/* Global common subexpression elimination/Partial redundancy elimination
and global constant/copy propagation for GNU compiler.
- Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003
+ Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005
Free Software Foundation, Inc.
This file is part of GCC.
#include "toplev.h"
#include "rtl.h"
+#include "tree.h"
#include "tm_p.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "ggc.h"
#include "params.h"
#include "cselib.h"
-
+#include "intl.h"
#include "obstack.h"
+#include "timevar.h"
/* Propagate flow information through back edges and thus enable PRE's
moving loop invariant calculations out of loops.
3) Perform copy/constant propagation.
- 4) Perform global cse.
+ 4) Perform global cse using lazy code motion if not optimizing
+ for size, or code hoisting if we are.
5) Perform another pass of copy/constant propagation.
substitutions.
PRE is quite expensive in complicated functions because the DFA can take
- awhile to converge. Hence we only perform one pass. The parameter max-gcse-passes can
- be modified if one wants to experiment.
+ a while to converge. Hence we only perform one pass. The parameter
+ max-gcse-passes can be modified if one wants to experiment.
**********************
the result of the expression is copied to a new register, and the redundant
expression is deleted by replacing it with this new register. Classic GCSE
doesn't have this problem as much as it computes the reaching defs of
- each register in each block and thus can try to use an existing register.
-
- **********************
-
- A fair bit of simplicity is created by creating small functions for simple
- tasks, even when the function is only called in one place. This may
- measurably slow things down [or may not] by creating more function call
- overhead than is necessary. The source is laid out so that it's trivial
- to make the affected functions inline so that one can measure what speed
- up, if any, can be achieved, and maybe later when things settle things can
- be rearranged.
-
- Help stamp out big monolithic functions! */
+ each register in each block and thus can try to use an existing
+ register. */
\f
/* GCSE global vars. */
* If we changed any jumps via cprop.
* If we added any labels via edge splitting. */
-
static int run_jump_opt_after_gcse;
/* Bitmaps are normally not included in debugging dumps.
/* Get the cuid of an insn. */
#ifdef ENABLE_CHECKING
-#define INSN_CUID(INSN) (INSN_UID (INSN) > max_uid ? (abort (), 0) : uid_cuid[INSN_UID (INSN)])
+#define INSN_CUID(INSN) \
+ (gcc_assert (INSN_UID (INSN) <= max_uid), uid_cuid[INSN_UID (INSN)])
#else
#define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
#endif
{
/* The next setting of this register. */
struct reg_set *next;
- /* The insn where it was set. */
- rtx insn;
+ /* The index of the block where it was set. */
+ int bb_index;
} reg_set;
static reg_set **reg_set_table;
/* This is a list of expressions which are MEMs and will be used by load
or store motion.
Load motion tracks MEMs which aren't killed by
- anything except itself. (ie, loads and stores to a single location).
+ anything except itself. (i.e., loads and stores to a single location).
We can then allow movement of these MEM refs with a little special
allowance. (all stores copy the same value to the reaching reg used
for the loads). This means all values used to store into memory must have
struct ls_expr * next; /* Next in the list. */
int invalid; /* Invalid for some reason. */
int index; /* If it maps to a bitmap index. */
- int hash_index; /* Index when in a hash table. */
+ unsigned int hash_index; /* Index when in a hash table. */
rtx reaching_reg; /* Register to use when re-writing. */
};
static regset reg_set_bitmap;
/* For each block, a bitmap of registers set in the block.
- This is used by expr_killed_p and compute_transp.
+ This is used by compute_transp.
It is computed during hash table computation and not by compute_sets
as it includes registers added since the last pass (or between cprop and
gcse) and it's currently not easy to realloc sbitmap vectors. */
/* Array, indexed by basic block number for a list of insns which modify
memory within that block. */
static rtx * modify_mem_list;
-bitmap modify_mem_list_set;
+static bitmap modify_mem_list_set;
/* This array parallels modify_mem_list, but is kept canonicalized. */
static rtx * canon_modify_mem_list;
-bitmap canon_modify_mem_list_set;
+
+/* Bitmap indexed by block numbers to record which blocks contain
+ function calls. */
+static bitmap blocks_with_calls;
+
/* Various variables for statistics gathering. */
/* Memory used in a pass.
static int gcse_subst_count;
/* Number of copy instructions created. */
static int gcse_create_count;
-/* Number of constants propagated. */
-static int const_prop_count;
-/* Number of copys propagated. */
-static int copy_prop_count;
+/* Number of local constants propagated. */
+static int local_const_prop_count;
+/* Number of local copys propagated. */
+static int local_copy_prop_count;
+/* Number of global constants propagated. */
+static int global_const_prop_count;
+/* Number of global copys propagated. */
+static int global_copy_prop_count;
\f
-/* These variables are used by classic GCSE.
- Normally they'd be defined a bit later, but `rd_gen' needs to
- be declared sooner. */
-
-/* Each block has a bitmap of each type.
- The length of each blocks bitmap is:
-
- max_cuid - for reaching definitions
- n_exprs - for available expressions
-
- Thus we view the bitmaps as 2 dimensional arrays. i.e.
- rd_kill[block_num][cuid_num]
- ae_kill[block_num][expr_num] */
-
-/* For reaching defs */
-static sbitmap *rd_kill, *rd_gen, *reaching_defs, *rd_out;
-
-/* for available exprs */
-static sbitmap *ae_kill, *ae_gen, *ae_in, *ae_out;
-
-/* Objects of this type are passed around by the null-pointer check
- removal routines. */
-struct null_pointer_info
-{
- /* The basic block being processed. */
- basic_block current_block;
- /* The first register to be handled in this pass. */
- unsigned int min_reg;
- /* One greater than the last register to be handled in this pass. */
- unsigned int max_reg;
- sbitmap *nonnull_local;
- sbitmap *nonnull_killed;
-};
+/* For available exprs */
+static sbitmap *ae_kill, *ae_gen;
\f
-static void compute_can_copy PARAMS ((void));
-static char *gmalloc PARAMS ((unsigned int));
-static char *grealloc PARAMS ((char *, unsigned int));
-static char *gcse_alloc PARAMS ((unsigned long));
-static void alloc_gcse_mem PARAMS ((rtx));
-static void free_gcse_mem PARAMS ((void));
-static void alloc_reg_set_mem PARAMS ((int));
-static void free_reg_set_mem PARAMS ((void));
-static int get_bitmap_width PARAMS ((int, int, int));
-static void record_one_set PARAMS ((int, rtx));
-static void record_set_info PARAMS ((rtx, rtx, void *));
-static void compute_sets PARAMS ((rtx));
-static void hash_scan_insn PARAMS ((rtx, struct hash_table *, int));
-static void hash_scan_set PARAMS ((rtx, rtx, struct hash_table *));
-static void hash_scan_clobber PARAMS ((rtx, rtx, struct hash_table *));
-static void hash_scan_call PARAMS ((rtx, rtx, struct hash_table *));
-static int want_to_gcse_p PARAMS ((rtx));
-static bool gcse_constant_p PARAMS ((rtx));
-static int oprs_unchanged_p PARAMS ((rtx, rtx, int));
-static int oprs_anticipatable_p PARAMS ((rtx, rtx));
-static int oprs_available_p PARAMS ((rtx, rtx));
-static void insert_expr_in_table PARAMS ((rtx, enum machine_mode, rtx,
- int, int, struct hash_table *));
-static void insert_set_in_table PARAMS ((rtx, rtx, struct hash_table *));
-static unsigned int hash_expr PARAMS ((rtx, enum machine_mode, int *, int));
-static unsigned int hash_expr_1 PARAMS ((rtx, enum machine_mode, int *));
-static unsigned int hash_string_1 PARAMS ((const char *));
-static unsigned int hash_set PARAMS ((int, int));
-static int expr_equiv_p PARAMS ((rtx, rtx));
-static void record_last_reg_set_info PARAMS ((rtx, int));
-static void record_last_mem_set_info PARAMS ((rtx));
-static void record_last_set_info PARAMS ((rtx, rtx, void *));
-static void compute_hash_table PARAMS ((struct hash_table *));
-static void alloc_hash_table PARAMS ((int, struct hash_table *, int));
-static void free_hash_table PARAMS ((struct hash_table *));
-static void compute_hash_table_work PARAMS ((struct hash_table *));
-static void dump_hash_table PARAMS ((FILE *, const char *,
- struct hash_table *));
-static struct expr *lookup_expr PARAMS ((rtx, struct hash_table *));
-static struct expr *lookup_set PARAMS ((unsigned int, struct hash_table *));
-static struct expr *next_set PARAMS ((unsigned int, struct expr *));
-static void reset_opr_set_tables PARAMS ((void));
-static int oprs_not_set_p PARAMS ((rtx, rtx));
-static void mark_call PARAMS ((rtx));
-static void mark_set PARAMS ((rtx, rtx));
-static void mark_clobber PARAMS ((rtx, rtx));
-static void mark_oprs_set PARAMS ((rtx));
-static void alloc_cprop_mem PARAMS ((int, int));
-static void free_cprop_mem PARAMS ((void));
-static void compute_transp PARAMS ((rtx, int, sbitmap *, int));
-static void compute_transpout PARAMS ((void));
-static void compute_local_properties PARAMS ((sbitmap *, sbitmap *, sbitmap *,
- struct hash_table *));
-static void compute_cprop_data PARAMS ((void));
-static void find_used_regs PARAMS ((rtx *, void *));
-static int try_replace_reg PARAMS ((rtx, rtx, rtx));
-static struct expr *find_avail_set PARAMS ((int, rtx));
-static int cprop_jump PARAMS ((basic_block, rtx, rtx, rtx, rtx));
-static void mems_conflict_for_gcse_p PARAMS ((rtx, rtx, void *));
-static int load_killed_in_block_p PARAMS ((basic_block, int, rtx, int));
-static void canon_list_insert PARAMS ((rtx, rtx, void *));
-static int cprop_insn PARAMS ((rtx, int));
-static int cprop PARAMS ((int));
-static void find_implicit_sets PARAMS ((void));
-static int one_cprop_pass PARAMS ((int, int, int));
-static bool constprop_register PARAMS ((rtx, rtx, rtx, int));
-static struct expr *find_bypass_set PARAMS ((int, int));
-static bool reg_killed_on_edge PARAMS ((rtx, edge));
-static int bypass_block PARAMS ((basic_block, rtx, rtx));
-static int bypass_conditional_jumps PARAMS ((void));
-static void alloc_pre_mem PARAMS ((int, int));
-static void free_pre_mem PARAMS ((void));
-static void compute_pre_data PARAMS ((void));
-static int pre_expr_reaches_here_p PARAMS ((basic_block, struct expr *,
- basic_block));
-static void insert_insn_end_bb PARAMS ((struct expr *, basic_block, int));
-static void pre_insert_copy_insn PARAMS ((struct expr *, rtx));
-static void pre_insert_copies PARAMS ((void));
-static int pre_delete PARAMS ((void));
-static int pre_gcse PARAMS ((void));
-static int one_pre_gcse_pass PARAMS ((int));
-static void add_label_notes PARAMS ((rtx, rtx));
-static void alloc_code_hoist_mem PARAMS ((int, int));
-static void free_code_hoist_mem PARAMS ((void));
-static void compute_code_hoist_vbeinout PARAMS ((void));
-static void compute_code_hoist_data PARAMS ((void));
-static int hoist_expr_reaches_here_p PARAMS ((basic_block, int, basic_block,
- char *));
-static void hoist_code PARAMS ((void));
-static int one_code_hoisting_pass PARAMS ((void));
-static void alloc_rd_mem PARAMS ((int, int));
-static void free_rd_mem PARAMS ((void));
-static void handle_rd_kill_set PARAMS ((rtx, int, basic_block));
-static void compute_kill_rd PARAMS ((void));
-static void compute_rd PARAMS ((void));
-static void alloc_avail_expr_mem PARAMS ((int, int));
-static void free_avail_expr_mem PARAMS ((void));
-static void compute_ae_gen PARAMS ((struct hash_table *));
-static int expr_killed_p PARAMS ((rtx, basic_block));
-static void compute_ae_kill PARAMS ((sbitmap *, sbitmap *, struct hash_table *));
-static int expr_reaches_here_p PARAMS ((struct occr *, struct expr *,
- basic_block, int));
-static rtx computing_insn PARAMS ((struct expr *, rtx));
-static int def_reaches_here_p PARAMS ((rtx, rtx));
-static int can_disregard_other_sets PARAMS ((struct reg_set **, rtx, int));
-static int handle_avail_expr PARAMS ((rtx, struct expr *));
-static int classic_gcse PARAMS ((void));
-static int one_classic_gcse_pass PARAMS ((int));
-static void invalidate_nonnull_info PARAMS ((rtx, rtx, void *));
-static int delete_null_pointer_checks_1 PARAMS ((unsigned int *,
- sbitmap *, sbitmap *,
- struct null_pointer_info *));
-static rtx process_insert_insn PARAMS ((struct expr *));
-static int pre_edge_insert PARAMS ((struct edge_list *, struct expr **));
-static int expr_reaches_here_p_work PARAMS ((struct occr *, struct expr *,
- basic_block, int, char *));
-static int pre_expr_reaches_here_p_work PARAMS ((basic_block, struct expr *,
- basic_block, char *));
-static struct ls_expr * ldst_entry PARAMS ((rtx));
-static void free_ldst_entry PARAMS ((struct ls_expr *));
-static void free_ldst_mems PARAMS ((void));
-static void print_ldst_list PARAMS ((FILE *));
-static struct ls_expr * find_rtx_in_ldst PARAMS ((rtx));
-static int enumerate_ldsts PARAMS ((void));
-static inline struct ls_expr * first_ls_expr PARAMS ((void));
-static inline struct ls_expr * next_ls_expr PARAMS ((struct ls_expr *));
-static int simple_mem PARAMS ((rtx));
-static void invalidate_any_buried_refs PARAMS ((rtx));
-static void compute_ld_motion_mems PARAMS ((void));
-static void trim_ld_motion_mems PARAMS ((void));
-static void update_ld_motion_stores PARAMS ((struct expr *));
-static void reg_set_info PARAMS ((rtx, rtx, void *));
-static bool store_ops_ok PARAMS ((rtx, int *));
-static rtx extract_mentioned_regs PARAMS ((rtx));
-static rtx extract_mentioned_regs_helper PARAMS ((rtx, rtx));
-static void find_moveable_store PARAMS ((rtx, int *, int *));
-static int compute_store_table PARAMS ((void));
-static bool load_kills_store PARAMS ((rtx, rtx));
-static bool find_loads PARAMS ((rtx, rtx));
-static bool store_killed_in_insn PARAMS ((rtx, rtx, rtx));
-static bool store_killed_after PARAMS ((rtx, rtx, rtx, basic_block,
- int *, rtx *));
-static bool store_killed_before PARAMS ((rtx, rtx, rtx, basic_block,
- int *));
-static void build_store_vectors PARAMS ((void));
-static void insert_insn_start_bb PARAMS ((rtx, basic_block));
-static int insert_store PARAMS ((struct ls_expr *, edge));
-static void replace_store_insn PARAMS ((rtx, rtx, basic_block));
-static void delete_store PARAMS ((struct ls_expr *,
- basic_block));
-static void free_store_memory PARAMS ((void));
-static void store_motion PARAMS ((void));
-static void free_insn_expr_list_list PARAMS ((rtx *));
-static void clear_modify_mem_tables PARAMS ((void));
-static void free_modify_mem_tables PARAMS ((void));
-static rtx gcse_emit_move_after PARAMS ((rtx, rtx, rtx));
-static void local_cprop_find_used_regs PARAMS ((rtx *, void *));
-static bool do_local_cprop PARAMS ((rtx, rtx, int, rtx*));
-static bool adjust_libcall_notes PARAMS ((rtx, rtx, rtx, rtx*));
-static void local_cprop_pass PARAMS ((int));
+static void compute_can_copy (void);
+static void *gmalloc (size_t) ATTRIBUTE_MALLOC;
+static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC;
+static void *grealloc (void *, size_t);
+static void *gcse_alloc (unsigned long);
+static void alloc_gcse_mem (rtx);
+static void free_gcse_mem (void);
+static void alloc_reg_set_mem (int);
+static void free_reg_set_mem (void);
+static void record_one_set (int, rtx);
+static void record_set_info (rtx, rtx, void *);
+static void compute_sets (rtx);
+static void hash_scan_insn (rtx, struct hash_table *, int);
+static void hash_scan_set (rtx, rtx, struct hash_table *);
+static void hash_scan_clobber (rtx, rtx, struct hash_table *);
+static void hash_scan_call (rtx, rtx, struct hash_table *);
+static int want_to_gcse_p (rtx);
+static bool can_assign_to_reg_p (rtx);
+static bool gcse_constant_p (rtx);
+static int oprs_unchanged_p (rtx, rtx, int);
+static int oprs_anticipatable_p (rtx, rtx);
+static int oprs_available_p (rtx, rtx);
+static void insert_expr_in_table (rtx, enum machine_mode, rtx, int, int,
+ struct hash_table *);
+static void insert_set_in_table (rtx, rtx, struct hash_table *);
+static unsigned int hash_expr (rtx, enum machine_mode, int *, int);
+static unsigned int hash_set (int, int);
+static int expr_equiv_p (rtx, rtx);
+static void record_last_reg_set_info (rtx, int);
+static void record_last_mem_set_info (rtx);
+static void record_last_set_info (rtx, rtx, void *);
+static void compute_hash_table (struct hash_table *);
+static void alloc_hash_table (int, struct hash_table *, int);
+static void free_hash_table (struct hash_table *);
+static void compute_hash_table_work (struct hash_table *);
+static void dump_hash_table (FILE *, const char *, struct hash_table *);
+static struct expr *lookup_set (unsigned int, struct hash_table *);
+static struct expr *next_set (unsigned int, struct expr *);
+static void reset_opr_set_tables (void);
+static int oprs_not_set_p (rtx, rtx);
+static void mark_call (rtx);
+static void mark_set (rtx, rtx);
+static void mark_clobber (rtx, rtx);
+static void mark_oprs_set (rtx);
+static void alloc_cprop_mem (int, int);
+static void free_cprop_mem (void);
+static void compute_transp (rtx, int, sbitmap *, int);
+static void compute_transpout (void);
+static void compute_local_properties (sbitmap *, sbitmap *, sbitmap *,
+ struct hash_table *);
+static void compute_cprop_data (void);
+static void find_used_regs (rtx *, void *);
+static int try_replace_reg (rtx, rtx, rtx);
+static struct expr *find_avail_set (int, rtx);
+static int cprop_jump (basic_block, rtx, rtx, rtx, rtx);
+static void mems_conflict_for_gcse_p (rtx, rtx, void *);
+static int load_killed_in_block_p (basic_block, int, rtx, int);
+static void canon_list_insert (rtx, rtx, void *);
+static int cprop_insn (rtx, int);
+static int cprop (int);
+static void find_implicit_sets (void);
+static int one_cprop_pass (int, int, int);
+static bool constprop_register (rtx, rtx, rtx, int);
+static struct expr *find_bypass_set (int, int);
+static bool reg_killed_on_edge (rtx, edge);
+static int bypass_block (basic_block, rtx, rtx);
+static int bypass_conditional_jumps (void);
+static void alloc_pre_mem (int, int);
+static void free_pre_mem (void);
+static void compute_pre_data (void);
+static int pre_expr_reaches_here_p (basic_block, struct expr *,
+ basic_block);
+static void insert_insn_end_bb (struct expr *, basic_block, int);
+static void pre_insert_copy_insn (struct expr *, rtx);
+static void pre_insert_copies (void);
+static int pre_delete (void);
+static int pre_gcse (void);
+static int one_pre_gcse_pass (int);
+static void add_label_notes (rtx, rtx);
+static void alloc_code_hoist_mem (int, int);
+static void free_code_hoist_mem (void);
+static void compute_code_hoist_vbeinout (void);
+static void compute_code_hoist_data (void);
+static int hoist_expr_reaches_here_p (basic_block, int, basic_block, char *);
+static void hoist_code (void);
+static int one_code_hoisting_pass (void);
+static rtx process_insert_insn (struct expr *);
+static int pre_edge_insert (struct edge_list *, struct expr **);
+static int pre_expr_reaches_here_p_work (basic_block, struct expr *,
+ basic_block, char *);
+static struct ls_expr * ldst_entry (rtx);
+static void free_ldst_entry (struct ls_expr *);
+static void free_ldst_mems (void);
+static void print_ldst_list (FILE *);
+static struct ls_expr * find_rtx_in_ldst (rtx);
+static int enumerate_ldsts (void);
+static inline struct ls_expr * first_ls_expr (void);
+static inline struct ls_expr * next_ls_expr (struct ls_expr *);
+static int simple_mem (rtx);
+static void invalidate_any_buried_refs (rtx);
+static void compute_ld_motion_mems (void);
+static void trim_ld_motion_mems (void);
+static void update_ld_motion_stores (struct expr *);
+static void reg_set_info (rtx, rtx, void *);
+static void reg_clear_last_set (rtx, rtx, void *);
+static bool store_ops_ok (rtx, int *);
+static rtx extract_mentioned_regs (rtx);
+static rtx extract_mentioned_regs_helper (rtx, rtx);
+static void find_moveable_store (rtx, int *, int *);
+static int compute_store_table (void);
+static bool load_kills_store (rtx, rtx, int);
+static bool find_loads (rtx, rtx, int);
+static bool store_killed_in_insn (rtx, rtx, rtx, int);
+static bool store_killed_after (rtx, rtx, rtx, basic_block, int *, rtx *);
+static bool store_killed_before (rtx, rtx, rtx, basic_block, int *);
+static void build_store_vectors (void);
+static void insert_insn_start_bb (rtx, basic_block);
+static int insert_store (struct ls_expr *, edge);
+static void remove_reachable_equiv_notes (basic_block, struct ls_expr *);
+static void replace_store_insn (rtx, rtx, basic_block, struct ls_expr *);
+static void delete_store (struct ls_expr *, basic_block);
+static void free_store_memory (void);
+static void store_motion (void);
+static void free_insn_expr_list_list (rtx *);
+static void clear_modify_mem_tables (void);
+static void free_modify_mem_tables (void);
+static rtx gcse_emit_move_after (rtx, rtx, rtx);
+static void local_cprop_find_used_regs (rtx *, void *);
+static bool do_local_cprop (rtx, rtx, int, rtx*);
+static bool adjust_libcall_notes (rtx, rtx, rtx, rtx*);
+static void local_cprop_pass (int);
+static bool is_too_expensive (const char *);
\f
+
/* Entry point for global common subexpression elimination.
- F is the first instruction in the function. */
+ F is the first instruction in the function. Return nonzero if a
+ change is mode. */
int
-gcse_main (f, file)
- rtx f;
- FILE *file;
+gcse_main (rtx f, FILE *file)
{
int changed, pass;
/* Bytes used at start of pass. */
if (file)
dump_flow_info (file);
- /* Return if there's nothing to do. */
- if (n_basic_blocks <= 1)
+ /* Return if there's nothing to do, or it is too expensive. */
+ if (n_basic_blocks <= 1 || is_too_expensive (_("GCSE disabled")))
return 0;
- /* Trying to perform global optimizations on flow graphs which have
- a high connectivity will take a long time and is unlikely to be
- particularly useful.
-
- In normal circumstances a cfg should have about twice as many edges
- as blocks. But we do not want to punish small functions which have
- a couple switch statements. So we require a relatively large number
- of basic blocks and the ratio of edges to blocks to be high. */
- if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
- {
- if (warn_disabled_optimization)
- warning ("GCSE disabled: %d > 1000 basic blocks and %d >= 20 edges/basic block",
- n_basic_blocks, n_edges / n_basic_blocks);
- return 0;
- }
-
- /* If allocating memory for the cprop bitmap would take up too much
- storage it's better just to disable the optimization. */
- if ((n_basic_blocks
- * SBITMAP_SET_SIZE (max_gcse_regno)
- * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
- {
- if (warn_disabled_optimization)
- warning ("GCSE disabled: %d basic blocks and %d registers",
- n_basic_blocks, max_gcse_regno);
-
- return 0;
- }
-
gcc_obstack_init (&gcse_obstack);
bytes_used = 0;
/* Don't allow constant propagation to modify jumps
during this pass. */
+ timevar_push (TV_CPROP1);
changed = one_cprop_pass (pass + 1, 0, 0);
+ timevar_pop (TV_CPROP1);
if (optimize_size)
- changed |= one_classic_gcse_pass (pass + 1);
+ /* Do nothing. */ ;
else
{
+ timevar_push (TV_PRE);
changed |= one_pre_gcse_pass (pass + 1);
/* We may have just created new basic blocks. Release and
recompute various things which are sized on the number of
if (changed)
{
free_modify_mem_tables ();
- modify_mem_list
- = (rtx *) gmalloc (last_basic_block * sizeof (rtx));
- canon_modify_mem_list
- = (rtx *) gmalloc (last_basic_block * sizeof (rtx));
- memset ((char *) modify_mem_list, 0, last_basic_block * sizeof (rtx));
- memset ((char *) canon_modify_mem_list, 0, last_basic_block * sizeof (rtx));
+ modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
+ canon_modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
}
free_reg_set_mem ();
alloc_reg_set_mem (max_reg_num ());
compute_sets (f);
run_jump_opt_after_gcse = 1;
+ timevar_pop (TV_PRE);
}
if (max_pass_bytes < bytes_used)
partial redundancy elimination. */
free_gcse_mem ();
- /* It does not make sense to run code hoisting unless we optimizing
+ /* It does not make sense to run code hoisting unless we are optimizing
for code size -- it rarely makes programs faster, and can make
them bigger if we did partial redundancy elimination (when optimizing
- for space, we use a classic gcse algorithm instead of partial
- redundancy algorithms). */
+ for space, we don't run the partial redundancy algorithms). */
if (optimize_size)
{
+ timevar_push (TV_HOIST);
max_gcse_regno = max_reg_num ();
alloc_gcse_mem (f);
changed |= one_code_hoisting_pass ();
if (max_pass_bytes < bytes_used)
max_pass_bytes = bytes_used;
+ timevar_pop (TV_HOIST);
}
if (file)
max_gcse_regno = max_reg_num ();
alloc_gcse_mem (f);
/* This time, go ahead and allow cprop to alter jumps. */
+ timevar_push (TV_CPROP2);
one_cprop_pass (pass + 1, 1, 0);
+ timevar_pop (TV_CPROP2);
free_gcse_mem ();
if (file)
{
fprintf (file, "GCSE of %s: %d basic blocks, ",
- current_function_name, n_basic_blocks);
+ current_function_name (), n_basic_blocks);
fprintf (file, "%d pass%s, %d bytes\n\n",
pass, pass > 1 ? "es" : "", max_pass_bytes);
}
obstack_free (&gcse_obstack, NULL);
free_reg_set_mem ();
+
/* We are finished with alias. */
end_alias_analysis ();
allocate_reg_info (max_reg_num (), FALSE, FALSE);
if (!optimize_size && flag_gcse_sm)
- store_motion ();
+ {
+ timevar_push (TV_LSM);
+ store_motion ();
+ timevar_pop (TV_LSM);
+ }
/* Record where pseudo-registers are set. */
return run_jump_opt_after_gcse;
/* Compute which modes support reg/reg copy operations. */
static void
-compute_can_copy ()
+compute_can_copy (void)
{
int i;
#ifndef AVOID_CCMODE_COPIES
/* Returns whether the mode supports reg/reg copy operations. */
bool
-can_copy_p (mode)
- enum machine_mode mode;
+can_copy_p (enum machine_mode mode)
{
static bool can_copy_init_p = false;
\f
/* Cover function to xmalloc to record bytes allocated. */
-static char *
-gmalloc (size)
- unsigned int size;
+static void *
+gmalloc (size_t size)
{
bytes_used += size;
return xmalloc (size);
}
+/* Cover function to xcalloc to record bytes allocated. */
+
+static void *
+gcalloc (size_t nelem, size_t elsize)
+{
+ bytes_used += nelem * elsize;
+ return xcalloc (nelem, elsize);
+}
+
/* Cover function to xrealloc.
We don't record the additional size since we don't know it.
It won't affect memory usage stats much anyway. */
-static char *
-grealloc (ptr, size)
- char *ptr;
- unsigned int size;
+static void *
+grealloc (void *ptr, size_t size)
{
return xrealloc (ptr, size);
}
/* Cover function to obstack_alloc. */
-static char *
-gcse_alloc (size)
- unsigned long size;
+static void *
+gcse_alloc (unsigned long size)
{
bytes_used += size;
- return (char *) obstack_alloc (&gcse_obstack, size);
+ return obstack_alloc (&gcse_obstack, size);
}
/* Allocate memory for the cuid mapping array,
This is called at the start of each pass. */
static void
-alloc_gcse_mem (f)
- rtx f;
+alloc_gcse_mem (rtx f)
{
- int i, n;
+ int i;
rtx insn;
/* Find the largest UID and create a mapping from UIDs to CUIDs.
and only apply to real insns. */
max_uid = get_max_uid ();
- n = (max_uid + 1) * sizeof (int);
- uid_cuid = (int *) gmalloc (n);
- memset ((char *) uid_cuid, 0, n);
+ uid_cuid = gcalloc (max_uid + 1, sizeof (int));
for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
{
if (INSN_P (insn))
/* Create a table mapping cuids to insns. */
max_cuid = i;
- n = (max_cuid + 1) * sizeof (rtx);
- cuid_insn = (rtx *) gmalloc (n);
- memset ((char *) cuid_insn, 0, n);
+ cuid_insn = gcalloc (max_cuid + 1, sizeof (rtx));
for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
if (INSN_P (insn))
CUID_INSN (i++) = insn;
/* Allocate vars to track sets of regs. */
- reg_set_bitmap = BITMAP_XMALLOC ();
+ reg_set_bitmap = BITMAP_ALLOC (NULL);
/* Allocate vars to track sets of regs, memory per block. */
- reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (last_basic_block,
- max_gcse_regno);
+ reg_set_in_block = sbitmap_vector_alloc (last_basic_block, max_gcse_regno);
/* Allocate array to keep a list of insns which modify memory in each
basic block. */
- modify_mem_list = (rtx *) gmalloc (last_basic_block * sizeof (rtx));
- canon_modify_mem_list = (rtx *) gmalloc (last_basic_block * sizeof (rtx));
- memset ((char *) modify_mem_list, 0, last_basic_block * sizeof (rtx));
- memset ((char *) canon_modify_mem_list, 0, last_basic_block * sizeof (rtx));
- modify_mem_list_set = BITMAP_XMALLOC ();
- canon_modify_mem_list_set = BITMAP_XMALLOC ();
+ modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
+ canon_modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
+ modify_mem_list_set = BITMAP_ALLOC (NULL);
+ blocks_with_calls = BITMAP_ALLOC (NULL);
}
/* Free memory allocated by alloc_gcse_mem. */
static void
-free_gcse_mem ()
+free_gcse_mem (void)
{
free (uid_cuid);
free (cuid_insn);
- BITMAP_XFREE (reg_set_bitmap);
+ BITMAP_FREE (reg_set_bitmap);
sbitmap_vector_free (reg_set_in_block);
free_modify_mem_tables ();
- BITMAP_XFREE (modify_mem_list_set);
- BITMAP_XFREE (canon_modify_mem_list_set);
-}
-
-/* Many of the global optimization algorithms work by solving dataflow
- equations for various expressions. Initially, some local value is
- computed for each expression in each block. Then, the values across the
- various blocks are combined (by following flow graph edges) to arrive at
- global values. Conceptually, each set of equations is independent. We
- may therefore solve all the equations in parallel, solve them one at a
- time, or pick any intermediate approach.
-
- When you're going to need N two-dimensional bitmaps, each X (say, the
- number of blocks) by Y (say, the number of expressions), call this
- function. It's not important what X and Y represent; only that Y
- correspond to the things that can be done in parallel. This function will
- return an appropriate chunking factor C; you should solve C sets of
- equations in parallel. By going through this function, we can easily
- trade space against time; by solving fewer equations in parallel we use
- less space. */
-
-static int
-get_bitmap_width (n, x, y)
- int n;
- int x;
- int y;
-{
- /* It's not really worth figuring out *exactly* how much memory will
- be used by a particular choice. The important thing is to get
- something approximately right. */
- size_t max_bitmap_memory = 10 * 1024 * 1024;
-
- /* The number of bytes we'd use for a single column of minimum
- width. */
- size_t column_size = n * x * sizeof (SBITMAP_ELT_TYPE);
-
- /* Often, it's reasonable just to solve all the equations in
- parallel. */
- if (column_size * SBITMAP_SET_SIZE (y) <= max_bitmap_memory)
- return y;
-
- /* Otherwise, pick the largest width we can, without going over the
- limit. */
- return SBITMAP_ELT_BITS * ((max_bitmap_memory + column_size - 1)
- / column_size);
+ BITMAP_FREE (modify_mem_list_set);
+ BITMAP_FREE (blocks_with_calls);
}
\f
/* Compute the local properties of each recorded expression.
ABSALTERED. */
static void
-compute_local_properties (transp, comp, antloc, table)
- sbitmap *transp;
- sbitmap *comp;
- sbitmap *antloc;
- struct hash_table *table;
+compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc,
+ struct hash_table *table)
{
unsigned int i;
static struct obstack reg_set_obstack;
static void
-alloc_reg_set_mem (n_regs)
- int n_regs;
+alloc_reg_set_mem (int n_regs)
{
- unsigned int n;
-
reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
- n = reg_set_table_size * sizeof (struct reg_set *);
- reg_set_table = (struct reg_set **) gmalloc (n);
- memset ((char *) reg_set_table, 0, n);
+ reg_set_table = gcalloc (reg_set_table_size, sizeof (struct reg_set *));
gcc_obstack_init (®_set_obstack);
}
static void
-free_reg_set_mem ()
+free_reg_set_mem (void)
{
free (reg_set_table);
obstack_free (®_set_obstack, NULL);
/* Record REGNO in the reg_set table. */
static void
-record_one_set (regno, insn)
- int regno;
- rtx insn;
+record_one_set (int regno, rtx insn)
{
/* Allocate a new reg_set element and link it onto the list. */
struct reg_set *new_reg_info;
{
int new_size = regno + REG_SET_TABLE_SLOP;
- reg_set_table
- = (struct reg_set **) grealloc ((char *) reg_set_table,
- new_size * sizeof (struct reg_set *));
- memset ((char *) (reg_set_table + reg_set_table_size), 0,
+ reg_set_table = grealloc (reg_set_table,
+ new_size * sizeof (struct reg_set *));
+ memset (reg_set_table + reg_set_table_size, 0,
(new_size - reg_set_table_size) * sizeof (struct reg_set *));
reg_set_table_size = new_size;
}
- new_reg_info = (struct reg_set *) obstack_alloc (®_set_obstack,
- sizeof (struct reg_set));
+ new_reg_info = obstack_alloc (®_set_obstack, sizeof (struct reg_set));
bytes_used += sizeof (struct reg_set);
- new_reg_info->insn = insn;
+ new_reg_info->bb_index = BLOCK_NUM (insn);
new_reg_info->next = reg_set_table[regno];
reg_set_table[regno] = new_reg_info;
}
occurring. */
static void
-record_set_info (dest, setter, data)
- rtx dest, setter ATTRIBUTE_UNUSED;
- void *data;
+record_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED, void *data)
{
rtx record_set_insn = (rtx) data;
- if (GET_CODE (dest) == REG && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
+ if (REG_P (dest) && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
record_one_set (REGNO (dest), record_set_insn);
}
`reg_set_table' for further documentation. */
static void
-compute_sets (f)
- rtx f;
+compute_sets (rtx f)
{
rtx insn;
/* See whether X, the source of a set, is something we want to consider for
GCSE. */
-static GTY(()) rtx test_insn;
static int
-want_to_gcse_p (x)
- rtx x;
+want_to_gcse_p (rtx x)
{
- int num_clobbers = 0;
- int icode;
-
switch (GET_CODE (x))
{
case REG:
case CONST_DOUBLE:
case CONST_VECTOR:
case CALL:
- case CONSTANT_P_RTX:
return 0;
default:
- break;
+ return can_assign_to_reg_p (x);
}
+}
+
+/* Used internally by can_assign_to_reg_p. */
+
+static GTY(()) rtx test_insn;
+
+/* Return true if we can assign X to a pseudo register. */
+
+static bool
+can_assign_to_reg_p (rtx x)
+{
+ int num_clobbers = 0;
+ int icode;
/* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
if (general_operand (x, GET_MODE (x)))
or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
static int
-oprs_unchanged_p (x, insn, avail_p)
- rtx x, insn;
- int avail_p;
+oprs_unchanged_p (rtx x, rtx insn, int avail_p)
{
int i, j;
enum rtx_code code;
gcse_mems_conflict_p to a nonzero value. */
static void
-mems_conflict_for_gcse_p (dest, setter, data)
- rtx dest, setter ATTRIBUTE_UNUSED;
- void *data ATTRIBUTE_UNUSED;
+mems_conflict_for_gcse_p (rtx dest, rtx setter ATTRIBUTE_UNUSED,
+ void *data ATTRIBUTE_UNUSED)
{
while (GET_CODE (dest) == SUBREG
|| GET_CODE (dest) == ZERO_EXTRACT
- || GET_CODE (dest) == SIGN_EXTRACT
|| GET_CODE (dest) == STRICT_LOW_PART)
dest = XEXP (dest, 0);
/* If DEST is not a MEM, then it will not conflict with the load. Note
that function calls are assumed to clobber memory, but are handled
elsewhere. */
- if (GET_CODE (dest) != MEM)
+ if (! MEM_P (dest))
return;
/* If we are setting a MEM in our list of specially recognized MEMs,
AVAIL_P to 0. */
static int
-load_killed_in_block_p (bb, uid_limit, x, avail_p)
- basic_block bb;
- int uid_limit;
- rtx x;
- int avail_p;
+load_killed_in_block_p (basic_block bb, int uid_limit, rtx x, int avail_p)
{
rtx list_entry = modify_mem_list[bb->index];
while (list_entry)
/* If SETTER is a call everything is clobbered. Note that calls
to pure functions are never put on the list, so we need not
worry about them. */
- if (GET_CODE (setter) == CALL_INSN)
+ if (CALL_P (setter))
return 1;
/* SETTER must be an INSN of some kind that sets memory. Call
the start of INSN's basic block up to but not including INSN. */
static int
-oprs_anticipatable_p (x, insn)
- rtx x, insn;
+oprs_anticipatable_p (rtx x, rtx insn)
{
return oprs_unchanged_p (x, insn, 0);
}
INSN to the end of INSN's basic block. */
static int
-oprs_available_p (x, insn)
- rtx x, insn;
+oprs_available_p (rtx x, rtx insn)
{
return oprs_unchanged_p (x, insn, 1);
}
MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
indicating if a volatile operand is found or if the expression contains
- something we don't want to insert in the table.
-
- ??? One might want to merge this with canon_hash. Later. */
+ something we don't want to insert in the table. HASH_TABLE_SIZE is
+ the current size of the hash table to be probed. */
static unsigned int
-hash_expr (x, mode, do_not_record_p, hash_table_size)
- rtx x;
- enum machine_mode mode;
- int *do_not_record_p;
- int hash_table_size;
+hash_expr (rtx x, enum machine_mode mode, int *do_not_record_p,
+ int hash_table_size)
{
unsigned int hash;
*do_not_record_p = 0;
- hash = hash_expr_1 (x, mode, do_not_record_p);
+ hash = hash_rtx (x, mode, do_not_record_p,
+ NULL, /*have_reg_qty=*/false);
return hash % hash_table_size;
}
-/* Hash a string. Just add its bytes up. */
-
-static inline unsigned
-hash_string_1 (ps)
- const char *ps;
-{
- unsigned hash = 0;
- const unsigned char *p = (const unsigned char *) ps;
-
- if (p)
- while (*p)
- hash += *p++;
-
- return hash;
-}
-
-/* Subroutine of hash_expr to do the actual work. */
-
-static unsigned int
-hash_expr_1 (x, mode, do_not_record_p)
- rtx x;
- enum machine_mode mode;
- int *do_not_record_p;
-{
- int i, j;
- unsigned hash = 0;
- enum rtx_code code;
- const char *fmt;
-
- /* Used to turn recursion into iteration. We can't rely on GCC's
- tail-recursion elimination since we need to keep accumulating values
- in HASH. */
-
- if (x == 0)
- return hash;
-
- repeat:
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- hash += ((unsigned int) REG << 7) + REGNO (x);
- return hash;
-
- case CONST_INT:
- hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
- + (unsigned int) INTVAL (x));
- return hash;
-
- case CONST_DOUBLE:
- /* This is like the general case, except that it only counts
- the integers representing the constant. */
- hash += (unsigned int) code + (unsigned int) GET_MODE (x);
- if (GET_MODE (x) != VOIDmode)
- for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
- hash += (unsigned int) XWINT (x, i);
- else
- hash += ((unsigned int) CONST_DOUBLE_LOW (x)
- + (unsigned int) CONST_DOUBLE_HIGH (x));
- return hash;
-
- case CONST_VECTOR:
- {
- int units;
- rtx elt;
-
- units = CONST_VECTOR_NUNITS (x);
-
- for (i = 0; i < units; ++i)
- {
- elt = CONST_VECTOR_ELT (x, i);
- hash += hash_expr_1 (elt, GET_MODE (elt), do_not_record_p);
- }
-
- return hash;
- }
-
- /* Assume there is only one rtx object for any given label. */
- case LABEL_REF:
- /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
- differences and differences between each stage's debugging dumps. */
- hash += (((unsigned int) LABEL_REF << 7)
- + CODE_LABEL_NUMBER (XEXP (x, 0)));
- return hash;
-
- case SYMBOL_REF:
- {
- /* Don't hash on the symbol's address to avoid bootstrap differences.
- Different hash values may cause expressions to be recorded in
- different orders and thus different registers to be used in the
- final assembler. This also avoids differences in the dump files
- between various stages. */
- unsigned int h = 0;
- const unsigned char *p = (const unsigned char *) XSTR (x, 0);
-
- while (*p)
- h += (h << 7) + *p++; /* ??? revisit */
-
- hash += ((unsigned int) SYMBOL_REF << 7) + h;
- return hash;
- }
-
- case MEM:
- if (MEM_VOLATILE_P (x))
- {
- *do_not_record_p = 1;
- return 0;
- }
-
- hash += (unsigned int) MEM;
- /* We used alias set for hashing, but this is not good, since the alias
- set may differ in -fprofile-arcs and -fbranch-probabilities compilation
- causing the profiles to fail to match. */
- x = XEXP (x, 0);
- goto repeat;
-
- case PRE_DEC:
- case PRE_INC:
- case POST_DEC:
- case POST_INC:
- case PC:
- case CC0:
- case CALL:
- case UNSPEC_VOLATILE:
- *do_not_record_p = 1;
- return 0;
-
- case ASM_OPERANDS:
- if (MEM_VOLATILE_P (x))
- {
- *do_not_record_p = 1;
- return 0;
- }
- else
- {
- /* We don't want to take the filename and line into account. */
- hash += (unsigned) code + (unsigned) GET_MODE (x)
- + hash_string_1 (ASM_OPERANDS_TEMPLATE (x))
- + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
- + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
-
- if (ASM_OPERANDS_INPUT_LENGTH (x))
- {
- for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
- {
- hash += (hash_expr_1 (ASM_OPERANDS_INPUT (x, i),
- GET_MODE (ASM_OPERANDS_INPUT (x, i)),
- do_not_record_p)
- + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
- (x, i)));
- }
-
- hash += hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
- x = ASM_OPERANDS_INPUT (x, 0);
- mode = GET_MODE (x);
- goto repeat;
- }
- return hash;
- }
-
- default:
- break;
- }
-
- hash += (unsigned) code + (unsigned) GET_MODE (x);
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- /* If we are about to do the last recursive call
- needed at this level, change it into iteration.
- This function is called enough to be worth it. */
- if (i == 0)
- {
- x = XEXP (x, i);
- goto repeat;
- }
-
- hash += hash_expr_1 (XEXP (x, i), 0, do_not_record_p);
- if (*do_not_record_p)
- return 0;
- }
-
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- {
- hash += hash_expr_1 (XVECEXP (x, i, j), 0, do_not_record_p);
- if (*do_not_record_p)
- return 0;
- }
-
- else if (fmt[i] == 's')
- hash += hash_string_1 (XSTR (x, i));
- else if (fmt[i] == 'i')
- hash += (unsigned int) XINT (x, i);
- else
- abort ();
- }
-
- return hash;
-}
-
/* Hash a set of register REGNO.
Sets are hashed on the register that is set. This simplifies the PRE copy
??? May need to make things more elaborate. Later, as necessary. */
static unsigned int
-hash_set (regno, hash_table_size)
- int regno;
- int hash_table_size;
+hash_set (int regno, int hash_table_size)
{
unsigned int hash;
return hash % hash_table_size;
}
-/* Return nonzero if exp1 is equivalent to exp2.
- ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
+/* Return nonzero if exp1 is equivalent to exp2. */
static int
-expr_equiv_p (x, y)
- rtx x, y;
+expr_equiv_p (rtx x, rtx y)
{
- int i, j;
- enum rtx_code code;
- const char *fmt;
-
- if (x == y)
- return 1;
-
- if (x == 0 || y == 0)
- return 0;
-
- code = GET_CODE (x);
- if (code != GET_CODE (y))
- return 0;
-
- /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
- if (GET_MODE (x) != GET_MODE (y))
- return 0;
-
- switch (code)
- {
- case PC:
- case CC0:
- case CONST_INT:
- return 0;
-
- case LABEL_REF:
- return XEXP (x, 0) == XEXP (y, 0);
-
- case SYMBOL_REF:
- return XSTR (x, 0) == XSTR (y, 0);
-
- case REG:
- return REGNO (x) == REGNO (y);
-
- case MEM:
- /* Can't merge two expressions in different alias sets, since we can
- decide that the expression is transparent in a block when it isn't,
- due to it being set with the different alias set. */
- if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
- return 0;
- break;
-
- /* For commutative operations, check both orders. */
- case PLUS:
- case MULT:
- case AND:
- case IOR:
- case XOR:
- case NE:
- case EQ:
- return ((expr_equiv_p (XEXP (x, 0), XEXP (y, 0))
- && expr_equiv_p (XEXP (x, 1), XEXP (y, 1)))
- || (expr_equiv_p (XEXP (x, 0), XEXP (y, 1))
- && expr_equiv_p (XEXP (x, 1), XEXP (y, 0))));
-
- case ASM_OPERANDS:
- /* We don't use the generic code below because we want to
- disregard filename and line numbers. */
-
- /* A volatile asm isn't equivalent to any other. */
- if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
- return 0;
-
- if (GET_MODE (x) != GET_MODE (y)
- || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
- || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
- ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
- || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
- || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
- return 0;
-
- if (ASM_OPERANDS_INPUT_LENGTH (x))
- {
- for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
- if (! expr_equiv_p (ASM_OPERANDS_INPUT (x, i),
- ASM_OPERANDS_INPUT (y, i))
- || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
- ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
- return 0;
- }
-
- return 1;
-
- default:
- break;
- }
-
- /* Compare the elements. If any pair of corresponding elements
- fail to match, return 0 for the whole thing. */
-
- fmt = GET_RTX_FORMAT (code);
- for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
- {
- switch (fmt[i])
- {
- case 'e':
- if (! expr_equiv_p (XEXP (x, i), XEXP (y, i)))
- return 0;
- break;
-
- case 'E':
- if (XVECLEN (x, i) != XVECLEN (y, i))
- return 0;
- for (j = 0; j < XVECLEN (x, i); j++)
- if (! expr_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j)))
- return 0;
- break;
-
- case 's':
- if (strcmp (XSTR (x, i), XSTR (y, i)))
- return 0;
- break;
-
- case 'i':
- if (XINT (x, i) != XINT (y, i))
- return 0;
- break;
-
- case 'w':
- if (XWINT (x, i) != XWINT (y, i))
- return 0;
- break;
-
- case '0':
- break;
-
- default:
- abort ();
- }
- }
-
- return 1;
+ return exp_equiv_p (x, y, 0, true);
}
/* Insert expression X in INSN in the hash TABLE.
AVAIL_P is nonzero if X is an available expression. */
static void
-insert_expr_in_table (x, mode, insn, antic_p, avail_p, table)
- rtx x;
- enum machine_mode mode;
- rtx insn;
- int antic_p, avail_p;
- struct hash_table *table;
+insert_expr_in_table (rtx x, enum machine_mode mode, rtx insn, int antic_p,
+ int avail_p, struct hash_table *table)
{
int found, do_not_record_p;
unsigned int hash;
struct expr *cur_expr, *last_expr = NULL;
struct occr *antic_occr, *avail_occr;
- struct occr *last_occr = NULL;
hash = hash_expr (x, mode, &do_not_record_p, table->size);
if (! found)
{
- cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
+ cur_expr = gcse_alloc (sizeof (struct expr));
bytes_used += sizeof (struct expr);
if (table->table[hash] == NULL)
/* This is the first pattern that hashed to this index. */
{
antic_occr = cur_expr->antic_occr;
- /* Search for another occurrence in the same basic block. */
- while (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
- {
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = antic_occr;
- antic_occr = antic_occr->next;
- }
+ if (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
+ antic_occr = NULL;
if (antic_occr)
/* Found another instance of the expression in the same basic block.
else
{
/* First occurrence of this expression in this basic block. */
- antic_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
+ antic_occr = gcse_alloc (sizeof (struct occr));
bytes_used += sizeof (struct occr);
- /* First occurrence of this expression in any block? */
- if (cur_expr->antic_occr == NULL)
- cur_expr->antic_occr = antic_occr;
- else
- last_occr->next = antic_occr;
-
antic_occr->insn = insn;
- antic_occr->next = NULL;
+ antic_occr->next = cur_expr->antic_occr;
+ antic_occr->deleted_p = 0;
+ cur_expr->antic_occr = antic_occr;
}
}
{
avail_occr = cur_expr->avail_occr;
- /* Search for another occurrence in the same basic block. */
- while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))
+ if (avail_occr && BLOCK_NUM (avail_occr->insn) == BLOCK_NUM (insn))
{
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = avail_occr;
- avail_occr = avail_occr->next;
+ /* Found another instance of the expression in the same basic block.
+ Prefer this occurrence to the currently recorded one. We want
+ the last one in the block and the block is scanned from start
+ to end. */
+ avail_occr->insn = insn;
}
-
- if (avail_occr)
- /* Found another instance of the expression in the same basic block.
- Prefer this occurrence to the currently recorded one. We want
- the last one in the block and the block is scanned from start
- to end. */
- avail_occr->insn = insn;
else
{
/* First occurrence of this expression in this basic block. */
- avail_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
+ avail_occr = gcse_alloc (sizeof (struct occr));
bytes_used += sizeof (struct occr);
-
- /* First occurrence of this expression in any block? */
- if (cur_expr->avail_occr == NULL)
- cur_expr->avail_occr = avail_occr;
- else
- last_occr->next = avail_occr;
-
avail_occr->insn = insn;
- avail_occr->next = NULL;
+ avail_occr->next = cur_expr->avail_occr;
+ avail_occr->deleted_p = 0;
+ cur_expr->avail_occr = avail_occr;
}
}
}
basic block. */
static void
-insert_set_in_table (x, insn, table)
- rtx x;
- rtx insn;
- struct hash_table *table;
+insert_set_in_table (rtx x, rtx insn, struct hash_table *table)
{
int found;
unsigned int hash;
struct expr *cur_expr, *last_expr = NULL;
- struct occr *cur_occr, *last_occr = NULL;
+ struct occr *cur_occr;
- if (GET_CODE (x) != SET
- || GET_CODE (SET_DEST (x)) != REG)
- abort ();
+ gcc_assert (GET_CODE (x) == SET && REG_P (SET_DEST (x)));
hash = hash_set (REGNO (SET_DEST (x)), table->size);
if (! found)
{
- cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
+ cur_expr = gcse_alloc (sizeof (struct expr));
bytes_used += sizeof (struct expr);
if (table->table[hash] == NULL)
/* This is the first pattern that hashed to this index. */
/* Now record the occurrence. */
cur_occr = cur_expr->avail_occr;
- /* Search for another occurrence in the same basic block. */
- while (cur_occr && BLOCK_NUM (cur_occr->insn) != BLOCK_NUM (insn))
+ if (cur_occr && BLOCK_NUM (cur_occr->insn) == BLOCK_NUM (insn))
{
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = cur_occr;
- cur_occr = cur_occr->next;
+ /* Found another instance of the expression in the same basic block.
+ Prefer this occurrence to the currently recorded one. We want
+ the last one in the block and the block is scanned from start
+ to end. */
+ cur_occr->insn = insn;
}
-
- if (cur_occr)
- /* Found another instance of the expression in the same basic block.
- Prefer this occurrence to the currently recorded one. We want the
- last one in the block and the block is scanned from start to end. */
- cur_occr->insn = insn;
else
{
/* First occurrence of this expression in this basic block. */
- cur_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
+ cur_occr = gcse_alloc (sizeof (struct occr));
bytes_used += sizeof (struct occr);
- /* First occurrence of this expression in any block? */
- if (cur_expr->avail_occr == NULL)
- cur_expr->avail_occr = cur_occr;
- else
- last_occr->next = cur_occr;
-
- cur_occr->insn = insn;
- cur_occr->next = NULL;
+ cur_occr->insn = insn;
+ cur_occr->next = cur_expr->avail_occr;
+ cur_occr->deleted_p = 0;
+ cur_expr->avail_occr = cur_occr;
}
}
the purposes of GCSE's constant propagation. */
static bool
-gcse_constant_p (x)
- rtx x;
+gcse_constant_p (rtx x)
{
/* Consider a COMPARE of two integers constant. */
if (GET_CODE (x) == COMPARE
&& GET_CODE (XEXP (x, 1)) == CONST_INT)
return true;
- if (GET_CODE (x) == CONSTANT_P_RTX)
- return false;
+ /* Consider a COMPARE of the same registers is a constant
+ if they are not floating point registers. */
+ if (GET_CODE(x) == COMPARE
+ && REG_P (XEXP (x, 0)) && REG_P (XEXP (x, 1))
+ && REGNO (XEXP (x, 0)) == REGNO (XEXP (x, 1))
+ && ! FLOAT_MODE_P (GET_MODE (XEXP (x, 0)))
+ && ! FLOAT_MODE_P (GET_MODE (XEXP (x, 1))))
+ return true;
return CONSTANT_P (x);
}
expression one). */
static void
-hash_scan_set (pat, insn, table)
- rtx pat, insn;
- struct hash_table *table;
+hash_scan_set (rtx pat, rtx insn, struct hash_table *table)
{
rtx src = SET_SRC (pat);
rtx dest = SET_DEST (pat);
if (GET_CODE (src) == CALL)
hash_scan_call (src, insn, table);
- else if (GET_CODE (dest) == REG)
+ else if (REG_P (dest))
{
unsigned int regno = REGNO (dest);
rtx tmp;
explicitly, it means address of parameter has been taken,
so we should not extend the lifetime of the pseudo. */
&& ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
- || GET_CODE (XEXP (note, 0)) != MEM))
+ || ! MEM_P (XEXP (note, 0))))
{
/* An expression is not anticipatable if its operands are
modified before this insn or if this is not the only SET in
/* Record sets for constant/copy propagation. */
else if (table->set_p
&& regno >= FIRST_PSEUDO_REGISTER
- && ((GET_CODE (src) == REG
+ && ((REG_P (src)
&& REGNO (src) >= FIRST_PSEUDO_REGISTER
&& can_copy_p (GET_MODE (dest))
&& REGNO (src) != regno)
/* A copy is not available if its src or dest is subsequently
modified. Here we want to search from INSN+1 on, but
oprs_available_p searches from INSN on. */
- && (insn == BLOCK_END (BLOCK_NUM (insn))
+ && (insn == BB_END (BLOCK_FOR_INSN (insn))
|| ((tmp = next_nonnote_insn (insn)) != NULL_RTX
&& oprs_available_p (pat, tmp))))
insert_set_in_table (pat, insn, table);
}
+ /* In case of store we want to consider the memory value as available in
+ the REG stored in that memory. This makes it possible to remove
+ redundant loads from due to stores to the same location. */
+ else if (flag_gcse_las && REG_P (src) && MEM_P (dest))
+ {
+ unsigned int regno = REGNO (src);
+
+ /* Do not do this for constant/copy propagation. */
+ if (! table->set_p
+ /* Only record sets of pseudo-regs in the hash table. */
+ && regno >= FIRST_PSEUDO_REGISTER
+ /* Don't GCSE something if we can't do a reg/reg copy. */
+ && can_copy_p (GET_MODE (src))
+ /* GCSE commonly inserts instruction after the insn. We can't
+ do that easily for EH_REGION notes so disable GCSE on these
+ for now. */
+ && ! find_reg_note (insn, REG_EH_REGION, NULL_RTX)
+ /* Is SET_DEST something we want to gcse? */
+ && want_to_gcse_p (dest)
+ /* Don't CSE a nop. */
+ && ! set_noop_p (pat)
+ /* Don't GCSE if it has attached REG_EQUIV note.
+ At this point this only function parameters should have
+ REG_EQUIV notes and if the argument slot is used somewhere
+ explicitly, it means address of parameter has been taken,
+ so we should not extend the lifetime of the pseudo. */
+ && ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
+ || ! MEM_P (XEXP (note, 0))))
+ {
+ /* Stores are never anticipatable. */
+ int antic_p = 0;
+ /* An expression is not available if its operands are
+ subsequently modified, including this insn. It's also not
+ available if this is a branch, because we can't insert
+ a set after the branch. */
+ int avail_p = oprs_available_p (dest, insn)
+ && ! JUMP_P (insn);
+
+ /* Record the memory expression (DEST) in the hash table. */
+ insert_expr_in_table (dest, GET_MODE (dest), insn,
+ antic_p, avail_p, table);
+ }
+ }
}
static void
-hash_scan_clobber (x, insn, table)
- rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
- struct hash_table *table ATTRIBUTE_UNUSED;
+hash_scan_clobber (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
+ struct hash_table *table ATTRIBUTE_UNUSED)
{
/* Currently nothing to do. */
}
static void
-hash_scan_call (x, insn, table)
- rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
- struct hash_table *table ATTRIBUTE_UNUSED;
+hash_scan_call (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
+ struct hash_table *table ATTRIBUTE_UNUSED)
{
/* Currently nothing to do. */
}
not record any expressions. */
static void
-hash_scan_insn (insn, table, in_libcall_block)
- rtx insn;
- struct hash_table *table;
- int in_libcall_block;
+hash_scan_insn (rtx insn, struct hash_table *table, int in_libcall_block)
{
rtx pat = PATTERN (insn);
int i;
}
static void
-dump_hash_table (file, name, table)
- FILE *file;
- const char *name;
- struct hash_table *table;
+dump_hash_table (FILE *file, const char *name, struct hash_table *table)
{
int i;
/* Flattened out table, so it's printed in proper order. */
unsigned int *hash_val;
struct expr *expr;
- flat_table
- = (struct expr **) xcalloc (table->n_elems, sizeof (struct expr *));
- hash_val = (unsigned int *) xmalloc (table->n_elems * sizeof (unsigned int));
+ flat_table = xcalloc (table->n_elems, sizeof (struct expr *));
+ hash_val = xmalloc (table->n_elems * sizeof (unsigned int));
for (i = 0; i < (int) table->size; i++)
for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
and is used to compute "transparency". */
static void
-record_last_reg_set_info (insn, regno)
- rtx insn;
- int regno;
+record_last_reg_set_info (rtx insn, int regno)
{
struct reg_avail_info *info = ®_avail_info[regno];
int cuid = INSN_CUID (insn);
taken off pairwise. */
static void
-canon_list_insert (dest, unused1, v_insn)
- rtx dest ATTRIBUTE_UNUSED;
- rtx unused1 ATTRIBUTE_UNUSED;
- void * v_insn;
+canon_list_insert (rtx dest ATTRIBUTE_UNUSED, rtx unused1 ATTRIBUTE_UNUSED,
+ void * v_insn)
{
rtx dest_addr, insn;
int bb;
while (GET_CODE (dest) == SUBREG
|| GET_CODE (dest) == ZERO_EXTRACT
- || GET_CODE (dest) == SIGN_EXTRACT
|| GET_CODE (dest) == STRICT_LOW_PART)
dest = XEXP (dest, 0);
that function calls are assumed to clobber memory, but are handled
elsewhere. */
- if (GET_CODE (dest) != MEM)
+ if (! MEM_P (dest))
return;
dest_addr = get_addr (XEXP (dest, 0));
alloc_EXPR_LIST (VOIDmode, dest_addr, canon_modify_mem_list[bb]);
canon_modify_mem_list[bb] =
alloc_EXPR_LIST (VOIDmode, dest, canon_modify_mem_list[bb]);
- bitmap_set_bit (canon_modify_mem_list_set, bb);
}
/* Record memory modification information for INSN. We do not actually care
a CALL_INSN). We merely need to record which insns modify memory. */
static void
-record_last_mem_set_info (insn)
- rtx insn;
+record_last_mem_set_info (rtx insn)
{
int bb = BLOCK_NUM (insn);
modify_mem_list[bb] = alloc_INSN_LIST (insn, modify_mem_list[bb]);
bitmap_set_bit (modify_mem_list_set, bb);
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
/* Note that traversals of this loop (other than for free-ing)
will break after encountering a CALL_INSN. So, there's no
need to insert a pair of items, as canon_list_insert does. */
canon_modify_mem_list[bb] =
alloc_INSN_LIST (insn, canon_modify_mem_list[bb]);
- bitmap_set_bit (canon_modify_mem_list_set, bb);
+ bitmap_set_bit (blocks_with_calls, bb);
}
else
note_stores (PATTERN (insn), canon_list_insert, (void*) insn);
the SET is taking place. */
static void
-record_last_set_info (dest, setter, data)
- rtx dest, setter ATTRIBUTE_UNUSED;
- void *data;
+record_last_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED, void *data)
{
rtx last_set_insn = (rtx) data;
if (GET_CODE (dest) == SUBREG)
dest = SUBREG_REG (dest);
- if (GET_CODE (dest) == REG)
+ if (REG_P (dest))
record_last_reg_set_info (last_set_insn, REGNO (dest));
- else if (GET_CODE (dest) == MEM
+ else if (MEM_P (dest)
/* Ignore pushes, they clobber nothing. */
&& ! push_operand (dest, GET_MODE (dest)))
record_last_mem_set_info (last_set_insn);
TABLE is the table computed. */
static void
-compute_hash_table_work (table)
- struct hash_table *table;
+compute_hash_table_work (struct hash_table *table)
{
unsigned int i;
/* re-Cache any INSN_LIST nodes we have allocated. */
clear_modify_mem_tables ();
/* Some working arrays used to track first and last set in each block. */
- reg_avail_info = (struct reg_avail_info*)
- gmalloc (max_gcse_regno * sizeof (struct reg_avail_info));
+ reg_avail_info = gmalloc (max_gcse_regno * sizeof (struct reg_avail_info));
for (i = 0; i < max_gcse_regno; ++i)
reg_avail_info[i].last_bb = NULL;
??? hard-reg reg_set_in_block computation
could be moved to compute_sets since they currently don't change. */
- for (insn = current_bb->head;
- insn && insn != NEXT_INSN (current_bb->end);
+ for (insn = BB_HEAD (current_bb);
+ insn && insn != NEXT_INSN (BB_END (current_bb));
insn = NEXT_INSN (insn))
{
if (! INSN_P (insn))
continue;
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
- bool clobbers_all = false;
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
-
for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
record_last_reg_set_info (insn, regno);
mark_call (insn);
if (table->set_p
&& implicit_sets[current_bb->index] != NULL_RTX)
hash_scan_set (implicit_sets[current_bb->index],
- current_bb->head, table);
+ BB_HEAD (current_bb), table);
/* The next pass builds the hash table. */
- for (insn = current_bb->head, in_libcall_block = 0;
- insn && insn != NEXT_INSN (current_bb->end);
+ for (insn = BB_HEAD (current_bb), in_libcall_block = 0;
+ insn && insn != NEXT_INSN (BB_END (current_bb));
insn = NEXT_INSN (insn))
if (INSN_P (insn))
{
be created. */
static void
-alloc_hash_table (n_insns, table, set_p)
- int n_insns;
- struct hash_table *table;
- int set_p;
+alloc_hash_table (int n_insns, struct hash_table *table, int set_p)
{
int n;
??? Later take some measurements. */
table->size |= 1;
n = table->size * sizeof (struct expr *);
- table->table = (struct expr **) gmalloc (n);
+ table->table = gmalloc (n);
table->set_p = set_p;
}
/* Free things allocated by alloc_hash_table. */
static void
-free_hash_table (table)
- struct hash_table *table;
+free_hash_table (struct hash_table *table)
{
free (table->table);
}
expression hash table. */
static void
-compute_hash_table (table)
- struct hash_table *table;
+compute_hash_table (struct hash_table *table)
{
/* Initialize count of number of entries in hash table. */
table->n_elems = 0;
- memset ((char *) table->table, 0,
- table->size * sizeof (struct expr *));
+ memset (table->table, 0, table->size * sizeof (struct expr *));
compute_hash_table_work (table);
}
\f
/* Expression tracking support. */
-/* Lookup pattern PAT in the expression TABLE.
- The result is a pointer to the table entry, or NULL if not found. */
+/* Lookup REGNO in the set TABLE. The result is a pointer to the
+ table entry, or NULL if not found. */
static struct expr *
-lookup_expr (pat, table)
- rtx pat;
- struct hash_table *table;
+lookup_set (unsigned int regno, struct hash_table *table)
{
- int do_not_record_p;
- unsigned int hash = hash_expr (pat, GET_MODE (pat), &do_not_record_p,
- table->size);
- struct expr *expr;
-
- if (do_not_record_p)
- return NULL;
-
- expr = table->table[hash];
-
- while (expr && ! expr_equiv_p (expr->expr, pat))
- expr = expr->next_same_hash;
-
- return expr;
-}
-
-/* Lookup REGNO in the set TABLE. The result is a pointer to the
- table entry, or NULL if not found. */
-
-static struct expr *
-lookup_set (regno, table)
- unsigned int regno;
- struct hash_table *table;
-{
- unsigned int hash = hash_set (regno, table->size);
+ unsigned int hash = hash_set (regno, table->size);
struct expr *expr;
expr = table->table[hash];
/* Return the next entry for REGNO in list EXPR. */
static struct expr *
-next_set (regno, expr)
- unsigned int regno;
- struct expr *expr;
+next_set (unsigned int regno, struct expr *expr)
{
do
expr = expr->next_same_hash;
types may be mixed. */
static void
-free_insn_expr_list_list (listp)
- rtx *listp;
+free_insn_expr_list_list (rtx *listp)
{
rtx list, next;
/* Clear canon_modify_mem_list and modify_mem_list tables. */
static void
-clear_modify_mem_tables ()
+clear_modify_mem_tables (void)
{
- int i;
+ unsigned i;
+ bitmap_iterator bi;
- EXECUTE_IF_SET_IN_BITMAP
- (modify_mem_list_set, 0, i, free_INSN_LIST_list (modify_mem_list + i));
+ EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set, 0, i, bi)
+ {
+ free_INSN_LIST_list (modify_mem_list + i);
+ free_insn_expr_list_list (canon_modify_mem_list + i);
+ }
bitmap_clear (modify_mem_list_set);
-
- EXECUTE_IF_SET_IN_BITMAP
- (canon_modify_mem_list_set, 0, i,
- free_insn_expr_list_list (canon_modify_mem_list + i));
- bitmap_clear (canon_modify_mem_list_set);
+ bitmap_clear (blocks_with_calls);
}
-/* Release memory used by modify_mem_list_set and canon_modify_mem_list_set. */
+/* Release memory used by modify_mem_list_set. */
static void
-free_modify_mem_tables ()
+free_modify_mem_tables (void)
{
clear_modify_mem_tables ();
free (modify_mem_list);
start of the block]. */
static void
-reset_opr_set_tables ()
+reset_opr_set_tables (void)
{
/* Maintain a bitmap of which regs have been set since beginning of
the block. */
INSN's basic block. */
static int
-oprs_not_set_p (x, insn)
- rtx x, insn;
+oprs_not_set_p (rtx x, rtx insn)
{
int i, j;
enum rtx_code code;
/* Mark things set by a CALL. */
static void
-mark_call (insn)
- rtx insn;
+mark_call (rtx insn)
{
if (! CONST_OR_PURE_CALL_P (insn))
record_last_mem_set_info (insn);
/* Mark things set by a SET. */
static void
-mark_set (pat, insn)
- rtx pat, insn;
+mark_set (rtx pat, rtx insn)
{
rtx dest = SET_DEST (pat);
while (GET_CODE (dest) == SUBREG
|| GET_CODE (dest) == ZERO_EXTRACT
- || GET_CODE (dest) == SIGN_EXTRACT
|| GET_CODE (dest) == STRICT_LOW_PART)
dest = XEXP (dest, 0);
- if (GET_CODE (dest) == REG)
+ if (REG_P (dest))
SET_REGNO_REG_SET (reg_set_bitmap, REGNO (dest));
- else if (GET_CODE (dest) == MEM)
+ else if (MEM_P (dest))
record_last_mem_set_info (insn);
if (GET_CODE (SET_SRC (pat)) == CALL)
- mark_call (insn);
-}
-
-/* Record things set by a CLOBBER. */
-
-static void
-mark_clobber (pat, insn)
- rtx pat, insn;
-{
- rtx clob = XEXP (pat, 0);
-
- while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
- clob = XEXP (clob, 0);
-
- if (GET_CODE (clob) == REG)
- SET_REGNO_REG_SET (reg_set_bitmap, REGNO (clob));
- else
- record_last_mem_set_info (insn);
-}
-
-/* Record things set by INSN.
- This data is used by oprs_not_set_p. */
-
-static void
-mark_oprs_set (insn)
- rtx insn;
-{
- rtx pat = PATTERN (insn);
- int i;
-
- if (GET_CODE (pat) == SET)
- mark_set (pat, insn);
- else if (GET_CODE (pat) == PARALLEL)
- for (i = 0; i < XVECLEN (pat, 0); i++)
- {
- rtx x = XVECEXP (pat, 0, i);
-
- if (GET_CODE (x) == SET)
- mark_set (x, insn);
- else if (GET_CODE (x) == CLOBBER)
- mark_clobber (x, insn);
- else if (GET_CODE (x) == CALL)
- mark_call (insn);
- }
-
- else if (GET_CODE (pat) == CLOBBER)
- mark_clobber (pat, insn);
- else if (GET_CODE (pat) == CALL)
- mark_call (insn);
-}
-
-\f
-/* Classic GCSE reaching definition support. */
-
-/* Allocate reaching def variables. */
-
-static void
-alloc_rd_mem (n_blocks, n_insns)
- int n_blocks, n_insns;
-{
- rd_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_kill, n_blocks);
-
- rd_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_gen, n_blocks);
-
- reaching_defs = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (reaching_defs, n_blocks);
-
- rd_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_out, n_blocks);
-}
-
-/* Free reaching def variables. */
-
-static void
-free_rd_mem ()
-{
- sbitmap_vector_free (rd_kill);
- sbitmap_vector_free (rd_gen);
- sbitmap_vector_free (reaching_defs);
- sbitmap_vector_free (rd_out);
-}
-
-/* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
-
-static void
-handle_rd_kill_set (insn, regno, bb)
- rtx insn;
- int regno;
- basic_block bb;
-{
- struct reg_set *this_reg;
-
- for (this_reg = reg_set_table[regno]; this_reg; this_reg = this_reg ->next)
- if (BLOCK_NUM (this_reg->insn) != BLOCK_NUM (insn))
- SET_BIT (rd_kill[bb->index], INSN_CUID (this_reg->insn));
-}
-
-/* Compute the set of kill's for reaching definitions. */
-
-static void
-compute_kill_rd ()
-{
- int cuid;
- unsigned int regno;
- int i;
- basic_block bb;
-
- /* For each block
- For each set bit in `gen' of the block (i.e each insn which
- generates a definition in the block)
- Call the reg set by the insn corresponding to that bit regx
- Look at the linked list starting at reg_set_table[regx]
- For each setting of regx in the linked list, which is not in
- this block
- Set the bit in `kill' corresponding to that insn. */
- FOR_EACH_BB (bb)
- for (cuid = 0; cuid < max_cuid; cuid++)
- if (TEST_BIT (rd_gen[bb->index], cuid))
- {
- rtx insn = CUID_INSN (cuid);
- rtx pat = PATTERN (insn);
-
- if (GET_CODE (insn) == CALL_INSN)
- {
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- handle_rd_kill_set (insn, regno, bb);
- }
-
- if (GET_CODE (pat) == PARALLEL)
- {
- for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
- {
- enum rtx_code code = GET_CODE (XVECEXP (pat, 0, i));
-
- if ((code == SET || code == CLOBBER)
- && GET_CODE (XEXP (XVECEXP (pat, 0, i), 0)) == REG)
- handle_rd_kill_set (insn,
- REGNO (XEXP (XVECEXP (pat, 0, i), 0)),
- bb);
- }
- }
- else if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == REG)
- /* Each setting of this register outside of this block
- must be marked in the set of kills in this block. */
- handle_rd_kill_set (insn, REGNO (SET_DEST (pat)), bb);
- }
-}
-
-/* Compute the reaching definitions as in
- Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
- Chapter 10. It is the same algorithm as used for computing available
- expressions but applied to the gens and kills of reaching definitions. */
-
-static void
-compute_rd ()
-{
- int changed, passes;
- basic_block bb;
-
- FOR_EACH_BB (bb)
- sbitmap_copy (rd_out[bb->index] /*dst*/, rd_gen[bb->index] /*src*/);
-
- passes = 0;
- changed = 1;
- while (changed)
- {
- changed = 0;
- FOR_EACH_BB (bb)
- {
- sbitmap_union_of_preds (reaching_defs[bb->index], rd_out, bb->index);
- changed |= sbitmap_union_of_diff_cg (rd_out[bb->index], rd_gen[bb->index],
- reaching_defs[bb->index], rd_kill[bb->index]);
- }
- passes++;
- }
-
- if (gcse_file)
- fprintf (gcse_file, "reaching def computation: %d passes\n", passes);
-}
-\f
-/* Classic GCSE available expression support. */
-
-/* Allocate memory for available expression computation. */
-
-static void
-alloc_avail_expr_mem (n_blocks, n_exprs)
- int n_blocks, n_exprs;
-{
- ae_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_kill, n_blocks);
-
- ae_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_gen, n_blocks);
-
- ae_in = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_in, n_blocks);
-
- ae_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_out, n_blocks);
-}
-
-static void
-free_avail_expr_mem ()
-{
- sbitmap_vector_free (ae_kill);
- sbitmap_vector_free (ae_gen);
- sbitmap_vector_free (ae_in);
- sbitmap_vector_free (ae_out);
-}
-
-/* Compute the set of available expressions generated in each basic block. */
-
-static void
-compute_ae_gen (expr_hash_table)
- struct hash_table *expr_hash_table;
-{
- unsigned int i;
- struct expr *expr;
- struct occr *occr;
-
- /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
- This is all we have to do because an expression is not recorded if it
- is not available, and the only expressions we want to work with are the
- ones that are recorded. */
- for (i = 0; i < expr_hash_table->size; i++)
- for (expr = expr_hash_table->table[i]; expr != 0; expr = expr->next_same_hash)
- for (occr = expr->avail_occr; occr != 0; occr = occr->next)
- SET_BIT (ae_gen[BLOCK_NUM (occr->insn)], expr->bitmap_index);
-}
-
-/* Return nonzero if expression X is killed in BB. */
-
-static int
-expr_killed_p (x, bb)
- rtx x;
- basic_block bb;
-{
- int i, j;
- enum rtx_code code;
- const char *fmt;
-
- if (x == 0)
- return 1;
-
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- return TEST_BIT (reg_set_in_block[bb->index], REGNO (x));
-
- case MEM:
- if (load_killed_in_block_p (bb, get_max_uid () + 1, x, 0))
- return 1;
- else
- return expr_killed_p (XEXP (x, 0), bb);
-
- case PC:
- case CC0: /*FIXME*/
- case CONST:
- case CONST_INT:
- case CONST_DOUBLE:
- case CONST_VECTOR:
- case SYMBOL_REF:
- case LABEL_REF:
- case ADDR_VEC:
- case ADDR_DIFF_VEC:
- return 0;
-
- default:
- break;
- }
-
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- /* If we are about to do the last recursive call
- needed at this level, change it into iteration.
- This function is called enough to be worth it. */
- if (i == 0)
- return expr_killed_p (XEXP (x, i), bb);
- else if (expr_killed_p (XEXP (x, i), bb))
- return 1;
- }
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- if (expr_killed_p (XVECEXP (x, i, j), bb))
- return 1;
- }
-
- return 0;
-}
-
-/* Compute the set of available expressions killed in each basic block. */
-
-static void
-compute_ae_kill (ae_gen, ae_kill, expr_hash_table)
- sbitmap *ae_gen, *ae_kill;
- struct hash_table *expr_hash_table;
-{
- basic_block bb;
- unsigned int i;
- struct expr *expr;
-
- FOR_EACH_BB (bb)
- for (i = 0; i < expr_hash_table->size; i++)
- for (expr = expr_hash_table->table[i]; expr; expr = expr->next_same_hash)
- {
- /* Skip EXPR if generated in this block. */
- if (TEST_BIT (ae_gen[bb->index], expr->bitmap_index))
- continue;
-
- if (expr_killed_p (expr->expr, bb))
- SET_BIT (ae_kill[bb->index], expr->bitmap_index);
- }
-}
-\f
-/* Actually perform the Classic GCSE optimizations. */
-
-/* Return nonzero if occurrence OCCR of expression EXPR reaches block BB.
-
- CHECK_SELF_LOOP is nonzero if we should consider a block reaching itself
- as a positive reach. We want to do this when there are two computations
- of the expression in the block.
-
- VISITED is a pointer to a working buffer for tracking which BB's have
- been visited. It is NULL for the top-level call.
-
- We treat reaching expressions that go through blocks containing the same
- reaching expression as "not reaching". E.g. if EXPR is generated in blocks
- 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
- 2 as not reaching. The intent is to improve the probability of finding
- only one reaching expression and to reduce register lifetimes by picking
- the closest such expression. */
-
-static int
-expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited)
- struct occr *occr;
- struct expr *expr;
- basic_block bb;
- int check_self_loop;
- char *visited;
-{
- edge pred;
-
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
- {
- basic_block pred_bb = pred->src;
-
- if (visited[pred_bb->index])
- /* This predecessor has already been visited. Nothing to do. */
- ;
- else if (pred_bb == bb)
- {
- /* BB loops on itself. */
- if (check_self_loop
- && TEST_BIT (ae_gen[pred_bb->index], expr->bitmap_index)
- && BLOCK_NUM (occr->insn) == pred_bb->index)
- return 1;
-
- visited[pred_bb->index] = 1;
- }
-
- /* Ignore this predecessor if it kills the expression. */
- else if (TEST_BIT (ae_kill[pred_bb->index], expr->bitmap_index))
- visited[pred_bb->index] = 1;
-
- /* Does this predecessor generate this expression? */
- else if (TEST_BIT (ae_gen[pred_bb->index], expr->bitmap_index))
- {
- /* Is this the occurrence we're looking for?
- Note that there's only one generating occurrence per block
- so we just need to check the block number. */
- if (BLOCK_NUM (occr->insn) == pred_bb->index)
- return 1;
-
- visited[pred_bb->index] = 1;
- }
-
- /* Neither gen nor kill. */
- else
- {
- visited[pred_bb->index] = 1;
- if (expr_reaches_here_p_work (occr, expr, pred_bb, check_self_loop,
- visited))
-
- return 1;
- }
- }
-
- /* All paths have been checked. */
- return 0;
-}
-
-/* This wrapper for expr_reaches_here_p_work() is to ensure that any
- memory allocated for that function is returned. */
-
-static int
-expr_reaches_here_p (occr, expr, bb, check_self_loop)
- struct occr *occr;
- struct expr *expr;
- basic_block bb;
- int check_self_loop;
-{
- int rval;
- char *visited = (char *) xcalloc (last_basic_block, 1);
-
- rval = expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited);
-
- free (visited);
- return rval;
-}
-
-/* Return the instruction that computes EXPR that reaches INSN's basic block.
- If there is more than one such instruction, return NULL.
-
- Called only by handle_avail_expr. */
-
-static rtx
-computing_insn (expr, insn)
- struct expr *expr;
- rtx insn;
-{
- basic_block bb = BLOCK_FOR_INSN (insn);
-
- if (expr->avail_occr->next == NULL)
- {
- if (BLOCK_FOR_INSN (expr->avail_occr->insn) == bb)
- /* The available expression is actually itself
- (i.e. a loop in the flow graph) so do nothing. */
- return NULL;
-
- /* (FIXME) Case that we found a pattern that was created by
- a substitution that took place. */
- return expr->avail_occr->insn;
- }
- else
- {
- /* Pattern is computed more than once.
- Search backwards from this insn to see how many of these
- computations actually reach this insn. */
- struct occr *occr;
- rtx insn_computes_expr = NULL;
- int can_reach = 0;
-
- for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
- {
- if (BLOCK_FOR_INSN (occr->insn) == bb)
- {
- /* The expression is generated in this block.
- The only time we care about this is when the expression
- is generated later in the block [and thus there's a loop].
- We let the normal cse pass handle the other cases. */
- if (INSN_CUID (insn) < INSN_CUID (occr->insn)
- && expr_reaches_here_p (occr, expr, bb, 1))
- {
- can_reach++;
- if (can_reach > 1)
- return NULL;
-
- insn_computes_expr = occr->insn;
- }
- }
- else if (expr_reaches_here_p (occr, expr, bb, 0))
- {
- can_reach++;
- if (can_reach > 1)
- return NULL;
-
- insn_computes_expr = occr->insn;
- }
- }
-
- if (insn_computes_expr == NULL)
- abort ();
-
- return insn_computes_expr;
- }
-}
-
-/* Return nonzero if the definition in DEF_INSN can reach INSN.
- Only called by can_disregard_other_sets. */
-
-static int
-def_reaches_here_p (insn, def_insn)
- rtx insn, def_insn;
-{
- rtx reg;
-
- if (TEST_BIT (reaching_defs[BLOCK_NUM (insn)], INSN_CUID (def_insn)))
- return 1;
-
- if (BLOCK_NUM (insn) == BLOCK_NUM (def_insn))
- {
- if (INSN_CUID (def_insn) < INSN_CUID (insn))
- {
- if (GET_CODE (PATTERN (def_insn)) == PARALLEL)
- return 1;
- else if (GET_CODE (PATTERN (def_insn)) == CLOBBER)
- reg = XEXP (PATTERN (def_insn), 0);
- else if (GET_CODE (PATTERN (def_insn)) == SET)
- reg = SET_DEST (PATTERN (def_insn));
- else
- abort ();
-
- return ! reg_set_between_p (reg, NEXT_INSN (def_insn), insn);
- }
- else
- return 0;
- }
-
- return 0;
-}
-
-/* Return nonzero if *ADDR_THIS_REG can only have one value at INSN. The
- value returned is the number of definitions that reach INSN. Returning a
- value of zero means that [maybe] more than one definition reaches INSN and
- the caller can't perform whatever optimization it is trying. i.e. it is
- always safe to return zero. */
-
-static int
-can_disregard_other_sets (addr_this_reg, insn, for_combine)
- struct reg_set **addr_this_reg;
- rtx insn;
- int for_combine;
-{
- int number_of_reaching_defs = 0;
- struct reg_set *this_reg;
-
- for (this_reg = *addr_this_reg; this_reg != 0; this_reg = this_reg->next)
- if (def_reaches_here_p (insn, this_reg->insn))
- {
- number_of_reaching_defs++;
- /* Ignore parallels for now. */
- if (GET_CODE (PATTERN (this_reg->insn)) == PARALLEL)
- return 0;
-
- if (!for_combine
- && (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER
- || ! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
- SET_SRC (PATTERN (insn)))))
- /* A setting of the reg to a different value reaches INSN. */
- return 0;
-
- if (number_of_reaching_defs > 1)
- {
- /* If in this setting the value the register is being set to is
- equal to the previous value the register was set to and this
- setting reaches the insn we are trying to do the substitution
- on then we are ok. */
- if (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER)
- return 0;
- else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
- SET_SRC (PATTERN (insn))))
- return 0;
- }
-
- *addr_this_reg = this_reg;
- }
-
- return number_of_reaching_defs;
-}
-
-/* Expression computed by insn is available and the substitution is legal,
- so try to perform the substitution.
-
- The result is nonzero if any changes were made. */
-
-static int
-handle_avail_expr (insn, expr)
- rtx insn;
- struct expr *expr;
-{
- rtx pat, insn_computes_expr, expr_set;
- rtx to;
- struct reg_set *this_reg;
- int found_setting, use_src;
- int changed = 0;
-
- /* We only handle the case where one computation of the expression
- reaches this instruction. */
- insn_computes_expr = computing_insn (expr, insn);
- if (insn_computes_expr == NULL)
- return 0;
- expr_set = single_set (insn_computes_expr);
- if (!expr_set)
- abort ();
-
- found_setting = 0;
- use_src = 0;
-
- /* At this point we know only one computation of EXPR outside of this
- block reaches this insn. Now try to find a register that the
- expression is computed into. */
- if (GET_CODE (SET_SRC (expr_set)) == REG)
- {
- /* This is the case when the available expression that reaches
- here has already been handled as an available expression. */
- unsigned int regnum_for_replacing
- = REGNO (SET_SRC (expr_set));
-
- /* If the register was created by GCSE we can't use `reg_set_table',
- however we know it's set only once. */
- if (regnum_for_replacing >= max_gcse_regno
- /* If the register the expression is computed into is set only once,
- or only one set reaches this insn, we can use it. */
- || (((this_reg = reg_set_table[regnum_for_replacing]),
- this_reg->next == NULL)
- || can_disregard_other_sets (&this_reg, insn, 0)))
- {
- use_src = 1;
- found_setting = 1;
- }
- }
-
- if (!found_setting)
- {
- unsigned int regnum_for_replacing
- = REGNO (SET_DEST (expr_set));
-
- /* This shouldn't happen. */
- if (regnum_for_replacing >= max_gcse_regno)
- abort ();
-
- this_reg = reg_set_table[regnum_for_replacing];
-
- /* If the register the expression is computed into is set only once,
- or only one set reaches this insn, use it. */
- if (this_reg->next == NULL
- || can_disregard_other_sets (&this_reg, insn, 0))
- found_setting = 1;
- }
-
- if (found_setting)
- {
- pat = PATTERN (insn);
- if (use_src)
- to = SET_SRC (expr_set);
- else
- to = SET_DEST (expr_set);
- changed = validate_change (insn, &SET_SRC (pat), to, 0);
-
- /* We should be able to ignore the return code from validate_change but
- to play it safe we check. */
- if (changed)
- {
- gcse_subst_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file, "GCSE: Replacing the source in insn %d with",
- INSN_UID (insn));
- fprintf (gcse_file, " reg %d %s insn %d\n",
- REGNO (to), use_src ? "from" : "set in",
- INSN_UID (insn_computes_expr));
- }
- }
- }
-
- /* The register that the expr is computed into is set more than once. */
- else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
- {
- /* Insert an insn after insnx that copies the reg set in insnx
- into a new pseudo register call this new register REGN.
- From insnb until end of basic block or until REGB is set
- replace all uses of REGB with REGN. */
- rtx new_insn;
-
- to = gen_reg_rtx (GET_MODE (SET_DEST (expr_set)));
-
- /* Generate the new insn. */
- /* ??? If the change fails, we return 0, even though we created
- an insn. I think this is ok. */
- new_insn
- = emit_insn_after (gen_rtx_SET (VOIDmode, to,
- SET_DEST (expr_set)),
- insn_computes_expr);
-
- /* Keep register set table up to date. */
- record_one_set (REGNO (to), new_insn);
-
- gcse_create_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file, "GCSE: Creating insn %d to copy value of reg %d",
- INSN_UID (NEXT_INSN (insn_computes_expr)),
- REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr)))));
- fprintf (gcse_file, ", computed in insn %d,\n",
- INSN_UID (insn_computes_expr));
- fprintf (gcse_file, " into newly allocated reg %d\n",
- REGNO (to));
- }
-
- pat = PATTERN (insn);
-
- /* Do register replacement for INSN. */
- changed = validate_change (insn, &SET_SRC (pat),
- SET_DEST (PATTERN
- (NEXT_INSN (insn_computes_expr))),
- 0);
-
- /* We should be able to ignore the return code from validate_change but
- to play it safe we check. */
- if (changed)
- {
- gcse_subst_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file,
- "GCSE: Replacing the source in insn %d with reg %d ",
- INSN_UID (insn),
- REGNO (SET_DEST (PATTERN (NEXT_INSN
- (insn_computes_expr)))));
- fprintf (gcse_file, "set in insn %d\n",
- INSN_UID (insn_computes_expr));
- }
- }
- }
-
- return changed;
-}
-
-/* Perform classic GCSE. This is called by one_classic_gcse_pass after all
- the dataflow analysis has been done.
-
- The result is nonzero if a change was made. */
-
-static int
-classic_gcse ()
-{
- int changed;
- rtx insn;
- basic_block bb;
-
- /* Note we start at block 1. */
-
- if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
- return 0;
-
- changed = 0;
- FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
- {
- /* Reset tables used to keep track of what's still valid [since the
- start of the block]. */
- reset_opr_set_tables ();
-
- for (insn = bb->head;
- insn != NULL && insn != NEXT_INSN (bb->end);
- insn = NEXT_INSN (insn))
- {
- /* Is insn of form (set (pseudo-reg) ...)? */
- if (GET_CODE (insn) == INSN
- && GET_CODE (PATTERN (insn)) == SET
- && GET_CODE (SET_DEST (PATTERN (insn))) == REG
- && REGNO (SET_DEST (PATTERN (insn))) >= FIRST_PSEUDO_REGISTER)
- {
- rtx pat = PATTERN (insn);
- rtx src = SET_SRC (pat);
- struct expr *expr;
-
- if (want_to_gcse_p (src)
- /* Is the expression recorded? */
- && ((expr = lookup_expr (src, &expr_hash_table)) != NULL)
- /* Is the expression available [at the start of the
- block]? */
- && TEST_BIT (ae_in[bb->index], expr->bitmap_index)
- /* Are the operands unchanged since the start of the
- block? */
- && oprs_not_set_p (src, insn))
- changed |= handle_avail_expr (insn, expr);
- }
-
- /* Keep track of everything modified by this insn. */
- /* ??? Need to be careful w.r.t. mods done to INSN. */
- if (INSN_P (insn))
- mark_oprs_set (insn);
- }
- }
-
- return changed;
+ mark_call (insn);
}
-/* Top level routine to perform one classic GCSE pass.
-
- Return nonzero if a change was made. */
+/* Record things set by a CLOBBER. */
-static int
-one_classic_gcse_pass (pass)
- int pass;
+static void
+mark_clobber (rtx pat, rtx insn)
{
- int changed = 0;
+ rtx clob = XEXP (pat, 0);
- gcse_subst_count = 0;
- gcse_create_count = 0;
+ while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
+ clob = XEXP (clob, 0);
- alloc_hash_table (max_cuid, &expr_hash_table, 0);
- alloc_rd_mem (last_basic_block, max_cuid);
- compute_hash_table (&expr_hash_table);
- if (gcse_file)
- dump_hash_table (gcse_file, "Expression", &expr_hash_table);
+ if (REG_P (clob))
+ SET_REGNO_REG_SET (reg_set_bitmap, REGNO (clob));
+ else
+ record_last_mem_set_info (insn);
+}
- if (expr_hash_table.n_elems > 0)
- {
- compute_kill_rd ();
- compute_rd ();
- alloc_avail_expr_mem (last_basic_block, expr_hash_table.n_elems);
- compute_ae_gen (&expr_hash_table);
- compute_ae_kill (ae_gen, ae_kill, &expr_hash_table);
- compute_available (ae_gen, ae_kill, ae_out, ae_in);
- changed = classic_gcse ();
- free_avail_expr_mem ();
- }
+/* Record things set by INSN.
+ This data is used by oprs_not_set_p. */
- free_rd_mem ();
- free_hash_table (&expr_hash_table);
+static void
+mark_oprs_set (rtx insn)
+{
+ rtx pat = PATTERN (insn);
+ int i;
- if (gcse_file)
- {
- fprintf (gcse_file, "\n");
- fprintf (gcse_file, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
- current_function_name, pass, bytes_used, gcse_subst_count);
- fprintf (gcse_file, "%d insns created\n", gcse_create_count);
- }
+ if (GET_CODE (pat) == SET)
+ mark_set (pat, insn);
+ else if (GET_CODE (pat) == PARALLEL)
+ for (i = 0; i < XVECLEN (pat, 0); i++)
+ {
+ rtx x = XVECEXP (pat, 0, i);
- return changed;
+ if (GET_CODE (x) == SET)
+ mark_set (x, insn);
+ else if (GET_CODE (x) == CLOBBER)
+ mark_clobber (x, insn);
+ else if (GET_CODE (x) == CALL)
+ mark_call (insn);
+ }
+
+ else if (GET_CODE (pat) == CLOBBER)
+ mark_clobber (pat, insn);
+ else if (GET_CODE (pat) == CALL)
+ mark_call (insn);
}
+
\f
/* Compute copy/constant propagation working variables. */
basic blocks. N_SETS is the number of sets. */
static void
-alloc_cprop_mem (n_blocks, n_sets)
- int n_blocks, n_sets;
+alloc_cprop_mem (int n_blocks, int n_sets)
{
cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
/* Free vars used by copy/const propagation. */
static void
-free_cprop_mem ()
+free_cprop_mem (void)
{
sbitmap_vector_free (cprop_pavloc);
sbitmap_vector_free (cprop_absaltered);
bit in BMAP. */
static void
-compute_transp (x, indx, bmap, set_p)
- rtx x;
- int indx;
- sbitmap *bmap;
- int set_p;
+compute_transp (rtx x, int indx, sbitmap *bmap, int set_p)
{
int i, j;
basic_block bb;
else
{
for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
- SET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
+ SET_BIT (bmap[r->bb_index], indx);
}
}
else
else
{
for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
- RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
+ RESET_BIT (bmap[r->bb_index], indx);
}
}
return;
case MEM:
- FOR_EACH_BB (bb)
- {
- rtx list_entry = canon_modify_mem_list[bb->index];
+ {
+ bitmap_iterator bi;
+ unsigned bb_index;
- while (list_entry)
- {
- rtx dest, dest_addr;
+ /* First handle all the blocks with calls. We don't need to
+ do any list walking for them. */
+ EXECUTE_IF_SET_IN_BITMAP (blocks_with_calls, 0, bb_index, bi)
+ {
+ if (set_p)
+ SET_BIT (bmap[bb_index], indx);
+ else
+ RESET_BIT (bmap[bb_index], indx);
+ }
- if (GET_CODE (XEXP (list_entry, 0)) == CALL_INSN)
- {
- if (set_p)
- SET_BIT (bmap[bb->index], indx);
- else
- RESET_BIT (bmap[bb->index], indx);
- break;
- }
- /* LIST_ENTRY must be an INSN of some kind that sets memory.
- Examine each hunk of memory that is modified. */
+ /* Now iterate over the blocks which have memory modifications
+ but which do not have any calls. */
+ EXECUTE_IF_AND_COMPL_IN_BITMAP (modify_mem_list_set, blocks_with_calls,
+ 0, bb_index, bi)
+ {
+ rtx list_entry = canon_modify_mem_list[bb_index];
- dest = XEXP (list_entry, 0);
- list_entry = XEXP (list_entry, 1);
- dest_addr = XEXP (list_entry, 0);
+ while (list_entry)
+ {
+ rtx dest, dest_addr;
- if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
- x, rtx_addr_varies_p))
- {
- if (set_p)
- SET_BIT (bmap[bb->index], indx);
- else
- RESET_BIT (bmap[bb->index], indx);
- break;
- }
- list_entry = XEXP (list_entry, 1);
- }
- }
+ /* LIST_ENTRY must be an INSN of some kind that sets memory.
+ Examine each hunk of memory that is modified. */
+
+ dest = XEXP (list_entry, 0);
+ list_entry = XEXP (list_entry, 1);
+ dest_addr = XEXP (list_entry, 0);
+
+ if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
+ x, rtx_addr_varies_p))
+ {
+ if (set_p)
+ SET_BIT (bmap[bb_index], indx);
+ else
+ RESET_BIT (bmap[bb_index], indx);
+ break;
+ }
+ list_entry = XEXP (list_entry, 1);
+ }
+ }
+ }
x = XEXP (x, 0);
goto repeat;
propagation. */
static void
-compute_cprop_data ()
+compute_cprop_data (void)
{
compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, &set_hash_table);
compute_available (cprop_pavloc, cprop_absaltered,
This doesn't hurt anything but it will slow things down. */
static void
-find_used_regs (xptr, data)
- rtx *xptr;
- void *data ATTRIBUTE_UNUSED;
+find_used_regs (rtx *xptr, void *data ATTRIBUTE_UNUSED)
{
int i, j;
enum rtx_code code;
Returns nonzero is successful. */
static int
-try_replace_reg (from, to, insn)
- rtx from, to, insn;
+try_replace_reg (rtx from, rtx to, rtx insn)
{
rtx note = find_reg_equal_equiv_note (insn);
rtx src = 0;
validate_change (insn, &SET_SRC (set), src, 0);
}
+ /* If there is already a NOTE, update the expression in it with our
+ replacement. */
+ if (note != 0)
+ XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
+
if (!success && set && reg_mentioned_p (from, SET_SRC (set)))
{
/* If above failed and this is a single set, try to simplify the source of
&& validate_change (insn, &SET_SRC (set), src, 0))
success = 1;
- /* If we've failed to do replacement, have a single SET, and don't already
- have a note, add a REG_EQUAL note to not lose information. */
- if (!success && note == 0 && set != 0)
+ /* If we've failed to do replacement, have a single SET, don't already
+ have a note, and have no special SET, add a REG_EQUAL note to not
+ lose information. */
+ if (!success && note == 0 && set != 0
+ && GET_CODE (SET_DEST (set)) != ZERO_EXTRACT)
note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
}
- /* If there is already a NOTE, update the expression in it with our
- replacement. */
- else if (note != 0)
- XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
-
/* REG_EQUAL may get simplified into register.
We don't allow that. Remove that note. This code ought
not to happen, because previous code ought to synthesize
NULL no such set is found. */
static struct expr *
-find_avail_set (regno, insn)
- int regno;
- rtx insn;
+find_avail_set (int regno, rtx insn)
{
/* SET1 contains the last set found that can be returned to the caller for
use in a substitution. */
struct expr *set1 = 0;
/* Loops are not possible here. To get a loop we would need two sets
- available at the start of the block containing INSN. ie we would
+ available at the start of the block containing INSN. i.e. we would
need two sets like this available at the start of the block:
(set (reg X) (reg Y))
if (set == 0)
break;
- if (GET_CODE (set->expr) != SET)
- abort ();
+ gcc_assert (GET_CODE (set->expr) == SET);
src = SET_SRC (set->expr);
/* If the source of the set is anything except a register, then
we have reached the end of the copy chain. */
- if (GET_CODE (src) != REG)
+ if (! REG_P (src))
break;
- /* Follow the copy chain, ie start another iteration of the loop
+ /* Follow the copy chain, i.e. start another iteration of the loop
and see if we have an available copy into SRC. */
regno = REGNO (src);
}
if a change was made. */
static int
-cprop_jump (bb, setcc, jump, from, src)
- basic_block bb;
- rtx setcc;
- rtx jump;
- rtx from;
- rtx src;
+cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
{
rtx new, set_src, note_src;
rtx set = pc_set (jump);
run_jump_opt_after_gcse = 1;
- const_prop_count++;
+ global_const_prop_count++;
if (gcse_file != NULL)
{
fprintf (gcse_file,
- "CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
+ "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
REGNO (from), INSN_UID (jump));
print_rtl (gcse_file, src);
fprintf (gcse_file, "\n");
}
static bool
-constprop_register (insn, from, to, alter_jumps)
- rtx insn;
- rtx from;
- rtx to;
- int alter_jumps;
+constprop_register (rtx insn, rtx from, rtx to, int alter_jumps)
{
rtx sset;
}
/* Handle normal insns next. */
- if (GET_CODE (insn) == INSN
+ if (NONJUMP_INSN_P (insn)
&& try_replace_reg (from, to, insn))
return 1;
The result is nonzero if a change was made. */
static int
-cprop_insn (insn, alter_jumps)
- rtx insn;
- int alter_jumps;
+cprop_insn (rtx insn, int alter_jumps)
{
struct reg_use *reg_used;
int changed = 0;
pat = set->expr;
/* ??? We might be able to handle PARALLELs. Later. */
- if (GET_CODE (pat) != SET)
- abort ();
+ gcc_assert (GET_CODE (pat) == SET);
src = SET_SRC (pat);
if (constprop_register (insn, reg_used->reg_rtx, src, alter_jumps))
{
changed = 1;
- const_prop_count++;
+ global_const_prop_count++;
if (gcse_file != NULL)
{
fprintf (gcse_file, "GLOBAL CONST-PROP: Replacing reg %d in ", regno);
return 1;
}
}
- else if (GET_CODE (src) == REG
+ else if (REG_P (src)
&& REGNO (src) >= FIRST_PSEUDO_REGISTER
&& REGNO (src) != regno)
{
if (try_replace_reg (reg_used->reg_rtx, src, insn))
{
changed = 1;
- copy_prop_count++;
+ global_copy_prop_count++;
if (gcse_file != NULL)
{
fprintf (gcse_file, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
can legitimately make replacements. */
static void
-local_cprop_find_used_regs (xptr, data)
- rtx *xptr;
- void *data;
+local_cprop_find_used_regs (rtx *xptr, void *data)
{
rtx x = *xptr;
find_used_regs (xptr, data);
}
-
+
/* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
their REG_EQUAL notes need updating. */
static bool
-do_local_cprop (x, insn, alter_jumps, libcall_sp)
- rtx x;
- rtx insn;
- int alter_jumps;
- rtx *libcall_sp;
+do_local_cprop (rtx x, rtx insn, int alter_jumps, rtx *libcall_sp)
{
rtx newreg = NULL, newcnst = NULL;
/* Rule out USE instructions and ASM statements as we don't want to
change the hard registers mentioned. */
- if (GET_CODE (x) == REG
+ if (REG_P (x)
&& (REGNO (x) >= FIRST_PSEUDO_REGISTER
|| (GET_CODE (PATTERN (insn)) != USE
&& asm_noperands (PATTERN (insn)) < 0)))
rtx this_rtx = l->loc;
rtx note;
- if (l->in_libcall)
+ /* Don't CSE non-constant values out of libcall blocks. */
+ if (l->in_libcall && ! CONSTANT_P (this_rtx))
continue;
if (gcse_constant_p (this_rtx))
explicitly, it means address of parameter has been taken,
so we should not extend the lifetime of the pseudo. */
&& (!(note = find_reg_note (l->setting_insn, REG_EQUIV, NULL_RTX))
- || GET_CODE (XEXP (note, 0)) != MEM))
+ || ! MEM_P (XEXP (note, 0))))
newreg = this_rtx;
}
if (newcnst && constprop_register (insn, x, newcnst, alter_jumps))
or fix delete_trivially_dead_insns to preserve the setting insn,
or make it delete the REG_EUAQL note, and fix up all passes that
require the REG_EQUAL note there. */
- if (!adjust_libcall_notes (x, newcnst, insn, libcall_sp))
- abort ();
+ bool adjusted;
+
+ adjusted = adjust_libcall_notes (x, newcnst, insn, libcall_sp);
+ gcc_assert (adjusted);
+
if (gcse_file != NULL)
{
fprintf (gcse_file, "LOCAL CONST-PROP: Replacing reg %d in ",
print_rtl (gcse_file, newcnst);
fprintf (gcse_file, "\n");
}
- const_prop_count++;
+ local_const_prop_count++;
return true;
}
else if (newreg && newreg != x && try_replace_reg (x, newreg, insn))
REGNO (x), INSN_UID (insn));
fprintf (gcse_file, " with reg %d\n", REGNO (newreg));
}
- copy_prop_count++;
+ local_copy_prop_count++;
return true;
}
}
replaced with NEWVAL in INSN. Return true if all substitutions could
be made. */
static bool
-adjust_libcall_notes (oldreg, newval, insn, libcall_sp)
- rtx oldreg, newval, insn, *libcall_sp;
+adjust_libcall_notes (rtx oldreg, rtx newval, rtx insn, rtx *libcall_sp)
{
rtx end;
return true;
}
}
- XEXP (note, 0) = replace_rtx (XEXP (note, 0), oldreg, newval);
+ XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), oldreg, newval);
insn = end;
}
return true;
#define MAX_NESTED_LIBCALLS 9
static void
-local_cprop_pass (alter_jumps)
- int alter_jumps;
+local_cprop_pass (int alter_jumps)
{
rtx insn;
struct reg_use *reg_used;
rtx libcall_stack[MAX_NESTED_LIBCALLS + 1], *libcall_sp;
bool changed = false;
- cselib_init ();
+ cselib_init (false);
libcall_sp = &libcall_stack[MAX_NESTED_LIBCALLS];
*libcall_sp = 0;
for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
if (note)
{
- if (libcall_sp == libcall_stack)
- abort ();
+ gcc_assert (libcall_sp != libcall_stack);
*--libcall_sp = XEXP (note, 0);
}
note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
nonzero if a change was made. */
static int
-cprop (alter_jumps)
- int alter_jumps;
+cprop (int alter_jumps)
{
int changed;
basic_block bb;
start of the block]. */
reset_opr_set_tables ();
- for (insn = bb->head;
- insn != NULL && insn != NEXT_INSN (bb->end);
+ for (insn = BB_HEAD (bb);
+ insn != NULL && insn != NEXT_INSN (BB_END (bb));
insn = NEXT_INSN (insn))
if (INSN_P (insn))
{
/* Keep track of everything modified by this insn. */
/* ??? Need to be careful w.r.t. mods done to INSN. Don't
call mark_oprs_set if we turned the insn into a NOTE. */
- if (GET_CODE (insn) != NOTE)
+ if (! NOTE_P (insn))
mark_oprs_set (insn);
}
}
settle for the condition variable in the jump instruction being integral.
We prefer to be able to record the value of a user variable, rather than
the value of a temporary used in a condition. This could be solved by
- recording the value of *every* register scaned by canonicalize_condition,
+ recording the value of *every* register scanned by canonicalize_condition,
but this would require some code reorganization. */
rtx
-fis_get_condition (jump)
- rtx jump;
+fis_get_condition (rtx jump)
{
- rtx cond, set, tmp, insn, earliest;
- bool reverse;
-
- if (! any_condjump_p (jump))
- return NULL_RTX;
-
- set = pc_set (jump);
- cond = XEXP (SET_SRC (set), 0);
-
- /* If this branches to JUMP_LABEL when the condition is false,
- reverse the condition. */
- reverse = (GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
- && XEXP (XEXP (SET_SRC (set), 2), 0) == JUMP_LABEL (jump));
-
- /* Use canonicalize_condition to do the dirty work of manipulating
- MODE_CC values and COMPARE rtx codes. */
- tmp = canonicalize_condition (jump, cond, reverse, &earliest, NULL_RTX);
- if (!tmp)
- return NULL_RTX;
+ return get_condition (jump, NULL, false, true);
+}
- /* Verify that the given condition is valid at JUMP by virtue of not
- having been modified since EARLIEST. */
- for (insn = earliest; insn != jump; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && modified_in_p (tmp, insn))
- break;
- if (insn == jump)
- return tmp;
+/* Check the comparison COND to see if we can safely form an implicit set from
+ it. COND is either an EQ or NE comparison. */
- /* The condition was modified. See if we can get a partial result
- that doesn't follow all the reversals. Perhaps combine can fold
- them together later. */
- tmp = XEXP (tmp, 0);
- if (!REG_P (tmp) || GET_MODE_CLASS (GET_MODE (tmp)) != MODE_INT)
- return NULL_RTX;
- tmp = canonicalize_condition (jump, cond, reverse, &earliest, tmp);
- if (!tmp)
- return NULL_RTX;
+static bool
+implicit_set_cond_p (rtx cond)
+{
+ enum machine_mode mode = GET_MODE (XEXP (cond, 0));
+ rtx cst = XEXP (cond, 1);
- /* For sanity's sake, re-validate the new result. */
- for (insn = earliest; insn != jump; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && modified_in_p (tmp, insn))
- return NULL_RTX;
+ /* We can't perform this optimization if either operand might be or might
+ contain a signed zero. */
+ if (HONOR_SIGNED_ZEROS (mode))
+ {
+ /* It is sufficient to check if CST is or contains a zero. We must
+ handle float, complex, and vector. If any subpart is a zero, then
+ the optimization can't be performed. */
+ /* ??? The complex and vector checks are not implemented yet. We just
+ always return zero for them. */
+ if (GET_CODE (cst) == CONST_DOUBLE)
+ {
+ REAL_VALUE_TYPE d;
+ REAL_VALUE_FROM_CONST_DOUBLE (d, cst);
+ if (REAL_VALUES_EQUAL (d, dconst0))
+ return 0;
+ }
+ else
+ return 0;
+ }
- return tmp;
+ return gcse_constant_p (cst);
}
/* Find the implicit sets of a function. An "implicit set" is a constraint
basic block. */
static void
-find_implicit_sets ()
+find_implicit_sets (void)
{
basic_block bb, dest;
unsigned int count;
count = 0;
FOR_EACH_BB (bb)
- /* Check for more than one sucessor. */
- if (bb->succ && bb->succ->succ_next)
+ /* Check for more than one successor. */
+ if (EDGE_COUNT (bb->succs) > 1)
{
- cond = fis_get_condition (bb->end);
+ cond = fis_get_condition (BB_END (bb));
if (cond
&& (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
- && GET_CODE (XEXP (cond, 0)) == REG
+ && REG_P (XEXP (cond, 0))
&& REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
- && gcse_constant_p (XEXP (cond, 1)))
+ && implicit_set_cond_p (cond))
{
dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
: FALLTHRU_EDGE (bb)->dest;
- if (dest && ! dest->pred->pred_next
+ if (dest && single_pred_p (dest)
&& dest != EXIT_BLOCK_PTR)
{
new = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
perform conditional jump bypassing optimizations. */
static int
-one_cprop_pass (pass, cprop_jumps, bypass_jumps)
- int pass;
- int cprop_jumps;
- int bypass_jumps;
+one_cprop_pass (int pass, int cprop_jumps, int bypass_jumps)
{
int changed = 0;
- const_prop_count = 0;
- copy_prop_count = 0;
+ global_const_prop_count = local_const_prop_count = 0;
+ global_copy_prop_count = local_copy_prop_count = 0;
local_cprop_pass (cprop_jumps);
/* Determine implicit sets. */
- implicit_sets = (rtx *) xcalloc (last_basic_block, sizeof (rtx));
+ implicit_sets = xcalloc (last_basic_block, sizeof (rtx));
find_implicit_sets ();
alloc_hash_table (max_cuid, &set_hash_table, 1);
if (gcse_file)
{
fprintf (gcse_file, "CPROP of %s, pass %d: %d bytes needed, ",
- current_function_name, pass, bytes_used);
- fprintf (gcse_file, "%d const props, %d copy props\n\n",
- const_prop_count, copy_prop_count);
+ current_function_name (), pass, bytes_used);
+ fprintf (gcse_file, "%d local const props, %d local copy props\n\n",
+ local_const_prop_count, local_copy_prop_count);
+ fprintf (gcse_file, "%d global const props, %d global copy props\n\n",
+ global_const_prop_count, global_copy_prop_count);
}
/* Global analysis may get into infinite loops for unreachable blocks. */
if (changed && cprop_jumps)
find_avail_set. */
static struct expr *
-find_bypass_set (regno, bb)
- int regno;
- int bb;
+find_bypass_set (int regno, int bb)
{
struct expr *result = 0;
if (set == 0)
break;
- if (GET_CODE (set->expr) != SET)
- abort ();
+ gcc_assert (GET_CODE (set->expr) == SET);
src = SET_SRC (set->expr);
if (gcse_constant_p (src))
result = set;
- if (GET_CODE (src) != REG)
+ if (! REG_P (src))
break;
regno = REGNO (src);
valid prior to commit_edge_insertions. */
static bool
-reg_killed_on_edge (reg, e)
- rtx reg;
- edge e;
+reg_killed_on_edge (rtx reg, edge e)
{
rtx insn;
- for (insn = e->insns; insn; insn = NEXT_INSN (insn))
+ for (insn = e->insns.r; insn; insn = NEXT_INSN (insn))
if (INSN_P (insn) && reg_set_p (reg, insn))
return true;
these inserted insns when performing its transformations. */
static int
-bypass_block (bb, setcc, jump)
- basic_block bb;
- rtx setcc, jump;
+bypass_block (basic_block bb, rtx setcc, rtx jump)
{
rtx insn, note;
- edge e, enext, edest;
+ edge e, edest;
int i, change;
int may_be_loop_header;
+ unsigned removed_p;
+ edge_iterator ei;
insn = (setcc != NULL) ? setcc : jump;
find_used_regs (&XEXP (note, 0), NULL);
may_be_loop_header = false;
- for (e = bb->pred; e; e = e->pred_next)
+ FOR_EACH_EDGE (e, ei, bb->preds)
if (e->flags & EDGE_DFS_BACK)
{
may_be_loop_header = true;
}
change = 0;
- for (e = bb->pred; e; e = enext)
+ for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
{
- enext = e->pred_next;
+ removed_p = 0;
+
if (e->flags & EDGE_COMPLEX)
- continue;
+ {
+ ei_next (&ei);
+ continue;
+ }
/* We can't redirect edges from new basic blocks. */
if (e->src->index >= bypass_last_basic_block)
- continue;
+ {
+ ei_next (&ei);
+ continue;
+ }
/* The irreducible loops created by redirecting of edges entering the
loop from outside would decrease effectiveness of some of the following
optimizations, so prevent this. */
if (may_be_loop_header
&& !(e->flags & EDGE_DFS_BACK))
- continue;
+ {
+ ei_next (&ei);
+ continue;
+ }
for (i = 0; i < reg_use_count; i++)
{
continue;
/* Check the data flow is valid after edge insertions. */
- if (e->insns && reg_killed_on_edge (reg_used->reg_rtx, e))
+ if (e->insns.r && reg_killed_on_edge (reg_used->reg_rtx, e))
continue;
src = SET_SRC (pc_set (jump));
new = simplify_replace_rtx (src, reg_used->reg_rtx,
SET_SRC (set->expr));
- /* Jump bypassing may have already placed instructions on
+ /* Jump bypassing may have already placed instructions on
edges of the CFG. We can't bypass an outgoing edge that
has instructions associated with it, as these insns won't
get executed if the incoming edge is redirected. */
if (new == pc_rtx)
{
edest = FALLTHRU_EDGE (bb);
- dest = edest->insns ? NULL : edest->dest;
+ dest = edest->insns.r ? NULL : edest->dest;
}
else if (GET_CODE (new) == LABEL_REF)
{
dest = BLOCK_FOR_INSN (XEXP (new, 0));
/* Don't bypass edges containing instructions. */
- for (edest = bb->succ; edest; edest = edest->succ_next)
- if (edest->dest == dest && edest->insns)
- {
- dest = NULL;
- break;
- }
+ edest = find_edge (bb, dest);
+ if (edest && edest->insns.r)
+ dest = NULL;
}
else
dest = NULL;
+ /* Avoid unification of the edge with other edges from original
+ branch. We would end up emitting the instruction on "both"
+ edges. */
+
+ if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc)))
+ && find_edge (e->src, dest))
+ dest = NULL;
+
old_dest = e->dest;
if (dest != NULL
&& dest != old_dest
if (gcse_file != NULL)
{
- fprintf (gcse_file, "JUMP-BYPASS: Proved reg %d in jump_insn %d equals constant ",
+ fprintf (gcse_file, "JUMP-BYPASS: Proved reg %d "
+ "in jump_insn %d equals constant ",
regno, INSN_UID (jump));
print_rtl (gcse_file, SET_SRC (set->expr));
fprintf (gcse_file, "\nBypass edge from %d->%d to %d\n",
e->src->index, old_dest->index, dest->index);
}
change = 1;
+ removed_p = 1;
break;
}
}
+ if (!removed_p)
+ ei_next (&ei);
}
return change;
}
This function is now mis-named, because we also handle indirect jumps. */
static int
-bypass_conditional_jumps ()
+bypass_conditional_jumps (void)
{
basic_block bb;
int changed;
EXIT_BLOCK_PTR, next_bb)
{
/* Check for more than one predecessor. */
- if (bb->pred && bb->pred->pred_next)
+ if (!single_pred_p (bb))
{
setcc = NULL_RTX;
- for (insn = bb->head;
- insn != NULL && insn != NEXT_INSN (bb->end);
+ for (insn = BB_HEAD (bb);
+ insn != NULL && insn != NEXT_INSN (BB_END (bb));
insn = NEXT_INSN (insn))
- if (GET_CODE (insn) == INSN)
+ if (NONJUMP_INSN_P (insn))
{
if (setcc)
break;
else
break;
}
- else if (GET_CODE (insn) == JUMP_INSN)
+ else if (JUMP_P (insn))
{
if ((any_condjump_p (insn) || computed_jump_p (insn))
&& onlyjump_p (insn))
/* Allocate vars used for PRE analysis. */
static void
-alloc_pre_mem (n_blocks, n_exprs)
- int n_blocks, n_exprs;
+alloc_pre_mem (int n_blocks, int n_exprs)
{
transp = sbitmap_vector_alloc (n_blocks, n_exprs);
comp = sbitmap_vector_alloc (n_blocks, n_exprs);
pre_redundant = NULL;
pre_insert_map = NULL;
pre_delete_map = NULL;
- ae_in = NULL;
- ae_out = NULL;
ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
/* pre_insert and pre_delete are allocated later. */
/* Free vars used for PRE analysis. */
static void
-free_pre_mem ()
+free_pre_mem (void)
{
sbitmap_vector_free (transp);
sbitmap_vector_free (comp);
sbitmap_vector_free (pre_insert_map);
if (pre_delete_map)
sbitmap_vector_free (pre_delete_map);
- if (ae_in)
- sbitmap_vector_free (ae_in);
- if (ae_out)
- sbitmap_vector_free (ae_out);
transp = comp = NULL;
pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
- ae_in = ae_out = NULL;
}
/* Top level routine to do the dataflow analysis needed by PRE. */
static void
-compute_pre_data ()
+compute_pre_data (void)
{
sbitmap trapping_expr;
basic_block bb;
/* Compute ae_kill for each basic block using:
~(TRANSP | COMP)
-
- This is significantly faster than compute_ae_kill. */
+ */
FOR_EACH_BB (bb)
{
edge e;
+ edge_iterator ei;
/* If the current block is the destination of an abnormal edge, we
kill all trapping expressions because we won't be able to properly
place the instruction on the edge. So make them neither
anticipatable nor transparent. This is fairly conservative. */
- for (e = bb->pred; e ; e = e->pred_next)
+ FOR_EACH_EDGE (e, ei, bb->preds)
if (e->flags & EDGE_ABNORMAL)
{
sbitmap_difference (antloc[bb->index], antloc[bb->index], trapping_expr);
the closest such expression. */
static int
-pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited)
- basic_block occr_bb;
- struct expr *expr;
- basic_block bb;
- char *visited;
+pre_expr_reaches_here_p_work (basic_block occr_bb, struct expr *expr, basic_block bb, char *visited)
{
edge pred;
-
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
+ edge_iterator ei;
+
+ FOR_EACH_EDGE (pred, ei, bb->preds)
{
basic_block pred_bb = pred->src;
memory allocated for that function is returned. */
static int
-pre_expr_reaches_here_p (occr_bb, expr, bb)
- basic_block occr_bb;
- struct expr *expr;
- basic_block bb;
+pre_expr_reaches_here_p (basic_block occr_bb, struct expr *expr, basic_block bb)
{
int rval;
- char *visited = (char *) xcalloc (last_basic_block, 1);
+ char *visited = xcalloc (last_basic_block, 1);
rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
the value of BB. */
static rtx
-process_insert_insn (expr)
- struct expr *expr;
+process_insert_insn (struct expr *expr)
{
rtx reg = expr->reaching_reg;
rtx exp = copy_rtx (expr->expr);
/* Otherwise, make a new insn to compute this expression and make sure the
insn will be recognized (this also adds any needed CLOBBERs). Copy the
expression to make sure we don't have any sharing issues. */
- else if (insn_invalid_p (emit_insn (gen_rtx_SET (VOIDmode, reg, exp))))
- abort ();
+ else
+ {
+ rtx insn = emit_insn (gen_rtx_SET (VOIDmode, reg, exp));
+
+ if (insn_invalid_p (insn))
+ gcc_unreachable ();
+ }
+
pat = get_insns ();
end_sequence ();
no sense for code hoisting. */
static void
-insert_insn_end_bb (expr, bb, pre)
- struct expr *expr;
- basic_block bb;
- int pre;
+insert_insn_end_bb (struct expr *expr, basic_block bb, int pre)
{
- rtx insn = bb->end;
+ rtx insn = BB_END (bb);
rtx new_insn;
rtx reg = expr->reaching_reg;
int regno = REGNO (reg);
rtx pat, pat_end;
pat = process_insert_insn (expr);
- if (pat == NULL_RTX || ! INSN_P (pat))
- abort ();
+ gcc_assert (pat && INSN_P (pat));
pat_end = pat;
while (NEXT_INSN (pat_end) != NULL_RTX)
pat_end = NEXT_INSN (pat_end);
/* If the last insn is a jump, insert EXPR in front [taking care to
- handle cc0, etc. properly]. Similary we need to care trapping
+ handle cc0, etc. properly]. Similarly we need to care trapping
instructions in presence of non-call exceptions. */
- if (GET_CODE (insn) == JUMP_INSN
- || (GET_CODE (insn) == INSN
- && (bb->succ->succ_next || (bb->succ->flags & EDGE_ABNORMAL))))
+ if (JUMP_P (insn)
+ || (NONJUMP_INSN_P (insn)
+ && (!single_succ_p (bb)
+ || single_succ_edge (bb)->flags & EDGE_ABNORMAL)))
{
#ifdef HAVE_cc0
rtx note;
/* It should always be the case that we can put these instructions
anywhere in the basic block with performing PRE optimizations.
Check this. */
- if (GET_CODE (insn) == INSN && pre
- && !TEST_BIT (antloc[bb->index], expr->bitmap_index)
- && !TEST_BIT (transp[bb->index], expr->bitmap_index))
- abort ();
+ gcc_assert (!NONJUMP_INSN_P (insn) || !pre
+ || TEST_BIT (antloc[bb->index], expr->bitmap_index)
+ || TEST_BIT (transp[bb->index], expr->bitmap_index));
/* If this is a jump table, then we can't insert stuff here. Since
we know the previous real insn must be the tablejump, we insert
}
#endif
/* FIXME: What if something in cc0/jump uses value set in new insn? */
- new_insn = emit_insn_before (pat, insn);
+ new_insn = emit_insn_before_noloc (pat, insn);
}
/* Likewise if the last insn is a call, as will happen in the presence
of exception handling. */
- else if (GET_CODE (insn) == CALL_INSN
- && (bb->succ->succ_next || (bb->succ->flags & EDGE_ABNORMAL)))
+ else if (CALL_P (insn)
+ && (!single_succ_p (bb)
+ || single_succ_edge (bb)->flags & EDGE_ABNORMAL))
{
/* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
we search backward and place the instructions before the first
anywhere in the basic block with performing PRE optimizations.
Check this. */
- if (pre
- && !TEST_BIT (antloc[bb->index], expr->bitmap_index)
- && !TEST_BIT (transp[bb->index], expr->bitmap_index))
- abort ();
+ gcc_assert (!pre
+ || TEST_BIT (antloc[bb->index], expr->bitmap_index)
+ || TEST_BIT (transp[bb->index], expr->bitmap_index));
/* Since different machines initialize their parameter registers
in different orders, assume nothing. Collect the set of all
parameter registers. */
- insn = find_first_parameter_load (insn, bb->head);
+ insn = find_first_parameter_load (insn, BB_HEAD (bb));
/* If we found all the parameter loads, then we want to insert
before the first parameter load.
If we inserted before the CODE_LABEL, then we would be putting
the insn in the wrong basic block. In that case, put the insn
after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
- while (GET_CODE (insn) == CODE_LABEL
+ while (LABEL_P (insn)
|| NOTE_INSN_BASIC_BLOCK_P (insn))
insn = NEXT_INSN (insn);
- new_insn = emit_insn_before (pat, insn);
+ new_insn = emit_insn_before_noloc (pat, insn);
}
else
- new_insn = emit_insn_after (pat, insn);
+ new_insn = emit_insn_after_noloc (pat, insn);
while (1)
{
the expressions fully redundant. */
static int
-pre_edge_insert (edge_list, index_map)
- struct edge_list *edge_list;
- struct expr **index_map;
+pre_edge_insert (struct edge_list *edge_list, struct expr **index_map)
{
int e, i, j, num_edges, set_size, did_insert = 0;
sbitmap *inserted;
if (! occr->deleted_p)
continue;
- /* Insert this expression on this edge if if it would
+ /* Insert this expression on this edge if it would
reach the deleted occurrence in BB. */
if (!TEST_BIT (inserted[e], j))
{
handling this situation. This one is easiest for
now. */
- if ((eg->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
+ if (eg->flags & EDGE_ABNORMAL)
insert_insn_end_bb (index_map[j], bb, 0);
else
{
return did_insert;
}
-/* Copy the result of INSN to REG. INDX is the expression number. */
+/* Copy the result of EXPR->EXPR generated by INSN to EXPR->REACHING_REG.
+ Given "old_reg <- expr" (INSN), instead of adding after it
+ reaching_reg <- old_reg
+ it's better to do the following:
+ reaching_reg <- expr
+ old_reg <- reaching_reg
+ because this way copy propagation can discover additional PRE
+ opportunities. But if this fails, we try the old way.
+ When "expr" is a store, i.e.
+ given "MEM <- old_reg", instead of adding after it
+ reaching_reg <- old_reg
+ it's better to add it before as follows:
+ reaching_reg <- old_reg
+ MEM <- reaching_reg. */
static void
-pre_insert_copy_insn (expr, insn)
- struct expr *expr;
- rtx insn;
+pre_insert_copy_insn (struct expr *expr, rtx insn)
{
rtx reg = expr->reaching_reg;
int regno = REGNO (reg);
int indx = expr->bitmap_index;
- rtx set = single_set (insn);
- rtx new_insn;
+ rtx pat = PATTERN (insn);
+ rtx set, new_insn;
+ rtx old_reg;
+ int i;
- if (!set)
- abort ();
+ /* This block matches the logic in hash_scan_insn. */
+ switch (GET_CODE (pat))
+ {
+ case SET:
+ set = pat;
+ break;
+
+ case PARALLEL:
+ /* Search through the parallel looking for the set whose
+ source was the expression that we're interested in. */
+ set = NULL_RTX;
+ for (i = 0; i < XVECLEN (pat, 0); i++)
+ {
+ rtx x = XVECEXP (pat, 0, i);
+ if (GET_CODE (x) == SET
+ && expr_equiv_p (SET_SRC (x), expr->expr))
+ {
+ set = x;
+ break;
+ }
+ }
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ if (REG_P (SET_DEST (set)))
+ {
+ old_reg = SET_DEST (set);
+ /* Check if we can modify the set destination in the original insn. */
+ if (validate_change (insn, &SET_DEST (set), reg, 0))
+ {
+ new_insn = gen_move_insn (old_reg, reg);
+ new_insn = emit_insn_after (new_insn, insn);
- new_insn = emit_insn_after (gen_move_insn (reg, copy_rtx (SET_DEST (set))), insn);
+ /* Keep register set table up to date. */
+ record_one_set (regno, insn);
+ }
+ else
+ {
+ new_insn = gen_move_insn (reg, old_reg);
+ new_insn = emit_insn_after (new_insn, insn);
+
+ /* Keep register set table up to date. */
+ record_one_set (regno, new_insn);
+ }
+ }
+ else /* This is possible only in case of a store to memory. */
+ {
+ old_reg = SET_SRC (set);
+ new_insn = gen_move_insn (reg, old_reg);
+
+ /* Check if we can modify the set source in the original insn. */
+ if (validate_change (insn, &SET_SRC (set), reg, 0))
+ new_insn = emit_insn_before (new_insn, insn);
+ else
+ new_insn = emit_insn_after (new_insn, insn);
- /* Keep register set table up to date. */
- record_one_set (regno, new_insn);
+ /* Keep register set table up to date. */
+ record_one_set (regno, new_insn);
+ }
gcse_create_count++;
"PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
BLOCK_NUM (insn), INSN_UID (new_insn), indx,
INSN_UID (insn), regno);
- update_ld_motion_stores (expr);
}
/* Copy available expressions that reach the redundant expression
to `reaching_reg'. */
static void
-pre_insert_copies ()
+pre_insert_copies (void)
{
- unsigned int i;
+ unsigned int i, added_copy;
struct expr *expr;
struct occr *occr;
struct occr *avail;
if (expr->reaching_reg == NULL)
continue;
+ /* Set when we add a copy for that expression. */
+ added_copy = 0;
+
for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
{
if (! occr->deleted_p)
BLOCK_FOR_INSN (occr->insn)))
continue;
+ added_copy = 1;
+
/* Copy the result of avail to reaching_reg. */
pre_insert_copy_insn (expr, insn);
avail->copied_p = 1;
}
}
+
+ if (added_copy)
+ update_ld_motion_stores (expr);
}
}
/* Emit move from SRC to DEST noting the equivalence with expression computed
in INSN. */
static rtx
-gcse_emit_move_after (src, dest, insn)
- rtx src, dest, insn;
+gcse_emit_move_after (rtx src, rtx dest, rtx insn)
{
rtx new;
rtx set = single_set (insn), set2;
Returns nonzero if a change is made. */
static int
-pre_delete ()
+pre_delete (void)
{
unsigned int i;
int changed;
changed = 0;
for (i = 0; i < expr_hash_table.size; i++)
- for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
+ for (expr = expr_hash_table.table[i];
+ expr != NULL;
+ expr = expr->next_same_hash)
{
int indx = expr->bitmap_index;
rtx set;
basic_block bb = BLOCK_FOR_INSN (insn);
- if (TEST_BIT (pre_delete_map[bb->index], indx))
+ /* We only delete insns that have a single_set. */
+ if (TEST_BIT (pre_delete_map[bb->index], indx)
+ && (set = single_set (insn)) != 0)
{
- set = single_set (insn);
- if (! set)
- abort ();
-
/* Create a pseudo-reg to store the result of reaching
expressions into. Get the mode for the new pseudo from
the mode of the original destination pseudo. */
redundancies. */
static int
-pre_gcse ()
+pre_gcse (void)
{
unsigned int i;
int did_insert, changed;
/* Compute a mapping from expression number (`bitmap_index') to
hash table entry. */
- index_map = (struct expr **) xcalloc (expr_hash_table.n_elems, sizeof (struct expr *));
+ index_map = xcalloc (expr_hash_table.n_elems, sizeof (struct expr *));
for (i = 0; i < expr_hash_table.size; i++)
for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
index_map[expr->bitmap_index] = expr;
Return nonzero if a change was made. */
static int
-one_pre_gcse_pass (pass)
- int pass;
+one_pre_gcse_pass (int pass)
{
int changed = 0;
}
free_ldst_mems ();
- remove_fake_edges ();
+ remove_fake_exit_edges ();
free_hash_table (&expr_hash_table);
if (gcse_file)
{
fprintf (gcse_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
- current_function_name, pass, bytes_used);
+ current_function_name (), pass, bytes_used);
fprintf (gcse_file, "%d substs, %d insns created\n",
gcse_subst_count, gcse_create_count);
}
necessary REG_LABEL notes. */
static void
-add_label_notes (x, insn)
- rtx x;
- rtx insn;
+add_label_notes (rtx x, rtx insn)
{
enum rtx_code code = GET_CODE (x);
int i, j;
EH table sizes, this may not be worthwhile. */
static void
-compute_transpout ()
-{
- basic_block bb;
- unsigned int i;
- struct expr *expr;
-
- sbitmap_vector_ones (transpout, last_basic_block);
-
- FOR_EACH_BB (bb)
- {
- /* Note that flow inserted a nop a the end of basic blocks that
- end in call instructions for reasons other than abnormal
- control flow. */
- if (GET_CODE (bb->end) != CALL_INSN)
- continue;
-
- for (i = 0; i < expr_hash_table.size; i++)
- for (expr = expr_hash_table.table[i]; expr ; expr = expr->next_same_hash)
- if (GET_CODE (expr->expr) == MEM)
- {
- if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
- && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
- continue;
-
- /* ??? Optimally, we would use interprocedural alias
- analysis to determine if this mem is actually killed
- by this call. */
- RESET_BIT (transpout[bb->index], expr->bitmap_index);
- }
- }
-}
-
-/* Removal of useless null pointer checks */
-
-/* Called via note_stores. X is set by SETTER. If X is a register we must
- invalidate nonnull_local and set nonnull_killed. DATA is really a
- `null_pointer_info *'.
-
- We ignore hard registers. */
-
-static void
-invalidate_nonnull_info (x, setter, data)
- rtx x;
- rtx setter ATTRIBUTE_UNUSED;
- void *data;
-{
- unsigned int regno;
- struct null_pointer_info *npi = (struct null_pointer_info *) data;
-
- while (GET_CODE (x) == SUBREG)
- x = SUBREG_REG (x);
-
- /* Ignore anything that is not a register or is a hard register. */
- if (GET_CODE (x) != REG
- || REGNO (x) < npi->min_reg
- || REGNO (x) >= npi->max_reg)
- return;
-
- regno = REGNO (x) - npi->min_reg;
-
- RESET_BIT (npi->nonnull_local[npi->current_block->index], regno);
- SET_BIT (npi->nonnull_killed[npi->current_block->index], regno);
-}
-
-/* Do null-pointer check elimination for the registers indicated in
- NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
- they are not our responsibility to free. */
-
-static int
-delete_null_pointer_checks_1 (block_reg, nonnull_avin,
- nonnull_avout, npi)
- unsigned int *block_reg;
- sbitmap *nonnull_avin;
- sbitmap *nonnull_avout;
- struct null_pointer_info *npi;
-{
- basic_block bb, current_block;
- sbitmap *nonnull_local = npi->nonnull_local;
- sbitmap *nonnull_killed = npi->nonnull_killed;
- int something_changed = 0;
-
- /* Compute local properties, nonnull and killed. A register will have
- the nonnull property if at the end of the current block its value is
- known to be nonnull. The killed property indicates that somewhere in
- the block any information we had about the register is killed.
-
- Note that a register can have both properties in a single block. That
- indicates that it's killed, then later in the block a new value is
- computed. */
- sbitmap_vector_zero (nonnull_local, last_basic_block);
- sbitmap_vector_zero (nonnull_killed, last_basic_block);
-
- FOR_EACH_BB (current_block)
- {
- rtx insn, stop_insn;
-
- /* Set the current block for invalidate_nonnull_info. */
- npi->current_block = current_block;
-
- /* Scan each insn in the basic block looking for memory references and
- register sets. */
- stop_insn = NEXT_INSN (current_block->end);
- for (insn = current_block->head;
- insn != stop_insn;
- insn = NEXT_INSN (insn))
- {
- rtx set;
- rtx reg;
-
- /* Ignore anything that is not a normal insn. */
- if (! INSN_P (insn))
- continue;
-
- /* Basically ignore anything that is not a simple SET. We do have
- to make sure to invalidate nonnull_local and set nonnull_killed
- for such insns though. */
- set = single_set (insn);
- if (!set)
- {
- note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
- continue;
- }
-
- /* See if we've got a usable memory load. We handle it first
- in case it uses its address register as a dest (which kills
- the nonnull property). */
- if (GET_CODE (SET_SRC (set)) == MEM
- && GET_CODE ((reg = XEXP (SET_SRC (set), 0))) == REG
- && REGNO (reg) >= npi->min_reg
- && REGNO (reg) < npi->max_reg)
- SET_BIT (nonnull_local[current_block->index],
- REGNO (reg) - npi->min_reg);
-
- /* Now invalidate stuff clobbered by this insn. */
- note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
-
- /* And handle stores, we do these last since any sets in INSN can
- not kill the nonnull property if it is derived from a MEM
- appearing in a SET_DEST. */
- if (GET_CODE (SET_DEST (set)) == MEM
- && GET_CODE ((reg = XEXP (SET_DEST (set), 0))) == REG
- && REGNO (reg) >= npi->min_reg
- && REGNO (reg) < npi->max_reg)
- SET_BIT (nonnull_local[current_block->index],
- REGNO (reg) - npi->min_reg);
- }
- }
-
- /* Now compute global properties based on the local properties. This
- is a classic global availability algorithm. */
- compute_available (nonnull_local, nonnull_killed,
- nonnull_avout, nonnull_avin);
-
- /* Now look at each bb and see if it ends with a compare of a value
- against zero. */
- FOR_EACH_BB (bb)
- {
- rtx last_insn = bb->end;
- rtx condition, earliest;
- int compare_and_branch;
-
- /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
- since BLOCK_REG[BB] is zero if this block did not end with a
- comparison against zero, this condition works. */
- if (block_reg[bb->index] < npi->min_reg
- || block_reg[bb->index] >= npi->max_reg)
- continue;
-
- /* LAST_INSN is a conditional jump. Get its condition. */
- condition = get_condition (last_insn, &earliest);
-
- /* If we can't determine the condition then skip. */
- if (! condition)
- continue;
-
- /* Is the register known to have a nonzero value? */
- if (!TEST_BIT (nonnull_avout[bb->index], block_reg[bb->index] - npi->min_reg))
- continue;
-
- /* Try to compute whether the compare/branch at the loop end is one or
- two instructions. */
- if (earliest == last_insn)
- compare_and_branch = 1;
- else if (earliest == prev_nonnote_insn (last_insn))
- compare_and_branch = 2;
- else
- continue;
-
- /* We know the register in this comparison is nonnull at exit from
- this block. We can optimize this comparison. */
- if (GET_CODE (condition) == NE)
- {
- rtx new_jump;
-
- new_jump = emit_jump_insn_after (gen_jump (JUMP_LABEL (last_insn)),
- last_insn);
- JUMP_LABEL (new_jump) = JUMP_LABEL (last_insn);
- LABEL_NUSES (JUMP_LABEL (new_jump))++;
- emit_barrier_after (new_jump);
- }
-
- something_changed = 1;
- delete_insn (last_insn);
- if (compare_and_branch == 2)
- delete_insn (earliest);
- purge_dead_edges (bb);
-
- /* Don't check this block again. (Note that BLOCK_END is
- invalid here; we deleted the last instruction in the
- block.) */
- block_reg[bb->index] = 0;
- }
-
- return something_changed;
-}
-
-/* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
- at compile time.
-
- This is conceptually similar to global constant/copy propagation and
- classic global CSE (it even uses the same dataflow equations as cprop).
-
- If a register is used as memory address with the form (mem (reg)), then we
- know that REG can not be zero at that point in the program. Any instruction
- which sets REG "kills" this property.
-
- So, if every path leading to a conditional branch has an available memory
- reference of that form, then we know the register can not have the value
- zero at the conditional branch.
-
- So we merely need to compute the local properties and propagate that data
- around the cfg, then optimize where possible.
-
- We run this pass two times. Once before CSE, then again after CSE. This
- has proven to be the most profitable approach. It is rare for new
- optimization opportunities of this nature to appear after the first CSE
- pass.
-
- This could probably be integrated with global cprop with a little work. */
-
-int
-delete_null_pointer_checks (f)
- rtx f ATTRIBUTE_UNUSED;
+compute_transpout (void)
{
- sbitmap *nonnull_avin, *nonnull_avout;
- unsigned int *block_reg;
basic_block bb;
- int reg;
- int regs_per_pass;
- int max_reg;
- struct null_pointer_info npi;
- int something_changed = 0;
-
- /* If we have only a single block, then there's nothing to do. */
- if (n_basic_blocks <= 1)
- return 0;
-
- /* Trying to perform global optimizations on flow graphs which have
- a high connectivity will take a long time and is unlikely to be
- particularly useful.
+ unsigned int i;
+ struct expr *expr;
- In normal circumstances a cfg should have about twice as many edges
- as blocks. But we do not want to punish small functions which have
- a couple switch statements. So we require a relatively large number
- of basic blocks and the ratio of edges to blocks to be high. */
- if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
- return 0;
+ sbitmap_vector_ones (transpout, last_basic_block);
- /* We need four bitmaps, each with a bit for each register in each
- basic block. */
- max_reg = max_reg_num ();
- regs_per_pass = get_bitmap_width (4, last_basic_block, max_reg);
-
- /* Allocate bitmaps to hold local and global properties. */
- npi.nonnull_local = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
- npi.nonnull_killed = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
- nonnull_avin = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
- nonnull_avout = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
-
- /* Go through the basic blocks, seeing whether or not each block
- ends with a conditional branch whose condition is a comparison
- against zero. Record the register compared in BLOCK_REG. */
- block_reg = (unsigned int *) xcalloc (last_basic_block, sizeof (int));
FOR_EACH_BB (bb)
{
- rtx last_insn = bb->end;
- rtx condition, earliest, reg;
-
- /* We only want conditional branches. */
- if (GET_CODE (last_insn) != JUMP_INSN
- || !any_condjump_p (last_insn)
- || !onlyjump_p (last_insn))
- continue;
-
- /* LAST_INSN is a conditional jump. Get its condition. */
- condition = get_condition (last_insn, &earliest);
-
- /* If we were unable to get the condition, or it is not an equality
- comparison against zero then there's nothing we can do. */
- if (!condition
- || (GET_CODE (condition) != NE && GET_CODE (condition) != EQ)
- || GET_CODE (XEXP (condition, 1)) != CONST_INT
- || (XEXP (condition, 1)
- != CONST0_RTX (GET_MODE (XEXP (condition, 0)))))
- continue;
-
- /* We must be checking a register against zero. */
- reg = XEXP (condition, 0);
- if (GET_CODE (reg) != REG)
+ /* Note that flow inserted a nop a the end of basic blocks that
+ end in call instructions for reasons other than abnormal
+ control flow. */
+ if (! CALL_P (BB_END (bb)))
continue;
- block_reg[bb->index] = REGNO (reg);
- }
+ for (i = 0; i < expr_hash_table.size; i++)
+ for (expr = expr_hash_table.table[i]; expr ; expr = expr->next_same_hash)
+ if (MEM_P (expr->expr))
+ {
+ if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
+ && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
+ continue;
- /* Go through the algorithm for each block of registers. */
- for (reg = FIRST_PSEUDO_REGISTER; reg < max_reg; reg += regs_per_pass)
- {
- npi.min_reg = reg;
- npi.max_reg = MIN (reg + regs_per_pass, max_reg);
- something_changed |= delete_null_pointer_checks_1 (block_reg,
- nonnull_avin,
- nonnull_avout,
- &npi);
+ /* ??? Optimally, we would use interprocedural alias
+ analysis to determine if this mem is actually killed
+ by this call. */
+ RESET_BIT (transpout[bb->index], expr->bitmap_index);
+ }
}
-
- /* Free the table of registers compared at the end of every block. */
- free (block_reg);
-
- /* Free bitmaps. */
- sbitmap_vector_free (npi.nonnull_local);
- sbitmap_vector_free (npi.nonnull_killed);
- sbitmap_vector_free (nonnull_avin);
- sbitmap_vector_free (nonnull_avout);
-
- return something_changed;
}
/* Code Hoisting variables and subroutines. */
/* Hoistable expressions. */
static sbitmap *hoist_exprs;
-/* Dominator bitmaps. */
-dominance_info dominators;
-
/* ??? We could compute post dominators and run this algorithm in
reverse to perform tail merging, doing so would probably be
more effective than the tail merging code in jump.c.
/* Allocate vars used for code hoisting analysis. */
static void
-alloc_code_hoist_mem (n_blocks, n_exprs)
- int n_blocks, n_exprs;
+alloc_code_hoist_mem (int n_blocks, int n_exprs)
{
antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
transp = sbitmap_vector_alloc (n_blocks, n_exprs);
/* Free vars used for code hoisting analysis. */
static void
-free_code_hoist_mem ()
+free_code_hoist_mem (void)
{
sbitmap_vector_free (antloc);
sbitmap_vector_free (transp);
sbitmap_vector_free (hoist_exprs);
sbitmap_vector_free (transpout);
- free_dominance_info (dominators);
+ free_dominance_info (CDI_DOMINATORS);
}
/* Compute the very busy expressions at entry/exit from each block.
compute the expression. */
static void
-compute_code_hoist_vbeinout ()
+compute_code_hoist_vbeinout (void)
{
int changed, passes;
basic_block bb;
/* Top level routine to do the dataflow analysis needed by code hoisting. */
static void
-compute_code_hoist_data ()
+compute_code_hoist_data (void)
{
compute_local_properties (transp, comp, antloc, &expr_hash_table);
compute_transpout ();
compute_code_hoist_vbeinout ();
- dominators = calculate_dominance_info (CDI_DOMINATORS);
+ calculate_dominance_info (CDI_DOMINATORS);
if (gcse_file)
fprintf (gcse_file, "\n");
}
paths. */
static int
-hoist_expr_reaches_here_p (expr_bb, expr_index, bb, visited)
- basic_block expr_bb;
- int expr_index;
- basic_block bb;
- char *visited;
+hoist_expr_reaches_here_p (basic_block expr_bb, int expr_index, basic_block bb, char *visited)
{
edge pred;
+ edge_iterator ei;
int visited_allocated_locally = 0;
visited = xcalloc (last_basic_block, 1);
}
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
+ FOR_EACH_EDGE (pred, ei, bb->preds)
{
basic_block pred_bb = pred->src;
/* Actually perform code hoisting. */
static void
-hoist_code ()
+hoist_code (void)
{
basic_block bb, dominated;
basic_block *domby;
/* Compute a mapping from expression number (`bitmap_index') to
hash table entry. */
- index_map = (struct expr **) xcalloc (expr_hash_table.n_elems, sizeof (struct expr *));
+ index_map = xcalloc (expr_hash_table.n_elems, sizeof (struct expr *));
for (i = 0; i < expr_hash_table.size; i++)
for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
index_map[expr->bitmap_index] = expr;
int found = 0;
int insn_inserted_p;
- domby_len = get_dominated_by (dominators, bb, &domby);
+ domby_len = get_dominated_by (CDI_DOMINATORS, bb, &domby);
/* Examine each expression that is very busy at the exit of this
block. These are the potentially hoistable expressions. */
for (i = 0; i < hoist_vbeout[bb->index]->n_bits; i++)
/* If we found nothing to hoist, then quit now. */
if (! found)
{
- free (domby);
+ free (domby);
continue;
}
while (BLOCK_FOR_INSN (occr->insn) != dominated && occr)
occr = occr->next;
- /* Should never happen. */
- if (!occr)
- abort ();
-
+ gcc_assert (occr);
insn = occr->insn;
-
set = single_set (insn);
- if (! set)
- abort ();
+ gcc_assert (set);
/* Create a pseudo-reg to store the result of reaching
expressions into. Get the mode for the new pseudo
Return nonzero if a change was made. */
static int
-one_code_hoisting_pass ()
+one_code_hoisting_pass (void)
{
int changed = 0;
doesn't find one, we create one and initialize it. */
static struct ls_expr *
-ldst_entry (x)
- rtx x;
+ldst_entry (rtx x)
{
+ int do_not_record_p = 0;
struct ls_expr * ptr;
+ unsigned int hash;
- for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
- if (expr_equiv_p (ptr->pattern, x))
- break;
+ hash = hash_rtx (x, GET_MODE (x), &do_not_record_p,
+ NULL, /*have_reg_qty=*/false);
- if (!ptr)
- {
- ptr = (struct ls_expr *) xmalloc (sizeof (struct ls_expr));
-
- ptr->next = pre_ldst_mems;
- ptr->expr = NULL;
- ptr->pattern = x;
- ptr->pattern_regs = NULL_RTX;
- ptr->loads = NULL_RTX;
- ptr->stores = NULL_RTX;
- ptr->reaching_reg = NULL_RTX;
- ptr->invalid = 0;
- ptr->index = 0;
- ptr->hash_index = 0;
- pre_ldst_mems = ptr;
- }
+ for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
+ if (ptr->hash_index == hash && expr_equiv_p (ptr->pattern, x))
+ return ptr;
+
+ ptr = xmalloc (sizeof (struct ls_expr));
+
+ ptr->next = pre_ldst_mems;
+ ptr->expr = NULL;
+ ptr->pattern = x;
+ ptr->pattern_regs = NULL_RTX;
+ ptr->loads = NULL_RTX;
+ ptr->stores = NULL_RTX;
+ ptr->reaching_reg = NULL_RTX;
+ ptr->invalid = 0;
+ ptr->index = 0;
+ ptr->hash_index = hash;
+ pre_ldst_mems = ptr;
return ptr;
}
/* Free up an individual ldst entry. */
static void
-free_ldst_entry (ptr)
- struct ls_expr * ptr;
+free_ldst_entry (struct ls_expr * ptr)
{
free_INSN_LIST_list (& ptr->loads);
free_INSN_LIST_list (& ptr->stores);
/* Free up all memory associated with the ldst list. */
static void
-free_ldst_mems ()
+free_ldst_mems (void)
{
while (pre_ldst_mems)
{
/* Dump debugging info about the ldst list. */
static void
-print_ldst_list (file)
- FILE * file;
+print_ldst_list (FILE * file)
{
struct ls_expr * ptr;
/* Returns 1 if X is in the list of ldst only expressions. */
static struct ls_expr *
-find_rtx_in_ldst (x)
- rtx x;
+find_rtx_in_ldst (rtx x)
{
struct ls_expr * ptr;
/* Assign each element of the list of mems a monotonically increasing value. */
static int
-enumerate_ldsts ()
+enumerate_ldsts (void)
{
struct ls_expr * ptr;
int n = 0;
/* Return first item in the list. */
static inline struct ls_expr *
-first_ls_expr ()
+first_ls_expr (void)
{
return pre_ldst_mems;
}
-/* Return the next item in ther list after the specified one. */
+/* Return the next item in the list after the specified one. */
static inline struct ls_expr *
-next_ls_expr (ptr)
- struct ls_expr * ptr;
+next_ls_expr (struct ls_expr * ptr)
{
return ptr->next;
}
ld_motion list, otherwise we let the usual aliasing take care of it. */
static int
-simple_mem (x)
- rtx x;
+simple_mem (rtx x)
{
- if (GET_CODE (x) != MEM)
+ if (! MEM_P (x))
return 0;
if (MEM_VOLATILE_P (x))
fix it up. */
static void
-invalidate_any_buried_refs (x)
- rtx x;
+invalidate_any_buried_refs (rtx x)
{
const char * fmt;
int i, j;
struct ls_expr * ptr;
/* Invalidate it in the list. */
- if (GET_CODE (x) == MEM && simple_mem (x))
+ if (MEM_P (x) && simple_mem (x))
{
ptr = ldst_entry (x);
ptr->invalid = 1;
later. */
static void
-compute_ld_motion_mems ()
+compute_ld_motion_mems (void)
{
struct ls_expr * ptr;
basic_block bb;
FOR_EACH_BB (bb)
{
- for (insn = bb->head;
- insn && insn != NEXT_INSN (bb->end);
+ for (insn = BB_HEAD (bb);
+ insn && insn != NEXT_INSN (BB_END (bb));
insn = NEXT_INSN (insn))
{
if (INSN_P (insn))
rtx dest = SET_DEST (PATTERN (insn));
/* Check for a simple LOAD... */
- if (GET_CODE (src) == MEM && simple_mem (src))
+ if (MEM_P (src) && simple_mem (src))
{
ptr = ldst_entry (src);
- if (GET_CODE (dest) == REG)
+ if (REG_P (dest))
ptr->loads = alloc_INSN_LIST (insn, ptr->loads);
else
ptr->invalid = 1;
will block any movement we might do later. We only care
about this exact pattern since those are the only
circumstance that we will ignore the aliasing info. */
- if (GET_CODE (dest) == MEM && simple_mem (dest))
+ if (MEM_P (dest) && simple_mem (dest))
{
ptr = ldst_entry (dest);
- if (GET_CODE (src) != MEM
+ if (! MEM_P (src)
&& GET_CODE (src) != ASM_OPERANDS
/* Check for REG manually since want_to_gcse_p
returns 0 for all REGs. */
- && (REG_P (src) || want_to_gcse_p (src)))
+ && can_assign_to_reg_p (src))
ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
else
ptr->invalid = 1;
expression list for pre gcse. */
static void
-trim_ld_motion_mems ()
+trim_ld_motion_mems (void)
{
- struct ls_expr * last = NULL;
- struct ls_expr * ptr = first_ls_expr ();
+ struct ls_expr * * last = & pre_ldst_mems;
+ struct ls_expr * ptr = pre_ldst_mems;
while (ptr != NULL)
{
- int del = ptr->invalid;
- struct expr * expr = NULL;
+ struct expr * expr;
/* Delete if entry has been made invalid. */
- if (!del)
+ if (! ptr->invalid)
{
- unsigned int i;
-
- del = 1;
/* Delete if we cannot find this mem in the expression list. */
- for (i = 0; i < expr_hash_table.size && del; i++)
- {
- for (expr = expr_hash_table.table[i];
- expr != NULL;
- expr = expr->next_same_hash)
- if (expr_equiv_p (expr->expr, ptr->pattern))
- {
- del = 0;
- break;
- }
- }
- }
+ unsigned int hash = ptr->hash_index % expr_hash_table.size;
- if (del)
- {
- if (last != NULL)
- {
- last->next = ptr->next;
- free_ldst_entry (ptr);
- ptr = last->next;
- }
- else
- {
- pre_ldst_mems = pre_ldst_mems->next;
- free_ldst_entry (ptr);
- ptr = pre_ldst_mems;
- }
+ for (expr = expr_hash_table.table[hash];
+ expr != NULL;
+ expr = expr->next_same_hash)
+ if (expr_equiv_p (expr->expr, ptr->pattern))
+ break;
}
else
+ expr = (struct expr *) 0;
+
+ if (expr)
{
/* Set the expression field if we are keeping it. */
- last = ptr;
ptr->expr = expr;
+ last = & ptr->next;
ptr = ptr->next;
}
+ else
+ {
+ *last = ptr->next;
+ free_ldst_entry (ptr);
+ ptr = * last;
+ }
}
/* Show the world what we've found. */
/* This routine will take an expression which we are replacing with
a reaching register, and update any stores that are needed if
that expression is in the ld_motion list. Stores are updated by
- copying their SRC to the reaching register, and then storeing
+ copying their SRC to the reaching register, and then storing
the reaching register into the store location. These keeps the
correct value in the reaching register for the loads. */
static void
-update_ld_motion_stores (expr)
- struct expr * expr;
+update_ld_motion_stores (struct expr * expr)
{
struct ls_expr * mem_ptr;
/* Global holding the number of store expressions we are dealing with. */
static int num_stores;
-/* Checks to set if we need to mark a register set. Called from note_stores. */
+/* Checks to set if we need to mark a register set. Called from
+ note_stores. */
+
+static void
+reg_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED,
+ void *data)
+{
+ sbitmap bb_reg = data;
+
+ if (GET_CODE (dest) == SUBREG)
+ dest = SUBREG_REG (dest);
+
+ if (REG_P (dest))
+ {
+ regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
+ if (bb_reg)
+ SET_BIT (bb_reg, REGNO (dest));
+ }
+}
+
+/* Clear any mark that says that this insn sets dest. Called from
+ note_stores. */
static void
-reg_set_info (dest, setter, data)
- rtx dest, setter ATTRIBUTE_UNUSED;
- void * data ATTRIBUTE_UNUSED;
+reg_clear_last_set (rtx dest, rtx setter ATTRIBUTE_UNUSED,
+ void *data)
{
+ int *dead_vec = data;
+
if (GET_CODE (dest) == SUBREG)
dest = SUBREG_REG (dest);
- if (GET_CODE (dest) == REG)
- regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
+ if (REG_P (dest) &&
+ dead_vec[REGNO (dest)] == INSN_UID (compute_store_table_current_insn))
+ dead_vec[REGNO (dest)] = 0;
}
/* Return zero if some of the registers in list X are killed
due to set of registers in bitmap REGS_SET. */
-
+
static bool
-store_ops_ok (x, regs_set)
- rtx x;
- int *regs_set;
+store_ops_ok (rtx x, int *regs_set)
{
rtx reg;
{
reg = XEXP (x, 0);
if (regs_set[REGNO(reg)])
- return false;
+ return false;
}
return true;
/* Returns a list of registers mentioned in X. */
static rtx
-extract_mentioned_regs (x)
- rtx x;
+extract_mentioned_regs (rtx x)
{
return extract_mentioned_regs_helper (x, NULL_RTX);
}
/* Helper for extract_mentioned_regs; ACCUM is used to accumulate used
registers. */
static rtx
-extract_mentioned_regs_helper (x, accum)
- rtx x;
- rtx accum;
+extract_mentioned_regs_helper (rtx x, rtx accum)
{
int i;
enum rtx_code code;
case POST_DEC:
case POST_INC:
/* We do not run this function with arguments having side effects. */
- abort ();
+ gcc_unreachable ();
case PC:
case CC0: /*FIXME*/
current insn, REGS_SET_AFTER is bitmap of registers set after (and
including) the insn in this basic block. We must be passing through BB from
head to end, as we are using this fact to speed things up.
-
+
The results are stored this way:
-- the first anticipatable expression is added into ANTIC_STORE_LIST
*/
static void
-find_moveable_store (insn, regs_set_before, regs_set_after)
- rtx insn;
- int *regs_set_before;
- int *regs_set_after;
+find_moveable_store (rtx insn, int *regs_set_before, int *regs_set_after)
{
struct ls_expr * ptr;
rtx dest, set, tmp;
dest = SET_DEST (set);
- if (GET_CODE (dest) != MEM || MEM_VOLATILE_P (dest)
+ if (! MEM_P (dest) || MEM_VOLATILE_P (dest)
|| GET_MODE (dest) == BLKmode)
return;
continue. */
if (flag_non_call_exceptions && may_trap_p (dest))
return;
-
+
+ /* Even if the destination cannot trap, the source may. In this case we'd
+ need to handle updating the REG_EH_REGION note. */
+ if (find_reg_note (insn, REG_EH_REGION, NULL_RTX))
+ return;
+
ptr = ldst_entry (dest);
if (!ptr->pattern_regs)
ptr->pattern_regs = extract_mentioned_regs (dest);
failed last time. */
if (LAST_AVAIL_CHECK_FAILURE (ptr))
{
- for (tmp = bb->end;
+ for (tmp = BB_END (bb);
tmp != insn && tmp != LAST_AVAIL_CHECK_FAILURE (ptr);
tmp = PREV_INSN (tmp))
continue;
if (!check_available)
AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn, AVAIL_STORE_LIST (ptr));
}
-
+
/* Find available and anticipatable stores. */
static int
-compute_store_table ()
+compute_store_table (void)
{
int ret;
basic_block bb;
max_gcse_regno = max_reg_num ();
- reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (last_basic_block,
+ reg_set_in_block = sbitmap_vector_alloc (last_basic_block,
max_gcse_regno);
sbitmap_vector_zero (reg_set_in_block, last_basic_block);
pre_ldst_mems = 0;
- last_set_in = xmalloc (sizeof (int) * max_gcse_regno);
+ last_set_in = xcalloc (max_gcse_regno, sizeof (int));
already_set = xmalloc (sizeof (int) * max_gcse_regno);
/* Find all the stores we care about. */
FOR_EACH_BB (bb)
{
/* First compute the registers set in this block. */
- memset (last_set_in, 0, sizeof (int) * max_gcse_regno);
regvec = last_set_in;
- for (insn = bb->head;
- insn != NEXT_INSN (bb->end);
+ for (insn = BB_HEAD (bb);
+ insn != NEXT_INSN (BB_END (bb));
insn = NEXT_INSN (insn))
{
if (! INSN_P (insn))
continue;
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
- bool clobbers_all = false;
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
-
for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- last_set_in[regno] = INSN_UID (insn);
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
+ {
+ last_set_in[regno] = INSN_UID (insn);
+ SET_BIT (reg_set_in_block[bb->index], regno);
+ }
}
pat = PATTERN (insn);
compute_store_table_current_insn = insn;
- note_stores (pat, reg_set_info, NULL);
+ note_stores (pat, reg_set_info, reg_set_in_block[bb->index]);
}
- /* Record the set registers. */
- for (regno = 0; regno < max_gcse_regno; regno++)
- if (last_set_in[regno])
- SET_BIT (reg_set_in_block[bb->index], regno);
-
/* Now find the stores. */
memset (already_set, 0, sizeof (int) * max_gcse_regno);
regvec = already_set;
- for (insn = bb->head;
- insn != NEXT_INSN (bb->end);
+ for (insn = BB_HEAD (bb);
+ insn != NEXT_INSN (BB_END (bb));
insn = NEXT_INSN (insn))
{
if (! INSN_P (insn))
continue;
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
- bool clobbers_all = false;
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
-
for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
already_set[regno] = 1;
}
find_moveable_store (insn, already_set, last_set_in);
/* Unmark regs that are no longer set. */
- for (regno = 0; regno < max_gcse_regno; regno++)
- if (last_set_in[regno] == INSN_UID (insn))
- last_set_in[regno] = 0;
+ compute_store_table_current_insn = insn;
+ note_stores (pat, reg_clear_last_set, last_set_in);
+ if (CALL_P (insn))
+ {
+ for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno)
+ && last_set_in[regno] == INSN_UID (insn))
+ last_set_in[regno] = 0;
+ }
}
+#ifdef ENABLE_CHECKING
+ /* last_set_in should now be all-zero. */
+ for (regno = 0; regno < max_gcse_regno; regno++)
+ gcc_assert (!last_set_in[regno]);
+#endif
+
/* Clear temporary marks. */
for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
{
return ret;
}
-/* Check to see if the load X is aliased with STORE_PATTERN. */
+/* Check to see if the load X is aliased with STORE_PATTERN.
+ AFTER is true if we are checking the case when STORE_PATTERN occurs
+ after the X. */
static bool
-load_kills_store (x, store_pattern)
- rtx x, store_pattern;
+load_kills_store (rtx x, rtx store_pattern, int after)
{
- if (true_dependence (x, GET_MODE (x), store_pattern, rtx_addr_varies_p))
- return true;
- return false;
+ if (after)
+ return anti_dependence (x, store_pattern);
+ else
+ return true_dependence (store_pattern, GET_MODE (store_pattern), x,
+ rtx_addr_varies_p);
}
/* Go through the entire insn X, looking for any loads which might alias
- STORE_PATTERN. Return true if found. */
+ STORE_PATTERN. Return true if found.
+ AFTER is true if we are checking the case when STORE_PATTERN occurs
+ after the insn X. */
static bool
-find_loads (x, store_pattern)
- rtx x, store_pattern;
+find_loads (rtx x, rtx store_pattern, int after)
{
const char * fmt;
int i, j;
if (GET_CODE (x) == SET)
x = SET_SRC (x);
- if (GET_CODE (x) == MEM)
+ if (MEM_P (x))
{
- if (load_kills_store (x, store_pattern))
+ if (load_kills_store (x, store_pattern, after))
return true;
}
for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0 && !ret; i--)
{
if (fmt[i] == 'e')
- ret |= find_loads (XEXP (x, i), store_pattern);
+ ret |= find_loads (XEXP (x, i), store_pattern, after);
else if (fmt[i] == 'E')
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
- ret |= find_loads (XVECEXP (x, i, j), store_pattern);
+ ret |= find_loads (XVECEXP (x, i, j), store_pattern, after);
}
return ret;
}
/* Check if INSN kills the store pattern X (is aliased with it).
- Return true if it it does. */
+ AFTER is true if we are checking the case when store X occurs
+ after the insn. Return true if it does. */
static bool
-store_killed_in_insn (x, x_regs, insn)
- rtx x, x_regs, insn;
+store_killed_in_insn (rtx x, rtx x_regs, rtx insn, int after)
{
- rtx reg, base;
+ rtx reg, base, note;
if (!INSN_P (insn))
return false;
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
/* A normal or pure call might read from pattern,
but a const call will not. */
if (GET_CODE (PATTERN (insn)) == SET)
{
rtx pat = PATTERN (insn);
+ rtx dest = SET_DEST (pat);
+
+ if (GET_CODE (dest) == ZERO_EXTRACT)
+ dest = XEXP (dest, 0);
+
/* Check for memory stores to aliased objects. */
- if (GET_CODE (SET_DEST (pat)) == MEM && !expr_equiv_p (SET_DEST (pat), x))
- /* pretend its a load and check for aliasing. */
- if (find_loads (SET_DEST (pat), x))
- return true;
- return find_loads (SET_SRC (pat), x);
+ if (MEM_P (dest)
+ && !expr_equiv_p (dest, x))
+ {
+ if (after)
+ {
+ if (output_dependence (dest, x))
+ return true;
+ }
+ else
+ {
+ if (output_dependence (x, dest))
+ return true;
+ }
+ }
+ if (find_loads (SET_SRC (pat), x, after))
+ return true;
}
- else
- return find_loads (PATTERN (insn), x);
+ else if (find_loads (PATTERN (insn), x, after))
+ return true;
+
+ /* If this insn has a REG_EQUAL or REG_EQUIV note referencing a memory
+ location aliased with X, then this insn kills X. */
+ note = find_reg_equal_equiv_note (insn);
+ if (! note)
+ return false;
+ note = XEXP (note, 0);
+
+ /* However, if the note represents a must alias rather than a may
+ alias relationship, then it does not kill X. */
+ if (expr_equiv_p (note, x))
+ return false;
+
+ /* See if there are any aliased loads in the note. */
+ return find_loads (note, x, after);
}
/* Returns true if the expression X is loaded or clobbered on or after INSN
is killed, return the last insn in that it occurs in FAIL_INSN. */
static bool
-store_killed_after (x, x_regs, insn, bb, regs_set_after, fail_insn)
- rtx x, x_regs, insn;
- basic_block bb;
- int *regs_set_after;
- rtx *fail_insn;
+store_killed_after (rtx x, rtx x_regs, rtx insn, basic_block bb,
+ int *regs_set_after, rtx *fail_insn)
{
- rtx last = bb->end, act;
+ rtx last = BB_END (bb), act;
if (!store_ops_ok (x_regs, regs_set_after))
- {
+ {
/* We do not know where it will happen. */
if (fail_insn)
*fail_insn = NULL_RTX;
/* Scan from the end, so that fail_insn is determined correctly. */
for (act = last; act != PREV_INSN (insn); act = PREV_INSN (act))
- if (store_killed_in_insn (x, x_regs, act))
+ if (store_killed_in_insn (x, x_regs, act, false))
{
if (fail_insn)
*fail_insn = act;
return false;
}
-
+
/* Returns true if the expression X is loaded or clobbered on or before INSN
within basic block BB. X_REGS is list of registers mentioned in X.
REGS_SET_BEFORE is bitmap of registers set before or in this insn. */
static bool
-store_killed_before (x, x_regs, insn, bb, regs_set_before)
- rtx x, x_regs, insn;
- basic_block bb;
- int *regs_set_before;
+store_killed_before (rtx x, rtx x_regs, rtx insn, basic_block bb,
+ int *regs_set_before)
{
- rtx first = bb->head;
+ rtx first = BB_HEAD (bb);
if (!store_ops_ok (x_regs, regs_set_before))
return true;
for ( ; insn != PREV_INSN (first); insn = PREV_INSN (insn))
- if (store_killed_in_insn (x, x_regs, insn))
+ if (store_killed_in_insn (x, x_regs, insn, true))
return true;
return false;
}
-
+
/* Fill in available, anticipatable, transparent and kill vectors in
STORE_DATA, based on lists of available and anticipatable stores. */
static void
-build_store_vectors ()
+build_store_vectors (void)
{
basic_block bb;
int *regs_set_in_block;
/* Build the gen_vector. This is any store in the table which is not killed
by aliasing later in its block. */
- ae_gen = (sbitmap *) sbitmap_vector_alloc (last_basic_block, num_stores);
+ ae_gen = sbitmap_vector_alloc (last_basic_block, num_stores);
sbitmap_vector_zero (ae_gen, last_basic_block);
- st_antloc = (sbitmap *) sbitmap_vector_alloc (last_basic_block, num_stores);
+ st_antloc = sbitmap_vector_alloc (last_basic_block, num_stores);
sbitmap_vector_zero (st_antloc, last_basic_block);
for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
rtx r = gen_reg_rtx (GET_MODE (ptr->pattern));
if (gcse_file)
fprintf (gcse_file, "Removing redundant store:\n");
- replace_store_insn (r, XEXP (st, 0), bb);
+ replace_store_insn (r, XEXP (st, 0), bb, ptr);
continue;
}
SET_BIT (ae_gen[bb->index], ptr->index);
}
}
- ae_kill = (sbitmap *) sbitmap_vector_alloc (last_basic_block, num_stores);
+ ae_kill = sbitmap_vector_alloc (last_basic_block, num_stores);
sbitmap_vector_zero (ae_kill, last_basic_block);
- transp = (sbitmap *) sbitmap_vector_alloc (last_basic_block, num_stores);
+ transp = sbitmap_vector_alloc (last_basic_block, num_stores);
sbitmap_vector_zero (transp, last_basic_block);
regs_set_in_block = xmalloc (sizeof (int) * max_gcse_regno);
for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
{
- if (store_killed_after (ptr->pattern, ptr->pattern_regs, bb->head,
+ if (store_killed_after (ptr->pattern, ptr->pattern_regs, BB_HEAD (bb),
bb, regs_set_in_block, NULL))
{
/* It should not be necessary to consider the expression
if (!TEST_BIT (st_antloc[bb->index], ptr->index)
|| !TEST_BIT (ae_gen[bb->index], ptr->index))
SET_BIT (ae_kill[bb->index], ptr->index);
- }
- else
- SET_BIT (transp[bb->index], ptr->index);
- }
+ }
+ else
+ SET_BIT (transp[bb->index], ptr->index);
+ }
}
free (regs_set_in_block);
}
/* Insert an instruction at the beginning of a basic block, and update
- the BLOCK_HEAD if needed. */
+ the BB_HEAD if needed. */
static void
-insert_insn_start_bb (insn, bb)
- rtx insn;
- basic_block bb;
+insert_insn_start_bb (rtx insn, basic_block bb)
{
/* Insert at start of successor block. */
- rtx prev = PREV_INSN (bb->head);
- rtx before = bb->head;
+ rtx prev = PREV_INSN (BB_HEAD (bb));
+ rtx before = BB_HEAD (bb);
while (before != 0)
{
- if (GET_CODE (before) != CODE_LABEL
- && (GET_CODE (before) != NOTE
+ if (! LABEL_P (before)
+ && (! NOTE_P (before)
|| NOTE_LINE_NUMBER (before) != NOTE_INSN_BASIC_BLOCK))
break;
prev = before;
- if (prev == bb->end)
+ if (prev == BB_END (bb))
break;
before = NEXT_INSN (before);
}
- insn = emit_insn_after (insn, prev);
+ insn = emit_insn_after_noloc (insn, prev);
if (gcse_file)
{
if an edge insertion was performed. */
static int
-insert_store (expr, e)
- struct ls_expr * expr;
- edge e;
+insert_store (struct ls_expr * expr, edge e)
{
rtx reg, insn;
basic_block bb;
edge tmp;
+ edge_iterator ei;
/* We did all the deleted before this insert, so if we didn't delete a
store, then we haven't set the reaching reg yet either. */
if (expr->reaching_reg == NULL_RTX)
return 0;
+ if (e->flags & EDGE_FAKE)
+ return 0;
+
reg = expr->reaching_reg;
insn = gen_move_insn (copy_rtx (expr->pattern), reg);
insert it at the start of the BB, and reset the insert bits on the other
edges so we don't try to insert it on the other edges. */
bb = e->dest;
- for (tmp = e->dest->pred; tmp ; tmp = tmp->pred_next)
- {
- int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
- if (index == EDGE_INDEX_NO_EDGE)
- abort ();
- if (! TEST_BIT (pre_insert_map[index], expr->index))
- break;
- }
+ FOR_EACH_EDGE (tmp, ei, e->dest->preds)
+ if (!(tmp->flags & EDGE_FAKE))
+ {
+ int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
+
+ gcc_assert (index != EDGE_INDEX_NO_EDGE);
+ if (! TEST_BIT (pre_insert_map[index], expr->index))
+ break;
+ }
/* If tmp is NULL, we found an insertion on every edge, blank the
insertion vector for these edges, and insert at the start of the BB. */
if (!tmp && bb != EXIT_BLOCK_PTR)
{
- for (tmp = e->dest->pred; tmp ; tmp = tmp->pred_next)
+ FOR_EACH_EDGE (tmp, ei, e->dest->preds)
{
int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
RESET_BIT (pre_insert_map[index], expr->index);
return 0;
}
- /* We can't insert on this edge, so we'll insert at the head of the
- successors block. See Morgan, sec 10.5. */
- if ((e->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
- {
- insert_insn_start_bb (insn, bb);
- return 0;
- }
+ /* We can't put stores in the front of blocks pointed to by abnormal
+ edges since that may put a store where one didn't used to be. */
+ gcc_assert (!(e->flags & EDGE_ABNORMAL));
insert_insn_on_edge (insn, e);
return 1;
}
+/* Remove any REG_EQUAL or REG_EQUIV notes containing a reference to the
+ memory location in SMEXPR set in basic block BB.
+
+ This could be rather expensive. */
+
+static void
+remove_reachable_equiv_notes (basic_block bb, struct ls_expr *smexpr)
+{
+ edge_iterator *stack, ei;
+ int sp;
+ edge act;
+ sbitmap visited = sbitmap_alloc (last_basic_block);
+ rtx last, insn, note;
+ rtx mem = smexpr->pattern;
+
+ stack = xmalloc (sizeof (edge_iterator) * n_basic_blocks);
+ sp = 0;
+ ei = ei_start (bb->succs);
+
+ sbitmap_zero (visited);
+
+ act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
+ while (1)
+ {
+ if (!act)
+ {
+ if (!sp)
+ {
+ free (stack);
+ sbitmap_free (visited);
+ return;
+ }
+ act = ei_edge (stack[--sp]);
+ }
+ bb = act->dest;
+
+ if (bb == EXIT_BLOCK_PTR
+ || TEST_BIT (visited, bb->index))
+ {
+ if (!ei_end_p (ei))
+ ei_next (&ei);
+ act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
+ continue;
+ }
+ SET_BIT (visited, bb->index);
+
+ if (TEST_BIT (st_antloc[bb->index], smexpr->index))
+ {
+ for (last = ANTIC_STORE_LIST (smexpr);
+ BLOCK_FOR_INSN (XEXP (last, 0)) != bb;
+ last = XEXP (last, 1))
+ continue;
+ last = XEXP (last, 0);
+ }
+ else
+ last = NEXT_INSN (BB_END (bb));
+
+ for (insn = BB_HEAD (bb); insn != last; insn = NEXT_INSN (insn))
+ if (INSN_P (insn))
+ {
+ note = find_reg_equal_equiv_note (insn);
+ if (!note || !expr_equiv_p (XEXP (note, 0), mem))
+ continue;
+
+ if (gcse_file)
+ fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
+ INSN_UID (insn));
+ remove_note (insn, note);
+ }
+
+ if (!ei_end_p (ei))
+ ei_next (&ei);
+ act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
+
+ if (EDGE_COUNT (bb->succs) > 0)
+ {
+ if (act)
+ stack[sp++] = ei;
+ ei = ei_start (bb->succs);
+ act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
+ }
+ }
+}
+
/* This routine will replace a store with a SET to a specified register. */
static void
-replace_store_insn (reg, del, bb)
- rtx reg, del;
- basic_block bb;
+replace_store_insn (rtx reg, rtx del, basic_block bb, struct ls_expr *smexpr)
{
- rtx insn;
+ rtx insn, mem, note, set, ptr, pair;
+ mem = smexpr->pattern;
insn = gen_move_insn (reg, SET_SRC (single_set (del)));
insn = emit_insn_after (insn, del);
fprintf (gcse_file, "\n");
}
+ for (ptr = ANTIC_STORE_LIST (smexpr); ptr; ptr = XEXP (ptr, 1))
+ if (XEXP (ptr, 0) == del)
+ {
+ XEXP (ptr, 0) = insn;
+ break;
+ }
+
+ /* Move the notes from the deleted insn to its replacement, and patch
+ up the LIBCALL notes. */
+ REG_NOTES (insn) = REG_NOTES (del);
+
+ note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
+ if (note)
+ {
+ pair = XEXP (note, 0);
+ note = find_reg_note (pair, REG_LIBCALL, NULL_RTX);
+ XEXP (note, 0) = insn;
+ }
+ note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
+ if (note)
+ {
+ pair = XEXP (note, 0);
+ note = find_reg_note (pair, REG_RETVAL, NULL_RTX);
+ XEXP (note, 0) = insn;
+ }
+
delete_insn (del);
+
+ /* Now we must handle REG_EQUAL notes whose contents is equal to the mem;
+ they are no longer accurate provided that they are reached by this
+ definition, so drop them. */
+ for (; insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn))
+ if (INSN_P (insn))
+ {
+ set = single_set (insn);
+ if (!set)
+ continue;
+ if (expr_equiv_p (SET_DEST (set), mem))
+ return;
+ note = find_reg_equal_equiv_note (insn);
+ if (!note || !expr_equiv_p (XEXP (note, 0), mem))
+ continue;
+
+ if (gcse_file)
+ fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
+ INSN_UID (insn));
+ remove_note (insn, note);
+ }
+ remove_reachable_equiv_notes (bb, smexpr);
}
the reaching_reg for later storing. */
static void
-delete_store (expr, bb)
- struct ls_expr * expr;
- basic_block bb;
+delete_store (struct ls_expr * expr, basic_block bb)
{
rtx reg, i, del;
{
/* We know there is only one since we deleted redundant
ones during the available computation. */
- replace_store_insn (reg, del, bb);
+ replace_store_insn (reg, del, bb, expr);
break;
}
}
/* Free memory used by store motion. */
static void
-free_store_memory ()
+free_store_memory (void)
{
free_ldst_mems ();
other way by looking at the flowgraph in reverse. */
static void
-store_motion ()
+store_motion (void)
{
basic_block bb;
int x;
/* Now compute kill & transp vectors. */
build_store_vectors ();
add_noreturn_fake_exit_edges ();
+ connect_infinite_loops_to_exit ();
edge_list = pre_edge_rev_lcm (gcse_file, num_stores, transp, ae_gen,
st_antloc, ae_kill, &pre_insert_map,
/* Now we want to insert the new stores which are going to be needed. */
for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
{
+ /* If any of the edges we have above are abnormal, we can't move this
+ store. */
+ for (x = NUM_EDGES (edge_list) - 1; x >= 0; x--)
+ if (TEST_BIT (pre_insert_map[x], ptr->index)
+ && (INDEX_EDGE (edge_list, x)->flags & EDGE_ABNORMAL))
+ break;
+
+ if (x >= 0)
+ {
+ if (gcse_file != NULL)
+ fprintf (gcse_file,
+ "Can't replace store %d: abnormal edge from %d to %d\n",
+ ptr->index, INDEX_EDGE (edge_list, x)->src->index,
+ INDEX_EDGE (edge_list, x)->dest->index);
+ continue;
+ }
+
+ /* Now we want to insert the new stores which are going to be needed. */
+
FOR_EACH_BB (bb)
if (TEST_BIT (pre_delete_map[bb->index], ptr->index))
delete_store (ptr, bb);
free_store_memory ();
free_edge_list (edge_list);
- remove_fake_edges ();
+ remove_fake_exit_edges ();
end_alias_analysis ();
}
/* Entry point for jump bypassing optimization pass. */
int
-bypass_jumps (file)
- FILE *file;
+bypass_jumps (FILE *file)
{
int changed;
if (file)
dump_flow_info (file);
- /* Return if there's nothing to do. */
- if (n_basic_blocks <= 1)
+ /* Return if there's nothing to do, or it is too expensive. */
+ if (n_basic_blocks <= 1 || is_too_expensive (_ ("jump bypassing disabled")))
return 0;
- /* Trying to perform global optimizations on flow graphs which have
- a high connectivity will take a long time and is unlikely to be
- particularly useful.
-
- In normal circumstances a cfg should have about twice as many edges
- as blocks. But we do not want to punish small functions which have
- a couple switch statements. So we require a relatively large number
- of basic blocks and the ratio of edges to blocks to be high. */
- if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
- {
- if (warn_disabled_optimization)
- warning ("BYPASS disabled: %d > 1000 basic blocks and %d >= 20 edges/basic block",
- n_basic_blocks, n_edges / n_basic_blocks);
- return 0;
- }
-
- /* If allocating memory for the cprop bitmap would take up too much
- storage it's better just to disable the optimization. */
- if ((n_basic_blocks
- * SBITMAP_SET_SIZE (max_gcse_regno)
- * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
- {
- if (warn_disabled_optimization)
- warning ("GCSE disabled: %d basic blocks and %d registers",
- n_basic_blocks, max_gcse_regno);
-
- return 0;
- }
-
gcc_obstack_init (&gcse_obstack);
bytes_used = 0;
max_gcse_regno = max_reg_num ();
alloc_gcse_mem (get_insns ());
- changed = one_cprop_pass (1, 1, 1);
+ changed = one_cprop_pass (MAX_GCSE_PASSES + 2, 1, 1);
free_gcse_mem ();
if (file)
{
fprintf (file, "BYPASS of %s: %d basic blocks, ",
- current_function_name, n_basic_blocks);
+ current_function_name (), n_basic_blocks);
fprintf (file, "%d bytes\n\n", bytes_used);
}
return changed;
}
+/* Return true if the graph is too expensive to optimize. PASS is the
+ optimization about to be performed. */
+
+static bool
+is_too_expensive (const char *pass)
+{
+ /* Trying to perform global optimizations on flow graphs which have
+ a high connectivity will take a long time and is unlikely to be
+ particularly useful.
+
+ In normal circumstances a cfg should have about twice as many
+ edges as blocks. But we do not want to punish small functions
+ which have a couple switch statements. Rather than simply
+ threshold the number of blocks, uses something with a more
+ graceful degradation. */
+ if (n_edges > 20000 + n_basic_blocks * 4)
+ {
+ if (warn_disabled_optimization)
+ warning ("%s: %d basic blocks and %d edges/basic block",
+ pass, n_basic_blocks, n_edges / n_basic_blocks);
+
+ return true;
+ }
+
+ /* If allocating memory for the cprop bitmap would take up too much
+ storage it's better just to disable the optimization. */
+ if ((n_basic_blocks
+ * SBITMAP_SET_SIZE (max_reg_num ())
+ * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
+ {
+ if (warn_disabled_optimization)
+ warning ("%s: %d basic blocks and %d registers",
+ pass, n_basic_blocks, max_reg_num ());
+
+ return true;
+ }
+
+ return false;
+}
+
#include "gt-gcse.h"
OSDN Git Service
