/* Static Single Assignment conversion routines for the GNU compiler.
- Copyright (C) 2000, 2001 Free Software Foundation, Inc.
+ Copyright (C) 2000, 2001, 2002 Free Software Foundation, Inc.
This file is part of GCC.
#include "config.h"
#include "system.h"
+#include "coretypes.h"
+#include "tm.h"
#include "rtl.h"
#include "expr.h"
#include "output.h"
#include "ssa.h"
-/* TODO:
+/* TODO:
Handle subregs better, maybe. For now, if a reg that's set in a
subreg expression is duplicated going into SSA form, an extra copy
the same hard register in the same machine mode are in the same
class. */
-/* If conservative_reg_partition is non-zero, use a conservative
+/* If conservative_reg_partition is nonzero, use a conservative
register partitioning algorithm (which leaves more regs after
emerging from SSA) instead of the coalescing one. This is being
left in for a limited time only, as a debugging tool until the
partition reg_partition;
};
+static rtx gen_sequence
+ PARAMS ((void));
static void ssa_rename_from_initialize
PARAMS ((void));
static rtx ssa_rename_from_lookup
static inline rtx * phi_alternative
PARAMS ((rtx, int));
static void compute_dominance_frontiers_1
- PARAMS ((sbitmap *frontiers, int *idom, int bb, sbitmap done));
+ PARAMS ((sbitmap *frontiers, dominance_info idom, int bb, sbitmap done));
static void find_evaluations_1
PARAMS ((rtx dest, rtx set, void *data));
static void find_evaluations
PARAMS ((int regno, int b));
static void insert_phi_nodes
PARAMS ((sbitmap *idfs, sbitmap *evals, int nregs));
-static void create_delayed_rename
+static void create_delayed_rename
PARAMS ((struct rename_context *, rtx *));
-static void apply_delayed_renames
+static void apply_delayed_renames
PARAMS ((struct rename_context *));
-static int rename_insn_1
+static int rename_insn_1
PARAMS ((rtx *ptr, void *data));
-static void rename_block
- PARAMS ((int b, int *idom));
-static void rename_registers
- PARAMS ((int nregs, int *idom));
+static void rename_block
+ PARAMS ((int b, dominance_info dom));
+static void rename_registers
+ PARAMS ((int nregs, dominance_info idom));
static inline int ephi_add_node
PARAMS ((rtx reg, rtx *nodes, int *n_nodes));
PARAMS ((int t, sbitmap visited, sbitmap *pred, sbitmap *succ, rtx *nodes));
static void eliminate_phi
PARAMS ((edge e, partition reg_partition));
-static int make_regs_equivalent_over_bad_edges
+static int make_regs_equivalent_over_bad_edges
PARAMS ((int bb, partition reg_partition));
/* These are used only in the conservative register partitioning
algorithms. */
-static int make_equivalent_phi_alternatives_equivalent
+static int make_equivalent_phi_alternatives_equivalent
PARAMS ((int bb, partition reg_partition));
-static partition compute_conservative_reg_partition
+static partition compute_conservative_reg_partition
PARAMS ((void));
static int record_canonical_element_1
PARAMS ((void **srfp, void *data));
static int check_hard_regs_in_partition
PARAMS ((partition reg_partition));
-static int rename_equivalent_regs_in_insn
+static int rename_equivalent_regs_in_insn
PARAMS ((rtx *ptr, void *data));
/* These are used in the register coalescing algorithm. */
PARAMS ((basic_block bb, partition p, conflict_graph conflicts));
static partition compute_coalesced_reg_partition
PARAMS ((void));
-static int mark_reg_in_phi
+static int mark_reg_in_phi
PARAMS ((rtx *ptr, void *data));
static void mark_phi_and_copy_regs
PARAMS ((regset phi_set));
-static int rename_equivalent_regs_in_insn
+static int rename_equivalent_regs_in_insn
PARAMS ((rtx *ptr, void *data));
-static void rename_equivalent_regs
+static void rename_equivalent_regs
PARAMS ((partition reg_partition));
/* Deal with hard registers. */
}
/* Given the SET of a phi node, remove the alternative for predecessor
- block C. Return non-zero on success, or zero if no alternative is
+ block C. Return nonzero on success, or zero if no alternative is
found for C. */
int
sbitmap *evals;
int nregs;
{
- int bb;
+ basic_block bb;
sbitmap_vector_zero (evals, nregs);
fe_evals = evals;
- for (bb = n_basic_blocks; --bb >= 0; )
+ FOR_EACH_BB_REVERSE (bb)
{
rtx p, last;
- fe_current_bb = bb;
- p = BLOCK_HEAD (bb);
- last = BLOCK_END (bb);
+ fe_current_bb = bb->index;
+ p = bb->head;
+ last = bb->end;
while (1)
{
if (INSN_P (p))
}
/* Computing the Dominance Frontier:
-
- As decribed in Morgan, section 3.5, this may be done simply by
+
+ As decribed in Morgan, section 3.5, this may be done simply by
walking the dominator tree bottom-up, computing the frontier for
the children before the parent. When considering a block B,
there are two cases:
static void
compute_dominance_frontiers_1 (frontiers, idom, bb, done)
sbitmap *frontiers;
- int *idom;
+ dominance_info idom;
int bb;
sbitmap done;
{
basic_block b = BASIC_BLOCK (bb);
edge e;
- int c;
+ basic_block c;
SET_BIT (done, bb);
sbitmap_zero (frontiers[bb]);
/* Do the frontier of the children first. Not all children in the
dominator tree (blocks dominated by this one) are children in the
CFG, so check all blocks. */
- for (c = 0; c < n_basic_blocks; ++c)
- if (idom[c] == bb && ! TEST_BIT (done, c))
- compute_dominance_frontiers_1 (frontiers, idom, c, done);
+ FOR_EACH_BB (c)
+ if (get_immediate_dominator (idom, c)->index == bb
+ && ! TEST_BIT (done, c->index))
+ compute_dominance_frontiers_1 (frontiers, idom, c->index, done);
/* Find blocks conforming to rule (1) above. */
for (e = b->succ; e; e = e->succ_next)
{
if (e->dest == EXIT_BLOCK_PTR)
continue;
- if (idom[e->dest->index] != bb)
+ if (get_immediate_dominator (idom, e->dest)->index != bb)
SET_BIT (frontiers[bb], e->dest->index);
}
/* Find blocks conforming to rule (2). */
- for (c = 0; c < n_basic_blocks; ++c)
- if (idom[c] == bb)
+ FOR_EACH_BB (c)
+ if (get_immediate_dominator (idom, c)->index == bb)
{
int x;
- EXECUTE_IF_SET_IN_SBITMAP (frontiers[c], 0, x,
+ EXECUTE_IF_SET_IN_SBITMAP (frontiers[c->index], 0, x,
{
- if (idom[x] != bb)
+ if (get_immediate_dominator (idom, BASIC_BLOCK (x))->index != bb)
SET_BIT (frontiers[bb], x);
});
}
void
compute_dominance_frontiers (frontiers, idom)
sbitmap *frontiers;
- int *idom;
+ dominance_info idom;
{
- sbitmap done = sbitmap_alloc (n_basic_blocks);
+ sbitmap done = sbitmap_alloc (last_basic_block);
sbitmap_zero (done);
compute_dominance_frontiers_1 (frontiers, idom, 0, done);
sbitmap worklist;
int reg, passes = 0;
- worklist = sbitmap_alloc (n_basic_blocks);
+ worklist = sbitmap_alloc (last_basic_block);
for (reg = 0; reg < nregs; ++reg)
{
if (rtl_dump_file)
{
- fprintf(rtl_dump_file,
- "Iterated dominance frontier: %d passes on %d regs.\n",
- passes, nregs);
+ fprintf (rtl_dump_file,
+ "Iterated dominance frontier: %d passes on %d regs.\n",
+ passes, nregs);
}
}
}
}
-/* Rename the registers to conform to SSA.
+/* Rename the registers to conform to SSA.
This is essentially the algorithm presented in Figure 7.8 of Morgan,
- with a few changes to reduce pattern search time in favour of a bit
+ with a few changes to reduce pattern search time in favor of a bit
more memory usage. */
/* One of these is created for each set. It will live in a list local
{
struct rename_set_data *r;
r = (struct rename_set_data *) xmalloc (sizeof(*r));
-
+
if (GET_CODE (*reg_loc) != REG
|| !CONVERT_REGISTER_TO_SSA_P (REGNO (*reg_loc)))
- abort();
+ abort ();
r->reg_loc = reg_loc;
r->old_reg = *reg_loc;
/* This is part of a rather ugly hack to allow the pre-ssa regno to be
reused. If, during processing, a register has not yet been touched,
ssa_rename_to[regno][machno] will be NULL. Now, in the course of pushing
- and popping values from ssa_rename_to, when we would ordinarily
+ and popping values from ssa_rename_to, when we would ordinarily
pop NULL back in, we pop RENAME_NO_RTX. We treat this exactly the
same as NULL, except that it signals that the original regno has
already been reused. */
for (r = c->new_renames; r != NULL; r = r->next)
{
int new_regno;
-
+
/* Failure here means that someone has a PARALLEL that sets
a register twice (bad!). */
if (ssa_rename_to_lookup (r->old_reg) != r->prev_reg)
- abort();
+ abort ();
/* Failure here means we have changed REG_LOC before applying
the rename. */
/* For the first set we come across, reuse the original regno. */
}
}
-/* Part one of the first step of rename_block, called through for_each_rtx.
+/* Part one of the first step of rename_block, called through for_each_rtx.
Mark pseudos that are set for later update. Transform uses of pseudos. */
static int
(set (subreg (reg foo)) ...)
into
(sequence [(set (reg foo_1) (reg foo))
- (set (subreg (reg foo_1)) ...)])
+ (set (subreg (reg foo_1)) ...)])
FIXME: Much of the time this is too much. For some constructs
we know that the output register is strictly an output
{
rtx i, reg;
reg = dest;
-
+
while (GET_CODE (reg) == STRICT_LOW_PART
|| GET_CODE (reg) == SUBREG
|| GET_CODE (reg) == SIGN_EXTRACT
|| GET_CODE (reg) == ZERO_EXTRACT)
reg = XEXP (reg, 0);
-
+
if (GET_CODE (reg) == REG
&& CONVERT_REGISTER_TO_SSA_P (REGNO (reg)))
{
}
case REG:
- if (CONVERT_REGISTER_TO_SSA_P (REGNO (x)) &&
- REGNO (x) < ssa_max_reg_num)
+ if (CONVERT_REGISTER_TO_SSA_P (REGNO (x))
+ && REGNO (x) < ssa_max_reg_num)
{
rtx new_reg = ssa_rename_to_lookup (x);
- if (new_reg != NULL_RTX && new_reg != RENAME_NO_RTX)
+ if (new_reg != RENAME_NO_RTX && new_reg != NULL_RTX)
{
if (GET_MODE (x) != GET_MODE (new_reg))
abort ();
*ptr = new_reg;
}
- /* Else this is a use before a set. Warn? */
+ else
+ {
+ /* Undefined value used, rename it to a new pseudo register so
+ that it cannot conflict with an existing register. */
+ *ptr = gen_reg_rtx (GET_MODE (x));
+ }
}
return -1;
}
/* Stop traversing. */
return -1;
- }
+ }
else
/* Continue traversing. */
return 0;
}
}
+static rtx
+gen_sequence ()
+{
+ rtx first_insn = get_insns ();
+ rtx result;
+ rtx tem;
+ int i;
+ int len;
+
+ /* Count the insns in the chain. */
+ len = 0;
+ for (tem = first_insn; tem; tem = NEXT_INSN (tem))
+ len++;
+
+ result = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (len));
+
+ for (i = 0, tem = first_insn; tem; tem = NEXT_INSN (tem), i++)
+ XVECEXP (result, 0, i) = tem;
+
+ return result;
+}
+
static void
rename_block (bb, idom)
int bb;
- int *idom;
+ dominance_info idom;
{
basic_block b = BASIC_BLOCK (bb);
edge e;
rtx insn, next, last;
struct rename_set_data *set_data = NULL;
- int c;
+ basic_block c;
/* Step One: Walk the basic block, adding new names for sets and
replacing uses. */
-
+
next = b->head;
last = b->end;
do
{
rtx seq;
int i;
-
+
emit (PATTERN (insn));
seq = gen_sequence ();
/* We really want a SEQUENCE of SETs, not a SEQUENCE
PATTERN (insn) = seq;
}
end_sequence ();
-
+
apply_delayed_renames (&context);
set_data = context.done_renames;
}
/* Step Three: Do the same to the children of this block in
dominator order. */
- for (c = 0; c < n_basic_blocks; ++c)
- if (idom[c] == bb)
- rename_block (c, idom);
+ FOR_EACH_BB (c)
+ if (get_immediate_dominator (idom, c)->index == bb)
+ rename_block (c->index, idom);
/* Step Four: Update the sets to refer to their new register,
and restore ssa_rename_to to its previous state. */
rtx old_reg = *set_data->reg_loc;
if (*set_data->reg_loc != set_data->old_reg)
- abort();
+ abort ();
*set_data->reg_loc = set_data->new_reg;
ssa_rename_to_insert (old_reg, set_data->prev_reg);
next = set_data->next;
free (set_data);
set_data = next;
- }
+ }
}
static void
rename_registers (nregs, idom)
int nregs;
- int *idom;
+ dominance_info idom;
{
VARRAY_RTX_INIT (ssa_definition, nregs * 3, "ssa_definition");
ssa_rename_from_initialize ();
ssa_rename_to_pseudo = (rtx *) alloca (nregs * sizeof(rtx));
memset ((char *) ssa_rename_to_pseudo, 0, nregs * sizeof(rtx));
- memset ((char *) ssa_rename_to_hard, 0,
+ memset ((char *) ssa_rename_to_hard, 0,
FIRST_PSEUDO_REGISTER * NUM_MACHINE_MODES * sizeof (rtx));
rename_block (0, idom);
- /* ??? Update basic_block_live_at_start, and other flow info
+ /* ??? Update basic_block_live_at_start, and other flow info
as needed. */
ssa_rename_to_pseudo = NULL;
sbitmap *idfs;
/* Element I is the immediate dominator of block I. */
- int *idom;
+ dominance_info idom;
int nregs;
+ basic_block bb;
+
/* Don't do it twice. */
if (in_ssa_form)
abort ();
dead code. We'll let the SSA optimizers do that. */
life_analysis (get_insns (), NULL, 0);
- idom = (int *) alloca (n_basic_blocks * sizeof (int));
- memset ((void *)idom, -1, (size_t)n_basic_blocks * sizeof (int));
- calculate_dominance_info (idom, NULL, CDI_DOMINATORS);
+ idom = calculate_dominance_info (CDI_DOMINATORS);
if (rtl_dump_file)
{
- int i;
fputs (";; Immediate Dominators:\n", rtl_dump_file);
- for (i = 0; i < n_basic_blocks; ++i)
- fprintf (rtl_dump_file, ";\t%3d = %3d\n", i, idom[i]);
+ FOR_EACH_BB (bb)
+ fprintf (rtl_dump_file, ";\t%3d = %3d\n", bb->index,
+ get_immediate_dominator (idom, bb)->index);
fflush (rtl_dump_file);
}
/* Compute dominance frontiers. */
- dfs = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
+ dfs = sbitmap_vector_alloc (last_basic_block, last_basic_block);
compute_dominance_frontiers (dfs, idom);
if (rtl_dump_file)
{
dump_sbitmap_vector (rtl_dump_file, ";; Dominance Frontiers:",
- "; Basic Block", dfs, n_basic_blocks);
+ "; Basic Block", dfs, last_basic_block);
fflush (rtl_dump_file);
}
/* Compute register evaluations. */
- ssa_max_reg_num = max_reg_num();
+ ssa_max_reg_num = max_reg_num ();
nregs = ssa_max_reg_num;
- evals = sbitmap_vector_alloc (nregs, n_basic_blocks);
+ evals = sbitmap_vector_alloc (nregs, last_basic_block);
find_evaluations (evals, nregs);
/* Compute the iterated dominance frontier for each register. */
- idfs = sbitmap_vector_alloc (nregs, n_basic_blocks);
+ idfs = sbitmap_vector_alloc (nregs, last_basic_block);
compute_iterated_dominance_frontiers (idfs, dfs, evals, nregs);
if (rtl_dump_file)
in_ssa_form = 1;
reg_scan (get_insns (), max_reg_num (), 1);
+ free_dominance_info (idom);
}
/* REG is the representative temporary of its partition. Add it to the
EXECUTE_IF_SET_IN_SBITMAP (succ[t], 0, s,
{
if (! TEST_BIT (visited, s))
- tstack = ephi_forward (s, visited, succ, tstack);
+ tstack = ephi_forward (s, visited, succ, tstack);
});
*tstack++ = t;
int p;
/* Iterate through the predecessor list looking for unvisited nodes.
- If there are any, we have a cycle, and must deal with that. At
+ If there are any, we have a cycle, and must deal with that. At
the same time, look for a visited predecessor. If there is one,
we won't need to create a temporary. */
emit_move_insn (nodes[p], reg_u);
}
});
- }
- else
+ }
+ else
{
/* No cycle. Just copy the value from a successor. */
if (n_nodes == 0)
return;
- /* Build the auxiliary graph R(B).
+ /* Build the auxiliary graph R(B).
The nodes of the graph are the members of the register partition
present in Phi(B). There is an edge from FIND(T0)->FIND(T1) for
reg = *preg;
if (GET_CODE (reg) != REG || GET_CODE (tgt) != REG)
- abort();
+ abort ();
reg = regno_reg_rtx[partition_find (reg_partition, REGNO (reg))];
tgt = regno_reg_rtx[partition_find (reg_partition, REGNO (tgt))];
- /* If the two registers are already in the same partition,
+ /* If the two registers are already in the same partition,
nothing will need to be done. */
if (reg != tgt)
{
sbitmap_zero (visited);
- /* As we find a solution to the tsort, collect the implementation
+ /* As we find a solution to the tsort, collect the implementation
insns in a sequence. */
start_sequence ();
-
+
while (tstack != stack)
{
i = *--tstack;
ephi_create (i, visited, pred, succ, nodes);
}
- insn = gen_sequence ();
+ insn = get_insns ();
end_sequence ();
insert_insn_on_edge (insn, e);
if (rtl_dump_file)
/* For basic block B, consider all phi insns which provide an
alternative corresponding to an incoming abnormal critical edge.
Place the phi alternative corresponding to that abnormal critical
- edge in the same register class as the destination of the set.
+ edge in the same register class as the destination of the set.
From Morgan, p. 178:
- For each abnormal critical edge (C, B),
- if T0 = phi (T1, ..., Ti, ..., Tm) is a phi node in B,
- and C is the ith predecessor of B,
- then T0 and Ti must be equivalent.
+ For each abnormal critical edge (C, B),
+ if T0 = phi (T1, ..., Ti, ..., Tm) is a phi node in B,
+ and C is the ith predecessor of B,
+ then T0 and Ti must be equivalent.
- Return non-zero iff any such cases were found for which the two
+ Return nonzero iff any such cases were found for which the two
regs were not already in the same class. */
static int
phi = first_insn_after_basic_block_note (b);
/* Scan all the phi nodes. */
- for (;
+ for (;
PHI_NODE_P (phi);
phi = next_nonnote_insn (phi))
{
rtx tgt = SET_DEST (set);
/* The set target is expected to be an SSA register. */
- if (GET_CODE (tgt) != REG
+ if (GET_CODE (tgt) != REG
|| !CONVERT_REGISTER_TO_SSA_P (REGNO (tgt)))
abort ();
tgt_regno = REGNO (tgt);
continue;
/* The phi alternative is expected to be an SSA register. */
- if (GET_CODE (*alt) != REG
+ if (GET_CODE (*alt) != REG
|| !CONVERT_REGISTER_TO_SSA_P (REGNO (*alt)))
abort ();
alt_regno = REGNO (*alt);
/* If the set destination and the phi alternative aren't
already in the same class... */
- if (partition_find (reg_partition, tgt_regno)
+ if (partition_find (reg_partition, tgt_regno)
!= partition_find (reg_partition, alt_regno))
{
/* ... make them such. */
/* It is illegal to unify a hard register with a
different register. */
abort ();
-
- partition_union (reg_partition,
+
+ partition_union (reg_partition,
tgt_regno, alt_regno);
++changed;
}
phi = first_insn_after_basic_block_note (b);
/* Scan all the phi nodes. */
- for (;
+ for (;
PHI_NODE_P (phi);
phi = next_nonnote_insn (phi))
{
rtx set2 = PATTERN (phi2);
/* The regno of the destination of the set. */
int tgt2_regno = REGNO (SET_DEST (set2));
-
+
/* Are the set destinations equivalent regs? */
if (partition_find (reg_partition, tgt_regno) ==
partition_find (reg_partition, tgt2_regno))
/* If the alternatives aren't already in the same
class ... */
- if (partition_find (reg_partition, REGNO (*alt))
+ if (partition_find (reg_partition, REGNO (*alt))
!= partition_find (reg_partition, REGNO (*alt2)))
{
/* ... make them so. */
a different register. */
abort ();
- partition_union (reg_partition,
+ partition_union (reg_partition,
REGNO (*alt), REGNO (*alt2));
++changed;
}
static partition
compute_conservative_reg_partition ()
{
- int bb;
+ basic_block bb;
int changed = 0;
/* We don't actually work with hard registers, but it's easier to
carry them around anyway rather than constantly doing register
number arithmetic. */
- partition p =
+ partition p =
partition_new (ssa_definition->num_elements);
/* The first priority is to make sure registers that might have to
be copied on abnormal critical edges are placed in the same
partition. This saves us from having to split abnormal critical
edges. */
- for (bb = n_basic_blocks; --bb >= 0; )
- changed += make_regs_equivalent_over_bad_edges (bb, p);
-
+ FOR_EACH_BB_REVERSE (bb)
+ changed += make_regs_equivalent_over_bad_edges (bb->index, p);
+
/* Now we have to insure that corresponding arguments of phi nodes
assigning to corresponding regs are equivalent. Iterate until
nothing changes. */
while (changed > 0)
{
changed = 0;
- for (bb = n_basic_blocks; --bb >= 0; )
- changed += make_equivalent_phi_alternatives_equivalent (bb, p);
+ FOR_EACH_BB_REVERSE (bb)
+ changed += make_equivalent_phi_alternatives_equivalent (bb->index, p);
}
return p;
abnormal critical edges (which isn't possible).
2. Figure out which regs are involved (in the LHS or RHS) of
- copies and phi nodes. Compute conflicts among these regs.
+ copies and phi nodes. Compute conflicts among these regs.
3. Walk around the instruction stream, placing two regs in the
same class of the partition if one appears on the LHS and the
other on the RHS of a copy or phi node and the two regs don't
conflict. The conflict information of course needs to be
- updated.
+ updated.
4. If anything has changed, there may be new opportunities to
coalesce regs, so go back to 2.
/* If REG1 and REG2 don't conflict in CONFLICTS, place them in the
same class of partition P, if they aren't already. Update
- CONFLICTS appropriately.
+ CONFLICTS appropriately.
Returns one if REG1 and REG2 were placed in the same class but were
- not previously; zero otherwise.
+ not previously; zero otherwise.
See Morgan figure 11.15. */
-static int
+static int
coalesce_if_unconflicting (p, conflicts, reg1, reg2)
partition p;
conflict_graph conflicts;
REG2. */
reg1 = partition_find (p, reg1);
reg2 = partition_find (p, reg2);
-
+
/* If they're already in the same class, there's nothing to do. */
if (reg1 == reg2)
return 0;
/* Find the new canonical reg for the merged class. */
reg = partition_find (p, reg1);
-
+
/* Merge conflicts from the two previous classes. */
conflict_graph_merge_regs (conflicts, reg, reg1);
conflict_graph_merge_regs (conflicts, reg, reg2);
/* Coalesce only if the reg modes are the same. As long as
each reg's rtx is unique, it can have only one mode, so two
- pseudos of different modes can't be coalesced into one.
+ pseudos of different modes can't be coalesced into one.
FIXME: We can probably get around this by inserting SUBREGs
where appropriate, but for now we don't bother. */
/* Found a copy; see if we can use the same reg for both the
source and destination (and thus eliminate the copy,
ultimately). */
- changed += coalesce_if_unconflicting (p, conflicts,
+ changed += coalesce_if_unconflicting (p, conflicts,
REGNO (src), REGNO (dest));
}
/* Callback function for for_each_successor_phi. If the set
destination and the phi alternative regs do not conflict, place
- them in the same paritition class. DATA is a pointer to a
+ them in the same partition class. DATA is a pointer to a
phi_coalesce_context struct. */
static int
int src_regno;
void *data;
{
- struct phi_coalesce_context *context =
+ struct phi_coalesce_context *context =
(struct phi_coalesce_context *) data;
-
+
/* Attempt to use the same reg, if they don't conflict. */
- context->changed
- += coalesce_if_unconflicting (context->p, context->conflicts,
+ context->changed
+ += coalesce_if_unconflicting (context->p, context->conflicts,
dest_regno, src_regno);
return 0;
}
/* For each alternative in a phi function corresponding to basic block
BB (in phi nodes in successor block to BB), place the reg in the
phi alternative and the reg to which the phi value is set into the
- same class in partition P, if allowed by CONFLICTS.
+ same class in partition P, if allowed by CONFLICTS.
Return the number of changes that were made to P.
-
+
See Morgan figure 11.14. */
static int
}
/* Compute and return a partition of pseudos. Where possible,
- non-conflicting pseudos are placed in the same class.
+ non-conflicting pseudos are placed in the same class.
The caller is responsible for deallocating the returned partition. */
static partition
compute_coalesced_reg_partition ()
{
- int bb;
+ basic_block bb;
int changed = 0;
+ regset_head phi_set_head;
+ regset phi_set = &phi_set_head;
- partition p =
+ partition p =
partition_new (ssa_definition->num_elements);
/* The first priority is to make sure registers that might have to
be copied on abnormal critical edges are placed in the same
partition. This saves us from having to split abnormal critical
edges (which can't be done). */
- for (bb = n_basic_blocks; --bb >= 0; )
- make_regs_equivalent_over_bad_edges (bb, p);
+ FOR_EACH_BB_REVERSE (bb)
+ make_regs_equivalent_over_bad_edges (bb->index, p);
+
+ INIT_REG_SET (phi_set);
do
{
- regset_head phi_set;
conflict_graph conflicts;
changed = 0;
/* Build the set of registers involved in phi nodes, either as
arguments to the phi function or as the target of a set. */
- INITIALIZE_REG_SET (phi_set);
- mark_phi_and_copy_regs (&phi_set);
+ CLEAR_REG_SET (phi_set);
+ mark_phi_and_copy_regs (phi_set);
/* Compute conflicts. */
- conflicts = conflict_graph_compute (&phi_set, p);
+ conflicts = conflict_graph_compute (phi_set, p);
/* FIXME: Better would be to process most frequently executed
blocks first, so that most frequently executed copies would
be more likely to be removed by register coalescing. But any
order will generate correct, if non-optimal, results. */
- for (bb = n_basic_blocks; --bb >= 0; )
+ FOR_EACH_BB_REVERSE (bb)
{
- basic_block block = BASIC_BLOCK (bb);
- changed += coalesce_regs_in_copies (block, p, conflicts);
- changed +=
- coalesce_regs_in_successor_phi_nodes (block, p, conflicts);
+ changed += coalesce_regs_in_copies (bb, p, conflicts);
+ changed +=
+ coalesce_regs_in_successor_phi_nodes (bb, p, conflicts);
}
conflict_graph_delete (conflicts);
}
while (changed > 0);
+ FREE_REG_SET (phi_set);
+
return p;
}
unsigned int new_regno = partition_find (reg_partition, regno);
rtx canonical_element_rtx = ssa_rename_from_lookup (new_regno);
- if (canonical_element_rtx != NULL_RTX &&
+ if (canonical_element_rtx != NULL_RTX &&
HARD_REGISTER_P (canonical_element_rtx))
{
if (REGNO (canonical_element_rtx) != regno)
((struct ssa_rename_from_hash_table_data *) data)->canonical_elements;
partition reg_partition =
((struct ssa_rename_from_hash_table_data *) data)->reg_partition;
-
+
SET_BIT (canonical_elements, partition_find (reg_partition, reg));
return 1;
}
rename_equivalent_regs (reg_partition)
partition reg_partition;
{
- int bb;
+ basic_block b;
- for (bb = n_basic_blocks; --bb >= 0; )
+ FOR_EACH_BB_REVERSE (b)
{
- basic_block b = BASIC_BLOCK (bb);
rtx next = b->head;
rtx last = b->end;
rtx insn;
insn = next;
if (INSN_P (insn))
{
- for_each_rtx (&PATTERN (insn),
- rename_equivalent_regs_in_insn,
+ for_each_rtx (&PATTERN (insn),
+ rename_equivalent_regs_in_insn,
reg_partition);
- for_each_rtx (®_NOTES (insn),
- rename_equivalent_regs_in_insn,
+ for_each_rtx (®_NOTES (insn),
+ rename_equivalent_regs_in_insn,
reg_partition);
if (GET_CODE (PATTERN (insn)) == SEQUENCE)
int i;
if (slen <= 1)
- abort();
+ abort ();
PATTERN (insn) = XVECEXP (s, 0, slen-1);
for (i = 0; i < slen - 1; i++)
/* The main entry point for moving from SSA. */
void
-convert_from_ssa()
+convert_from_ssa ()
{
- int bb;
+ basic_block b, bb;
partition reg_partition;
rtx insns = get_insns ();
/* Need global_live_at_{start,end} up to date. There should not be
any significant dead code at this point, except perhaps dead
- stores. So do not take the time to perform dead code elimination.
+ stores. So do not take the time to perform dead code elimination.
Register coalescing needs death notes, so generate them. */
life_analysis (insns, NULL, PROP_DEATH_NOTES);
rename_equivalent_regs (reg_partition);
/* Eliminate the PHI nodes. */
- for (bb = n_basic_blocks; --bb >= 0; )
+ FOR_EACH_BB_REVERSE (b)
{
- basic_block b = BASIC_BLOCK (bb);
edge e;
for (e = b->pred; e; e = e->pred_next)
partition_delete (reg_partition);
/* Actually delete the PHI nodes. */
- for (bb = n_basic_blocks; --bb >= 0; )
+ FOR_EACH_BB_REVERSE (bb)
{
- rtx insn = BLOCK_HEAD (bb);
+ rtx insn = bb->head;
while (1)
{
/* If this is a PHI node delete it. */
if (PHI_NODE_P (insn))
{
- if (insn == BLOCK_END (bb))
- BLOCK_END (bb) = PREV_INSN (insn);
+ if (insn == bb->end)
+ bb->end = PREV_INSN (insn);
insn = delete_insn (insn);
}
/* Since all the phi nodes come at the beginning of the
else if (INSN_P (insn))
break;
/* If we've reached the end of the block, stop. */
- else if (insn == BLOCK_END (bb))
+ else if (insn == bb->end)
break;
- else
+ else
insn = NEXT_INSN (insn);
}
}
count_or_remove_death_notes (NULL, 1);
/* Deallocate the data structures. */
- VARRAY_FREE (ssa_definition);
+ ssa_definition = 0;
ssa_rename_from_free ();
}
destination, the regno of the phi argument corresponding to BB,
and DATA.
- If FN ever returns non-zero, stops immediately and returns this
+ If FN ever returns nonzero, stops immediately and returns this
value. Otherwise, returns zero. */
int
void *data;
{
edge e;
-
+
if (bb == EXIT_BLOCK_PTR)
return 0;
rtx insn;
basic_block successor = e->dest;
- if (successor == ENTRY_BLOCK_PTR
+ if (successor == ENTRY_BLOCK_PTR
|| successor == EXIT_BLOCK_PTR)
continue;
rtx phi_set = PATTERN (insn);
rtx *alternative = phi_alternative (phi_set, bb->index);
rtx phi_src;
-
+
/* This phi function may not have an alternative
corresponding to the incoming edge, indicating the
assigned variable is not defined along the edge. */
phi_src = *alternative;
/* Invoke the callback. */
- result = (*fn) (insn, REGNO (SET_DEST (phi_set)),
+ result = (*fn) (insn, REGNO (SET_DEST (phi_set)),
REGNO (phi_src), data);
/* Terminate if requested. */
return 1;
if (!HARD_REGISTER_NUM_P (orig_reg1) && HARD_REGISTER_NUM_P (orig_reg2))
return 1;
-
+
return 0;
}
OSDN Git Service
