg_index_2 = arch_register_get_index(op2_reg);
+ vfp_liveness live = vfp_live_args_after(sim, n, REGMASK(out));
+ int op1_live_after;
+ int op2_live_after;
+
+ DB((dbg, LEVEL_1, ">>> %+F %s, %s -> %s\n", n,
+ arch_register_get_name(op1_reg), arch_register_get_name(op2_reg),
+ arch_register_get_name(out)));
+ DEBUG_ONLY(vfp_dump_live(live));
+ DB((dbg, LEVEL_1, "Stack before: "));
+ DEBUG_ONLY(x87_dump_stack(state));
+
+ if (reg_index_1 == REG_VFP_UKNWN) {
+ op1_idx = 0;
+ op1_live_after = 1;
+ } else {
+ op1_idx = x87_on_stack(state, reg_index_1);
+ assert(op1_idx >= 0);
+ op1_live_after = is_vfp_live(arch_register_get_index(op1_reg), live);
+ }
+
+ attr = get_ia32_x87_attr(n);
+ permuted = attr->attr.data.ins_permuted;
+
+ if (reg_index_2 != REG_VFP_NOREG) {
+ assert(!permuted);
+
+ if (reg_index_2 == REG_VFP_UKNWN) {
+ op2_idx = 0;
+ op2_live_after = 1;
+ } else {
+ /* second operand is a vfp register */
+ op2_idx = x87_on_stack(state, reg_index_2);
+ assert(op2_idx >= 0);
+ op2_live_after
+ = is_vfp_live(arch_register_get_index(op2_reg), live);
+ }
+
+ if (op2_live_after) {
+ /* Second operand is live. */
- if (op1_idx == 0)
- op1_idx = out_idx;
+ if (op1_live_after) {
+ /* Both operands are live: push the first one.
+ This works even for op1 == op2. */
+ x87_create_fpush(state, n, op1_idx, n_ia32_binary_right);
+ /* now do fxxx (tos=tos X op) */
+ op1_idx = 0;
+ op2_idx += 1;
+ out_idx = 0;
+ dst = tmpl->normal_op;
+ } else {
+ /* Second live, first operand is dead here, bring it to tos. */
+ if (op1_idx != 0) {
+ x87_create_fxch(state, n, op1_idx);
+ if (op2_idx == 0)
+ op2_idx = op1_idx;
+ op1_idx = 0;
+ }
+ /* now do fxxx (tos=tos X op) */
+ out_idx = 0;
+ dst = tmpl->normal_op;
+ }
+ } else {
+ /* Second operand is dead. */
+ if (op1_live_after) {
+ /* First operand is live: bring second to tos. */
+ if (op2_idx != 0) {
+ x87_create_fxch(state, n, op2_idx);
+ if (op1_idx == 0)
+ op1_idx = op2_idx;
+ op2_idx = 0;
+ }
+ /* 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) {
+ if (op1_idx == 0) {
+ /* Both are identically and on tos, no pop needed. */
+ /* here fxxx (tos = tos X tos) */
+ dst = tmpl->normal_op;
+ 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 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);
+ if (op1_idx == op2_idx) {
+ /* Both are identically and on tos now, no pop needed. */
+ op1_idx = 0;
+ 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;
+ }
+ }
}
- n->op = tmpl->reverse_op;
}
- else {
- /* bad, both numbers did not match, push op1 */
- x87_push(state, op1);
- if (op2->index != REG_VFP_NOREG)
- ++op2_idx;
- else
- op2_idx = REG_ST_NOREG;
+ } else {
+ /* second operand is an address mode */
+ if (op1_live_after) {
+ /* first operand is live: push it here */
+ x87_create_fpush(state, n, op1_idx, n_ia32_binary_left);
op1_idx = 0;
- out_idx = 0;
-
- n->op = tmpl->normal_op;
+ } else {
+ /* first operand is dead: bring it to tos */
+ if (op1_idx != 0) {
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ }
}
+
+ /* use fxxx (tos = tos X mem) */
+ dst = permuted ? tmpl->reverse_op : tmpl->normal_op;
+ out_idx = 0;
+ }
+
+ 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->x87[0] = op1_reg = &ia32_st_regs[op1_idx];
+ if (reg_index_2 != REG_VFP_NOREG) {
+ attr->x87[1] = op2_reg = &ia32_st_regs[op2_idx];
}
- x87_set_tos_reg(state, out);
- attr->x87[0] = op1 = &ia32_st_regs[op1_idx];
- attr->x87[1] = op2 = &ia32_st_regs[op2_idx];
attr->x87[2] = out = &ia32_st_regs[out_idx];
- printf("<<< %s %s, %s -> %s\n", get_irn_opname(n), op1->name, op2->name, out->name);
-}
+ 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_reg), arch_register_get_name(op2_reg),
+ arch_register_get_name(out)));
+ } else {
+ DB((dbg, LEVEL_1, "<<< %s %s, [AM] -> %s\n", get_irn_opname(n),
+ arch_register_get_name(op1_reg),
+ arch_register_get_name(out)));
+ }
+
+ return NO_NODE_ADDED;
+} /* sim_binop */
/**
- * Simulate a virtual Unop
+ * Simulate a virtual Unop.
+ *
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ * @param op the x87 opcode that will replace n's opcode
+ *
+ * @return NO_NODE_ADDED
*/
-static void 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);
- ia32_attr_t *attr = get_ia32_attr(n);
+ x87_simulator *sim = state->sim;
+ const arch_register_t *op1 = x87_get_irn_register(get_irn_n(n, UNOP_IDX));
+ const arch_register_t *out = x87_get_irn_register(n);
+ ia32_x87_attr_t *attr;
+ unsigned live = vfp_live_args_after(sim, n, REGMASK(out));
- out_idx = x87_on_stack(state, out);
- op1_idx = x87_on_stack(state, op1);
+ DB((dbg, LEVEL_1, ">>> %+F -> %s\n", n, out->name));
+ DEBUG_ONLY(vfp_dump_live(live));
- if (out_idx == op1_idx) {
- /* good, operands match, bring it to top */
- if (out_idx != 0) {
- x87_create_fxch(state, n, out_idx, UNOP_IDX);
- op1_idx = 0;
- out_idx = 0;
- }
+ op1_idx = x87_on_stack(state, arch_register_get_index(op1));
+
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
+ /* push the operand here */
+ x87_create_fpush(state, n, op1_idx, UNOP_IDX);
+ op1_idx = 0;
}
else {
- /* create a push or an exchange here */
- if (1)
- x87_create_fpush(env, state, n, op1_idx, UNOP_IDX);
- else
- x87_create_fxch(state, n, op1_idx, UNOP_IDX);
+ /* operand is dead, bring it to tos */
+ if (op1_idx != 0) {
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ }
}
- printf(">>> %s -> %s\n", get_irn_opname(n), out->name);
- n->op = op;
+
+ x87_set_tos(state, arch_register_get_index(out), x87_patch_insn(n, op));
+ out_idx = 0;
+ attr = get_ia32_x87_attr(n);
attr->x87[0] = op1 = &ia32_st_regs[0];
attr->x87[2] = out = &ia32_st_regs[0];
- printf("<<< %s -> %s\n", get_irn_opname(n), out->name);
-}
+ DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), out->name));
+
+ return NO_NODE_ADDED;
+} /* sim_unop */
/**
- * Simulate a virtual Load
+ * Simulate a virtual Load instruction.
+ *
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ * @param op the x87 opcode that will replace n's opcode
+ *
+ * @return NO_NODE_ADDED
*/
-static void 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);
- ia32_attr_t *attr = get_ia32_attr(n);
-
- printf(">>> %s -> %s\n", get_irn_opname(n), out->name);
- x87_push(state, out);
- n->op = op;
+static int sim_load(x87_state *state, ir_node *n, ir_op *op)
+{
+ const arch_register_t *out = x87_get_irn_register(n);
+ ia32_x87_attr_t *attr;
+
+ 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(n));
+ attr = get_ia32_x87_attr(n);
attr->x87[2] = out = &ia32_st_regs[0];
- printf("<<< %s -> %s\n", get_irn_opname(n), out->name);
-}
+ DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), arch_register_get_name(out)));
+
+ return NO_NODE_ADDED;
+} /* sim_load */
+
+/**
+ * Rewire all users of @p old_val to @new_val iff they are scheduled after @p store.
+ *
+ * @param store The store
+ * @param old_val The former value
+ * @param new_val The new value
+ */
+static void collect_and_rewire_users(ir_node *store, ir_node *old_val, ir_node *new_val)
+{
+ const ir_edge_t *edge, *ne;
+
+ foreach_out_edge_safe(old_val, edge, ne) {
+ ir_node *user = get_edge_src_irn(edge);
+
+ if (! user || user == store)
+ continue;
+
+ /* if the user is scheduled after the store: rewire */
+ if (sched_is_scheduled(user) && sched_comes_after(store, user)) {
+ int i;
+ /* find the input of the user pointing to the old value */
+ for (i = get_irn_arity(user) - 1; i >= 0; i--) {
+ if (get_irn_n(user, i) == old_val)
+ set_irn_n(user, i, new_val);
+ }
+ }
+ }
+} /* collect_and_rewire_users */
+
+/**
+ * Simulate a virtual Store.
+ *
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ * @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, ir_op *op, ir_op *op_p)
+{
+ ir_node *val = get_irn_n(n, n_ia32_vfst_val);
+ const arch_register_t *op2 = x87_get_irn_register(val);
+ unsigned live = vfp_live_args_after(state->sim, n, 0);
+ int insn = NO_NODE_ADDED;
+ ia32_x87_attr_t *attr;
+ int op2_reg_idx, op2_idx, depth;
+ int live_after_node;
+ ir_mode *mode;
+
+ 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 immediately */
+ 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);
+ DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
+ assert(op2_idx >= 0);
+ }
+
+ mode = get_ia32_ls_mode(n);
+ depth = x87_get_depth(state);
+
+ if (live_after_node) {
+ /*
+ Problem: fst doesn't support mode_E (spills), only fstp does
+ Solution:
+ - stack not full: push value and fstp
+ - stack full: fstp value and load again
+ Note that we cannot test on mode_E, because floats might be 96bit ...
+ */
+ if (get_mode_size_bits(mode) > 64 || mode == mode_Ls) {
+ if (depth < N_x87_REGS) {
+ /* ok, we have a free register: push + fstp */
+ x87_create_fpush(state, n, op2_idx, n_ia32_vfst_val);
+ x87_pop(state);
+ x87_patch_insn(n, op_p);
+ } else {
+ ir_node *vfld, *mem, *block, *rproj, *mproj;
+ ir_graph *irg;
+
+ /* stack full here: need fstp + load */
+ x87_pop(state);
+ x87_patch_insn(n, op_p);
+
+ block = get_nodes_block(n);
+ irg = get_irn_irg(n);
+ vfld = new_rd_ia32_vfld(NULL, irg, block, get_irn_n(n, 0), get_irn_n(n, 1), new_rd_NoMem(irg), get_ia32_ls_mode(n));
+
+ /* copy all attributes */
+ set_ia32_frame_ent(vfld, get_ia32_frame_ent(n));
+ if (is_ia32_use_frame(n))
+ set_ia32_use_frame(vfld);
+ set_ia32_op_type(vfld, ia32_AddrModeS);
+ 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));
+
+ rproj = new_r_Proj(irg, block, vfld, get_ia32_ls_mode(vfld), pn_ia32_vfld_res);
+ mproj = new_r_Proj(irg, block, vfld, mode_M, pn_ia32_vfld_M);
+ mem = get_irn_Proj_for_mode(n, mode_M);
+
+ assert(mem && "Store memory not found");
+
+ arch_set_irn_register(rproj, op2);
+
+ /* reroute all former users of the store memory to the load memory */
+ edges_reroute(mem, mproj, irg);
+ /* set the memory input of the load to the store memory */
+ set_irn_n(vfld, n_ia32_vfld_mem, mem);
+
+ sched_add_after(n, vfld);
+ sched_add_after(vfld, rproj);
+
+ /* rewire all users, scheduled after the store, to the loaded value */
+ collect_and_rewire_users(n, val, rproj);
+
+ insn = NODE_ADDED;
+ }
+ } else {
+ /* we can only store the tos to memory */
+ if (op2_idx != 0)
+ x87_create_fxch(state, n, op2_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);
+
+ x87_pop(state);
+ x87_patch_insn(n, op_p);
+ }
+
+ attr = get_ia32_x87_attr(n);
+ attr->x87[1] = op2 = &ia32_st_regs[0];
+ DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2)));
+
+ return insn;
+} /* sim_store */
#define _GEN_BINOP(op, rev) \
-static void sim_##op(x87_state *state, ir_node *n, const arch_env_t *env) { \
- exchange_tmpl tmpl = { op_ia32_##op, op_ia32_##rev }; \
- sim_binop(state, n, env, &tmpl); \
+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, &tmpl); \
}
-#define GEN_BINOP(op) _GEN_BINOP(op, op)
+#define GEN_BINOP(op) _GEN_BINOP(op, op)
#define GEN_BINOPR(op) _GEN_BINOP(op, op##r)
#define GEN_LOAD2(op, nop) \
-static void sim_##op(x87_state *state, ir_node *n, const arch_env_t *env) { \
- 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 void sim_##op(x87_state *state, ir_node *n, const arch_env_t *env) { \
- 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) { \
+ return sim_store(state, n, op_ia32_##op, op_ia32_##op##p); \
+}
+/* all stubs */
GEN_BINOP(fadd)
GEN_BINOPR(fsub)
GEN_BINOP(fmul)
GEN_BINOPR(fdiv)
+GEN_BINOP(fprem)
+
+GEN_UNOP(fabs)
+GEN_UNOP(fchs)
GEN_LOAD(fld)
+GEN_LOAD(fild)
GEN_LOAD(fldz)
GEN_LOAD(fld1)
-GEN_LOAD2(fConst, fldConst)
-GEN_UNOP(fabs)
-GEN_UNOP(fchs)
-GEN_UNOP(fsin)
-GEN_UNOP(fcos)
-GEN_UNOP(fsqrt)
+GEN_STORE(fst)
+GEN_STORE(fist)
+
+/**
+* Simulate a virtual fisttp.
+*
+* @param state the x87 state
+* @param n the node that should be simulated (and patched)
+*/
+static int sim_fisttp(x87_state *state, ir_node *n)
+{
+ ir_node *val = get_irn_n(n, n_ia32_vfst_val);
+ const arch_register_t *op2 = x87_get_irn_register(val);
+ int insn = NO_NODE_ADDED;
+ ia32_x87_attr_t *attr;
+ int op2_reg_idx, op2_idx, depth;
+
+ 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);
+ } else {
+ /* produce a new value which we will consume immediately */
+ x87_create_fldz(state, n, op2_reg_idx);
+ op2_idx = x87_on_stack(state, op2_reg_idx);
+ assert(op2_idx >= 0);
+ }
+ } else {
+ op2_idx = x87_on_stack(state, op2_reg_idx);
+ DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
+ assert(op2_idx >= 0);
+ }
+
+ depth = x87_get_depth(state);
+
+ /* Note: although the value is still live here, it is destroyed because
+ of the pop. The register allocator is aware of that and introduced a copy
+ if the value must be alive. */
+
+ /* we can only store the tos to memory */
+ if (op2_idx != 0)
+ x87_create_fxch(state, n, op2_idx);
+
+ x87_pop(state);
+ x87_patch_insn(n, op_ia32_fisttp);
+
+ attr = get_ia32_x87_attr(n);
+ attr->x87[1] = op2 = &ia32_st_regs[0];
+ DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2)));
+
+ return insn;
+} /* sim_fisttp */
+
+static int sim_FtstFnstsw(x87_state *state, ir_node *n)
+{
+ x87_simulator *sim = state->sim;
+ ia32_x87_attr_t *attr = get_ia32_x87_attr(n);
+ ir_node *op1_node = get_irn_n(n, n_ia32_vFtstFnstsw_left);
+ const arch_register_t *reg1 = x87_get_irn_register(op1_node);
+ int reg_index_1 = arch_register_get_index(reg1);
+ int op1_idx = x87_on_stack(state, reg_index_1);
+ unsigned live = vfp_live_args_after(sim, n, 0);
+
+ DB((dbg, LEVEL_1, ">>> %+F %s\n", n, arch_register_get_name(reg1)));
+ DEBUG_ONLY(vfp_dump_live(live));
+ DB((dbg, LEVEL_1, "Stack before: "));
+ DEBUG_ONLY(x87_dump_stack(state));
+ assert(op1_idx >= 0);
+
+ if (op1_idx != 0) {
+ /* bring the value to tos */
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ }
+
+ /* patch the operation */
+ x87_patch_insn(n, op_ia32_FtstFnstsw);
+ reg1 = &ia32_st_regs[op1_idx];
+ attr->x87[0] = reg1;
+ attr->x87[1] = NULL;
+ attr->x87[2] = NULL;
+
+ if (!is_vfp_live(reg_index_1, live)) {
+ x87_create_fpop(state, sched_next(n), 1);
+ return NODE_ADDED;
+ }
+
+ return NO_NODE_ADDED;
+}
+
+/**
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ */
+static int sim_Fucom(x87_state *state, ir_node *n)
+{
+ int op1_idx;
+ int op2_idx = -1;
+ ia32_x87_attr_t *attr = get_ia32_x87_attr(n);
+ ir_op *dst;
+ x87_simulator *sim = state->sim;
+ ir_node *op1_node = get_irn_n(n, n_ia32_vFucomFnstsw_left);
+ ir_node *op2_node = get_irn_n(n, n_ia32_vFucomFnstsw_right);
+ const arch_register_t *op1 = x87_get_irn_register(op1_node);
+ const arch_register_t *op2 = x87_get_irn_register(op2_node);
+ 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);
+ int permuted = attr->attr.data.ins_permuted;
+ int xchg = 0;
+ int pops = 0;
+ int node_added = NO_NODE_ADDED;
+
+ 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, reg_index_1);
+ assert(op1_idx >= 0);
+
+ /* BEWARE: check for comp a,a cases, they might happen */
+ 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(reg_index_2, live)) {
+ /* second operand is live */
+
+ if (is_vfp_live(reg_index_1, live)) {
+ /* both operands are live */
+
+ if (op1_idx == 0) {
+ /* res = tos X op */
+ } else if (op2_idx == 0) {
+ /* res = op X tos */
+ permuted = !permuted;
+ xchg = 1;
+ } else {
+ /* bring the first one to tos */
+ x87_create_fxch(state, n, op1_idx);
+ if (op2_idx == 0)
+ op2_idx = op1_idx;
+ op1_idx = 0;
+ /* res = tos X op */
+ }
+ } 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);
+ if (op2_idx == 0)
+ op2_idx = op1_idx;
+ op1_idx = 0;
+ }
+ /* res = tos X op, pop */
+ pops = 1;
+ }
+ } else {
+ /* second operand is dead */
+ if (is_vfp_live(reg_index_1, 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);
+ if (op1_idx == 0)
+ op1_idx = op2_idx;
+ op2_idx = 0;
+ }
+ /* res = op X tos, pop */
+ pops = 1;
+ permuted = !permuted;
+ xchg = 1;
+ } else {
+ /* both operands are dead here, check first for identity. */
+ if (op1_idx == op2_idx) {
+ /* identically, one pop needed */
+ if (op1_idx != 0) {
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ op2_idx = 0;
+ }
+ /* res = tos X op, pop */
+ pops = 1;
+ }
+ /* different, move them to st and st(1) and pop both.
+ The tricky part is to get one into st(1).*/
+ else if (op2_idx == 1) {
+ /* good, second operand is already in the right place, move the first */
+ if (op1_idx != 0) {
+ /* bring the first on top */
+ x87_create_fxch(state, n, op1_idx);
+ assert(op2_idx != 0);
+ op1_idx = 0;
+ }
+ /* res = tos X op, pop, pop */
+ pops = 2;
+ } 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);
+ assert(op1_idx != 0);
+ op2_idx = 0;
+ }
+ /* res = op X tos, pop, pop */
+ permuted = !permuted;
+ xchg = 1;
+ pops = 2;
+ } 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);
+ assert(op2_idx != 1);
+ op1_idx = 1;
+ x87_create_fxch(state, n, op2_idx);
+ op2_idx = 0;
+ /* res = op X tos, pop, pop */
+ pops = 2;
+ permuted = !permuted;
+ xchg = 1;
+ } else if (op2_idx == 0) {
+ /* second one is TOS, move to st(1) */
+ x87_create_fxch(state, n, 1);
+ assert(op1_idx != 1);
+ op2_idx = 1;
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ /* res = tos X op, pop, pop */
+ pops = 2;
+ } else {
+ /* none of them is either TOS or st(1), 3 fxch needed */
+ x87_create_fxch(state, n, op2_idx);
+ assert(op1_idx != 0);
+ x87_create_fxch(state, n, 1);
+ op2_idx = 1;
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ /* res = tos X op, pop, pop */
+ pops = 2;
+ }
+ }
+ }
+ }
+ } else {
+ /* second operand is an address mode */
+ if (is_vfp_live(reg_index_1, live)) {
+ /* first operand is live: bring it to TOS */
+ if (op1_idx != 0) {
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ }
+ } else {
+ /* first operand is dead: bring it to tos */
+ if (op1_idx != 0) {
+ x87_create_fxch(state, n, op1_idx);
+ op1_idx = 0;
+ }
+ pops = 1;
+ }
+ }
+
+ /* patch the operation */
+ if (is_ia32_vFucomFnstsw(n)) {
+ int i;
+
+ switch (pops) {
+ case 0: dst = op_ia32_FucomFnstsw; break;
+ case 1: dst = op_ia32_FucompFnstsw; break;
+ case 2: dst = op_ia32_FucomppFnstsw; break;
+ default: panic("invalid popcount in sim_Fucom");
+ }
+
+ for (i = 0; i < pops; ++i) {
+ x87_pop(state);
+ }
+ } else if (is_ia32_vFucomi(n)) {
+ switch (pops) {
+ case 0: dst = op_ia32_Fucomi; break;
+ case 1: dst = op_ia32_Fucompi; x87_pop(state); break;
+ case 2:
+ dst = op_ia32_Fucompi;
+ x87_pop(state);
+ x87_create_fpop(state, sched_next(n), 1);
+ node_added = NODE_ADDED;
+ break;
+ default: panic("invalid popcount in sim_Fucom");
+ }
+ } else {
+ panic("invalid operation %+F in sim_FucomFnstsw", n);
+ }
+
+ x87_patch_insn(n, dst);
+ if (xchg) {
+ int tmp = op1_idx;
+ op1_idx = op2_idx;
+ op2_idx = tmp;
+ }
+
+ 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;
+ attr->attr.data.ins_permuted = permuted;
+
+ if (op2_idx >= 0) {
+ DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(n),
+ arch_register_get_name(op1), arch_register_get_name(op2)));
+ } else {
+ DB((dbg, LEVEL_1, "<<< %s %s, [AM]\n", get_irn_opname(n),
+ arch_register_get_name(op1)));
+ }
+
+ return node_added;
+}
+
+static int sim_Keep(x87_state *state, ir_node *node)
+{
+ const ir_node *op;
+ const arch_register_t *op_reg;
+ int reg_id;
+ int op_stack_idx;
+ unsigned live;
+ int i, arity;
+ int node_added = NO_NODE_ADDED;
+
+ DB((dbg, LEVEL_1, ">>> %+F\n", node));
+
+ arity = get_irn_arity(node);
+ for (i = 0; i < arity; ++i) {
+ op = get_irn_n(node, i);
+ op_reg = arch_get_irn_register(op);
+ if (arch_register_get_class(op_reg) != &ia32_reg_classes[CLASS_ia32_vfp])
+ continue;
+
+ reg_id = arch_register_get_index(op_reg);
+ live = vfp_live_args_after(state->sim, node, 0);
+
+ op_stack_idx = x87_on_stack(state, reg_id);
+ if (op_stack_idx >= 0 && !is_vfp_live(reg_id, live)) {
+ x87_create_fpop(state, sched_next(node), 1);
+ node_added = NODE_ADDED;
+ }
+ }
+
+ DB((dbg, LEVEL_1, "Stack after: "));
+ DEBUG_ONLY(x87_dump_stack(state));
+
+ return node_added;
+}
+
+static void keep_float_node_alive(ir_node *node)
+{
+ ir_graph *irg;
+ ir_node *block;
+ ir_node *in[1];
+ ir_node *keep;
+ const arch_register_class_t *cls;
+
+ irg = get_irn_irg(node);
+ block = get_nodes_block(node);
+ cls = arch_get_irn_reg_class(node, -1);
+ in[0] = node;
+ keep = be_new_Keep(cls, irg, block, 1, in);
+
+ assert(sched_is_scheduled(node));
+ sched_add_after(node, keep);
+}
+
+/**
+ * Create a copy of a node. Recreate the node if it's a constant.
+ *
+ * @param state the x87 state
+ * @param n the node to be copied
+ *
+ * @return the copy of n
+ */
+static ir_node *create_Copy(x87_state *state, ir_node *n)
+{
+ 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_x87_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;
+ default:
+ break;
+ }
+
+ out = x87_get_irn_register(n);
+ op1 = x87_get_irn_register(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_x87_attr(res);
+ attr->x87[2] = &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_x87_attr(res);
+ attr->x87[0] = &ia32_st_regs[op1_idx];
+ attr->x87[2] = &ia32_st_regs[0];
+ }
+ arch_set_irn_register(res, out);
+
+ return res;
+} /* create_Copy */
+
+/**
+ * Simulate a be_Copy.
+ *
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ *
+ * @return NO_NODE_ADDED
+ */
+static int sim_Copy(x87_state *state, ir_node *n)
+{
+ ir_node *pred;
+ const arch_register_t *out;
+ const arch_register_t *op1;
+ const arch_register_class_t *cls;
+ ir_node *node, *next;
+ ia32_x87_attr_t *attr;
+ int op1_idx, out_idx;
+ unsigned live;
+
+ cls = arch_get_irn_reg_class(n, -1);
+ if (cls->regs != ia32_vfp_regs)
+ return 0;
+
+ pred = get_irn_n(n, 0);
+ out = x87_get_irn_register(n);
+ op1 = x87_get_irn_register(pred);
+ live = vfp_live_args_after(state->sim, n, REGMASK(out));
+
+ 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));
+
+ /* 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);
+
+ assert(is_ia32_fldz(node));
+ 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(node)->name));
+ return NO_NODE_ADDED;
+ }
+
+ op1_idx = x87_on_stack(state, arch_register_get_index(op1));
+
+ if (is_vfp_live(arch_register_get_index(op1), live)) {
+ ir_node *pred = get_irn_n(n, 0);
+
+ /* 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);
+
+ /* We have to make sure the old value doesn't go dead (which can happen
+ * when we recreate constants). As the simulator expected that value in
+ * the pred blocks. This is unfortunate as removing it would save us 1
+ * instruction, but we would have to rerun all the simulation to get
+ * this correct...
+ */
+ next = sched_next(n);
+ sched_remove(n);
+ exchange(n, node);
+ sched_add_before(next, node);
+
+ if (get_irn_n_edges(pred) == 0) {
+ keep_float_node_alive(pred);
+ }
+
+ DB((dbg, LEVEL_1, "<<< %+F %s -> ?\n", node, op1->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_x87_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);
+ op1_idx = 0;
+ }
+ x87_patch_insn(n, op_ia32_Pop);
+ attr = get_ia32_x87_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);
+ /* don't remove the node to keep the verifier quiet :),
+ the emitter won't emit any code for the node */
+#if 0
+ sched_remove(n);
+ DB((dbg, LEVEL_1, "<<< KILLED %s\n", get_irn_opname(n)));
+ exchange(n, get_unop_op(n));
+#endif
+ }
+ }
+ return NO_NODE_ADDED;
+} /* sim_Copy */
+
+/**
+ * Returns the result proj of the call
+ */
+static ir_node *get_call_result_proj(ir_node *call)
+{
+ const ir_edge_t *edge;
+
+ /* search the result proj */
+ foreach_out_edge(call, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ long pn = get_Proj_proj(proj);
+
+ if (pn == pn_ia32_Call_vf0) {
+ return proj;
+ }
+ }
+
+ return NULL;
+} /* get_call_result_proj */
+
+/**
+ * Simulate a ia32_Call.
+ *
+ * @param state the x87 state
+ * @param n the node that should be simulated
+ *
+ * @return NO_NODE_ADDED
+ */
+static int sim_Call(x87_state *state, ir_node *n)
+{
+ ir_type *call_tp = get_ia32_call_attr_const(n)->call_tp;
+ ir_type *res_type;
+ ir_mode *mode;
+ ir_node *resproj;
+ const arch_register_t *reg;
+
+ DB((dbg, LEVEL_1, ">>> %+F\n", n));
+
+ /* at the begin of a call the x87 state should be empty */
+ assert(state->depth == 0 && "stack not empty before call");
+
+ if (get_method_n_ress(call_tp) <= 0)
+ goto end_call;
+
+ /*
+ * If the called function returns a float, it is returned in st(0).
+ * This even happens if the return value is NOT used.
+ * Moreover, only one return result is supported.
+ */
+ res_type = get_method_res_type(call_tp, 0);
+ mode = get_type_mode(res_type);
+
+ if (mode == NULL || !mode_is_float(mode))
+ goto end_call;
+
+ resproj = get_call_result_proj(n);
+ assert(resproj != NULL);
+
+ reg = x87_get_irn_register(resproj);
+ x87_push(state, arch_register_get_index(reg), resproj);
+
+end_call:
+ DB((dbg, LEVEL_1, "Stack after: "));
+ DEBUG_ONLY(x87_dump_stack(state));
+
+ return NO_NODE_ADDED;
+} /* sim_Call */
+
+/**
+ * Simulate a be_Spill.
+ *
+ * @param state the x87 state
+ * @param n the node that should be simulated (and patched)
+ *
+ * Should not happen, spills are lowered before x87 simulator see them.
+ */
+static int sim_Spill(x87_state *state, ir_node *n)
+{
+ assert(0 && "Spill not lowered");
+ 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)
+ *
+ * Should not happen, reloads are lowered before x87 simulator see them.
+ */
+static int sim_Reload(x87_state *state, ir_node *n)
+{
+ assert(0 && "Reload not lowered");
+ 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)
+ *
+ * @return NO_NODE_ADDED
+ */
+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;
+
+ /* only floating point return values must reside on stack */
+ for (i = 0; i < n_res; ++i) {
+ ir_node *res = get_irn_n(n, be_pos_Return_val + i);
+
+ if (mode_is_float(get_irn_mode(res)))
+ ++n_float_res;
+ }
+ assert(x87_get_depth(state) == n_float_res);
+
+ /* pop them virtually */
+ for (i = n_float_res - 1; i >= 0; --i)
+ x87_pop(state);
+
+ return NO_NODE_ADDED;
+} /* 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)
+ *
+ * @return NO_NODE_ADDED
+ */
+static int sim_Perm(x87_state *state, ir_node *irn)
+{
+ int i, n;
+ 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 NO_NODE_ADDED;
+
+ 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(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(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 NO_NODE_ADDED;
+} /* sim_Perm */
+
+static int sim_Barrier(x87_state *state, ir_node *node)
+{
+ int i, arity;
+
+ /* materialize unknown if needed */
+ arity = get_irn_arity(node);
+ for (i = 0; i < arity; ++i) {
+ const arch_register_t *reg;
+ ir_node *zero;
+ ir_node *block;
+ ia32_x87_attr_t *attr;
+ ir_node *in = get_irn_n(node, i);
+
+ if (!is_ia32_Unknown_VFP(in))
+ continue;
+
+ /* TODO: not completely correct... */
+ reg = &ia32_vfp_regs[REG_VFP_UKNWN];
+
+ /* create a zero */
+ block = get_nodes_block(node);
+ zero = new_rd_ia32_fldz(NULL, current_ir_graph, block, mode_E);
+ x87_push(state, arch_register_get_index(reg), zero);
+
+ attr = get_ia32_x87_attr(zero);
+ attr->x87[2] = &ia32_st_regs[0];
+
+ sched_add_before(node, zero);
+
+ set_irn_n(node, i, zero);
+ }
+
+ return NO_NODE_ADDED;
+}
+
+
+/**
+ * Kill any dead registers at block start by popping them from the stack.
+ *
+ * @param sim the simulator handle
+ * @param block the current block
+ * @param start_state the x87 state at the begin of the block
+ *
+ * @return the x87 state after dead register killed
+ */
+static x87_state *x87_kill_deads(x87_simulator *sim, ir_node *block, x87_state *start_state)
+{
+ x87_state *state = start_state;
+ ir_node *first_insn = sched_first(block);
+ ir_node *keep = NULL;
+ unsigned live = vfp_live_args_after(sim, block, 0);
+ unsigned kill_mask;
+ int i, depth, num_pop;
+
+ kill_mask = 0;
+ depth = x87_get_depth(state);
+ for (i = depth - 1; i >= 0; --i) {
+ int reg = x87_get_st_reg(state, i);
+
+ if (! is_vfp_live(reg, live))
+ kill_mask |= (1 << i);
+ }
+
+ if (kill_mask) {
+ /* create a new state, will be changed */
+ state = x87_clone_state(sim, state);
+
+ DB((dbg, LEVEL_1, "Killing deads:\n"));
+ DEBUG_ONLY(vfp_dump_live(live));
+ DEBUG_ONLY(x87_dump_stack(state));
+
+ if (kill_mask != 0 && live == 0) {
+ /* special case: kill all registers */
+ if (ia32_cg_config.use_femms || ia32_cg_config.use_emms) {
+ if (ia32_cg_config.use_femms) {
+ /* use FEMMS on AMD processors to clear all */
+ keep = new_rd_ia32_femms(NULL, get_irn_irg(block), block);
+ } else {
+ /* use EMMS to clear all */
+ keep = new_rd_ia32_emms(NULL, get_irn_irg(block), block);
+ }
+ sched_add_before(first_insn, keep);
+ keep_alive(keep);
+ x87_emms(state);
+ return state;
+ }
+ }
+ /* now kill registers */
+ while (kill_mask) {
+ /* we can only kill from TOS, so bring them up */
+ if (! (kill_mask & 1)) {
+ /* search from behind, because we can to a double-pop */
+ for (i = depth - 1; i >= 0; --i) {
+ if (kill_mask & (1 << i)) {
+ kill_mask &= ~(1 << i);
+ kill_mask |= 1;
+ break;
+ }
+ }
+
+ if (keep)
+ x87_set_st(state, -1, keep, i);
+ x87_create_fxch(state, first_insn, i);
+ }
+
+ if ((kill_mask & 3) == 3) {
+ /* we can do a double-pop */
+ num_pop = 2;
+ }
+ else {
+ /* only a single pop */
+ num_pop = 1;
+ }
+
+ depth -= num_pop;
+ kill_mask >>= num_pop;
+ keep = x87_create_fpop(state, first_insn, num_pop);
+ }
+ keep_alive(keep);
+ }
+ return state;
+} /* x87_kill_deads */
+
+/**
+ * If we have PhiEs with unknown operands then we have to make sure that some
+ * value is actually put onto the stack.
+ */
+static void fix_unknown_phis(x87_state *state, ir_node *block,
+ ir_node *pred_block, int pos)
+{
+ ir_node *node, *op;
+
+ sched_foreach(block, node) {
+ ir_node *zero;
+ const arch_register_t *reg;
+ ia32_x87_attr_t *attr;
+
+ if (!is_Phi(node))
+ break;
+
+ op = get_Phi_pred(node, pos);
+ if (!is_ia32_Unknown_VFP(op))
+ continue;
+
+ reg = arch_get_irn_register(node);
+
+ /* create a zero at end of pred block */
+ zero = new_rd_ia32_fldz(NULL, current_ir_graph, pred_block, mode_E);
+ x87_push(state, arch_register_get_index(reg), zero);
+
+ attr = get_ia32_x87_attr(zero);
+ attr->x87[2] = &ia32_st_regs[0];
+
+ assert(is_ia32_fldz(zero));
+ sched_add_before(sched_last(pred_block), zero);
+
+ set_Phi_pred(node, pos, zero);
+ }
+}
/**
* 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
+ * @param sim the simulator handle
+ * @param block the current block
*/
-static int x87_simulate_block(x87_simulator *sim, ir_node *block) {
- ir_node *n;
+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);
+ DB((dbg, LEVEL_1, "Simulate %+F\n", block));
+ DB((dbg, LEVEL_2, "State at Block begin:\n "));
+ DEBUG_ONLY(x87_dump_stack(state));
- sched_foreach(block, n) {
+ /* at block begin, kill all dead registers */
+ state = x87_kill_deads(sim, block, state);
+ /* create a new state, will be changed */
+ state = x87_clone_state(sim, state);
+
+ /* 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);
- if (op->ops.generic) {
- sim_func func = (sim_func)op->ops.generic;
+ next = sched_next(n);
+ if (op->ops.generic == NULL)
+ continue;
- /* have work to do */
- if (state == bl_state->begin) {
- /* create a new state, will be changed */
- state = x87_clone_state(sim, state);
- }
+ func = (sim_func)op->ops.generic;
- /* simulate it */
- (*func)(state, n, sim->env);
- }
+ /* simulate it */
+ node_inserted = (*func)(state, n);
+
+ /*
+ sim_func might have added an additional node 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 != NO_NODE_ADDED)
+ next = sched_next(n);
}
start_block = get_irg_start_block(get_irn_irg(block));
+ DB((dbg, LEVEL_2, "State at Block end:\n ")); DEBUG_ONLY(x87_dump_stack(state));
+
/* 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_state->begin && succ != start_block) {
- /* There is already a begin state for this block, bad.
- Do the necessary permutations.
- Note that critical edges are removed, so this is always possible. */
- x87_shuffle(sim, block, state, succ_state->begin);
+ if (succ == start_block)
+ continue;
- /* Note further, that there can be only one such situation,
- so we can break here. */
- break;
- }
- }
- bl_state->end = state;
+ succ_state = x87_get_bl_state(sim, succ);
- /* 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);
+ fix_unknown_phis(state, succ, block, get_edge_src_pos(edge));
- if (! succ_state->begin)
+ if (succ_state->begin == NULL) {
+ DB((dbg, LEVEL_2, "Set begin state for succ %+F:\n", succ));
+ DEBUG_ONLY(x87_dump_stack(state));
succ_state->begin = state;
+
+ waitq_put(sim->worklist, succ);
+ } else {
+ DB((dbg, LEVEL_2, "succ %+F already has a state, shuffling\n", succ));
+ /* 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:
+ If the successor has more than one possible input, then it must
+ be the only one.
+ */
+ x87_shuffle(sim, block, state, succ, succ_state->begin);
+ }
}
+ bl_state->end = state;
+} /* x87_simulate_block */
- return 1;
+static void register_sim(ir_op *op, sim_func func)
+{
+ assert(op->ops.generic == NULL);
+ op->ops.generic = (op_func) func;
}
/**
* Create a new x87 simulator.
+ *
+ * @param sim a simulator handle, will be initialized
+ * @param irg the current graph
*/
-static void x87_init_simulator(x87_simulator *sim, const arch_env_t *env) {
+static void x87_init_simulator(x87_simulator *sim, ir_graph *irg)
+{
obstack_init(&sim->obst);
sim->blk_states = pmap_create();
- sim->env = env;
+ sim->n_idx = get_irg_last_idx(irg);
+ sim->live = obstack_alloc(&sim->obst, sizeof(*sim->live) * sim->n_idx);
+
+ DB((dbg, LEVEL_1, "--------------------------------\n"
+ "x87 Simulator started for %+F\n", irg));
+ /* set the generic function pointer of instruction we must simulate */
clear_irp_opcodes_generic_func();
-#define ASSOC(op) (op_ia32_v ## op)->ops.generic = (op_func)(sim_##op)
- ASSOC(fConst);
- ASSOC(fld);
- ASSOC(fld1);
- ASSOC(fldz);
- ASSOC(fadd);
- ASSOC(fsub);
- ASSOC(fmul);
- ASSOC(fdiv);
- ASSOC(fldz);
- ASSOC(fabs);
- ASSOC(fchs);
- ASSOC(fsin);
- ASSOC(fcos);
- ASSOC(fsqrt);
-#undef ASSOC
-}
+ register_sim(op_ia32_Call, sim_Call);
+ register_sim(op_ia32_vfld, sim_fld);
+ register_sim(op_ia32_vfild, sim_fild);
+ register_sim(op_ia32_vfld1, sim_fld1);
+ register_sim(op_ia32_vfldz, sim_fldz);
+ register_sim(op_ia32_vfadd, sim_fadd);
+ register_sim(op_ia32_vfsub, sim_fsub);
+ register_sim(op_ia32_vfmul, sim_fmul);
+ register_sim(op_ia32_vfdiv, sim_fdiv);
+ register_sim(op_ia32_vfprem, sim_fprem);
+ register_sim(op_ia32_vfabs, sim_fabs);
+ register_sim(op_ia32_vfchs, sim_fchs);
+ register_sim(op_ia32_vfist, sim_fist);
+ register_sim(op_ia32_vfisttp, sim_fisttp);
+ register_sim(op_ia32_vfst, sim_fst);
+ register_sim(op_ia32_vFtstFnstsw, sim_FtstFnstsw);
+ register_sim(op_ia32_vFucomFnstsw, sim_Fucom);
+ register_sim(op_ia32_vFucomi, sim_Fucom);
+ register_sim(op_be_Copy, sim_Copy);
+ register_sim(op_be_Spill, sim_Spill);
+ register_sim(op_be_Reload, sim_Reload);
+ register_sim(op_be_Return, sim_Return);
+ register_sim(op_be_Perm, sim_Perm);
+ register_sim(op_be_Keep, sim_Keep);
+ register_sim(op_be_Barrier, sim_Barrier);
+} /* x87_init_simulator */
/**
* Destroy a x87 simulator.
+ *
+ * @param sim the simulator handle
*/
-static void x87_destroy_simulator(x87_simulator *sim) {
+static void x87_destroy_simulator(x87_simulator *sim)
+{
pmap_destroy(sim->blk_states);
obstack_free(&sim->obst, NULL);
-}
+ DB((dbg, LEVEL_1, "x87 Simulator stopped\n\n"));
+} /* x87_destroy_simulator */
/**
- * Run a simulation and fix all virtual instructions for a graph.
- *
- * Needs a block-schedule.
+ * Pre-block walker: calculate the liveness information for the block
+ * and store it into the sim->live cache.
*/
-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;
+static void update_liveness_walker(ir_node *block, void *data)
+{
+ x87_simulator *sim = data;
+ update_liveness(sim, block);
+} /* update_liveness_walker */
+
+void x87_simulate_graph(be_irg_t *birg)
+{
+ /* TODO improve code quality (less executed fxch) by using execfreqs */
+
+ ir_node *block, *start_block;
+ blk_state *bl_state;
x87_simulator sim;
- int i;
-
- be_liveness(irg);
- be_liveness_dumpto(irg, "-x87-live");
+ ir_graph *irg = be_get_birg_irg(birg);
- x87_init_simulator(&sim, env);
+ /* create the simulator */
+ x87_init_simulator(&sim, irg);
start_block = get_irg_start_block(irg);
- bl_state = x87_get_bl_state(&sim, start_block);
+ bl_state = x87_get_bl_state(&sim, start_block);
/* start with the empty state */
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_assure_liveness(birg);
+ sim.lv = be_get_birg_liveness(birg);
+// sim.lv = be_liveness(be_get_birg_irg(birg));
+ be_liveness_assure_sets(sim.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 (! waitq_empty(sim.worklist));
+ /* kill it */
+ del_waitq(sim.worklist);
x87_destroy_simulator(&sim);
-}
+} /* x87_simulate_graph */
+
+void ia32_init_x87(void)
+{
+ FIRM_DBG_REGISTER(dbg, "firm.be.ia32.x87");
+} /* ia32_init_x87 */