/* Global common subexpression elimination/Partial redundancy elimination
and global constant/copy propagation for GNU compiler.
- Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002
- Free Software Foundation, Inc.
+ Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005,
+ 2006, 2007, 2008, 2009 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
+Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
/* TODO
- reordering of memory allocation and freeing to be more space efficient
#include "config.h"
#include "system.h"
+#include "coretypes.h"
+#include "tm.h"
#include "toplev.h"
#include "rtl.h"
+#include "tree.h"
#include "tm_p.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "basic-block.h"
#include "output.h"
#include "function.h"
-#include "expr.h"
+#include "expr.h"
#include "except.h"
#include "ggc.h"
#include "params.h"
-
+#include "cselib.h"
+#include "intl.h"
#include "obstack.h"
-#define obstack_chunk_alloc gmalloc
-#define obstack_chunk_free free
+#include "timevar.h"
+#include "tree-pass.h"
+#include "hashtab.h"
+#include "df.h"
+#include "dbgcnt.h"
/* Propagate flow information through back edges and thus enable PRE's
moving loop invariant calculations out of loops.
be done by loop.c, which has more heuristics for when to move invariants
out of loops. At some point we might need to move some of those
heuristics into gcse.c. */
-#define FOLLOW_BACK_EDGES 1
/* We support GCSE via Partial Redundancy Elimination. PRE optimizations
are a superset of those done by GCSE.
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. */
-/* -dG dump file. */
-static FILE *gcse_file;
-
/* Note whether or not we should run jump optimization after gcse. We
want to do this for two cases.
* 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.
- However it's useful to be able to print them from GDB.
- We could create special functions for this, but it's simpler to
- just allow passing stderr to the dump_foo fns. Since stderr can
- be a macro, we store a copy here. */
-static FILE *debug_stderr;
-
/* An obstack for our working variables. */
static struct obstack gcse_obstack;
-/* Non-zero for each mode that supports (set (reg) (reg)).
- This is trivially true for integer and floating point values.
- It may or may not be true for condition codes. */
-static char can_copy_p[(int) NUM_MACHINE_MODES];
-
-/* Non-zero if can_copy_p has been initialized. */
-static int can_copy_init_p;
-
struct reg_use {rtx reg_rtx; };
/* Hash table of expressions. */
struct occr *next;
/* The insn that computes the expression. */
rtx insn;
- /* Non-zero if this [anticipatable] occurrence has been deleted. */
+ /* Nonzero if this [anticipatable] occurrence has been deleted. */
char deleted_p;
- /* Non-zero if this [available] occurrence has been copied to
+ /* Nonzero if this [available] occurrence has been copied to
reaching_reg. */
/* ??? This is mutually exclusive with deleted_p, so they could share
the same byte. */
[one could build a mapping table without holes afterwards though].
Someday I'll perform the computation and figure it out. */
-/* Total size of the expression hash table, in elements. */
-static unsigned int expr_hash_table_size;
+struct hash_table
+{
+ /* The table itself.
+ This is an array of `expr_hash_table_size' elements. */
+ struct expr **table;
+
+ /* Size of the hash table, in elements. */
+ unsigned int size;
-/* The table itself.
- This is an array of `expr_hash_table_size' elements. */
-static struct expr **expr_hash_table;
+ /* Number of hash table elements. */
+ unsigned int n_elems;
+
+ /* Whether the table is expression of copy propagation one. */
+ int set_p;
+};
-/* Total size of the copy propagation hash table, in elements. */
-static unsigned int set_hash_table_size;
+/* Expression hash table. */
+static struct hash_table expr_hash_table;
-/* The table itself.
- This is an array of `set_hash_table_size' elements. */
-static struct expr **set_hash_table;
+/* Copy propagation hash table. */
+static struct hash_table set_hash_table;
/* Mapping of uids to cuids.
Only real insns get cuids. */
/* 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
/* Number of cuids. */
static int max_cuid;
-/* Mapping of cuids to insns. */
-static rtx *cuid_insn;
-
-/* Get insn from cuid. */
-#define CUID_INSN(CUID) (cuid_insn[CUID])
-
/* Maximum register number in function prior to doing gcse + 1.
Registers created during this pass have regno >= max_gcse_regno.
This is named with "gcse" to not collide with global of same name. */
static unsigned int max_gcse_regno;
-/* Maximum number of cse-able expressions found. */
-static int n_exprs;
-
-/* Maximum number of assignments for copy propagation found. */
-static int n_sets;
-
/* Table of registers that are modified.
For each register, each element is a list of places where the pseudo-reg
{
/* 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;
#define REG_SET_TABLE_SLOP 100
/* This is a list of expressions which are MEMs and will be used by load
- or store motion.
+ or store motion.
Load motion tracks MEMs which aren't killed by
- anything except itself. (ie, loads and stores to a single location).
- We can then allow movement of these MEM refs with a little special
+ 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
- no side effects so we can re-issue the setter value.
+ no side effects so we can re-issue the setter value.
Store Motion uses this structure as an expression table to track stores
which look interesting, and might be moveable towards the exit block. */
{
struct expr * expr; /* Gcse expression reference for LM. */
rtx pattern; /* Pattern of this mem. */
+ rtx pattern_regs; /* List of registers mentioned by the mem. */
rtx loads; /* INSN list of loads seen. */
rtx stores; /* INSN list of stores seen. */
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. */
};
+/* Array of implicit set patterns indexed by basic block index. */
+static rtx *implicit_sets;
+
/* Head of the list of load/store memory refs. */
static struct ls_expr * pre_ldst_mems = NULL;
+/* Hashtable for the load/store memory refs. */
+static htab_t pre_ldst_table = NULL;
+
/* Bitmap containing one bit for each register in the program.
Used when performing GCSE to track which registers have been set since
the start of the basic block. */
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 copies propagated. */
+static int local_copy_prop_count;
+/* Number of global constants propagated. */
+static int global_const_prop_count;
+/* Number of global copies 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. */
- int 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, int, int));
-static void hash_scan_set PARAMS ((rtx, rtx, int));
-static void hash_scan_clobber PARAMS ((rtx, rtx));
-static void hash_scan_call PARAMS ((rtx, rtx));
-static int want_to_gcse_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));
-static void insert_set_in_table PARAMS ((rtx, rtx));
-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 ((int));
-static void alloc_set_hash_table PARAMS ((int));
-static void free_set_hash_table PARAMS ((void));
-static void compute_set_hash_table PARAMS ((void));
-static void alloc_expr_hash_table PARAMS ((unsigned int));
-static void free_expr_hash_table PARAMS ((void));
-static void compute_expr_hash_table PARAMS ((void));
-static void dump_hash_table PARAMS ((FILE *, const char *, struct expr **,
- int, int));
-static struct expr *lookup_expr PARAMS ((rtx));
-static struct expr *lookup_set PARAMS ((unsigned int, rtx));
-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 *,
- int));
-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));
-#ifdef HAVE_cc0
-static int cprop_cc0_jump PARAMS ((basic_block, rtx, struct reg_use *, rtx));
-#endif
-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 ((basic_block, rtx, int));
-static int cprop PARAMS ((int));
-static int one_cprop_pass PARAMS ((int, int));
-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 ((void));
-static int expr_killed_p PARAMS ((rtx, basic_block));
-static void compute_ae_kill PARAMS ((sbitmap *, sbitmap *));
-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 void 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 int store_ops_ok PARAMS ((rtx, basic_block));
-static void find_moveable_store PARAMS ((rtx));
-static int compute_store_table PARAMS ((void));
-static int load_kills_store PARAMS ((rtx, rtx));
-static int find_loads PARAMS ((rtx, rtx));
-static int store_killed_in_insn PARAMS ((rtx, rtx));
-static int store_killed_after PARAMS ((rtx, rtx, basic_block));
-static int store_killed_before PARAMS ((rtx, rtx, basic_block));
-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 void local_cprop_find_used_regs PARAMS ((rtx *, void *));
+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 (void);
+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, const_rtx, void *);
+static void compute_sets (void);
+static void hash_scan_insn (rtx, struct hash_table *);
+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 (const_rtx);
+static int oprs_unchanged_p (const_rtx, const_rtx, int);
+static int oprs_anticipatable_p (const_rtx, const_rtx);
+static int oprs_available_p (const_rtx, const_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 (const_rtx, enum machine_mode, int *, int);
+static unsigned int hash_set (int, int);
+static int expr_equiv_p (const_rtx, const_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, const_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 (const_rtx, const_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 (const_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, const_rtx, void *);
+static int load_killed_in_block_p (const_basic_block, int, const_rtx, int);
+static void canon_list_insert (rtx, const_rtx, void *);
+static int cprop_insn (rtx, int);
+static int cprop (int);
+static void find_implicit_sets (void);
+static int one_cprop_pass (int, bool, bool);
+static bool constprop_register (rtx, rtx, rtx, bool);
+static struct expr *find_bypass_set (int, int);
+static bool reg_killed_on_edge (const_rtx, const_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_basic_block (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 (const_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, const_rtx, void *);
+static void reg_clear_last_set (rtx, const_rtx, void *);
+static bool store_ops_ok (const_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 (const_rtx, const_rtx, int);
+static bool find_loads (const_rtx, const_rtx, int);
+static bool store_killed_in_insn (const_rtx, const_rtx, const_rtx, int);
+static bool store_killed_after (const_rtx, const_rtx, const_rtx, const_basic_block, int *, rtx *);
+static bool store_killed_before (const_rtx, const_rtx, const_rtx, const_basic_block, int *);
+static void build_store_vectors (void);
+static void insert_insn_start_basic_block (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, bool);
+static void local_cprop_pass (bool);
+static bool is_too_expensive (const char *);
+
+#define GNEW(T) ((T *) gmalloc (sizeof (T)))
+#define GCNEW(T) ((T *) gcalloc (1, sizeof (T)))
+
+#define GNEWVEC(T, N) ((T *) gmalloc (sizeof (T) * (N)))
+#define GCNEWVEC(T, N) ((T *) gcalloc ((N), sizeof (T)))
+#define GRESIZEVEC(T, P, N) ((T *) grealloc ((void *) (P), sizeof (T) * (N)))
+
+#define GNEWVAR(T, S) ((T *) gmalloc ((S)))
+#define GCNEWVAR(T, S) ((T *) gcalloc (1, (S)))
+#define GRESIZEVAR(T, P, S) ((T *) grealloc ((P), (S)))
+
+#define GOBNEW(T) ((T *) gcse_alloc (sizeof (T)))
+#define GOBNEWVAR(T, S) ((T *) gcse_alloc ((S)))
\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;
+static int
+gcse_main (rtx f ATTRIBUTE_UNUSED)
{
int changed, pass;
/* Bytes used at start of pass. */
/* Point to release obstack data from for each pass. */
char *gcse_obstack_bottom;
- /* Insertion of instructions on edges can create new basic blocks; we
- need the original basic block count so that we can properly deallocate
- arrays sized on the number of basic blocks originally in the cfg. */
- int orig_bb_count;
/* We do not construct an accurate cfg in functions which call
setjmp, so just punt to be safe. */
- if (current_function_calls_setjmp)
+ if (cfun->calls_setjmp)
return 0;
-
+
/* Assume that we do not need to run jump optimizations after gcse. */
run_jump_opt_after_gcse = 0;
- /* For calling dump_foo fns from gdb. */
- debug_stderr = stderr;
- gcse_file = file;
-
/* Identify the basic block information for this function, including
successors and predecessors. */
max_gcse_regno = max_reg_num ();
- if (file)
- dump_flow_info (file);
+ df_note_add_problem ();
+ df_analyze ();
- orig_bb_count = n_basic_blocks;
- /* Return if 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.
-
- 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;
- }
+ if (dump_file)
+ dump_flow_info (dump_file, dump_flags);
- /* See what modes support reg/reg copy operations. */
- if (! can_copy_init_p)
- {
- compute_can_copy ();
- can_copy_init_p = 1;
- }
+ /* Return if there's nothing to do, or it is too expensive. */
+ if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
+ || is_too_expensive (_("GCSE disabled")))
+ return 0;
gcc_obstack_init (&gcse_obstack);
bytes_used = 0;
information about memory sets when we build the hash tables. */
alloc_reg_set_mem (max_gcse_regno);
- compute_sets (f);
+ compute_sets ();
pass = 0;
initial_bytes_used = bytes_used;
max_pass_bytes = 0;
- gcse_obstack_bottom = gcse_alloc (1);
+ gcse_obstack_bottom = GOBNEWVAR (char, 1);
changed = 1;
while (changed && pass < MAX_GCSE_PASSES)
{
changed = 0;
- if (file)
- fprintf (file, "GCSE pass %d\n\n", pass + 1);
+ if (dump_file)
+ fprintf (dump_file, "GCSE pass %d\n\n", pass + 1);
/* Initialize bytes_used to the space for the pred/succ lists,
and the reg_set_table data. */
/* Each pass may create new registers, so recalculate each time. */
max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (f);
+ alloc_gcse_mem ();
/* Don't allow constant propagation to modify jumps
during this pass. */
- changed = one_cprop_pass (pass + 1, 0);
+ if (dbg_cnt (cprop1))
+ {
+ timevar_push (TV_CPROP1);
+ changed = one_cprop_pass (pass + 1, false, false);
+ timevar_pop (TV_CPROP1);
+ }
- if (optimize_size)
- changed |= one_classic_gcse_pass (pass + 1);
- else
- {
+ if (optimize_function_for_speed_p (cfun))
+ {
+ 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 (n_basic_blocks * sizeof (rtx));
- canon_modify_mem_list
- = (rtx *) gmalloc (n_basic_blocks * sizeof (rtx));
- memset ((char *) modify_mem_list, 0, n_basic_blocks * sizeof (rtx));
- memset ((char *) canon_modify_mem_list, 0, n_basic_blocks * sizeof (rtx));
- orig_bb_count = n_basic_blocks;
+ modify_mem_list = GCNEWVEC (rtx, last_basic_block);
+ canon_modify_mem_list = GCNEWVEC (rtx, last_basic_block);
}
free_reg_set_mem ();
alloc_reg_set_mem (max_reg_num ());
- compute_sets (f);
+ compute_sets ();
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). */
- if (optimize_size)
- {
+ for space, we don't run the partial redundancy algorithms). */
+ if (optimize_function_for_size_p (cfun))
+ {
+ timevar_push (TV_HOIST);
max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (f);
+ alloc_gcse_mem ();
changed |= one_code_hoisting_pass ();
free_gcse_mem ();
if (max_pass_bytes < bytes_used)
max_pass_bytes = bytes_used;
- }
+ timevar_pop (TV_HOIST);
+ }
- if (file)
+ if (dump_file)
{
- fprintf (file, "\n");
- fflush (file);
+ fprintf (dump_file, "\n");
+ fflush (dump_file);
}
obstack_free (&gcse_obstack, gcse_obstack_bottom);
/* Do one last pass of copy propagation, including cprop into
conditional jumps. */
- max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (f);
- /* This time, go ahead and allow cprop to alter jumps. */
- one_cprop_pass (pass + 1, 1);
- free_gcse_mem ();
+ if (dbg_cnt (cprop2))
+ {
+ max_gcse_regno = max_reg_num ();
+ alloc_gcse_mem ();
+
+ /* This time, go ahead and allow cprop to alter jumps. */
+ timevar_push (TV_CPROP2);
+ one_cprop_pass (pass + 1, true, true);
+ timevar_pop (TV_CPROP2);
+ free_gcse_mem ();
+ }
- if (file)
+ if (dump_file)
{
- fprintf (file, "GCSE of %s: %d basic blocks, ",
- current_function_name, n_basic_blocks);
- fprintf (file, "%d pass%s, %d bytes\n\n",
+ fprintf (dump_file, "GCSE of %s: %d basic blocks, ",
+ current_function_name (), n_basic_blocks);
+ fprintf (dump_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);
- /* Store motion disabled until it is fixed. */
- if (0 && !optimize_size && flag_gcse_sm)
- store_motion ();
+ if (optimize_function_for_speed_p (cfun) && flag_gcse_sm)
+ {
+ timevar_push (TV_LSM);
+ store_motion ();
+ timevar_pop (TV_LSM);
+ }
+
/* Record where pseudo-registers are set. */
return run_jump_opt_after_gcse;
}
\f
/* Misc. utilities. */
+/* Nonzero for each mode that supports (set (reg) (reg)).
+ This is trivially true for integer and floating point values.
+ It may or may not be true for condition codes. */
+static char can_copy[(int) NUM_MACHINE_MODES];
+
/* Compute which modes support reg/reg copy operations. */
static void
-compute_can_copy ()
+compute_can_copy (void)
{
int i;
#ifndef AVOID_CCMODE_COPIES
rtx reg, insn;
#endif
- memset (can_copy_p, 0, NUM_MACHINE_MODES);
+ memset (can_copy, 0, NUM_MACHINE_MODES);
start_sequence ();
for (i = 0; i < NUM_MACHINE_MODES; i++)
if (GET_MODE_CLASS (i) == MODE_CC)
{
#ifdef AVOID_CCMODE_COPIES
- can_copy_p[i] = 0;
+ can_copy[i] = 0;
#else
reg = gen_rtx_REG ((enum machine_mode) i, LAST_VIRTUAL_REGISTER + 1);
insn = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
if (recog (PATTERN (insn), insn, NULL) >= 0)
- can_copy_p[i] = 1;
+ can_copy[i] = 1;
#endif
}
else
- can_copy_p[i] = 1;
+ can_copy[i] = 1;
end_sequence ();
}
+
+/* Returns whether the mode supports reg/reg copy operations. */
+
+bool
+can_copy_p (enum machine_mode mode)
+{
+ static bool can_copy_init_p = false;
+
+ if (! can_copy_init_p)
+ {
+ compute_can_copy ();
+ can_copy_init_p = true;
+ }
+
+ return can_copy[mode] != 0;
+}
\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.
- We don't need to record the bytes allocated here since
- obstack_chunk_alloc is set to gmalloc. */
+/* Cover function to obstack_alloc. */
-static char *
-gcse_alloc (size)
- unsigned long size;
+static void *
+gcse_alloc (unsigned long size)
{
- return (char *) obstack_alloc (&gcse_obstack, size);
+ bytes_used += 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 (void)
{
- int i, n;
+ int i;
+ basic_block bb;
rtx insn;
/* Find the largest UID and create a mapping from UIDs to CUIDs.
CUIDs are like UIDs except they increase monotonically, have no gaps,
- and only apply to real insns. */
+ and only apply to real insns.
+ (Actually, there are gaps, for insn that are not inside a basic block.
+ but we should never see those anyway, so this is OK.) */
max_uid = get_max_uid ();
- n = (max_uid + 1) * sizeof (int);
- uid_cuid = (int *) gmalloc (n);
- memset ((char *) uid_cuid, 0, n);
- for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
- {
- if (INSN_P (insn))
- uid_cuid[INSN_UID (insn)] = i++;
- else
- uid_cuid[INSN_UID (insn)] = i;
- }
-
- /* Create a table mapping cuids to insns. */
+ uid_cuid = GCNEWVEC (int, max_uid + 1);
+ i = 0;
+ FOR_EACH_BB (bb)
+ FOR_BB_INSNS (bb, insn)
+ {
+ if (INSN_P (insn))
+ uid_cuid[INSN_UID (insn)] = i++;
+ else
+ uid_cuid[INSN_UID (insn)] = i;
+ }
max_cuid = i;
- n = (max_cuid + 1) * sizeof (rtx);
- cuid_insn = (rtx *) gmalloc (n);
- memset ((char *) cuid_insn, 0, n);
- 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 (n_basic_blocks,
- 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 (n_basic_blocks * sizeof (rtx));
- canon_modify_mem_list = (rtx *) gmalloc (n_basic_blocks * sizeof (rtx));
- memset ((char *) modify_mem_list, 0, n_basic_blocks * sizeof (rtx));
- memset ((char *) canon_modify_mem_list, 0, n_basic_blocks * sizeof (rtx));
- modify_mem_list_set = BITMAP_XMALLOC ();
- canon_modify_mem_list_set = BITMAP_XMALLOC ();
+ modify_mem_list = GCNEWVEC (rtx, last_basic_block);
+ canon_modify_mem_list = GCNEWVEC (rtx, last_basic_block);
+ 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.
properties. If NULL, then it is not necessary to compute or record that
particular property.
- SETP controls which hash table to look at. If zero, this routine looks at
- the expr hash table; if nonzero this routine looks at the set hash table.
- Additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
+ TABLE controls which hash table to look at. If it is set hash table,
+ additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
ABSALTERED. */
-
+
static void
-compute_local_properties (transp, comp, antloc, setp)
- sbitmap *transp;
- sbitmap *comp;
- sbitmap *antloc;
- int setp;
-{
- unsigned int i, hash_table_size;
- struct expr **hash_table;
-
+compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc,
+ struct hash_table *table)
+{
+ unsigned int i;
+
/* Initialize any bitmaps that were passed in. */
if (transp)
{
- if (setp)
- sbitmap_vector_zero (transp, n_basic_blocks);
+ if (table->set_p)
+ sbitmap_vector_zero (transp, last_basic_block);
else
- sbitmap_vector_ones (transp, n_basic_blocks);
+ sbitmap_vector_ones (transp, last_basic_block);
}
if (comp)
- sbitmap_vector_zero (comp, n_basic_blocks);
+ sbitmap_vector_zero (comp, last_basic_block);
if (antloc)
- sbitmap_vector_zero (antloc, n_basic_blocks);
+ sbitmap_vector_zero (antloc, last_basic_block);
- /* We use the same code for cprop, pre and hoisting. For cprop
- we care about the set hash table, for pre and hoisting we
- care about the expr hash table. */
- hash_table_size = setp ? set_hash_table_size : expr_hash_table_size;
- hash_table = setp ? set_hash_table : expr_hash_table;
-
- for (i = 0; i < hash_table_size; i++)
+ for (i = 0; i < table->size; i++)
{
struct expr *expr;
- for (expr = hash_table[i]; expr != NULL; expr = expr->next_same_hash)
+ for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
{
int indx = expr->bitmap_index;
struct occr *occr;
We start by assuming all are transparent [none are killed], and
then reset the bits for those that are. */
if (transp)
- compute_transp (expr->expr, indx, transp, setp);
+ compute_transp (expr->expr, indx, transp, table->set_p);
/* The occurrences recorded in antic_occr are exactly those that
- we want to set to non-zero in ANTLOC. */
+ we want to set to nonzero in ANTLOC. */
if (antloc)
for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
{
}
/* The occurrences recorded in avail_occr are exactly those that
- we want to set to non-zero in COMP. */
+ we want to set to nonzero in COMP. */
if (comp)
for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
{
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 = GCNEWVEC (struct reg_set *, reg_set_table_size);
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 = GRESIZEVEC (struct reg_set *, reg_set_table, new_size);
+ 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 = XOBNEW (®_set_obstack, 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, const_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);
}
/* Scan the function and record each set of each pseudo-register.
This is called once, at the start of the gcse pass. See the comments for
- `reg_set_table' for further documenation. */
+ `reg_set_table' for further documentation. */
static void
-compute_sets (f)
- rtx f;
+compute_sets (void)
{
+ basic_block bb;
rtx insn;
- for (insn = f; insn != 0; insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- note_stores (PATTERN (insn), record_set_info, insn);
+ FOR_EACH_BB (bb)
+ FOR_BB_INSNS (bb, insn)
+ if (INSN_P (insn))
+ note_stores (PATTERN (insn), record_set_info, insn);
}
\f
/* Hash table support. */
-/* For each register, the cuid of the first/last insn in the block
- that set it, or -1 if not set. */
-#define NEVER_SET -1
-
struct reg_avail_info
{
- int last_bb;
+ basic_block last_bb;
int first_set;
int last_set;
};
static struct reg_avail_info *reg_avail_info;
-static int current_bb;
+static basic_block current_bb;
/* See whether X, the source of a set, is something we want to consider for
GCSE. */
static int
-want_to_gcse_p (x)
- rtx x;
-{
- static rtx test_insn = 0;
- int num_clobbers = 0;
- int icode;
+want_to_gcse_p (rtx x)
+{
+#ifdef STACK_REGS
+ /* On register stack architectures, don't GCSE constants from the
+ constant pool, as the benefits are often swamped by the overhead
+ of shuffling the register stack between basic blocks. */
+ if (IS_STACK_MODE (GET_MODE (x)))
+ x = avoid_constant_pool_reference (x);
+#endif
switch (GET_CODE (x))
{
case SUBREG:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case CALL:
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)))
FIRST_PSEUDO_REGISTER * 2),
const0_rtx));
NEXT_INSN (test_insn) = PREV_INSN (test_insn) = 0;
- ggc_add_rtx_root (&test_insn, 1);
}
/* Now make an insn like the one we would make when GCSE'ing and see if
&& (num_clobbers == 0 || ! added_clobbers_hard_reg_p (icode)));
}
-/* Return non-zero if the operands of expression X are unchanged from the
+/* Return nonzero if the operands of expression X are unchanged from the
start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
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 (const_rtx x, const_rtx insn, int avail_p)
{
int i, j;
enum rtx_code code;
if (info->last_bb != current_bb)
return 1;
- if (avail_p)
+ if (avail_p)
return info->last_set < INSN_CUID (insn);
else
return info->first_set >= INSN_CUID (insn);
}
case MEM:
- if (load_killed_in_block_p (BASIC_BLOCK (current_bb), INSN_CUID (insn),
+ if (load_killed_in_block_p (current_bb, INSN_CUID (insn),
x, avail_p))
return 0;
else
case CONST:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case SYMBOL_REF:
case LABEL_REF:
load_killed_in_block_p. A memory reference for a load instruction,
mems_conflict_for_gcse_p will see if a memory store conflicts with
this memory load. */
-static rtx gcse_mem_operand;
+static const_rtx gcse_mem_operand;
/* DEST is the output of an instruction. If it is a memory reference, and
possibly conflicts with the load found in gcse_mem_operand, then set
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, const_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,
- don't mark as killed this time. */
-
- if (dest == gcse_mem_operand && pre_ldst_mems != NULL)
+ don't mark as killed this time. */
+
+ if (expr_equiv_p (dest, gcse_mem_operand) && pre_ldst_mems != NULL)
{
if (!find_rtx_in_ldst (dest))
gcse_mems_conflict_p = 1;
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 (const_basic_block bb, int uid_limit, const_rtx x, int avail_p)
{
rtx list_entry = modify_mem_list[bb->index];
+
+ /* If this is a readonly then we aren't going to be changing it. */
+ if (MEM_READONLY_P (x))
+ return 0;
+
while (list_entry)
{
rtx setter;
/* 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
- note_stores to examine each hunk of memory that is modified.
+ note_stores to examine each hunk of memory that is modified.
The note_stores interface is pretty limited, so we have to
communicate via global variables. Yuk. */
return 0;
}
-/* Return non-zero if the operands of expression X are unchanged from
+/* Return nonzero if the operands of expression X are unchanged from
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 (const_rtx x, const_rtx insn)
{
return oprs_unchanged_p (x, insn, 0);
}
-/* Return non-zero if the operands of expression X are unchanged from
+/* Return nonzero if the operands of expression X are unchanged from
INSN to the end of INSN's basic block. */
static int
-oprs_available_p (x, insn)
- rtx x, insn;
+oprs_available_p (const_rtx x, const_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 (const_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. */
+/* Hash a set of register REGNO.
+
+ Sets are hashed on the register that is set. This simplifies the PRE copy
+ propagation code.
+
+ ??? May need to make things more elaborate. Later, as necessary. */
-static inline unsigned
-hash_string_1 (ps)
- const char *ps;
+static unsigned int
+hash_set (int regno, int hash_table_size)
{
- unsigned hash = 0;
- const unsigned char *p = (const unsigned char *) ps;
-
- if (p)
- while (*p)
- hash += *p++;
+ unsigned int hash;
- return hash;
+ hash = regno;
+ return hash % hash_table_size;
}
-/* Subroutine of hash_expr to do the actual work. */
+/* Return nonzero if exp1 is equivalent to exp2. */
-static unsigned int
-hash_expr_1 (x, mode, do_not_record_p)
- rtx x;
- enum machine_mode mode;
- int *do_not_record_p;
+static int
+expr_equiv_p (const_rtx x, const_rtx y)
{
- int i, j;
- unsigned hash = 0;
- enum rtx_code code;
- const char *fmt;
+ return exp_equiv_p (x, y, 0, true);
+}
- /* Used to turn recursion into iteration. We can't rely on GCC's
- tail-recursion eliminatio since we need to keep accumulating values
- in HASH. */
+/* Insert expression X in INSN in the hash TABLE.
+ If it is already present, record it as the last occurrence in INSN's
+ basic block.
- if (x == 0)
- return hash;
+ MODE is the mode of the value X is being stored into.
+ It is only used if X is a CONST_INT.
- repeat:
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- hash += ((unsigned int) REG << 7) + REGNO (x);
- return hash;
+ ANTIC_P is nonzero if X is an anticipatable expression.
+ AVAIL_P is nonzero if X is an available expression. */
- case CONST_INT:
- hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
- + (unsigned int) INTVAL (x));
- return hash;
+static void
+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;
- 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;
+ hash = hash_expr (x, mode, &do_not_record_p, table->size);
- case CONST_VECTOR:
- {
- int units;
- rtx elt;
+ /* Do not insert expression in table if it contains volatile operands,
+ or if hash_expr determines the expression is something we don't want
+ to or can't handle. */
+ if (do_not_record_p)
+ return;
- units = CONST_VECTOR_NUNITS (x);
+ cur_expr = table->table[hash];
+ found = 0;
- for (i = 0; i < units; ++i)
- {
- elt = CONST_VECTOR_ELT (x, i);
- hash += hash_expr_1 (elt, GET_MODE (elt), do_not_record_p);
- }
+ while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
+ {
+ /* If the expression isn't found, save a pointer to the end of
+ the list. */
+ last_expr = cur_expr;
+ cur_expr = cur_expr->next_same_hash;
+ }
- return hash;
- }
+ if (! found)
+ {
+ cur_expr = GOBNEW (struct expr);
+ bytes_used += sizeof (struct expr);
+ if (table->table[hash] == NULL)
+ /* This is the first pattern that hashed to this index. */
+ table->table[hash] = cur_expr;
+ else
+ /* Add EXPR to end of this hash chain. */
+ last_expr->next_same_hash = cur_expr;
- /* 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;
+ /* Set the fields of the expr element. */
+ cur_expr->expr = x;
+ cur_expr->bitmap_index = table->n_elems++;
+ cur_expr->next_same_hash = NULL;
+ cur_expr->antic_occr = NULL;
+ cur_expr->avail_occr = NULL;
+ }
- 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;
- }
+ /* Now record the occurrence(s). */
+ if (antic_p)
+ {
+ antic_occr = cur_expr->antic_occr;
- case MEM:
- if (MEM_VOLATILE_P (x))
+ 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.
+ Prefer the currently recorded one. We want the first one in the
+ block and the block is scanned from start to end. */
+ ; /* nothing to do */
+ else
{
- *do_not_record_p = 1;
- return 0;
+ /* First occurrence of this expression in this basic block. */
+ antic_occr = GOBNEW (struct occr);
+ bytes_used += sizeof (struct occr);
+ antic_occr->insn = insn;
+ antic_occr->next = cur_expr->antic_occr;
+ antic_occr->deleted_p = 0;
+ cur_expr->antic_occr = antic_occr;
}
+ }
- hash += (unsigned int) MEM;
- hash += MEM_ALIAS_SET (x);
- 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;
+ if (avail_p)
+ {
+ avail_occr = cur_expr->avail_occr;
- case ASM_OPERANDS:
- if (MEM_VOLATILE_P (x))
+ if (avail_occr && BLOCK_NUM (avail_occr->insn) == BLOCK_NUM (insn))
{
- *do_not_record_p = 1;
- return 0;
+ /* 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
{
- /* 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;
+ /* First occurrence of this expression in this basic block. */
+ avail_occr = GOBNEW (struct occr);
+ bytes_used += sizeof (struct occr);
+ avail_occr->insn = insn;
+ avail_occr->next = cur_expr->avail_occr;
+ avail_occr->deleted_p = 0;
+ cur_expr->avail_occr = avail_occr;
}
-
- 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
- propagation code.
-
- ??? 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;
-{
- unsigned int hash;
-
- hash = regno;
- return hash % hash_table_size;
-}
-
-/* Return non-zero if exp1 is equivalent to exp2.
- ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
-
-static int
-expr_equiv_p (x, y)
- rtx x, y;
-{
- int i, j;
- enum rtx_code code;
- const char *fmt;
-
- if (x == y)
- return 1;
-
- if (x == 0 || y == 0)
- return x == y;
-
- 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:
- return x == y;
-
- case CONST_INT:
- return INTVAL (x) == INTVAL (y);
-
- 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;
-}
-
-/* Insert expression X in INSN in the hash table.
- If it is already present, record it as the last occurrence in INSN's
- basic block.
-
- MODE is the mode of the value X is being stored into.
- It is only used if X is a CONST_INT.
-
- ANTIC_P is non-zero if X is an anticipatable expression.
- AVAIL_P is non-zero if X is an available expression. */
-
-static void
-insert_expr_in_table (x, mode, insn, antic_p, avail_p)
- rtx x;
- enum machine_mode mode;
- rtx insn;
- int antic_p, avail_p;
-{
- 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, expr_hash_table_size);
-
- /* Do not insert expression in table if it contains volatile operands,
- or if hash_expr determines the expression is something we don't want
- to or can't handle. */
- if (do_not_record_p)
- return;
-
- cur_expr = expr_hash_table[hash];
- found = 0;
-
- while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
- {
- /* If the expression isn't found, save a pointer to the end of
- the list. */
- last_expr = cur_expr;
- cur_expr = cur_expr->next_same_hash;
- }
-
- if (! found)
- {
- cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
- bytes_used += sizeof (struct expr);
- if (expr_hash_table[hash] == NULL)
- /* This is the first pattern that hashed to this index. */
- expr_hash_table[hash] = cur_expr;
- else
- /* Add EXPR to end of this hash chain. */
- last_expr->next_same_hash = cur_expr;
-
- /* Set the fields of the expr element. */
- cur_expr->expr = x;
- cur_expr->bitmap_index = n_exprs++;
- cur_expr->next_same_hash = NULL;
- cur_expr->antic_occr = NULL;
- cur_expr->avail_occr = NULL;
- }
-
- /* Now record the occurrence(s). */
- if (antic_p)
- {
- 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)
- /* Found another instance of the expression in the same basic block.
- Prefer the currently recorded one. We want the first one in the
- block and the block is scanned from start to end. */
- ; /* nothing to do */
- else
- {
- /* First occurrence of this expression in this basic block. */
- antic_occr = (struct 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;
- }
- }
-
- if (avail_p)
- {
- 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 an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = avail_occr;
- avail_occr = avail_occr->next;
- }
-
- 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));
- 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;
- }
- }
-}
+}
/* Insert pattern X in INSN in the hash table.
X is a SET of a reg to either another reg or a constant.
basic block. */
static void
-insert_set_in_table (x, insn)
- rtx x;
- rtx insn;
+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)), set_hash_table_size);
+ hash = hash_set (REGNO (SET_DEST (x)), table->size);
- cur_expr = set_hash_table[hash];
+ cur_expr = table->table[hash];
found = 0;
while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
if (! found)
{
- cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
+ cur_expr = GOBNEW (struct expr);
bytes_used += sizeof (struct expr);
- if (set_hash_table[hash] == NULL)
+ if (table->table[hash] == NULL)
/* This is the first pattern that hashed to this index. */
- set_hash_table[hash] = cur_expr;
+ table->table[hash] = cur_expr;
else
/* Add EXPR to end of this hash chain. */
last_expr->next_same_hash = cur_expr;
We must copy X because it can be modified when copy propagation is
performed on its operands. */
cur_expr->expr = copy_rtx (x);
- cur_expr->bitmap_index = n_sets++;
+ cur_expr->bitmap_index = table->n_elems++;
cur_expr->next_same_hash = NULL;
cur_expr->antic_occr = NULL;
cur_expr->avail_occr = NULL;
/* 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 = GOBNEW (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;
}
}
-/* Scan pattern PAT of INSN and add an entry to the hash table. If SET_P is
- non-zero, this is for the assignment hash table, otherwise it is for the
- expression hash table. */
+/* Determine whether the rtx X should be treated as a constant for
+ the purposes of GCSE's constant propagation. */
+
+static bool
+gcse_constant_p (const_rtx x)
+{
+ /* Consider a COMPARE of two integers constant. */
+ if (GET_CODE (x) == COMPARE
+ && GET_CODE (XEXP (x, 0)) == CONST_INT
+ && GET_CODE (XEXP (x, 1)) == CONST_INT)
+ return true;
+
+ /* 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);
+}
+
+/* Scan pattern PAT of INSN and add an entry to the hash TABLE (set or
+ expression one). */
static void
-hash_scan_set (pat, insn, set_p)
- rtx pat, insn;
- int set_p;
+hash_scan_set (rtx pat, rtx insn, struct hash_table *table)
{
rtx src = SET_SRC (pat);
rtx dest = SET_DEST (pat);
rtx note;
if (GET_CODE (src) == CALL)
- hash_scan_call (src, insn);
+ hash_scan_call (src, insn, table);
- else if (GET_CODE (dest) == REG)
+ else if (REG_P (dest))
{
unsigned int regno = REGNO (dest);
rtx tmp;
- /* If this is a single set and we are doing constant propagation,
- see if a REG_NOTE shows this equivalent to a constant. */
- if (set_p && (note = find_reg_equal_equiv_note (insn)) != 0
- && CONSTANT_P (XEXP (note, 0)))
+ /* See if a REG_EQUAL note shows this equivalent to a simpler expression.
+
+ This allows us to do a single GCSE pass and still eliminate
+ redundant constants, addresses or other expressions that are
+ constructed with multiple instructions.
+
+ However, keep the original SRC if INSN is a simple reg-reg move. In
+ In this case, there will almost always be a REG_EQUAL note on the
+ insn that sets SRC. By recording the REG_EQUAL value here as SRC
+ for INSN, we miss copy propagation opportunities and we perform the
+ same PRE GCSE operation repeatedly on the same REG_EQUAL value if we
+ do more than one PRE GCSE pass.
+
+ Note that this does not impede profitable constant propagations. We
+ "look through" reg-reg sets in lookup_avail_set. */
+ note = find_reg_equal_equiv_note (insn);
+ if (note != 0
+ && REG_NOTE_KIND (note) == REG_EQUAL
+ && !REG_P (src)
+ && (table->set_p
+ ? gcse_constant_p (XEXP (note, 0))
+ : want_to_gcse_p (XEXP (note, 0))))
src = XEXP (note, 0), pat = gen_rtx_SET (VOIDmode, dest, src);
/* Only record sets of pseudo-regs in the hash table. */
- if (! set_p
+ if (! table->set_p
&& regno >= FIRST_PSEUDO_REGISTER
/* Don't GCSE something if we can't do a reg/reg copy. */
- && can_copy_p [GET_MODE (dest)]
+ && can_copy_p (GET_MODE (dest))
/* GCSE commonly inserts instruction after the insn. We can't
do that easily for EH_REGION notes so disable GCSE on these
for now. */
- && !can_throw_internal (insn)
+ && !find_reg_note (insn, REG_EH_REGION, NULL_RTX)
/* Is SET_SRC something we want to gcse? */
&& want_to_gcse_p (src)
/* Don't CSE a nop. */
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
- || GET_CODE (XEXP (note, 0)) != MEM))
+ && (note == NULL_RTX || ! 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
- this insn. */
- int antic_p = oprs_anticipatable_p (src, insn) && single_set (insn);
+ this insn. The latter condition does not have to mean that
+ SRC itself is not anticipatable, but we just will not be
+ able to handle code motion of insns with multiple sets. */
+ int antic_p = oprs_anticipatable_p (src, insn)
+ && !multiple_sets (insn);
/* 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
int avail_p = (oprs_available_p (src, insn)
&& ! JUMP_P (insn));
- insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p);
+ insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p, table);
}
/* Record sets for constant/copy propagation. */
- else if (set_p
+ 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)]
+ && can_copy_p (GET_MODE (dest))
&& REGNO (src) != regno)
- || CONSTANT_P (src))
+ || gcse_constant_p (src))
/* 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))
- || ((tmp = next_nonnote_insn (insn)) != NULL_RTX
- && oprs_available_p (pat, tmp))))
- insert_set_in_table (pat, insn);
+ && (insn == BB_END (BLOCK_FOR_INSN (insn))
+ || (tmp = next_nonnote_insn (insn)) == NULL_RTX
+ || BLOCK_FOR_INSN (tmp) != BLOCK_FOR_INSN (insn)
+ || 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)
- rtx x ATTRIBUTE_UNUSED, insn 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)
- rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
+hash_scan_call (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
+ struct hash_table *table ATTRIBUTE_UNUSED)
{
/* Currently nothing to do. */
}
that isn't dealt with right now. The trick is handling the CLOBBERs that
are also in the PARALLEL. Later.
- If SET_P is non-zero, this is for the assignment hash table,
- otherwise it is for the expression hash table.
- If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
- not record any expressions. */
+ If SET_P is nonzero, this is for the assignment hash table,
+ otherwise it is for the expression hash table. */
static void
-hash_scan_insn (insn, set_p, in_libcall_block)
- rtx insn;
- int set_p;
- int in_libcall_block;
+hash_scan_insn (rtx insn, struct hash_table *table)
{
rtx pat = PATTERN (insn);
int i;
- if (in_libcall_block)
- return;
-
/* Pick out the sets of INSN and for other forms of instructions record
what's been modified. */
if (GET_CODE (pat) == SET)
- hash_scan_set (pat, insn, set_p);
+ hash_scan_set (pat, insn, table);
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)
- hash_scan_set (x, insn, set_p);
+ hash_scan_set (x, insn, table);
else if (GET_CODE (x) == CLOBBER)
- hash_scan_clobber (x, insn);
+ hash_scan_clobber (x, insn, table);
else if (GET_CODE (x) == CALL)
- hash_scan_call (x, insn);
+ hash_scan_call (x, insn, table);
}
else if (GET_CODE (pat) == CLOBBER)
- hash_scan_clobber (pat, insn);
+ hash_scan_clobber (pat, insn, table);
else if (GET_CODE (pat) == CALL)
- hash_scan_call (pat, insn);
+ hash_scan_call (pat, insn, table);
}
static void
-dump_hash_table (file, name, table, table_size, total_size)
- FILE *file;
- const char *name;
- struct expr **table;
- int table_size, total_size;
+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 (total_size, sizeof (struct expr *));
- hash_val = (unsigned int *) xmalloc (total_size * sizeof (unsigned int));
+ flat_table = XCNEWVEC (struct expr *, table->n_elems);
+ hash_val = XNEWVEC (unsigned int, table->n_elems);
- for (i = 0; i < table_size; i++)
- for (expr = table[i]; expr != NULL; expr = expr->next_same_hash)
+ for (i = 0; i < (int) table->size; i++)
+ for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
{
flat_table[expr->bitmap_index] = expr;
hash_val[expr->bitmap_index] = i;
}
fprintf (file, "%s hash table (%d buckets, %d entries)\n",
- name, table_size, total_size);
+ name, table->size, table->n_elems);
- for (i = 0; i < total_size; i++)
+ for (i = 0; i < (int) table->n_elems; i++)
if (flat_table[i] != 0)
{
expr = flat_table[i];
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);
{
info->last_bb = current_bb;
info->first_set = cuid;
- SET_BIT (reg_set_in_block[current_bb], regno);
+ SET_BIT (reg_set_in_block[current_bb->index], regno);
}
}
Note we store a pair of elements in the list, so they have to be
taken off pairwise. */
-static void
-canon_list_insert (dest, unused1, v_insn)
- rtx dest ATTRIBUTE_UNUSED;
- rtx unused1 ATTRIBUTE_UNUSED;
- void * v_insn;
+static void
+canon_list_insert (rtx dest ATTRIBUTE_UNUSED, const_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));
dest_addr = canon_rtx (dest_addr);
- insn = (rtx) v_insn;
+ insn = (rtx) v_insn;
bb = BLOCK_NUM (insn);
- canon_modify_mem_list[bb] =
+ canon_modify_mem_list[bb] =
alloc_EXPR_LIST (VOIDmode, dest_addr, canon_modify_mem_list[bb]);
- 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);
+ canon_modify_mem_list[bb] =
+ alloc_INSN_LIST (insn, canon_modify_mem_list[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, const_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);
Currently src must be a pseudo-reg or a const_int.
- F is the first insn.
- SET_P is non-zero for computing the assignment hash table. */
+ TABLE is the table computed. */
static void
-compute_hash_table (set_p)
- int set_p;
+compute_hash_table_work (struct hash_table *table)
{
unsigned int i;
registers are set in which blocks.
??? This isn't needed during const/copy propagation, but it's cheap to
compute. Later. */
- sbitmap_vector_zero (reg_set_in_block, n_basic_blocks);
+ sbitmap_vector_zero (reg_set_in_block, last_basic_block);
/* 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 = GNEWVEC (struct reg_avail_info, max_gcse_regno);
for (i = 0; i < max_gcse_regno; ++i)
- reg_avail_info[i].last_bb = NEVER_SET;
+ reg_avail_info[i].last_bb = NULL;
- for (current_bb = 0; current_bb < n_basic_blocks; current_bb++)
+ FOR_EACH_BB (current_bb)
{
rtx insn;
unsigned int regno;
- int in_libcall_block;
/* First pass over the instructions records information used to
determine when registers and memory are first and last set.
??? hard-reg reg_set_in_block computation
could be moved to compute_sets since they currently don't change. */
- for (insn = BLOCK_HEAD (current_bb);
- insn && insn != NEXT_INSN (BLOCK_END (current_bb));
- insn = NEXT_INSN (insn))
+ FOR_BB_INSNS (current_bb, 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);
note_stores (PATTERN (insn), record_last_set_info, insn);
}
- /* The next pass builds the hash table. */
+ /* Insert implicit sets in the hash table. */
+ if (table->set_p
+ && implicit_sets[current_bb->index] != NULL_RTX)
+ hash_scan_set (implicit_sets[current_bb->index],
+ BB_HEAD (current_bb), table);
- for (insn = BLOCK_HEAD (current_bb), in_libcall_block = 0;
- insn && insn != NEXT_INSN (BLOCK_END (current_bb));
- insn = NEXT_INSN (insn))
+ /* The next pass builds the hash table. */
+ FOR_BB_INSNS (current_bb, insn)
if (INSN_P (insn))
- {
- if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
- in_libcall_block = 1;
- else if (set_p && find_reg_note (insn, REG_RETVAL, NULL_RTX))
- in_libcall_block = 0;
- hash_scan_insn (insn, set_p, in_libcall_block);
- if (!set_p && find_reg_note (insn, REG_RETVAL, NULL_RTX))
- in_libcall_block = 0;
- }
+ hash_scan_insn (insn, table);
}
free (reg_avail_info);
reg_avail_info = NULL;
}
-/* Allocate space for the set hash table.
- N_INSNS is the number of instructions in the function.
- It is used to determine the number of buckets to use. */
-
-static void
-alloc_set_hash_table (n_insns)
- int n_insns;
-{
- int n;
-
- set_hash_table_size = n_insns / 4;
- if (set_hash_table_size < 11)
- set_hash_table_size = 11;
-
- /* Attempt to maintain efficient use of hash table.
- Making it an odd number is simplest for now.
- ??? Later take some measurements. */
- set_hash_table_size |= 1;
- n = set_hash_table_size * sizeof (struct expr *);
- set_hash_table = (struct expr **) gmalloc (n);
-}
-
-/* Free things allocated by alloc_set_hash_table. */
-
-static void
-free_set_hash_table ()
-{
- free (set_hash_table);
-}
-
-/* Compute the hash table for doing copy/const propagation. */
-
-static void
-compute_set_hash_table ()
-{
- /* Initialize count of number of entries in hash table. */
- n_sets = 0;
- memset ((char *) set_hash_table, 0,
- set_hash_table_size * sizeof (struct expr *));
-
- compute_hash_table (1);
-}
-
-/* Allocate space for the expression hash table.
+/* Allocate space for the set/expr hash TABLE.
N_INSNS is the number of instructions in the function.
- It is used to determine the number of buckets to use. */
+ It is used to determine the number of buckets to use.
+ SET_P determines whether set or expression table will
+ be created. */
static void
-alloc_expr_hash_table (n_insns)
- unsigned int n_insns;
+alloc_hash_table (int n_insns, struct hash_table *table, int set_p)
{
int n;
- expr_hash_table_size = n_insns / 2;
- /* Make sure the amount is usable. */
- if (expr_hash_table_size < 11)
- expr_hash_table_size = 11;
+ table->size = n_insns / 4;
+ if (table->size < 11)
+ table->size = 11;
/* Attempt to maintain efficient use of hash table.
Making it an odd number is simplest for now.
??? Later take some measurements. */
- expr_hash_table_size |= 1;
- n = expr_hash_table_size * sizeof (struct expr *);
- expr_hash_table = (struct expr **) gmalloc (n);
+ table->size |= 1;
+ n = table->size * sizeof (struct expr *);
+ table->table = GNEWVAR (struct expr *, n);
+ table->set_p = set_p;
}
-/* Free things allocated by alloc_expr_hash_table. */
+/* Free things allocated by alloc_hash_table. */
static void
-free_expr_hash_table ()
+free_hash_table (struct hash_table *table)
{
- free (expr_hash_table);
+ free (table->table);
}
-/* Compute the hash table for doing GCSE. */
+/* Compute the hash TABLE for doing copy/const propagation or
+ expression hash table. */
static void
-compute_expr_hash_table ()
+compute_hash_table (struct hash_table *table)
{
/* Initialize count of number of entries in hash table. */
- n_exprs = 0;
- memset ((char *) expr_hash_table, 0,
- expr_hash_table_size * sizeof (struct expr *));
+ table->n_elems = 0;
+ memset (table->table, 0, table->size * sizeof (struct expr *));
- compute_hash_table (0);
+ 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)
- rtx pat;
+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,
- expr_hash_table_size);
+ unsigned int hash = hash_set (regno, table->size);
struct expr *expr;
- if (do_not_record_p)
- return NULL;
-
- expr = expr_hash_table[hash];
+ expr = table->table[hash];
- while (expr && ! expr_equiv_p (expr->expr, pat))
+ while (expr && REGNO (SET_DEST (expr->expr)) != regno)
expr = expr->next_same_hash;
return expr;
}
-/* Lookup REGNO in the set table. If PAT is non-NULL look for the entry that
- matches it, otherwise return the first entry for REGNO. The result is a
- pointer to the table entry, or NULL if not found. */
+/* Return the next entry for REGNO in list EXPR. */
static struct expr *
-lookup_set (regno, pat)
- unsigned int regno;
- rtx pat;
-{
- unsigned int hash = hash_set (regno, set_hash_table_size);
- struct expr *expr;
-
- expr = set_hash_table[hash];
-
- if (pat)
- {
- while (expr && ! expr_equiv_p (expr->expr, pat))
- expr = expr->next_same_hash;
- }
- else
- {
- while (expr && REGNO (SET_DEST (expr->expr)) != regno)
- expr = expr->next_same_hash;
- }
-
- return expr;
-}
-
-/* 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. */
clear_modify_mem_tables ();
}
-/* Return non-zero if the operands of X are not set before INSN in
+/* Return nonzero if the operands of X are not set before INSN in
INSN's basic block. */
static int
-oprs_not_set_p (x, insn)
- rtx x, insn;
+oprs_not_set_p (const_rtx x, const_rtx insn)
{
int i, j;
enum rtx_code code;
case CONST:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case SYMBOL_REF:
case LABEL_REF:
return 1;
case MEM:
- if (load_killed_in_block_p (BLOCK_FOR_INSN (insn),
+ if (load_killed_in_block_p (BLOCK_FOR_INSN (insn),
INSN_CUID (insn), x, 0))
return 0;
else
/* Mark things set by a CALL. */
static void
-mark_call (insn)
- rtx insn;
+mark_call (rtx insn)
{
- if (! CONST_OR_PURE_CALL_P (insn))
+ if (! RTL_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)
/* Record things set by a CLOBBER. */
static void
-mark_clobber (pat, insn)
- rtx pat, insn;
+mark_clobber (rtx pat, rtx 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)
+ if (REG_P (clob))
SET_REGNO_REG_SET (reg_set_bitmap, REGNO (clob));
else
record_last_mem_set_info (insn);
This data is used by oprs_not_set_p. */
static void
-mark_oprs_set (insn)
- rtx insn;
+mark_oprs_set (rtx insn)
{
rtx pat = PATTERN (insn);
int i;
}
\f
-/* Classic GCSE reaching definition support. */
+/* Compute copy/constant propagation working variables. */
-/* Allocate reaching def variables. */
+/* Local properties of assignments. */
+static sbitmap *cprop_pavloc;
+static sbitmap *cprop_absaltered;
-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_basic_blocks);
+/* Global properties of assignments (computed from the local properties). */
+static sbitmap *cprop_avin;
+static sbitmap *cprop_avout;
- rd_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_gen, n_basic_blocks);
+/* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
+ basic blocks. N_SETS is the number of sets. */
- reaching_defs = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (reaching_defs, n_basic_blocks);
+static void
+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);
- rd_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_out, n_basic_blocks);
+ cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
+ cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
}
-/* Free reaching def variables. */
+/* Free vars used by copy/const propagation. */
static void
-free_rd_mem ()
+free_cprop_mem (void)
{
- sbitmap_vector_free (rd_kill);
- sbitmap_vector_free (rd_gen);
- sbitmap_vector_free (reaching_defs);
- sbitmap_vector_free (rd_out);
+ sbitmap_vector_free (cprop_pavloc);
+ sbitmap_vector_free (cprop_absaltered);
+ sbitmap_vector_free (cprop_avin);
+ sbitmap_vector_free (cprop_avout);
}
-/* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
+/* For each block, compute whether X is transparent. X is either an
+ expression or an assignment [though we don't care which, for this context
+ an assignment is treated as an expression]. For each block where an
+ element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
+ bit in BMAP. */
static void
-handle_rd_kill_set (insn, regno, bb)
- rtx insn;
- int regno;
- basic_block bb;
+compute_transp (const_rtx x, int indx, sbitmap *bmap, int set_p)
{
- 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));
-}
+ int i, j;
+ basic_block bb;
+ enum rtx_code code;
+ reg_set *r;
+ const char *fmt;
-/* Compute the set of kill's for reaching definitions. */
+ /* repeat is used to turn tail-recursion into iteration since GCC
+ can't do it when there's no return value. */
+ repeat:
-static void
-compute_kill_rd ()
-{
- int bb, cuid;
- unsigned int regno;
- int i;
+ if (x == 0)
+ return;
- /* 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 (bb = 0; bb < n_basic_blocks; bb++)
- for (cuid = 0; cuid < max_cuid; cuid++)
- if (TEST_BIT (rd_gen[bb], cuid))
+ code = GET_CODE (x);
+ switch (code)
+ {
+ case REG:
+ if (set_p)
{
- rtx insn = CUID_INSN (cuid);
- rtx pat = PATTERN (insn);
-
- if (GET_CODE (insn) == CALL_INSN)
+ if (REGNO (x) < FIRST_PSEUDO_REGISTER)
{
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- handle_rd_kill_set (insn, regno, BASIC_BLOCK (bb));
+ FOR_EACH_BB (bb)
+ if (TEST_BIT (reg_set_in_block[bb->index], REGNO (x)))
+ SET_BIT (bmap[bb->index], indx);
}
-
- if (GET_CODE (pat) == PARALLEL)
+ else
{
- 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)),
- BASIC_BLOCK (bb));
- }
+ for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
+ SET_BIT (bmap[r->bb_index], indx);
}
- 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)), BASIC_BLOCK (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 bb, changed, passes;
-
- for (bb = 0; bb < n_basic_blocks; bb++)
- sbitmap_copy (rd_out[bb] /*dst*/, rd_gen[bb] /*src*/);
-
- passes = 0;
- changed = 1;
- while (changed)
- {
- changed = 0;
- for (bb = 0; bb < n_basic_blocks; bb++)
+ else
{
- sbitmap_union_of_preds (reaching_defs[bb], rd_out, bb);
- changed |= sbitmap_union_of_diff (rd_out[bb], rd_gen[bb],
- reaching_defs[bb], rd_kill[bb]);
+ if (REGNO (x) < FIRST_PSEUDO_REGISTER)
+ {
+ FOR_EACH_BB (bb)
+ if (TEST_BIT (reg_set_in_block[bb->index], REGNO (x)))
+ RESET_BIT (bmap[bb->index], indx);
+ }
+ else
+ {
+ for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
+ RESET_BIT (bmap[r->bb_index], indx);
+ }
}
- 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_basic_blocks);
-
- ae_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_gen, n_basic_blocks);
-
- ae_in = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_in, n_basic_blocks);
-
- ae_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_out, n_basic_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);
-}
+ return;
-/* Compute the set of available expressions generated in each basic block. */
+ case MEM:
+ if (! MEM_READONLY_P (x))
+ {
+ bitmap_iterator bi;
+ unsigned bb_index;
-static void
-compute_ae_gen ()
-{
- unsigned int i;
- struct expr *expr;
- struct occr *occr;
+ /* 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);
+ }
- /* 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[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);
-}
+ /* 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];
-/* Return non-zero if expression X is killed in BB. */
+ while (list_entry)
+ {
+ rtx dest, dest_addr;
-static int
-expr_killed_p (x, bb)
- rtx x;
- basic_block bb;
-{
- int i, j;
- enum rtx_code code;
- const char *fmt;
+ /* LIST_ENTRY must be an INSN of some kind that sets memory.
+ Examine each hunk of memory that is modified. */
- if (x == 0)
- return 1;
+ dest = XEXP (list_entry, 0);
+ list_entry = XEXP (list_entry, 1);
+ dest_addr = XEXP (list_entry, 0);
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- return TEST_BIT (reg_set_in_block[bb->index], REGNO (x));
+ if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
+ x, NULL_RTX, 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);
+ }
+ }
+ }
- 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);
+ x = XEXP (x, 0);
+ goto repeat;
case PC:
case CC0: /*FIXME*/
case CONST:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case SYMBOL_REF:
case LABEL_REF:
case ADDR_VEC:
case ADDR_DIFF_VEC:
- return 0;
+ return;
default:
break;
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;
+ {
+ x = XEXP (x, i);
+ goto repeat;
+ }
+
+ compute_transp (XEXP (x, i), indx, bmap, set_p);
}
else if (fmt[i] == 'E')
for (j = 0; j < XVECLEN (x, i); j++)
- if (expr_killed_p (XVECEXP (x, i, j), bb))
- return 1;
+ compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
}
-
- return 0;
}
-/* Compute the set of available expressions killed in each basic block. */
+/* Top level routine to do the dataflow analysis needed by copy/const
+ propagation. */
static void
-compute_ae_kill (ae_gen, ae_kill)
- sbitmap *ae_gen, *ae_kill;
+compute_cprop_data (void)
{
- int bb;
- unsigned int i;
- struct expr *expr;
-
- for (bb = 0; bb < n_basic_blocks; bb++)
- for (i = 0; i < expr_hash_table_size; i++)
- for (expr = expr_hash_table[i]; expr; expr = expr->next_same_hash)
- {
- /* Skip EXPR if generated in this block. */
- if (TEST_BIT (ae_gen[bb], expr->bitmap_index))
- continue;
-
- if (expr_killed_p (expr->expr, BASIC_BLOCK (bb)))
- SET_BIT (ae_kill[bb], expr->bitmap_index);
- }
+ compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, &set_hash_table);
+ compute_available (cprop_pavloc, cprop_absaltered,
+ cprop_avout, cprop_avin);
}
\f
-/* Actually perform the Classic GCSE optimizations. */
-
-/* Return non-zero if occurrence OCCR of expression EXPR reaches block BB.
+/* Copy/constant propagation. */
- CHECK_SELF_LOOP is non-zero 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.
+/* Maximum number of register uses in an insn that we handle. */
+#define MAX_USES 8
- VISITED is a pointer to a working buffer for tracking which BB's have
- been visited. It is NULL for the top-level call.
+/* Table of uses found in an insn.
+ Allocated statically to avoid alloc/free complexity and overhead. */
+static struct reg_use reg_use_table[MAX_USES];
- 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. */
+/* Index into `reg_use_table' while building it. */
+static int reg_use_count;
-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;
+/* Set up a list of register numbers used in INSN. The found uses are stored
+ in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
+ and contains the number of uses in the table upon exit.
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
- {
- basic_block pred_bb = pred->src;
+ ??? If a register appears multiple times we will record it multiple times.
+ This doesn't hurt anything but it will slow things down. */
- 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;
+static void
+find_used_regs (rtx *xptr, void *data ATTRIBUTE_UNUSED)
+{
+ int i, j;
+ enum rtx_code code;
+ const char *fmt;
+ rtx x = *xptr;
- visited[pred_bb->index] = 1;
- }
+ /* repeat is used to turn tail-recursion into iteration since GCC
+ can't do it when there's no return value. */
+ repeat:
+ if (x == 0)
+ return;
- /* 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;
+ code = GET_CODE (x);
+ if (REG_P (x))
+ {
+ if (reg_use_count == MAX_USES)
+ return;
- /* 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;
+ reg_use_table[reg_use_count].reg_rtx = x;
+ reg_use_count++;
+ }
- visited[pred_bb->index] = 1;
- }
+ /* Recursively scan the operands of this expression. */
- /* Neither gen nor kill. */
- else
+ for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
+ {
+ if (fmt[i] == 'e')
{
- visited[pred_bb->index] = 1;
- if (expr_reaches_here_p_work (occr, expr, pred_bb, check_self_loop,
- visited))
+ /* 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, 0);
+ goto repeat;
+ }
- return 1;
+ find_used_regs (&XEXP (x, i), data);
}
- }
-
- /* 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 (n_basic_blocks, 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 non-zero 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 non-zero 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 non-zero 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 non-zero if a change was made. */
-
-static int
-classic_gcse ()
-{
- int bb, changed;
- rtx insn;
-
- /* Note we start at block 1. */
-
- changed = 0;
- for (bb = 1; bb < n_basic_blocks; bb++)
- {
- /* Reset tables used to keep track of what's still valid [since the
- start of the block]. */
- reset_opr_set_tables ();
-
- for (insn = BLOCK_HEAD (bb);
- insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
- 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)) != NULL)
- /* Is the expression available [at the start of the
- block]? */
- && TEST_BIT (ae_in[bb], 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;
-}
-
-/* Top level routine to perform one classic GCSE pass.
-
- Return non-zero if a change was made. */
-
-static int
-one_classic_gcse_pass (pass)
- int pass;
-{
- int changed = 0;
-
- gcse_subst_count = 0;
- gcse_create_count = 0;
-
- alloc_expr_hash_table (max_cuid);
- alloc_rd_mem (n_basic_blocks, max_cuid);
- compute_expr_hash_table ();
- if (gcse_file)
- dump_hash_table (gcse_file, "Expression", expr_hash_table,
- expr_hash_table_size, n_exprs);
-
- if (n_exprs > 0)
- {
- compute_kill_rd ();
- compute_rd ();
- alloc_avail_expr_mem (n_basic_blocks, n_exprs);
- compute_ae_gen ();
- compute_ae_kill (ae_gen, ae_kill);
- compute_available (ae_gen, ae_kill, ae_out, ae_in);
- changed = classic_gcse ();
- free_avail_expr_mem ();
- }
-
- free_rd_mem ();
- free_expr_hash_table ();
-
- 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);
- }
-
- return changed;
-}
-\f
-/* Compute copy/constant propagation working variables. */
-
-/* Local properties of assignments. */
-static sbitmap *cprop_pavloc;
-static sbitmap *cprop_absaltered;
-
-/* Global properties of assignments (computed from the local properties). */
-static sbitmap *cprop_avin;
-static sbitmap *cprop_avout;
-
-/* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
- basic blocks. N_SETS is the number of sets. */
-
-static void
-alloc_cprop_mem (n_blocks, n_sets)
- int n_blocks, n_sets;
-{
- cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
- cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
-
- cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
- cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
-}
-
-/* Free vars used by copy/const propagation. */
-
-static void
-free_cprop_mem ()
-{
- sbitmap_vector_free (cprop_pavloc);
- sbitmap_vector_free (cprop_absaltered);
- sbitmap_vector_free (cprop_avin);
- sbitmap_vector_free (cprop_avout);
-}
-
-/* For each block, compute whether X is transparent. X is either an
- expression or an assignment [though we don't care which, for this context
- an assignment is treated as an expression]. For each block where an
- element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
- bit in BMAP. */
-
-static void
-compute_transp (x, indx, bmap, set_p)
- rtx x;
- int indx;
- sbitmap *bmap;
- int set_p;
-{
- int bb, i, j;
- enum rtx_code code;
- reg_set *r;
- const char *fmt;
-
- /* repeat is used to turn tail-recursion into iteration since GCC
- can't do it when there's no return value. */
- repeat:
-
- if (x == 0)
- return;
-
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- if (set_p)
- {
- if (REGNO (x) < FIRST_PSEUDO_REGISTER)
- {
- for (bb = 0; bb < n_basic_blocks; bb++)
- if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
- SET_BIT (bmap[bb], indx);
- }
- else
- {
- for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
- SET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
- }
- }
- else
- {
- if (REGNO (x) < FIRST_PSEUDO_REGISTER)
- {
- for (bb = 0; bb < n_basic_blocks; bb++)
- if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
- RESET_BIT (bmap[bb], indx);
- }
- else
- {
- for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
- RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
- }
- }
-
- return;
-
- case MEM:
- for (bb = 0; bb < n_basic_blocks; bb++)
- {
- rtx list_entry = canon_modify_mem_list[bb];
-
- while (list_entry)
- {
- rtx dest, dest_addr;
-
- if (GET_CODE (XEXP (list_entry, 0)) == CALL_INSN)
- {
- if (set_p)
- SET_BIT (bmap[bb], indx);
- else
- RESET_BIT (bmap[bb], indx);
- break;
- }
- /* 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], indx);
- else
- RESET_BIT (bmap[bb], indx);
- break;
- }
- list_entry = XEXP (list_entry, 1);
- }
- }
-
- x = XEXP (x, 0);
- goto repeat;
-
- 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;
-
- 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)
- {
- x = XEXP (x, i);
- goto repeat;
- }
-
- compute_transp (XEXP (x, i), indx, bmap, set_p);
- }
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
- }
-}
-
-/* Top level routine to do the dataflow analysis needed by copy/const
- propagation. */
-
-static void
-compute_cprop_data ()
-{
- compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, 1);
- compute_available (cprop_pavloc, cprop_absaltered,
- cprop_avout, cprop_avin);
-}
-\f
-/* Copy/constant propagation. */
-
-/* Maximum number of register uses in an insn that we handle. */
-#define MAX_USES 8
-
-/* Table of uses found in an insn.
- Allocated statically to avoid alloc/free complexity and overhead. */
-static struct reg_use reg_use_table[MAX_USES];
-
-/* Index into `reg_use_table' while building it. */
-static int reg_use_count;
-
-/* Set up a list of register numbers used in INSN. The found uses are stored
- in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
- and contains the number of uses in the table upon exit.
-
- ??? If a register appears multiple times we will record it multiple times.
- This doesn't hurt anything but it will slow things down. */
-
-static void
-find_used_regs (xptr, data)
- rtx *xptr;
- void *data ATTRIBUTE_UNUSED;
-{
- int i, j;
- enum rtx_code code;
- const char *fmt;
- rtx x = *xptr;
-
- /* repeat is used to turn tail-recursion into iteration since GCC
- can't do it when there's no return value. */
- repeat:
- if (x == 0)
- return;
-
- code = GET_CODE (x);
- if (REG_P (x))
- {
- if (reg_use_count == MAX_USES)
- return;
-
- reg_use_table[reg_use_count].reg_rtx = x;
- reg_use_count++;
- }
-
- /* Recursively scan the operands of this expression. */
-
- 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, 0);
- goto repeat;
- }
-
- find_used_regs (&XEXP (x, i), data);
- }
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- find_used_regs (&XVECEXP (x, i, j), data);
+ else if (fmt[i] == 'E')
+ for (j = 0; j < XVECLEN (x, i); j++)
+ find_used_regs (&XVECEXP (x, i, j), data);
}
}
/* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
- Returns non-zero is successful. */
+ 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;
int success = 0;
rtx set = single_set (insn);
- success = validate_replace_src (from, to, insn);
+ /* Usually we substitute easy stuff, so we won't copy everything.
+ We however need to take care to not duplicate non-trivial CONST
+ expressions. */
+ to = copy_rtx (to);
+
+ validate_replace_src_group (from, to, insn);
+ if (num_changes_pending () && apply_change_group ())
+ success = 1;
- /* If above failed and this is a single set, try to simplify the source of
- the set given our substitution. We could perhaps try this for multiple
- SETs, but it probably won't buy us anything. */
- if (!success && set != 0)
+ /* Try to simplify SET_SRC if we have substituted a constant. */
+ if (success && set && CONSTANT_P (to))
{
+ src = simplify_rtx (SET_SRC (set));
+
+ if (src)
+ validate_change (insn, &SET_SRC (set), src, 0);
+ }
+
+ /* If there is already a REG_EQUAL note, update the expression in it
+ with our replacement. */
+ if (note != 0 && REG_NOTE_KIND (note) == REG_EQUAL)
+ set_unique_reg_note (insn, REG_EQUAL,
+ simplify_replace_rtx (XEXP (note, 0), from,
+ copy_rtx (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
+ the set given our substitution. We could perhaps try this for multiple
+ SETs, but it probably won't buy us anything. */
src = simplify_replace_rtx (SET_SRC (set), from, to);
if (!rtx_equal_p (src, SET_SRC (set))
&& 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)
- 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);
+ /* 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
+ && GET_CODE (SET_DEST (set)) != STRICT_LOW_PART)
+ note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
+ }
/* REG_EQUAL may get simplified into register.
We don't allow that. Remove that note. This code ought
- not to hapen, because previous code ought to syntetize
+ not to happen, because previous code ought to synthesize
reg-reg move, but be on the safe side. */
- if (note && REG_P (XEXP (note, 0)))
+ if (note && REG_NOTE_KIND (note) == REG_EQUAL && REG_P (XEXP (note, 0)))
remove_note (insn, note);
return success;
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))
while (1)
{
rtx src;
- struct expr *set = lookup_set (regno, NULL_RTX);
+ struct expr *set = lookup_set (regno, &set_hash_table);
/* Find a set that is available at the start of the block
which contains INSN. */
/* If no available set was found we've reached the end of the
(possibly empty) copy chain. */
if (set == 0)
- break;
+ break;
- if (GET_CODE (set->expr) != SET)
- abort ();
+ gcc_assert (GET_CODE (set->expr) == SET);
src = SET_SRC (set->expr);
/* We know the set is available.
Now check that SRC is ANTLOC (i.e. none of the source operands
- have changed since the start of the block).
+ have changed since the start of the block).
If the source operand changed, we may still use it for the next
iteration of this loop, but we may not use it for substitutions. */
- if (CONSTANT_P (src) || oprs_not_set_p (src, insn))
+ if (gcse_constant_p (src) || oprs_not_set_p (src, insn))
set1 = set;
/* 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);
}
return set1;
}
-/* Subroutine of cprop_insn that tries to propagate constants into
- JUMP_INSNS. INSN must be a conditional jump. FROM is what we will try to
- replace, SRC is the constant we will try to substitute for it. Returns
- nonzero if a change was made. We know INSN has just a SET. */
+/* Subroutine of cprop_insn that tries to propagate constants into
+ JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
+ it is the instruction that immediately precedes JUMP, and must be a
+ single SET of a register. FROM is what we will try to replace,
+ SRC is the constant we will try to substitute for it. Returns nonzero
+ if a change was made. */
+
+static int
+cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
+{
+ rtx new_rtx, set_src, note_src;
+ rtx set = pc_set (jump);
+ rtx note = find_reg_equal_equiv_note (jump);
+
+ if (note)
+ {
+ note_src = XEXP (note, 0);
+ if (GET_CODE (note_src) == EXPR_LIST)
+ note_src = NULL_RTX;
+ }
+ else note_src = NULL_RTX;
+
+ /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
+ set_src = note_src ? note_src : SET_SRC (set);
-static int
-cprop_jump (bb, insn, from, src)
- rtx insn;
- rtx from;
- rtx src;
- basic_block bb;
-{
- rtx set = PATTERN (insn);
- rtx new = simplify_replace_rtx (SET_SRC (set), from, src);
-
- /* If no simplification can be made, then try the next
- register. */
- if (rtx_equal_p (new, SET_SRC (set)))
+ /* First substitute the SETCC condition into the JUMP instruction,
+ then substitute that given values into this expanded JUMP. */
+ if (setcc != NULL_RTX
+ && !modified_between_p (from, setcc, jump)
+ && !modified_between_p (src, setcc, jump))
+ {
+ rtx setcc_src;
+ rtx setcc_set = single_set (setcc);
+ rtx setcc_note = find_reg_equal_equiv_note (setcc);
+ setcc_src = (setcc_note && GET_CODE (XEXP (setcc_note, 0)) != EXPR_LIST)
+ ? XEXP (setcc_note, 0) : SET_SRC (setcc_set);
+ set_src = simplify_replace_rtx (set_src, SET_DEST (setcc_set),
+ setcc_src);
+ }
+ else
+ setcc = NULL_RTX;
+
+ new_rtx = simplify_replace_rtx (set_src, from, src);
+
+ /* If no simplification can be made, then try the next register. */
+ if (rtx_equal_p (new_rtx, SET_SRC (set)))
return 0;
-
+
/* If this is now a no-op delete it, otherwise this must be a valid insn. */
- if (new == pc_rtx)
- delete_insn (insn);
+ if (new_rtx == pc_rtx)
+ delete_insn (jump);
else
{
- if (! validate_change (insn, &SET_SRC (set), new, 0))
+ /* Ensure the value computed inside the jump insn to be equivalent
+ to one computed by setcc. */
+ if (setcc && modified_in_p (new_rtx, setcc))
return 0;
+ if (! validate_unshare_change (jump, &SET_SRC (set), new_rtx, 0))
+ {
+ /* When (some) constants are not valid in a comparison, and there
+ are two registers to be replaced by constants before the entire
+ comparison can be folded into a constant, we need to keep
+ intermediate information in REG_EQUAL notes. For targets with
+ separate compare insns, such notes are added by try_replace_reg.
+ When we have a combined compare-and-branch instruction, however,
+ we need to attach a note to the branch itself to make this
+ optimization work. */
+
+ if (!rtx_equal_p (new_rtx, note_src))
+ set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new_rtx));
+ return 0;
+ }
- /* If this has turned into an unconditional jump,
- then put a barrier after it so that the unreachable
- code will be deleted. */
- if (GET_CODE (SET_SRC (set)) == LABEL_REF)
- emit_barrier_after (insn);
+ /* Remove REG_EQUAL note after simplification. */
+ if (note_src)
+ remove_note (jump, note);
}
+#ifdef HAVE_cc0
+ /* Delete the cc0 setter. */
+ if (setcc != NULL && CC0_P (SET_DEST (single_set (setcc))))
+ delete_insn (setcc);
+#endif
+
run_jump_opt_after_gcse = 1;
- const_prop_count++;
- if (gcse_file != NULL)
+ global_const_prop_count++;
+ if (dump_file != NULL)
{
- fprintf (gcse_file,
- "CONST-PROP: Replacing reg %d in insn %d with constant ",
- REGNO (from), INSN_UID (insn));
- print_rtl (gcse_file, src);
- fprintf (gcse_file, "\n");
+ fprintf (dump_file,
+ "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
+ REGNO (from), INSN_UID (jump));
+ print_rtl (dump_file, src);
+ fprintf (dump_file, "\n");
}
purge_dead_edges (bb);
+ /* If a conditional jump has been changed into unconditional jump, remove
+ the jump and make the edge fallthru - this is always called in
+ cfglayout mode. */
+ if (new_rtx != pc_rtx && simplejump_p (jump))
+ {
+ edge e;
+ edge_iterator ei;
+
+ for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ei_next (&ei))
+ if (e->dest != EXIT_BLOCK_PTR
+ && BB_HEAD (e->dest) == JUMP_LABEL (jump))
+ {
+ e->flags |= EDGE_FALLTHRU;
+ break;
+ }
+ delete_insn (jump);
+ }
+
return 1;
}
-#ifdef HAVE_cc0
+static bool
+constprop_register (rtx insn, rtx from, rtx to, bool alter_jumps)
+{
+ rtx sset;
-/* Subroutine of cprop_insn that tries to propagate constants into JUMP_INSNS
- for machines that have CC0. INSN is a single set that stores into CC0;
- the insn following it is a conditional jump. REG_USED is the use we will
- try to replace, SRC is the constant we will try to substitute for it.
- Returns nonzero if a change was made. */
+ /* Check for reg or cc0 setting instructions followed by
+ conditional branch instructions first. */
+ if (alter_jumps
+ && (sset = single_set (insn)) != NULL
+ && NEXT_INSN (insn)
+ && any_condjump_p (NEXT_INSN (insn)) && onlyjump_p (NEXT_INSN (insn)))
+ {
+ rtx dest = SET_DEST (sset);
+ if ((REG_P (dest) || CC0_P (dest))
+ && cprop_jump (BLOCK_FOR_INSN (insn), insn, NEXT_INSN (insn), from, to))
+ return 1;
+ }
-static int
-cprop_cc0_jump (bb, insn, reg_used, src)
- basic_block bb;
- rtx insn;
- struct reg_use *reg_used;
- rtx src;
-{
- /* First substitute in the SET_SRC of INSN, then substitute that for
- CC0 in JUMP. */
- rtx jump = NEXT_INSN (insn);
- rtx new_src = simplify_replace_rtx (SET_SRC (PATTERN (insn)),
- reg_used->reg_rtx, src);
-
- if (! cprop_jump (bb, jump, cc0_rtx, new_src))
- return 0;
+ /* Handle normal insns next. */
+ if (NONJUMP_INSN_P (insn)
+ && try_replace_reg (from, to, insn))
+ return 1;
- /* If we succeeded, delete the cc0 setter. */
- delete_insn (insn);
+ /* Try to propagate a CONST_INT into a conditional jump.
+ We're pretty specific about what we will handle in this
+ code, we can extend this as necessary over time.
- return 1;
+ Right now the insn in question must look like
+ (set (pc) (if_then_else ...)) */
+ else if (alter_jumps && any_condjump_p (insn) && onlyjump_p (insn))
+ return cprop_jump (BLOCK_FOR_INSN (insn), NULL, insn, from, to);
+ return 0;
}
-#endif
-
+
/* Perform constant and copy propagation on INSN.
- The result is non-zero if a change was made. */
+ The result is nonzero if a change was made. */
static int
-cprop_insn (bb, insn, alter_jumps)
- basic_block bb;
- rtx insn;
- int alter_jumps;
+cprop_insn (rtx insn, int alter_jumps)
{
struct reg_use *reg_used;
int changed = 0;
return 0;
reg_use_count = 0;
- note_uses (&PATTERN (insn), local_cprop_find_used_regs, NULL);
-
+ note_uses (&PATTERN (insn), find_used_regs, NULL);
+
note = find_reg_equal_equiv_note (insn);
/* We may win even when propagating constants into notes. */
if (note)
- local_cprop_find_used_regs (&XEXP (note, 0), NULL);
+ find_used_regs (&XEXP (note, 0), NULL);
for (reg_used = ®_use_table[0]; reg_use_count > 0;
reg_used++, reg_use_count--)
set = find_avail_set (regno, insn);
if (! set)
continue;
-
+
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);
/* Constant propagation. */
- if (CONSTANT_P (src))
+ if (gcse_constant_p (src))
{
- /* Handle normal insns first. */
- if (GET_CODE (insn) == INSN
- && try_replace_reg (reg_used->reg_rtx, src, insn))
+ if (constprop_register (insn, reg_used->reg_rtx, src, alter_jumps))
{
changed = 1;
- const_prop_count++;
- if (gcse_file != NULL)
+ global_const_prop_count++;
+ if (dump_file != NULL)
{
- fprintf (gcse_file, "CONST-PROP: Replacing reg %d in ",
- regno);
- fprintf (gcse_file, "insn %d with constant ",
- INSN_UID (insn));
- print_rtl (gcse_file, src);
- fprintf (gcse_file, "\n");
+ fprintf (dump_file, "GLOBAL CONST-PROP: Replacing reg %d in ", regno);
+ fprintf (dump_file, "insn %d with constant ", INSN_UID (insn));
+ print_rtl (dump_file, src);
+ fprintf (dump_file, "\n");
}
-
- /* The original insn setting reg_used may or may not now be
- deletable. We leave the deletion to flow. */
- }
-
- /* Try to propagate a CONST_INT into a conditional jump.
- We're pretty specific about what we will handle in this
- code, we can extend this as necessary over time.
-
- Right now the insn in question must look like
- (set (pc) (if_then_else ...)) */
- else if (alter_jumps
- && GET_CODE (insn) == JUMP_INSN
- && condjump_p (insn)
- && ! simplejump_p (insn))
- changed |= cprop_jump (bb, insn, reg_used->reg_rtx, src);
-
-#ifdef HAVE_cc0
- /* Similar code for machines that use a pair of CC0 setter and
- conditional jump insn. */
- else if (alter_jumps
- && GET_CODE (PATTERN (insn)) == SET
- && SET_DEST (PATTERN (insn)) == cc0_rtx
- && GET_CODE (NEXT_INSN (insn)) == JUMP_INSN
- && condjump_p (NEXT_INSN (insn))
- && ! simplejump_p (NEXT_INSN (insn))
- && cprop_cc0_jump (bb, insn, reg_used, src))
- {
- changed = 1;
- break;
+ if (INSN_DELETED_P (insn))
+ return 1;
}
-#endif
}
- 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++;
- if (gcse_file != NULL)
+ global_copy_prop_count++;
+ if (dump_file != NULL)
{
- fprintf (gcse_file, "COPY-PROP: Replacing reg %d in insn %d",
+ fprintf (dump_file, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
regno, INSN_UID (insn));
- fprintf (gcse_file, " with reg %d\n", REGNO (src));
+ fprintf (dump_file, " with reg %d\n", REGNO (src));
}
/* The original insn setting reg_used may or may not now be
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;
case PRE_MODIFY:
case POST_MODIFY:
/* Can only legitimately appear this early in the context of
- stack pushes for function arguments, but handle all of the
- codes nonetheless. */
+ stack pushes for function arguments, but handle all of the
+ codes nonetheless. */
return;
- case SUBREG:
- /* Setting a subreg of a register larger than word_mode leaves
- the non-written words unchanged. */
- if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) > BITS_PER_WORD)
- return;
- break;
+ case SUBREG:
+ /* Setting a subreg of a register larger than word_mode leaves
+ the non-written words unchanged. */
+ if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) > BITS_PER_WORD)
+ return;
+ break;
+
+ default:
+ break;
+ }
+
+ find_used_regs (xptr, data);
+}
+
+/* Try to perform local const/copy propagation on X in INSN.
+ If ALTER_JUMPS is false, changing jump insns is not allowed. */
+
+static bool
+do_local_cprop (rtx x, rtx insn, bool alter_jumps)
+{
+ rtx newreg = NULL, newcnst = NULL;
+
+ /* Rule out USE instructions and ASM statements as we don't want to
+ change the hard registers mentioned. */
+ if (REG_P (x)
+ && (REGNO (x) >= FIRST_PSEUDO_REGISTER
+ || (GET_CODE (PATTERN (insn)) != USE
+ && asm_noperands (PATTERN (insn)) < 0)))
+ {
+ cselib_val *val = cselib_lookup (x, GET_MODE (x), 0);
+ struct elt_loc_list *l;
+
+ if (!val)
+ return false;
+ for (l = val->locs; l; l = l->next)
+ {
+ rtx this_rtx = l->loc;
+ rtx note;
+
+ if (gcse_constant_p (this_rtx))
+ newcnst = this_rtx;
+ if (REG_P (this_rtx) && REGNO (this_rtx) >= FIRST_PSEUDO_REGISTER
+ /* Don't copy propagate 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 (l->setting_insn, REG_EQUIV, NULL_RTX))
+ || ! MEM_P (XEXP (note, 0))))
+ newreg = this_rtx;
+ }
+ if (newcnst && constprop_register (insn, x, newcnst, alter_jumps))
+ {
+ if (dump_file != NULL)
+ {
+ fprintf (dump_file, "LOCAL CONST-PROP: Replacing reg %d in ",
+ REGNO (x));
+ fprintf (dump_file, "insn %d with constant ",
+ INSN_UID (insn));
+ print_rtl (dump_file, newcnst);
+ fprintf (dump_file, "\n");
+ }
+ local_const_prop_count++;
+ return true;
+ }
+ else if (newreg && newreg != x && try_replace_reg (x, newreg, insn))
+ {
+ if (dump_file != NULL)
+ {
+ fprintf (dump_file,
+ "LOCAL COPY-PROP: Replacing reg %d in insn %d",
+ REGNO (x), INSN_UID (insn));
+ fprintf (dump_file, " with reg %d\n", REGNO (newreg));
+ }
+ local_copy_prop_count++;
+ return true;
+ }
+ }
+ return false;
+}
+
+/* Do local const/copy propagation (i.e. within each basic block).
+ If ALTER_JUMPS is true, allow propagating into jump insns, which
+ could modify the CFG. */
+
+static void
+local_cprop_pass (bool alter_jumps)
+{
+ basic_block bb;
+ rtx insn;
+ struct reg_use *reg_used;
+ bool changed = false;
+
+ cselib_init (false);
+ FOR_EACH_BB (bb)
+ {
+ FOR_BB_INSNS (bb, insn)
+ {
+ if (INSN_P (insn))
+ {
+ rtx note = find_reg_equal_equiv_note (insn);
+ do
+ {
+ reg_use_count = 0;
+ note_uses (&PATTERN (insn), local_cprop_find_used_regs,
+ NULL);
+ if (note)
+ local_cprop_find_used_regs (&XEXP (note, 0), NULL);
+
+ for (reg_used = ®_use_table[0]; reg_use_count > 0;
+ reg_used++, reg_use_count--)
+ {
+ if (do_local_cprop (reg_used->reg_rtx, insn, alter_jumps))
+ {
+ changed = true;
+ break;
+ }
+ }
+ if (INSN_DELETED_P (insn))
+ break;
+ }
+ while (reg_use_count);
+ }
+ cselib_process_insn (insn);
+ }
+
+ /* Forget everything at the end of a basic block. */
+ cselib_clear_table ();
+ }
+
+ cselib_finish ();
+
+ /* Global analysis may get into infinite loops for unreachable blocks. */
+ if (changed && alter_jumps)
+ {
+ delete_unreachable_blocks ();
+ free_reg_set_mem ();
+ alloc_reg_set_mem (max_reg_num ());
+ compute_sets ();
+ }
+}
+
+/* Forward propagate copies. This includes copies and constants. Return
+ nonzero if a change was made. */
+
+static int
+cprop (int alter_jumps)
+{
+ int changed;
+ basic_block bb;
+ rtx insn;
+
+ /* Note we start at block 1. */
+ if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
+ {
+ if (dump_file != NULL)
+ fprintf (dump_file, "\n");
+ 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_BB_INSNS (bb, insn)
+ if (INSN_P (insn))
+ {
+ changed |= cprop_insn (insn, alter_jumps);
+
+ /* 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 (! NOTE_P (insn))
+ mark_oprs_set (insn);
+ }
+ }
+
+ if (dump_file != NULL)
+ fprintf (dump_file, "\n");
+
+ return changed;
+}
+
+/* Similar to get_condition, only the resulting condition must be
+ valid at JUMP, instead of at EARLIEST.
+
+ This differs from noce_get_condition in ifcvt.c in that we prefer not to
+ 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 scanned by canonicalize_condition,
+ but this would require some code reorganization. */
+
+rtx
+fis_get_condition (rtx jump)
+{
+ return get_condition (jump, NULL, false, true);
+}
+
+/* Check the comparison COND to see if we can safely form an implicit set from
+ it. COND is either an EQ or NE comparison. */
+
+static bool
+implicit_set_cond_p (const_rtx cond)
+{
+ const enum machine_mode mode = GET_MODE (XEXP (cond, 0));
+ const_rtx cst = XEXP (cond, 1);
+
+ /* 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 gcse_constant_p (cst);
+}
+
+/* Find the implicit sets of a function. An "implicit set" is a constraint
+ on the value of a variable, implied by a conditional jump. For example,
+ following "if (x == 2)", the then branch may be optimized as though the
+ conditional performed an "explicit set", in this example, "x = 2". This
+ function records the set patterns that are implicit at the start of each
+ basic block. */
+
+static void
+find_implicit_sets (void)
+{
+ basic_block bb, dest;
+ unsigned int count;
+ rtx cond, new_rtx;
+
+ count = 0;
+ FOR_EACH_BB (bb)
+ /* Check for more than one successor. */
+ if (EDGE_COUNT (bb->succs) > 1)
+ {
+ cond = fis_get_condition (BB_END (bb));
+
+ if (cond
+ && (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
+ && REG_P (XEXP (cond, 0))
+ && REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
+ && implicit_set_cond_p (cond))
+ {
+ dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
+ : FALLTHRU_EDGE (bb)->dest;
+
+ if (dest && single_pred_p (dest)
+ && dest != EXIT_BLOCK_PTR)
+ {
+ new_rtx = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
+ XEXP (cond, 1));
+ implicit_sets[dest->index] = new_rtx;
+ if (dump_file)
+ {
+ fprintf(dump_file, "Implicit set of reg %d in ",
+ REGNO (XEXP (cond, 0)));
+ fprintf(dump_file, "basic block %d\n", dest->index);
+ }
+ count++;
+ }
+ }
+ }
+
+ if (dump_file)
+ fprintf (dump_file, "Found %d implicit sets\n", count);
+}
+
+/* Perform one copy/constant propagation pass.
+ PASS is the pass count. If CPROP_JUMPS is true, perform constant
+ propagation into conditional jumps. If BYPASS_JUMPS is true,
+ perform conditional jump bypassing optimizations. */
+
+static int
+one_cprop_pass (int pass, bool cprop_jumps, bool bypass_jumps)
+{
+ int changed = 0;
+
+ global_const_prop_count = local_const_prop_count = 0;
+ global_copy_prop_count = local_copy_prop_count = 0;
+
+ if (cprop_jumps)
+ local_cprop_pass (cprop_jumps);
+
+ /* Determine implicit sets. */
+ implicit_sets = XCNEWVEC (rtx, last_basic_block);
+ find_implicit_sets ();
+
+ alloc_hash_table (max_cuid, &set_hash_table, 1);
+ compute_hash_table (&set_hash_table);
+
+ /* Free implicit_sets before peak usage. */
+ free (implicit_sets);
+ implicit_sets = NULL;
+
+ if (dump_file)
+ dump_hash_table (dump_file, "SET", &set_hash_table);
+ if (set_hash_table.n_elems > 0)
+ {
+ alloc_cprop_mem (last_basic_block, set_hash_table.n_elems);
+ compute_cprop_data ();
+ changed = cprop (cprop_jumps);
+ if (bypass_jumps)
+ changed |= bypass_conditional_jumps ();
+ free_cprop_mem ();
+ }
+
+ free_hash_table (&set_hash_table);
+
+ if (dump_file)
+ {
+ fprintf (dump_file, "CPROP of %s, pass %d: %d bytes needed, ",
+ current_function_name (), pass, bytes_used);
+ fprintf (dump_file, "%d local const props, %d local copy props, ",
+ local_const_prop_count, local_copy_prop_count);
+ fprintf (dump_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)
+ delete_unreachable_blocks ();
+
+ return changed;
+}
+\f
+/* Bypass conditional jumps. */
+
+/* The value of last_basic_block at the beginning of the jump_bypass
+ pass. The use of redirect_edge_and_branch_force may introduce new
+ basic blocks, but the data flow analysis is only valid for basic
+ block indices less than bypass_last_basic_block. */
+
+static int bypass_last_basic_block;
+
+/* Find a set of REGNO to a constant that is available at the end of basic
+ block BB. Returns NULL if no such set is found. Based heavily upon
+ find_avail_set. */
+
+static struct expr *
+find_bypass_set (int regno, int bb)
+{
+ struct expr *result = 0;
+
+ for (;;)
+ {
+ rtx src;
+ struct expr *set = lookup_set (regno, &set_hash_table);
+
+ while (set)
+ {
+ if (TEST_BIT (cprop_avout[bb], set->bitmap_index))
+ break;
+ set = next_set (regno, set);
+ }
+
+ if (set == 0)
+ break;
+
+ gcc_assert (GET_CODE (set->expr) == SET);
+
+ src = SET_SRC (set->expr);
+ if (gcse_constant_p (src))
+ result = set;
+
+ if (! REG_P (src))
+ break;
+
+ regno = REGNO (src);
+ }
+ return result;
+}
+
+
+/* Subroutine of bypass_block that checks whether a pseudo is killed by
+ any of the instructions inserted on an edge. Jump bypassing places
+ condition code setters on CFG edges using insert_insn_on_edge. This
+ function is required to check that our data flow analysis is still
+ valid prior to commit_edge_insertions. */
+
+static bool
+reg_killed_on_edge (const_rtx reg, const_edge e)
+{
+ rtx insn;
+
+ for (insn = e->insns.r; insn; insn = NEXT_INSN (insn))
+ if (INSN_P (insn) && reg_set_p (reg, insn))
+ return true;
+
+ return false;
+}
+
+/* Subroutine of bypass_conditional_jumps that attempts to bypass the given
+ basic block BB which has more than one predecessor. If not NULL, SETCC
+ is the first instruction of BB, which is immediately followed by JUMP_INSN
+ JUMP. Otherwise, SETCC is NULL, and JUMP is the first insn of BB.
+ Returns nonzero if a change was made.
+
+ During the jump bypassing pass, we may place copies of SETCC instructions
+ on CFG edges. The following routine must be careful to pay attention to
+ these inserted insns when performing its transformations. */
+
+static int
+bypass_block (basic_block bb, rtx setcc, rtx jump)
+{
+ rtx insn, note;
+ edge e, edest;
+ int i, change;
+ int may_be_loop_header;
+ unsigned removed_p;
+ edge_iterator ei;
+
+ insn = (setcc != NULL) ? setcc : jump;
+
+ /* Determine set of register uses in INSN. */
+ reg_use_count = 0;
+ note_uses (&PATTERN (insn), find_used_regs, NULL);
+ note = find_reg_equal_equiv_note (insn);
+ if (note)
+ find_used_regs (&XEXP (note, 0), NULL);
+
+ may_be_loop_header = false;
+ FOR_EACH_EDGE (e, ei, bb->preds)
+ if (e->flags & EDGE_DFS_BACK)
+ {
+ may_be_loop_header = true;
+ break;
+ }
+
+ change = 0;
+ for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
+ {
+ removed_p = 0;
+
+ if (e->flags & EDGE_COMPLEX)
+ {
+ ei_next (&ei);
+ continue;
+ }
+
+ /* We can't redirect edges from new basic blocks. */
+ if (e->src->index >= bypass_last_basic_block)
+ {
+ ei_next (&ei);
+ continue;
+ }
- default:
- break;
- }
+ /* 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))
+ {
+ ei_next (&ei);
+ continue;
+ }
- find_used_regs (xptr, data);
-}
+ for (i = 0; i < reg_use_count; i++)
+ {
+ struct reg_use *reg_used = ®_use_table[i];
+ unsigned int regno = REGNO (reg_used->reg_rtx);
+ basic_block dest, old_dest;
+ struct expr *set;
+ rtx src, new_rtx;
+ if (regno >= max_gcse_regno)
+ continue;
-/* Forward propagate copies. This includes copies and constants. Return
- non-zero if a change was made. */
+ set = find_bypass_set (regno, e->src->index);
-static int
-cprop (alter_jumps)
- int alter_jumps;
-{
- int bb, changed;
- rtx insn;
+ if (! set)
+ continue;
- /* Note we start at block 1. */
+ /* Check the data flow is valid after edge insertions. */
+ if (e->insns.r && reg_killed_on_edge (reg_used->reg_rtx, e))
+ continue;
- changed = 0;
- for (bb = 1; bb < n_basic_blocks; bb++)
- {
- /* Reset tables used to keep track of what's still valid [since the
- start of the block]. */
- reset_opr_set_tables ();
+ src = SET_SRC (pc_set (jump));
- for (insn = BLOCK_HEAD (bb);
- insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
- insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- {
- changed |= cprop_insn (BASIC_BLOCK (bb), insn, alter_jumps);
+ if (setcc != NULL)
+ src = simplify_replace_rtx (src,
+ SET_DEST (PATTERN (setcc)),
+ SET_SRC (PATTERN (setcc)));
- /* 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)
- mark_oprs_set (insn);
- }
- }
+ new_rtx = simplify_replace_rtx (src, reg_used->reg_rtx,
+ SET_SRC (set->expr));
- if (gcse_file != NULL)
- fprintf (gcse_file, "\n");
+ /* 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. */
- return changed;
+ if (new_rtx == pc_rtx)
+ {
+ edest = FALLTHRU_EDGE (bb);
+ dest = edest->insns.r ? NULL : edest->dest;
+ }
+ else if (GET_CODE (new_rtx) == LABEL_REF)
+ {
+ dest = BLOCK_FOR_INSN (XEXP (new_rtx, 0));
+ /* Don't bypass edges containing instructions. */
+ 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
+ && dest != EXIT_BLOCK_PTR)
+ {
+ redirect_edge_and_branch_force (e, dest);
+
+ /* Copy the register setter to the redirected edge.
+ Don't copy CC0 setters, as CC0 is dead after jump. */
+ if (setcc)
+ {
+ rtx pat = PATTERN (setcc);
+ if (!CC0_P (SET_DEST (pat)))
+ insert_insn_on_edge (copy_insn (pat), e);
+ }
+
+ if (dump_file != NULL)
+ {
+ fprintf (dump_file, "JUMP-BYPASS: Proved reg %d "
+ "in jump_insn %d equals constant ",
+ regno, INSN_UID (jump));
+ print_rtl (dump_file, SET_SRC (set->expr));
+ fprintf (dump_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;
}
-/* Perform one copy/constant propagation pass.
- F is the first insn in the function.
- PASS is the pass count. */
+/* Find basic blocks with more than one predecessor that only contain a
+ single conditional jump. If the result of the comparison is known at
+ compile-time from any incoming edge, redirect that edge to the
+ appropriate target. Returns nonzero if a change was made.
+
+ This function is now mis-named, because we also handle indirect jumps. */
static int
-one_cprop_pass (pass, alter_jumps)
- int pass;
- int alter_jumps;
+bypass_conditional_jumps (void)
{
- int changed = 0;
-
- const_prop_count = 0;
- copy_prop_count = 0;
+ basic_block bb;
+ int changed;
+ rtx setcc;
+ rtx insn;
+ rtx dest;
- alloc_set_hash_table (max_cuid);
- compute_set_hash_table ();
- if (gcse_file)
- dump_hash_table (gcse_file, "SET", set_hash_table, set_hash_table_size,
- n_sets);
- if (n_sets > 0)
- {
- alloc_cprop_mem (n_basic_blocks, n_sets);
- compute_cprop_data ();
- changed = cprop (alter_jumps);
- free_cprop_mem ();
- }
+ /* Note we start at block 1. */
+ if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
+ return 0;
- free_set_hash_table ();
+ bypass_last_basic_block = last_basic_block;
+ mark_dfs_back_edges ();
- if (gcse_file)
+ changed = 0;
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb,
+ EXIT_BLOCK_PTR, next_bb)
{
- 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);
+ /* Check for more than one predecessor. */
+ if (!single_pred_p (bb))
+ {
+ setcc = NULL_RTX;
+ FOR_BB_INSNS (bb, insn)
+ if (NONJUMP_INSN_P (insn))
+ {
+ if (setcc)
+ break;
+ if (GET_CODE (PATTERN (insn)) != SET)
+ break;
+
+ dest = SET_DEST (PATTERN (insn));
+ if (REG_P (dest) || CC0_P (dest))
+ setcc = insn;
+ else
+ break;
+ }
+ else if (JUMP_P (insn))
+ {
+ if ((any_condjump_p (insn) || computed_jump_p (insn))
+ && onlyjump_p (insn))
+ changed |= bypass_block (bb, setcc, insn);
+ break;
+ }
+ else if (INSN_P (insn))
+ break;
+ }
}
+ /* If we bypassed any register setting insns, we inserted a
+ copy on the redirected edge. These need to be committed. */
+ if (changed)
+ commit_edge_insertions ();
+
return changed;
}
\f
/* 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;
- int i;
+ basic_block bb;
unsigned int ui;
- compute_local_properties (transp, comp, antloc, 0);
- sbitmap_vector_zero (ae_kill, n_basic_blocks);
+ compute_local_properties (transp, comp, antloc, &expr_hash_table);
+ sbitmap_vector_zero (ae_kill, last_basic_block);
/* Collect expressions which might trap. */
- trapping_expr = sbitmap_alloc (n_exprs);
+ trapping_expr = sbitmap_alloc (expr_hash_table.n_elems);
sbitmap_zero (trapping_expr);
- for (ui = 0; ui < expr_hash_table_size; ui++)
+ for (ui = 0; ui < expr_hash_table.size; ui++)
{
struct expr *e;
- for (e = expr_hash_table[ui]; e != NULL; e = e->next_same_hash)
+ for (e = expr_hash_table.table[ui]; e != NULL; e = e->next_same_hash)
if (may_trap_p (e->expr))
SET_BIT (trapping_expr, e->bitmap_index);
}
/* Compute ae_kill for each basic block using:
~(TRANSP | COMP)
+ */
- This is significantly faster than compute_ae_kill. */
-
- for (i = 0; i < n_basic_blocks; i++)
+ 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 = BASIC_BLOCK (i)->pred; e ; e = e->pred_next)
+ FOR_EACH_EDGE (e, ei, bb->preds)
if (e->flags & EDGE_ABNORMAL)
{
- sbitmap_difference (antloc[i], antloc[i], trapping_expr);
- sbitmap_difference (transp[i], transp[i], trapping_expr);
+ sbitmap_difference (antloc[bb->index], antloc[bb->index], trapping_expr);
+ sbitmap_difference (transp[bb->index], transp[bb->index], trapping_expr);
break;
}
- sbitmap_a_or_b (ae_kill[i], transp[i], comp[i]);
- sbitmap_not (ae_kill[i], ae_kill[i]);
+ sbitmap_a_or_b (ae_kill[bb->index], transp[bb->index], comp[bb->index]);
+ sbitmap_not (ae_kill[bb->index], ae_kill[bb->index]);
}
- edge_list = pre_edge_lcm (gcse_file, n_exprs, transp, comp, antloc,
+ edge_list = pre_edge_lcm (expr_hash_table.n_elems, transp, comp, antloc,
ae_kill, &pre_insert_map, &pre_delete_map);
sbitmap_vector_free (antloc);
antloc = NULL;
sbitmap_vector_free (ae_kill);
- ae_kill = NULL;
+ ae_kill = NULL;
sbitmap_free (trapping_expr);
}
\f
/* PRE utilities */
-/* Return non-zero if an occurrence of expression EXPR in OCCR_BB would reach
+/* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
block BB.
VISITED is a pointer to a working buffer for tracking which BB's have
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 (n_basic_blocks, 1);
+ char *visited = XCNEWVEC (char, last_basic_block);
rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
\f
/* Given an expr, generate RTL which we can insert at the end of a BB,
- or on an edge. Set the block number of any insns generated to
+ or on an edge. Set the block number of any insns generated to
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 = gen_sequence ();
+
+ pat = get_insns ();
end_sequence ();
return pat;
}
-
+
/* Add EXPR to the end of basic block BB.
This is used by both the PRE and code hoisting.
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_basic_block (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;
- int i;
+ rtx pat, pat_end;
pat = process_insert_insn (expr);
+ 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, bb);
}
/* 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
parameter is loaded. Do this for everyone for consistency and a
- presumtion that we'll get better code elsewhere as well.
+ presumption that we'll get better code elsewhere as well.
It should always be the case that we can put these instructions
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, bb);
}
else
- new_insn = emit_insn_after (pat, insn);
+ new_insn = emit_insn_after_noloc (pat, insn, bb);
- /* Keep block number table up to date.
- Note, PAT could be a multiple insn sequence, we have to make
- sure that each insn in the sequence is handled. */
- if (GET_CODE (pat) == SEQUENCE)
+ while (1)
{
- for (i = 0; i < XVECLEN (pat, 0); i++)
+ if (INSN_P (pat))
{
- rtx insn = XVECEXP (pat, 0, i);
- if (INSN_P (insn))
- add_label_notes (PATTERN (insn), new_insn);
-
- note_stores (PATTERN (insn), record_set_info, insn);
+ add_label_notes (PATTERN (pat), new_insn);
+ note_stores (PATTERN (pat), record_set_info, pat);
}
- }
- else
- {
- add_label_notes (pat, new_insn);
-
- /* Keep register set table up to date. */
- record_one_set (regno, new_insn);
+ if (pat == pat_end)
+ break;
+ pat = NEXT_INSN (pat);
}
gcse_create_count++;
- if (gcse_file)
+ if (dump_file)
{
- fprintf (gcse_file, "PRE/HOIST: end of bb %d, insn %d, ",
+ fprintf (dump_file, "PRE/HOIST: end of bb %d, insn %d, ",
bb->index, INSN_UID (new_insn));
- fprintf (gcse_file, "copying expression %d to reg %d\n",
+ fprintf (dump_file, "copying expression %d to reg %d\n",
expr->bitmap_index, regno);
}
}
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;
set_size = pre_insert_map[0]->size;
num_edges = NUM_EDGES (edge_list);
- inserted = sbitmap_vector_alloc (num_edges, n_exprs);
+ inserted = sbitmap_vector_alloc (num_edges, expr_hash_table.n_elems);
sbitmap_vector_zero (inserted, num_edges);
for (e = 0; e < num_edges; e++)
{
SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
- for (j = indx; insert && j < n_exprs; j++, insert >>= 1)
+ for (j = indx; insert && j < (int) expr_hash_table.n_elems; j++, insert >>= 1)
if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
{
struct expr *expr = index_map[j];
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)
- insert_insn_end_bb (index_map[j], bb, 0);
+ if (eg->flags & EDGE_ABNORMAL)
+ insert_insn_end_basic_block (index_map[j], bb, 0);
else
{
insn = process_insert_insn (index_map[j]);
insert_insn_on_edge (insn, eg);
}
- if (gcse_file)
+ if (dump_file)
{
- fprintf (gcse_file, "PRE/HOIST: edge (%d,%d), ",
+ fprintf (dump_file, "PRE/HOIST: edge (%d,%d), ",
bb->index,
INDEX_EDGE_SUCC_BB (edge_list, e)->index);
- fprintf (gcse_file, "copy expression %d\n",
+ fprintf (dump_file, "copy expression %d\n",
expr->bitmap_index);
}
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, first_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. */
+ first_set = NULL_RTX;
+ set = NULL_RTX;
+ for (i = 0; i < XVECLEN (pat, 0); i++)
+ {
+ rtx x = XVECEXP (pat, 0, i);
+ if (GET_CODE (x) == SET)
+ {
+ /* If the source was a REG_EQUAL or REG_EQUIV note, we
+ may not find an equivalent expression, but in this
+ case the PARALLEL will have a single set. */
+ if (first_set == NULL_RTX)
+ first_set = x;
+ if (expr_equiv_p (SET_SRC (x), expr->expr))
+ {
+ set = x;
+ break;
+ }
+ }
+ }
+
+ gcc_assert (first_set);
+ if (set == NULL_RTX)
+ set = first_set;
+ 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, 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++;
- if (gcse_file)
- fprintf (gcse_file,
+ if (dump_file)
+ fprintf (dump_file,
"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;
??? The current algorithm is rather brute force.
Need to do some profiling. */
- for (i = 0; i < expr_hash_table_size; i++)
- for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
+ for (i = 0; i < expr_hash_table.size; i++)
+ for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
{
/* If the basic block isn't reachable, PPOUT will be TRUE. However,
we don't want to insert a copy here because the expression may not
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)
continue;
/* Or if the expression doesn't reach the deleted one. */
- if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail->insn),
+ if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail->insn),
expr,
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 (rtx src, rtx dest, rtx insn)
+{
+ rtx new_rtx;
+ rtx set = single_set (insn), set2;
+ rtx note;
+ rtx eqv;
+
+ /* This should never fail since we're creating a reg->reg copy
+ we've verified to be valid. */
+
+ new_rtx = emit_insn_after (gen_move_insn (dest, src), insn);
+
+ /* Note the equivalence for local CSE pass. */
+ set2 = single_set (new_rtx);
+ if (!set2 || !rtx_equal_p (SET_DEST (set2), dest))
+ return new_rtx;
+ if ((note = find_reg_equal_equiv_note (insn)))
+ eqv = XEXP (note, 0);
+ else
+ eqv = SET_SRC (set);
+
+ set_unique_reg_note (new_rtx, REG_EQUAL, copy_insn_1 (eqv));
+
+ return new_rtx;
+}
+
/* Delete redundant computations.
Deletion is done by changing the insn to copy the `reaching_reg' of
the expression into the result of the SET. It is left to later passes
(cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
- Returns non-zero if a change is made. */
+ Returns nonzero if a change is made. */
static int
-pre_delete ()
+pre_delete (void)
{
unsigned int i;
int changed;
struct occr *occr;
changed = 0;
- for (i = 0; i < expr_hash_table_size; i++)
- for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
+ for (i = 0; i < expr_hash_table.size; i++)
+ 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
+ && dbg_cnt (pre_insn))
{
- 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. */
if (expr->reaching_reg == NULL)
- expr->reaching_reg
- = gen_reg_rtx (GET_MODE (SET_DEST (set)));
-
- /* In theory this should never fail since we're creating
- a reg->reg copy.
-
- However, on the x86 some of the movXX patterns actually
- contain clobbers of scratch regs. This may cause the
- insn created by validate_change to not match any pattern
- and thus cause validate_change to fail. */
- if (validate_change (insn, &SET_SRC (set),
- expr->reaching_reg, 0))
- {
- occr->deleted_p = 1;
- SET_BIT (pre_redundant_insns, INSN_CUID (insn));
- changed = 1;
- gcse_subst_count++;
- }
+ expr->reaching_reg = gen_reg_rtx_and_attrs (SET_DEST (set));
- if (gcse_file)
+ gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
+ delete_insn (insn);
+ occr->deleted_p = 1;
+ SET_BIT (pre_redundant_insns, INSN_CUID (insn));
+ changed = 1;
+ gcse_subst_count++;
+
+ if (dump_file)
{
- fprintf (gcse_file,
+ fprintf (dump_file,
"PRE: redundant insn %d (expression %d) in ",
INSN_UID (insn), indx);
- fprintf (gcse_file, "bb %d, reaching reg is %d\n",
+ fprintf (dump_file, "bb %d, reaching reg is %d\n",
bb->index, REGNO (expr->reaching_reg));
}
}
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 (n_exprs, sizeof (struct expr *));
- for (i = 0; i < expr_hash_table_size; i++)
- for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
+ index_map = XCNEWVEC (struct expr *, expr_hash_table.n_elems);
+ 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;
/* Reset bitmap used to track which insns are redundant. */
- we know which insns are redundant when we go to create copies */
changed = pre_delete ();
-
did_insert = pre_edge_insert (edge_list, index_map);
/* In other places with reaching expressions, copy the expression to the
/* Top level routine to perform one PRE GCSE pass.
- Return non-zero if a change was made. */
+ 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;
gcse_subst_count = 0;
gcse_create_count = 0;
- alloc_expr_hash_table (max_cuid);
+ alloc_hash_table (max_cuid, &expr_hash_table, 0);
add_noreturn_fake_exit_edges ();
if (flag_gcse_lm)
compute_ld_motion_mems ();
- compute_expr_hash_table ();
+ compute_hash_table (&expr_hash_table);
trim_ld_motion_mems ();
- if (gcse_file)
- dump_hash_table (gcse_file, "Expression", expr_hash_table,
- expr_hash_table_size, n_exprs);
+ if (dump_file)
+ dump_hash_table (dump_file, "Expression", &expr_hash_table);
- if (n_exprs > 0)
+ if (expr_hash_table.n_elems > 0)
{
- alloc_pre_mem (n_basic_blocks, n_exprs);
+ alloc_pre_mem (last_basic_block, expr_hash_table.n_elems);
compute_pre_data ();
changed |= pre_gcse ();
free_edge_list (edge_list);
}
free_ldst_mems ();
- remove_fake_edges ();
- free_expr_hash_table ();
+ remove_fake_exit_edges ();
+ free_hash_table (&expr_hash_table);
- if (gcse_file)
+ if (dump_file)
{
- fprintf (gcse_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
- current_function_name, pass, bytes_used);
- fprintf (gcse_file, "%d substs, %d insns created\n",
+ fprintf (dump_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
+ current_function_name (), pass, bytes_used);
+ fprintf (dump_file, "%d substs, %d insns created\n",
gcse_subst_count, gcse_create_count);
}
return changed;
}
\f
-/* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
- If notes are added to an insn which references a CODE_LABEL, the
- LABEL_NUSES count is incremented. We have to add REG_LABEL notes,
- because the following loop optimization pass requires them. */
-
-/* ??? This is very similar to the loop.c add_label_notes function. We
- could probably share code here. */
+/* If X contains any LABEL_REF's, add REG_LABEL_OPERAND notes for them
+ to INSN. If such notes are added to an insn which references a
+ CODE_LABEL, the LABEL_NUSES count is incremented. We have to add
+ that note, because the following loop optimization pass requires
+ them. */
/* ??? If there was a jump optimization pass after gcse and before loop,
then we would not need to do this here, because jump would add the
- necessary REG_LABEL notes. */
+ necessary REG_LABEL_OPERAND and REG_LABEL_TARGET 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;
if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
{
/* This code used to ignore labels that referred to dispatch tables to
- avoid flow generating (slighly) worse code.
+ avoid flow generating (slightly) worse code.
We no longer ignore such label references (see LABEL_REF handling in
mark_jump_label for additional information). */
- REG_NOTES (insn) = gen_rtx_INSN_LIST (REG_LABEL, XEXP (x, 0),
- REG_NOTES (insn));
+ /* There's no reason for current users to emit jump-insns with
+ such a LABEL_REF, so we don't have to handle REG_LABEL_TARGET
+ notes. */
+ gcc_assert (!JUMP_P (insn));
+ add_reg_note (insn, REG_LABEL_OPERAND, XEXP (x, 0));
+
if (LABEL_P (XEXP (x, 0)))
- LABEL_NUSES (XEXP (x, 0))++;
+ LABEL_NUSES (XEXP (x, 0))++;
+
return;
}
EH table sizes, this may not be worthwhile. */
static void
-compute_transpout ()
+compute_transpout (void)
{
- int bb;
+ basic_block bb;
unsigned int i;
struct expr *expr;
- sbitmap_vector_ones (transpout, n_basic_blocks);
+ sbitmap_vector_ones (transpout, last_basic_block);
- for (bb = 0; bb < n_basic_blocks; ++bb)
+ FOR_EACH_BB (bb)
{
- /* Note that flow inserted a nop a the end of basic blocks that
+ /* Note that flow inserted a nop at the end of basic blocks that
end in call instructions for reasons other than abnormal
control flow. */
- if (GET_CODE (BLOCK_END (bb)) != CALL_INSN)
- continue;
-
- for (i = 0; i < expr_hash_table_size; i++)
- for (expr = expr_hash_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], 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], regno);
- SET_BIT (npi->nonnull_killed[npi->current_block], 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 void
-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;
-{
- int bb;
- int current_block;
- sbitmap *nonnull_local = npi->nonnull_local;
- sbitmap *nonnull_killed = npi->nonnull_killed;
-
- /* 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, n_basic_blocks);
- sbitmap_vector_zero (nonnull_killed, n_basic_blocks);
-
- for (current_block = 0; current_block < n_basic_blocks; 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 (BLOCK_END (current_block));
- for (insn = BLOCK_HEAD (current_block);
- 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],
- 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],
- REGNO (reg) - npi->min_reg);
- }
- }
-
- /* Now compute global properties based on the local properties. This
- is a classic global availablity 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 (bb = 0; bb < n_basic_blocks; bb++)
- {
- rtx last_insn = BLOCK_END (bb);
- 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] < npi->min_reg
- || block_reg[bb] >= 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], block_reg[bb] - 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_before (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);
- }
-
- delete_insn (last_insn);
- if (compare_and_branch == 2)
- delete_insn (earliest);
- purge_dead_edges (BASIC_BLOCK (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] = 0;
- }
-}
-
-/* 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 properies 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. */
-
-void
-delete_null_pointer_checks (f)
- rtx f ATTRIBUTE_UNUSED;
-{
- sbitmap *nonnull_avin, *nonnull_avout;
- unsigned int *block_reg;
- int bb;
- int reg;
- int regs_per_pass;
- int max_reg;
- struct null_pointer_info npi;
-
- /* If we have only a single block, then there's nothing to do. */
- if (n_basic_blocks <= 1)
- return;
-
- /* 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)
- return;
-
- /* 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, n_basic_blocks, max_reg);
-
- /* Allocate bitmaps to hold local and global properties. */
- npi.nonnull_local = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
- npi.nonnull_killed = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
- nonnull_avin = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
- nonnull_avout = sbitmap_vector_alloc (n_basic_blocks, 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 (n_basic_blocks, sizeof (int));
- for (bb = 0; bb < n_basic_blocks; bb++)
- {
- rtx last_insn = BLOCK_END (bb);
- 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)
+ if (! CALL_P (BB_END (bb)))
continue;
- block_reg[bb] = 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);
- 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);
}
/* Code Hoisting variables and subroutines. */
/* Hoistable expressions. */
static sbitmap *hoist_exprs;
-/* Dominator bitmaps. */
-static sbitmap *dominators;
-
/* ??? We could compute post dominators and run this algorithm in
- reverse to to perform tail merging, doing so would probably be
+ reverse to perform tail merging, doing so would probably be
more effective than the tail merging code in jump.c.
It's unclear if tail merging could be run in parallel with
/* 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);
hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
-
- dominators = sbitmap_vector_alloc (n_blocks, n_blocks);
}
/* 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);
- sbitmap_vector_free (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 bb, changed, passes;
+ int changed, passes;
+ basic_block bb;
- sbitmap_vector_zero (hoist_vbeout, n_basic_blocks);
- sbitmap_vector_zero (hoist_vbein, n_basic_blocks);
+ sbitmap_vector_zero (hoist_vbeout, last_basic_block);
+ sbitmap_vector_zero (hoist_vbein, last_basic_block);
passes = 0;
changed = 1;
/* We scan the blocks in the reverse order to speed up
the convergence. */
- for (bb = n_basic_blocks - 1; bb >= 0; bb--)
+ FOR_EACH_BB_REVERSE (bb)
{
- changed |= sbitmap_a_or_b_and_c (hoist_vbein[bb], antloc[bb],
- hoist_vbeout[bb], transp[bb]);
- if (bb != n_basic_blocks - 1)
- sbitmap_intersection_of_succs (hoist_vbeout[bb], hoist_vbein, bb);
+ if (bb->next_bb != EXIT_BLOCK_PTR)
+ sbitmap_intersection_of_succs (hoist_vbeout[bb->index],
+ hoist_vbein, bb->index);
+
+ changed |= sbitmap_a_or_b_and_c_cg (hoist_vbein[bb->index],
+ antloc[bb->index],
+ hoist_vbeout[bb->index],
+ transp[bb->index]);
}
passes++;
}
- if (gcse_file)
- fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
+ if (dump_file)
+ fprintf (dump_file, "hoisting vbeinout computation: %d passes\n", passes);
}
/* 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, 0);
+ compute_local_properties (transp, comp, antloc, &expr_hash_table);
compute_transpout ();
compute_code_hoist_vbeinout ();
- calculate_dominance_info (NULL, dominators, CDI_DOMINATORS);
- if (gcse_file)
- fprintf (gcse_file, "\n");
+ calculate_dominance_info (CDI_DOMINATORS);
+ if (dump_file)
+ fprintf (dump_file, "\n");
}
/* Determine if the expression identified by EXPR_INDEX would
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;
-
+
if (visited == NULL)
{
visited_allocated_locally = 1;
- visited = xcalloc (n_basic_blocks, 1);
+ visited = XCNEWVEC (char, last_basic_block);
}
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
+ FOR_EACH_EDGE (pred, ei, bb->preds)
{
basic_block pred_bb = pred->src;
if (pred->src == ENTRY_BLOCK_PTR)
break;
+ else if (pred_bb == expr_bb)
+ continue;
else if (visited[pred_bb->index])
continue;
break;
}
}
- if (visited_allocated_locally)
+ if (visited_allocated_locally)
free (visited);
return (pred == NULL);
/* Actually perform code hoisting. */
static void
-hoist_code ()
+hoist_code (void)
{
- int bb, dominated;
- unsigned int i;
+ basic_block bb, dominated;
+ VEC (basic_block, heap) *domby;
+ unsigned int i,j;
struct expr **index_map;
struct expr *expr;
- sbitmap_vector_zero (hoist_exprs, n_basic_blocks);
+ sbitmap_vector_zero (hoist_exprs, last_basic_block);
/* Compute a mapping from expression number (`bitmap_index') to
hash table entry. */
- index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
- for (i = 0; i < expr_hash_table_size; i++)
- for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
+ index_map = XCNEWVEC (struct expr *, expr_hash_table.n_elems);
+ 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;
/* Walk over each basic block looking for potentially hoistable
expressions, nothing gets hoisted from the entry block. */
- for (bb = 0; bb < n_basic_blocks; bb++)
+ FOR_EACH_BB (bb)
{
int found = 0;
int insn_inserted_p;
+ domby = get_dominated_by (CDI_DOMINATORS, bb);
/* 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]->n_bits; i++)
+ for (i = 0; i < hoist_vbeout[bb->index]->n_bits; i++)
{
int hoistable = 0;
- if (TEST_BIT (hoist_vbeout[bb], i) && TEST_BIT (transpout[bb], i))
+ if (TEST_BIT (hoist_vbeout[bb->index], i)
+ && TEST_BIT (transpout[bb->index], i))
{
/* We've found a potentially hoistable expression, now
we look at every block BB dominates to see if it
computes the expression. */
- for (dominated = 0; dominated < n_basic_blocks; dominated++)
+ for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
{
/* Ignore self dominance. */
- if (bb == dominated
- || ! TEST_BIT (dominators[dominated], bb))
+ if (bb == dominated)
continue;
-
/* We've found a dominated block, now see if it computes
the busy expression and whether or not moving that
expression to the "beginning" of that block is safe. */
- if (!TEST_BIT (antloc[dominated], i))
+ if (!TEST_BIT (antloc[dominated->index], i))
continue;
/* Note if the expression would reach the dominated block
- unimpared if it was placed at the end of BB.
+ unimpared if it was placed at the end of BB.
Keep track of how many times this expression is hoistable
from a dominated block into BB. */
- if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb), i,
- BASIC_BLOCK (dominated), NULL))
+ if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
hoistable++;
}
to avoid any possible code expansion due to register
allocation issues; however experiments have shown that
the vast majority of hoistable expressions are only movable
- from two successors, so raising this threshhold is likely
+ from two successors, so raising this threshold is likely
to nullify any benefit we get from code hoisting. */
if (hoistable > 1)
{
- SET_BIT (hoist_exprs[bb], i);
+ SET_BIT (hoist_exprs[bb->index], i);
found = 1;
}
}
}
-
/* If we found nothing to hoist, then quit now. */
if (! found)
- continue;
+ {
+ VEC_free (basic_block, heap, domby);
+ continue;
+ }
/* Loop over all the hoistable expressions. */
- for (i = 0; i < hoist_exprs[bb]->n_bits; i++)
+ for (i = 0; i < hoist_exprs[bb->index]->n_bits; i++)
{
/* We want to insert the expression into BB only once, so
note when we've inserted it. */
insn_inserted_p = 0;
/* These tests should be the same as the tests above. */
- if (TEST_BIT (hoist_vbeout[bb], i))
+ if (TEST_BIT (hoist_exprs[bb->index], i))
{
/* We've found a potentially hoistable expression, now
we look at every block BB dominates to see if it
computes the expression. */
- for (dominated = 0; dominated < n_basic_blocks; dominated++)
+ for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
{
/* Ignore self dominance. */
- if (bb == dominated
- || ! TEST_BIT (dominators[dominated], bb))
+ if (bb == dominated)
continue;
/* We've found a dominated block, now see if it computes
the busy expression and whether or not moving that
expression to the "beginning" of that block is safe. */
- if (!TEST_BIT (antloc[dominated], i))
+ if (!TEST_BIT (antloc[dominated->index], i))
continue;
/* The expression is computed in the dominated block and
dominated block. Now we have to determine if the
expression would reach the dominated block if it was
placed at the end of BB. */
- if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb), i,
- BASIC_BLOCK (dominated), NULL))
+ if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
{
struct expr *expr = index_map[i];
struct occr *occr = expr->antic_occr;
rtx set;
/* Find the right occurrence of this expression. */
- while (BLOCK_NUM (occr->insn) != dominated && occr)
+ 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
from the mode of the original destination pseudo. */
if (expr->reaching_reg == NULL)
expr->reaching_reg
- = gen_reg_rtx (GET_MODE (SET_DEST (set)));
-
- /* In theory this should never fail since we're creating
- a reg->reg copy.
-
- However, on the x86 some of the movXX patterns
- actually contain clobbers of scratch regs. This may
- cause the insn created by validate_change to not
- match any pattern and thus cause validate_change to
- fail. */
- if (validate_change (insn, &SET_SRC (set),
- expr->reaching_reg, 0))
+ = gen_reg_rtx_and_attrs (SET_DEST (set));
+
+ gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
+ delete_insn (insn);
+ occr->deleted_p = 1;
+ if (!insn_inserted_p)
{
- occr->deleted_p = 1;
- if (!insn_inserted_p)
- {
- insert_insn_end_bb (index_map[i],
- BASIC_BLOCK (bb), 0);
- insn_inserted_p = 1;
- }
+ insert_insn_end_basic_block (index_map[i], bb, 0);
+ insn_inserted_p = 1;
}
}
}
}
}
+ VEC_free (basic_block, heap, domby);
}
free (index_map);
/* Top level routine to perform one code hoisting (aka unification) pass
- Return non-zero if a change was made. */
+ Return nonzero if a change was made. */
static int
-one_code_hoisting_pass ()
+one_code_hoisting_pass (void)
{
int changed = 0;
- alloc_expr_hash_table (max_cuid);
- compute_expr_hash_table ();
- if (gcse_file)
- dump_hash_table (gcse_file, "Code Hosting Expressions", expr_hash_table,
- expr_hash_table_size, n_exprs);
+ alloc_hash_table (max_cuid, &expr_hash_table, 0);
+ compute_hash_table (&expr_hash_table);
+ if (dump_file)
+ dump_hash_table (dump_file, "Code Hosting Expressions", &expr_hash_table);
- if (n_exprs > 0)
+ if (expr_hash_table.n_elems > 0)
{
- alloc_code_hoist_mem (n_basic_blocks, n_exprs);
+ alloc_code_hoist_mem (last_basic_block, expr_hash_table.n_elems);
compute_code_hoist_data ();
hoist_code ();
free_code_hoist_mem ();
}
- free_expr_hash_table ();
+ free_hash_table (&expr_hash_table);
return changed;
}
}
'i' is both loaded and stored to in the loop. Normally, gcse cannot move
- the load out since its live around the loop, and stored at the bottom
- of the loop.
+ the load out since its live around the loop, and stored at the bottom
+ of the loop.
- The 'Load Motion' referred to and implemented in this file is
+ The 'Load Motion' referred to and implemented in this file is
an enhancement to gcse which when using edge based lcm, recognizes
this situation and allows gcse to move the load out of the loop.
load towards the exit, and we end up with no loads or stores of 'i'
in the loop. */
+static hashval_t
+pre_ldst_expr_hash (const void *p)
+{
+ int do_not_record_p = 0;
+ const struct ls_expr *const x = (const struct ls_expr *) p;
+ return hash_rtx (x->pattern, GET_MODE (x->pattern), &do_not_record_p, NULL, false);
+}
+
+static int
+pre_ldst_expr_eq (const void *p1, const void *p2)
+{
+ const struct ls_expr *const ptr1 = (const struct ls_expr *) p1,
+ *const ptr2 = (const struct ls_expr *) p2;
+ return expr_equiv_p (ptr1->pattern, ptr2->pattern);
+}
+
/* This will search the ldst list for a matching expression. If it
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;
+ void **slot;
+ struct ls_expr e;
+
+ hash = hash_rtx (x, GET_MODE (x), &do_not_record_p,
+ NULL, /*have_reg_qty=*/false);
+
+ e.pattern = x;
+ slot = htab_find_slot_with_hash (pre_ldst_table, &e, hash, INSERT);
+ if (*slot)
+ return (struct ls_expr *)*slot;
+
+ ptr = XNEW (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;
+ *slot = ptr;
- for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
- if (expr_equiv_p (ptr->pattern, x))
- break;
-
- if (!ptr)
- {
- ptr = (struct ls_expr *) xmalloc (sizeof (struct ls_expr));
-
- ptr->next = pre_ldst_mems;
- ptr->expr = NULL;
- ptr->pattern = x;
- 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;
- }
-
return ptr;
}
/* Free up an individual ldst entry. */
-static void
-free_ldst_entry (ptr)
- struct ls_expr * ptr;
+static void
+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)
+ if (pre_ldst_table)
+ htab_delete (pre_ldst_table);
+ pre_ldst_table = NULL;
+
+ while (pre_ldst_mems)
{
struct ls_expr * tmp = 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;
fprintf (file, "LDST list: \n");
- for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
+ for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
{
fprintf (file, " Pattern (%3d): ", ptr->index);
/* 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;
-
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- if (expr_equiv_p (ptr->pattern, x) && ! ptr->invalid)
- return ptr;
-
- return NULL;
+ struct ls_expr e;
+ void **slot;
+ if (!pre_ldst_table)
+ return NULL;
+ e.pattern = x;
+ slot = htab_find_slot (pre_ldst_table, &e, NO_INSERT);
+ if (!slot || ((struct ls_expr *)*slot)->invalid)
+ return NULL;
+ return (struct ls_expr *) *slot;
}
/* 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;
}
side effects. These are the types of loads we consider for the
ld_motion list, otherwise we let the usual aliasing take care of it. */
-static int
-simple_mem (x)
- rtx x;
+static int
+simple_mem (const_rtx x)
{
- if (GET_CODE (x) != MEM)
+ if (! MEM_P (x))
return 0;
-
+
if (MEM_VOLATILE_P (x))
return 0;
-
+
if (GET_MODE (x) == BLKmode)
return 0;
- if (!rtx_varies_p (XEXP (x, 0), 0))
- return 1;
-
- return 0;
+ /* If we are handling exceptions, we must be careful with memory references
+ that may trap. If we are not, the behavior is undefined, so we may just
+ continue. */
+ if (flag_non_call_exceptions && may_trap_p (x))
+ return 0;
+
+ if (side_effects_p (x))
+ return 0;
+
+ /* Do not consider function arguments passed on stack. */
+ if (reg_mentioned_p (stack_pointer_rtx, x))
+ return 0;
+
+ if (flag_float_store && FLOAT_MODE_P (GET_MODE (x)))
+ return 0;
+
+ return 1;
}
-/* Make sure there isn't a buried reference in this pattern anywhere.
- If there is, invalidate the entry for it since we're not capable
- of fixing it up just yet.. We have to be sure we know about ALL
+/* Make sure there isn't a buried reference in this pattern anywhere.
+ If there is, invalidate the entry for it since we're not capable
+ of fixing it up just yet.. We have to be sure we know about ALL
loads since the aliasing code will allow all entries in the
ld_motion list to not-alias itself. If we miss a load, we will get
- the wrong value since gcse might common it and we won't know to
+ the wrong value since gcse might common it and we won't know to
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;
/* Recursively process the insn. */
fmt = GET_RTX_FORMAT (GET_CODE (x));
-
+
for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
{
if (fmt[i] == 'e')
}
}
-/* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
- being defined as MEM loads and stores to symbols, with no
- side effects and no registers in the expression. If there are any
- uses/defs which don't match this criteria, it is invalidated and
- trimmed out later. */
+/* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
+ being defined as MEM loads and stores to symbols, with no side effects
+ and no registers in the expression. For a MEM destination, we also
+ check that the insn is still valid if we replace the destination with a
+ REG, as is done in update_ld_motion_stores. If there are any uses/defs
+ which don't match this criteria, they are invalidated and trimmed out
+ later. */
-static void
-compute_ld_motion_mems ()
+static void
+compute_ld_motion_mems (void)
{
struct ls_expr * ptr;
- int bb;
+ basic_block bb;
rtx insn;
-
+
pre_ldst_mems = NULL;
+ pre_ldst_table = htab_create (13, pre_ldst_expr_hash,
+ pre_ldst_expr_eq, NULL);
- for (bb = 0; bb < n_basic_blocks; bb++)
+ FOR_EACH_BB (bb)
{
- for (insn = BLOCK_HEAD (bb);
- insn && insn != NEXT_INSN (BLOCK_END (bb));
- insn = NEXT_INSN (insn))
+ FOR_BB_INSNS (bb, insn)
{
- if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
+ if (INSN_P (insn))
{
if (GET_CODE (PATTERN (insn)) == SET)
{
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;
/* Make sure there isn't a buried load somewhere. */
invalidate_any_buried_refs (src);
}
-
+
/* Check for stores. Don't worry about aliased ones, they
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
- && GET_CODE (src) != ASM_OPERANDS)
+
+ if (! MEM_P (src)
+ && GET_CODE (src) != ASM_OPERANDS
+ /* Check for REG manually since want_to_gcse_p
+ returns 0 for all REGs. */
+ && can_assign_to_reg_p (src))
ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
else
ptr->invalid = 1;
}
}
-/* Remove any references that have been either invalidated or are not in the
+/* Remove any references that have been either invalidated or are not in the
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[i];
- expr != NULL;
- expr = expr->next_same_hash)
- if (expr_equiv_p (expr->expr, ptr->pattern))
- {
- del = 0;
- break;
- }
- }
- }
-
- 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;
- }
+ unsigned int hash = ptr->hash_index % expr_hash_table.size;
+
+ 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;
+ htab_remove_elt_with_hash (pre_ldst_table, ptr, ptr->hash_index);
+ free_ldst_entry (ptr);
+ ptr = * last;
+ }
}
/* Show the world what we've found. */
- if (gcse_file && pre_ldst_mems != NULL)
- print_ldst_list (gcse_file);
+ if (dump_file && pre_ldst_mems != NULL)
+ print_ldst_list (dump_file);
}
/* 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;
if ((mem_ptr = find_rtx_in_ldst (expr->expr)))
{
- /* We can try to find just the REACHED stores, but is shouldn't
- matter to set the reaching reg everywhere... some might be
+ /* We can try to find just the REACHED stores, but is shouldn't
+ matter to set the reaching reg everywhere... some might be
dead and should be eliminated later. */
- /* We replace SET mem = expr with
- SET reg = expr
- SET mem = reg , where reg is the
- reaching reg used in the load. */
+ /* We replace (set mem expr) with (set reg expr) (set mem reg)
+ where reg is the reaching reg used in the load. We checked in
+ compute_ld_motion_mems that we can replace (set mem expr) with
+ (set reg expr) in that insn. */
rtx list = mem_ptr->stores;
-
+
for ( ; list != NULL_RTX; list = XEXP (list, 1))
{
rtx insn = XEXP (list, 0);
rtx pat = PATTERN (insn);
rtx src = SET_SRC (pat);
rtx reg = expr->reaching_reg;
- rtx copy, new;
+ rtx copy, new_rtx;
/* If we've already copied it, continue. */
if (expr->reaching_reg == src)
continue;
-
- if (gcse_file)
+
+ if (dump_file)
{
- fprintf (gcse_file, "PRE: store updated with reaching reg ");
- print_rtl (gcse_file, expr->reaching_reg);
- fprintf (gcse_file, ":\n ");
- print_inline_rtx (gcse_file, insn, 8);
- fprintf (gcse_file, "\n");
+ fprintf (dump_file, "PRE: store updated with reaching reg ");
+ print_rtl (dump_file, expr->reaching_reg);
+ fprintf (dump_file, ":\n ");
+ print_inline_rtx (dump_file, insn, 8);
+ fprintf (dump_file, "\n");
}
-
- copy = gen_move_insn ( reg, SET_SRC (pat));
- new = emit_insn_before (copy, insn);
- record_one_set (REGNO (reg), new);
+
+ copy = gen_move_insn ( reg, copy_rtx (SET_SRC (pat)));
+ new_rtx = emit_insn_before (copy, insn);
+ record_one_set (REGNO (reg), new_rtx);
SET_SRC (pat) = reg;
+ df_insn_rescan (insn);
/* un-recognize this pattern since it's probably different now. */
INSN_CODE (insn) = -1;
\f
/* Store motion code. */
-/* This is used to communicate the target bitvector we want to use in the
+#define ANTIC_STORE_LIST(x) ((x)->loads)
+#define AVAIL_STORE_LIST(x) ((x)->stores)
+#define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
+
+/* This is used to communicate the target bitvector we want to use in the
reg_set_info routine when called via the note_stores mechanism. */
-static sbitmap * regvec;
+static int * regvec;
+
+/* And current insn, for the same routine. */
+static rtx compute_store_table_current_insn;
/* Used in computing the reverse edge graph bit vectors. */
static sbitmap * st_antloc;
/* 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 (dest, setter, data)
- rtx dest, setter ATTRIBUTE_UNUSED;
- void * data ATTRIBUTE_UNUSED;
+reg_set_info (rtx dest, const_rtx setter ATTRIBUTE_UNUSED,
+ void *data)
{
+ sbitmap bb_reg = (sbitmap) data;
+
if (GET_CODE (dest) == SUBREG)
dest = SUBREG_REG (dest);
- if (GET_CODE (dest) == REG)
- SET_BIT (*regvec, REGNO (dest));
+ if (REG_P (dest))
+ {
+ regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
+ if (bb_reg)
+ SET_BIT (bb_reg, REGNO (dest));
+ }
}
-/* Return non-zero if the register operands of expression X are killed
- anywhere in basic block BB. */
+/* Clear any mark that says that this insn sets dest. Called from
+ note_stores. */
-static int
-store_ops_ok (x, bb)
- rtx x;
- basic_block bb;
+static void
+reg_clear_last_set (rtx dest, const_rtx setter ATTRIBUTE_UNUSED,
+ void *data)
+{
+ int *dead_vec = (int *) data;
+
+ if (GET_CODE (dest) == SUBREG)
+ dest = SUBREG_REG (dest);
+
+ 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 (const_rtx x, int *regs_set)
+{
+ const_rtx reg;
+
+ for (; x; x = XEXP (x, 1))
+ {
+ reg = XEXP (x, 0);
+ if (regs_set[REGNO(reg)])
+ return false;
+ }
+
+ return true;
+}
+
+/* Returns a list of registers mentioned in X. */
+static rtx
+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 (rtx x, rtx accum)
{
int i;
enum rtx_code code;
repeat:
if (x == 0)
- return 1;
+ return accum;
code = GET_CODE (x);
switch (code)
{
case REG:
- /* If a reg has changed after us in this
- block, the operand has been killed. */
- return TEST_BIT (reg_set_in_block[bb->index], REGNO (x));
+ return alloc_EXPR_LIST (0, x, accum);
case MEM:
x = XEXP (x, 0);
case PRE_DEC:
case PRE_INC:
+ case PRE_MODIFY:
case POST_DEC:
case POST_INC:
- return 0;
+ case POST_MODIFY:
+ /* We do not run this function with arguments having side effects. */
+ gcc_unreachable ();
case PC:
case CC0: /*FIXME*/
case CONST:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case SYMBOL_REF:
case LABEL_REF:
case ADDR_VEC:
case ADDR_DIFF_VEC:
- return 1;
+ return accum;
default:
break;
i = GET_RTX_LENGTH (code) - 1;
fmt = GET_RTX_FORMAT (code);
-
+
for (; i >= 0; i--)
{
if (fmt[i] == 'e')
rtx tem = XEXP (x, i);
/* 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. */
+ needed at this level, change it into iteration. */
if (i == 0)
{
x = tem;
goto repeat;
}
-
- if (! store_ops_ok (tem, bb))
- return 0;
+
+ accum = extract_mentioned_regs_helper (tem, accum);
}
else if (fmt[i] == 'E')
{
int j;
-
+
for (j = 0; j < XVECLEN (x, i); j++)
- {
- if (! store_ops_ok (XVECEXP (x, i, j), bb))
- return 0;
- }
+ accum = extract_mentioned_regs_helper (XVECEXP (x, i, j), accum);
}
}
- return 1;
+ return accum;
}
-/* Determine whether insn is MEM store pattern that we will consider moving. */
+/* Determine whether INSN is MEM store pattern that we will consider moving.
+ REGS_SET_BEFORE is bitmap of registers set before (and including) the
+ 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
+ -- if the processed expression is not anticipatable, NULL_RTX is added
+ there instead, so that we can use it as indicator that no further
+ expression of this type may be anticipatable
+ -- if the expression is available, it is added as head of AVAIL_STORE_LIST;
+ consequently, all of them but this head are dead and may be deleted.
+ -- if the expression is not available, the insn due to that it fails to be
+ available is stored in reaching_reg.
+
+ The things are complicated a bit by fact that there already may be stores
+ to the same MEM from other blocks; also caller must take care of the
+ necessary cleanup of the temporary markers after end of the basic block.
+ */
static void
-find_moveable_store (insn)
- rtx insn;
+find_moveable_store (rtx insn, int *regs_set_before, int *regs_set_after)
{
struct ls_expr * ptr;
- rtx dest = PATTERN (insn);
+ rtx dest, set, tmp;
+ int check_anticipatable, check_available;
+ basic_block bb = BLOCK_FOR_INSN (insn);
- if (GET_CODE (dest) != SET
- || GET_CODE (SET_SRC (dest)) == ASM_OPERANDS)
+ set = single_set (insn);
+ if (!set)
return;
- dest = SET_DEST (dest);
-
- if (GET_CODE (dest) != MEM || MEM_VOLATILE_P (dest)
+ dest = SET_DEST (set);
+
+ if (! MEM_P (dest) || MEM_VOLATILE_P (dest)
|| GET_MODE (dest) == BLKmode)
return;
- if (GET_CODE (XEXP (dest, 0)) != SYMBOL_REF)
- return;
+ if (side_effects_p (dest))
+ return;
+
+ /* If we are handling exceptions, we must be careful with memory references
+ that may trap. If we are not, the behavior is undefined, so we may just
+ continue. */
+ if (flag_non_call_exceptions && may_trap_p (dest))
+ return;
- if (rtx_varies_p (XEXP (dest, 0), 0))
+ /* 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;
+
+ /* Make sure that the SET_SRC of this store insns can be assigned to
+ a register, or we will fail later on in replace_store_insn, which
+ assumes that we can do this. But sometimes the target machine has
+ oddities like MEM read-modify-write instruction. See for example
+ PR24257. */
+ if (!can_assign_to_reg_p (SET_SRC (set)))
return;
ptr = ldst_entry (dest);
- ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
+ if (!ptr->pattern_regs)
+ ptr->pattern_regs = extract_mentioned_regs (dest);
+
+ /* Do not check for anticipatability if we either found one anticipatable
+ store already, or tested for one and found out that it was killed. */
+ check_anticipatable = 0;
+ if (!ANTIC_STORE_LIST (ptr))
+ check_anticipatable = 1;
+ else
+ {
+ tmp = XEXP (ANTIC_STORE_LIST (ptr), 0);
+ if (tmp != NULL_RTX
+ && BLOCK_FOR_INSN (tmp) != bb)
+ check_anticipatable = 1;
+ }
+ if (check_anticipatable)
+ {
+ if (store_killed_before (dest, ptr->pattern_regs, insn, bb, regs_set_before))
+ tmp = NULL_RTX;
+ else
+ tmp = insn;
+ ANTIC_STORE_LIST (ptr) = alloc_INSN_LIST (tmp,
+ ANTIC_STORE_LIST (ptr));
+ }
+
+ /* It is not necessary to check whether store is available if we did
+ it successfully before; if we failed before, do not bother to check
+ until we reach the insn that caused us to fail. */
+ check_available = 0;
+ if (!AVAIL_STORE_LIST (ptr))
+ check_available = 1;
+ else
+ {
+ tmp = XEXP (AVAIL_STORE_LIST (ptr), 0);
+ if (BLOCK_FOR_INSN (tmp) != bb)
+ check_available = 1;
+ }
+ if (check_available)
+ {
+ /* Check that we have already reached the insn at that the check
+ failed last time. */
+ if (LAST_AVAIL_CHECK_FAILURE (ptr))
+ {
+ for (tmp = BB_END (bb);
+ tmp != insn && tmp != LAST_AVAIL_CHECK_FAILURE (ptr);
+ tmp = PREV_INSN (tmp))
+ continue;
+ if (tmp == insn)
+ check_available = 0;
+ }
+ else
+ check_available = store_killed_after (dest, ptr->pattern_regs, insn,
+ bb, regs_set_after,
+ &LAST_AVAIL_CHECK_FAILURE (ptr));
+ }
+ if (!check_available)
+ AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn, AVAIL_STORE_LIST (ptr));
}
-/* Perform store motion. Much like gcse, except we move expressions the
- other way by looking at the flowgraph in reverse. */
+/* Find available and anticipatable stores. */
static int
-compute_store_table ()
+compute_store_table (void)
{
- int bb, ret;
+ int ret;
+ basic_block bb;
unsigned regno;
- rtx insn, pat;
+ rtx insn, pat, tmp;
+ int *last_set_in, *already_set;
+ struct ls_expr * ptr, **prev_next_ptr_ptr;
max_gcse_regno = max_reg_num ();
- reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks,
+ reg_set_in_block = sbitmap_vector_alloc (last_basic_block,
max_gcse_regno);
- sbitmap_vector_zero (reg_set_in_block, n_basic_blocks);
+ sbitmap_vector_zero (reg_set_in_block, last_basic_block);
pre_ldst_mems = 0;
+ pre_ldst_table = htab_create (13, pre_ldst_expr_hash,
+ pre_ldst_expr_eq, NULL);
+ last_set_in = XCNEWVEC (int, max_gcse_regno);
+ already_set = XNEWVEC (int, max_gcse_regno);
/* Find all the stores we care about. */
- for (bb = 0; bb < n_basic_blocks; bb++)
+ FOR_EACH_BB (bb)
{
- regvec = & (reg_set_in_block[bb]);
- for (insn = BLOCK_END (bb);
- insn && insn != PREV_INSN (BLOCK_HEAD (bb));
- insn = PREV_INSN (insn))
+ /* First compute the registers set in this block. */
+ regvec = last_set_in;
+
+ FOR_BB_INSNS (bb, insn)
{
- /* Ignore anything that is not a normal 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 (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, reg_set_in_block[bb->index]);
+ }
+
+ /* Now find the stores. */
+ memset (already_set, 0, sizeof (int) * max_gcse_regno);
+ regvec = already_set;
+ FOR_BB_INSNS (bb, insn)
+ {
+ if (! INSN_P (insn))
+ continue;
+
+ if (CALL_P (insn))
+ {
for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- SET_BIT (reg_set_in_block[bb], regno);
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
+ already_set[regno] = 1;
}
-
+
pat = PATTERN (insn);
note_stores (pat, reg_set_info, NULL);
-
+
/* Now that we've marked regs, look for stores. */
- if (GET_CODE (pat) == SET)
- find_moveable_store (insn);
+ find_moveable_store (insn, already_set, last_set_in);
+
+ /* Unmark regs that are no longer set. */
+ 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))
+ {
+ LAST_AVAIL_CHECK_FAILURE(ptr) = NULL_RTX;
+ if (ANTIC_STORE_LIST (ptr)
+ && (tmp = XEXP (ANTIC_STORE_LIST (ptr), 0)) == NULL_RTX)
+ ANTIC_STORE_LIST (ptr) = XEXP (ANTIC_STORE_LIST (ptr), 1);
+ }
+ }
+
+ /* Remove the stores that are not available anywhere, as there will
+ be no opportunity to optimize them. */
+ for (ptr = pre_ldst_mems, prev_next_ptr_ptr = &pre_ldst_mems;
+ ptr != NULL;
+ ptr = *prev_next_ptr_ptr)
+ {
+ if (!AVAIL_STORE_LIST (ptr))
+ {
+ *prev_next_ptr_ptr = ptr->next;
+ htab_remove_elt_with_hash (pre_ldst_table, ptr, ptr->hash_index);
+ free_ldst_entry (ptr);
}
+ else
+ prev_next_ptr_ptr = &ptr->next;
}
ret = enumerate_ldsts ();
-
- if (gcse_file)
+
+ if (dump_file)
{
- fprintf (gcse_file, "Store Motion Expressions.\n");
- print_ldst_list (gcse_file);
+ fprintf (dump_file, "ST_avail and ST_antic (shown under loads..)\n");
+ print_ldst_list (dump_file);
}
-
+
+ free (last_set_in);
+ free (already_set);
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 int
-load_kills_store (x, store_pattern)
- rtx x, store_pattern;
+static bool
+load_kills_store (const_rtx x, const_rtx store_pattern, int after)
{
- if (true_dependence (x, GET_MODE (x), store_pattern, rtx_addr_varies_p))
- return 1;
- return 0;
+ 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 1 if found. */
+/* Go through the entire insn X, looking for any loads which might alias
+ STORE_PATTERN. Return true if found.
+ AFTER is true if we are checking the case when STORE_PATTERN occurs
+ after the insn X. */
-static int
-find_loads (x, store_pattern)
- rtx x, store_pattern;
+static bool
+find_loads (const_rtx x, const_rtx store_pattern, int after)
{
const char * fmt;
int i, j;
- int ret = 0;
+ int ret = false;
if (!x)
- return 0;
+ return false;
- if (GET_CODE (x) == SET)
+ 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))
- return 1;
+ if (load_kills_store (x, store_pattern, after))
+ return true;
}
/* Recursively process the insn. */
fmt = GET_RTX_FORMAT (GET_CODE (x));
-
+
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 1 if it it does. */
+static inline bool
+store_killed_in_pat (const_rtx x, const_rtx pat, int after)
+{
+ if (GET_CODE (pat) == SET)
+ {
+ rtx dest = SET_DEST (pat);
+
+ if (GET_CODE (dest) == ZERO_EXTRACT)
+ dest = XEXP (dest, 0);
+
+ /* Check for memory stores to aliased objects. */
+ 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 (pat, x, after))
+ return true;
+
+ return false;
+}
+
+/* Check if INSN kills the store pattern X (is aliased with it).
+ AFTER is true if we are checking the case when store X occurs
+ after the insn. Return true if it does. */
-static int
-store_killed_in_insn (x, insn)
- rtx x, insn;
+static bool
+store_killed_in_insn (const_rtx x, const_rtx x_regs, const_rtx insn, int after)
{
- if (GET_RTX_CLASS (GET_CODE (insn)) != 'i')
- return 0;
-
- if (GET_CODE (insn) == CALL_INSN)
+ const_rtx reg, base, note, pat;
+
+ if (!INSN_P (insn))
+ return false;
+
+ if (CALL_P (insn))
{
/* A normal or pure call might read from pattern,
but a const call will not. */
- return ! CONST_OR_PURE_CALL_P (insn) || pure_call_p (insn);
+ if (!RTL_CONST_CALL_P (insn))
+ return true;
+
+ /* But even a const call reads its parameters. Check whether the
+ base of some of registers used in mem is stack pointer. */
+ for (reg = x_regs; reg; reg = XEXP (reg, 1))
+ {
+ base = find_base_term (XEXP (reg, 0));
+ if (!base
+ || (GET_CODE (base) == ADDRESS
+ && GET_MODE (base) == Pmode
+ && XEXP (base, 0) == stack_pointer_rtx))
+ return true;
+ }
+
+ return false;
+ }
+
+ pat = PATTERN (insn);
+ if (GET_CODE (pat) == SET)
+ {
+ if (store_killed_in_pat (x, pat, after))
+ return true;
}
-
- if (GET_CODE (PATTERN (insn)) == SET)
+ else if (GET_CODE (pat) == PARALLEL)
{
- rtx pat = PATTERN (insn);
- /* 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 1;
- return find_loads (SET_SRC (pat), x);
+ int i;
+
+ for (i = 0; i < XVECLEN (pat, 0); i++)
+ if (store_killed_in_pat (x, XVECEXP (pat, 0, i), 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 1 if the expression X is loaded or clobbered on or after INSN
- within basic block BB. */
+/* Returns true if the expression X is loaded or clobbered on or after INSN
+ within basic block BB. REGS_SET_AFTER is bitmap of registers set in
+ or after the insn. X_REGS is list of registers mentioned in X. If the store
+ is killed, return the last insn in that it occurs in FAIL_INSN. */
-static int
-store_killed_after (x, insn, bb)
- rtx x, insn;
- basic_block bb;
+static bool
+store_killed_after (const_rtx x, const_rtx x_regs, const_rtx insn, const_basic_block bb,
+ int *regs_set_after, rtx *fail_insn)
{
- rtx last = bb->end;
-
- if (insn == last)
- return 0;
+ rtx last = BB_END (bb), act;
- /* Check if the register operands of the store are OK in this block.
- Note that if registers are changed ANYWHERE in the block, we'll
- decide we can't move it, regardless of whether it changed above
- or below the store. This could be improved by checking the register
- operands while lookinng for aliasing in each insn. */
- if (!store_ops_ok (XEXP (x, 0), bb))
- return 1;
+ if (!store_ops_ok (x_regs, regs_set_after))
+ {
+ /* We do not know where it will happen. */
+ if (fail_insn)
+ *fail_insn = NULL_RTX;
+ return true;
+ }
- for ( ; insn && insn != NEXT_INSN (last); insn = NEXT_INSN (insn))
- if (store_killed_in_insn (x, insn))
- return 1;
-
- return 0;
+ /* 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, false))
+ {
+ if (fail_insn)
+ *fail_insn = act;
+ return true;
+ }
+
+ return false;
}
-/* Returns 1 if the expression X is loaded or clobbered on or before INSN
- within basic block BB. */
-static int
-store_killed_before (x, insn, bb)
- rtx x, insn;
- basic_block bb;
-{
- rtx first = bb->head;
-
- if (insn == first)
- return store_killed_in_insn (x, insn);
-
- /* Check if the register operands of the store are OK in this block.
- Note that if registers are changed ANYWHERE in the block, we'll
- decide we can't move it, regardless of whether it changed above
- or below the store. This could be improved by checking the register
- operands while lookinng for aliasing in each insn. */
- if (!store_ops_ok (XEXP (x, 0), bb))
- return 1;
+/* 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 (const_rtx x, const_rtx x_regs, const_rtx insn, const_basic_block bb,
+ int *regs_set_before)
+{
+ rtx first = BB_HEAD (bb);
- for ( ; insn && insn != PREV_INSN (first); insn = PREV_INSN (insn))
- if (store_killed_in_insn (x, insn))
- return 1;
-
- return 0;
-}
+ if (!store_ops_ok (x_regs, regs_set_before))
+ return true;
-#define ANTIC_STORE_LIST(x) ((x)->loads)
-#define AVAIL_STORE_LIST(x) ((x)->stores)
+ for ( ; insn != PREV_INSN (first); insn = PREV_INSN (insn))
+ if (store_killed_in_insn (x, x_regs, insn, true))
+ return true;
+
+ return false;
+}
-/* Given the table of available store insns at the end of blocks,
- determine which ones are not killed by aliasing, and generate
- the appropriate vectors for gen and killed. */
+/* 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 b;
+ int *regs_set_in_block;
rtx insn, st;
struct ls_expr * ptr;
+ unsigned regno;
/* 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 (n_basic_blocks, num_stores);
- sbitmap_vector_zero (ae_gen, n_basic_blocks);
+ ae_gen = sbitmap_vector_alloc (last_basic_block, num_stores);
+ sbitmap_vector_zero (ae_gen, last_basic_block);
- st_antloc = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks, num_stores);
- sbitmap_vector_zero (st_antloc, n_basic_blocks);
+ 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))
- {
- /* Put all the stores into either the antic list, or the avail list,
- or both. */
- rtx store_list = ptr->stores;
- ptr->stores = NULL_RTX;
-
- for (st = store_list; st != NULL; st = XEXP (st, 1))
+ {
+ for (st = AVAIL_STORE_LIST (ptr); st != NULL; st = XEXP (st, 1))
{
insn = XEXP (st, 0);
bb = BLOCK_FOR_INSN (insn);
-
- if (!store_killed_after (ptr->pattern, insn, bb))
+
+ /* If we've already seen an available expression in this block,
+ we can delete this one (It occurs earlier in the block). We'll
+ copy the SRC expression to an unused register in case there
+ are any side effects. */
+ if (TEST_BIT (ae_gen[bb->index], ptr->index))
{
- /* If we've already seen an availale expression in this block,
- we can delete the one we saw already (It occurs earlier in
- the block), and replace it with this one). We'll copy the
- old SRC expression to an unused register in case there
- are any side effects. */
- if (TEST_BIT (ae_gen[bb->index], ptr->index))
- {
- /* Find previous store. */
- rtx st;
- for (st = AVAIL_STORE_LIST (ptr); st ; st = XEXP (st, 1))
- if (BLOCK_FOR_INSN (XEXP (st, 0)) == bb)
- break;
- if (st)
- {
- 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);
- XEXP (st, 0) = insn;
- continue;
- }
- }
- SET_BIT (ae_gen[bb->index], ptr->index);
- AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn,
- AVAIL_STORE_LIST (ptr));
+ rtx r = gen_reg_rtx_and_attrs (ptr->pattern);
+ if (dump_file)
+ fprintf (dump_file, "Removing redundant store:\n");
+ replace_store_insn (r, XEXP (st, 0), bb, ptr);
+ continue;
}
-
- if (!store_killed_before (ptr->pattern, insn, bb))
+ SET_BIT (ae_gen[bb->index], ptr->index);
+ }
+
+ for (st = ANTIC_STORE_LIST (ptr); st != NULL; st = XEXP (st, 1))
+ {
+ insn = XEXP (st, 0);
+ bb = BLOCK_FOR_INSN (insn);
+ SET_BIT (st_antloc[bb->index], ptr->index);
+ }
+ }
+
+ ae_kill = sbitmap_vector_alloc (last_basic_block, num_stores);
+ sbitmap_vector_zero (ae_kill, last_basic_block);
+
+ transp = sbitmap_vector_alloc (last_basic_block, num_stores);
+ sbitmap_vector_zero (transp, last_basic_block);
+ regs_set_in_block = XNEWVEC (int, max_gcse_regno);
+
+ FOR_EACH_BB (bb)
+ {
+ for (regno = 0; regno < max_gcse_regno; regno++)
+ regs_set_in_block[regno] = TEST_BIT (reg_set_in_block[bb->index], regno);
+
+ for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
+ {
+ if (store_killed_after (ptr->pattern, ptr->pattern_regs, BB_HEAD (bb),
+ bb, regs_set_in_block, NULL))
{
- SET_BIT (st_antloc[BLOCK_NUM (insn)], ptr->index);
- ANTIC_STORE_LIST (ptr) = alloc_INSN_LIST (insn,
- ANTIC_STORE_LIST (ptr));
+ /* It should not be necessary to consider the expression
+ killed if it is both anticipatable and available. */
+ 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);
}
-
- /* Free the original list of store insns. */
- free_INSN_LIST_list (&store_list);
}
-
- ae_kill = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks, num_stores);
- sbitmap_vector_zero (ae_kill, n_basic_blocks);
-
- transp = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks, num_stores);
- sbitmap_vector_zero (transp, n_basic_blocks);
- for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
- for (b = 0; b < n_basic_blocks; b++)
- {
- if (store_killed_after (ptr->pattern, BLOCK_HEAD (b), BASIC_BLOCK (b)))
- {
- /* The anticipatable expression is not killed if it's gen'd. */
- /*
- We leave this check out for now. If we have a code sequence
- in a block which looks like:
- ST MEMa = x
- L y = MEMa
- ST MEMa = z
- We should flag this as having an ANTIC expression, NOT
- transparent, NOT killed, and AVAIL.
- Unfortunately, since we haven't re-written all loads to
- use the reaching reg, we'll end up doing an incorrect
- Load in the middle here if we push the store down. It happens in
- gcc.c-torture/execute/960311-1.c with -O3
- If we always kill it in this case, we'll sometimes do
- uneccessary work, but it shouldn't actually hurt anything.
- if (!TEST_BIT (ae_gen[b], ptr->index)). */
- SET_BIT (ae_kill[b], ptr->index);
- }
- else
- SET_BIT (transp[b], ptr->index);
- }
+ free (regs_set_in_block);
- /* Any block with no exits calls some non-returning function, so
- we better mark the store killed here, or we might not store to
- it at all. If we knew it was abort, we wouldn't have to store,
- but we don't know that for sure. */
- if (gcse_file)
+ if (dump_file)
{
- fprintf (gcse_file, "ST_avail and ST_antic (shown under loads..)\n");
- print_ldst_list (gcse_file);
- dump_sbitmap_vector (gcse_file, "st_antloc", "", st_antloc, n_basic_blocks);
- dump_sbitmap_vector (gcse_file, "st_kill", "", ae_kill, n_basic_blocks);
- dump_sbitmap_vector (gcse_file, "Transpt", "", transp, n_basic_blocks);
- dump_sbitmap_vector (gcse_file, "st_avloc", "", ae_gen, n_basic_blocks);
+ dump_sbitmap_vector (dump_file, "st_antloc", "", st_antloc, last_basic_block);
+ dump_sbitmap_vector (dump_file, "st_kill", "", ae_kill, last_basic_block);
+ dump_sbitmap_vector (dump_file, "Transpt", "", transp, last_basic_block);
+ dump_sbitmap_vector (dump_file, "st_avloc", "", ae_gen, last_basic_block);
}
}
-/* Insert an instruction at the begining of a basic block, and update
- the BLOCK_HEAD if needed. */
+/* Insert an instruction at the beginning of a basic block, and update
+ the BB_HEAD if needed. */
-static void
-insert_insn_start_bb (insn, bb)
- rtx insn;
- basic_block bb;
+static void
+insert_insn_start_basic_block (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
- || NOTE_LINE_NUMBER (before) != NOTE_INSN_BASIC_BLOCK))
+ if (! LABEL_P (before)
+ && !NOTE_INSN_BASIC_BLOCK_P (before))
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, bb);
- if (gcse_file)
+ if (dump_file)
{
- fprintf (gcse_file, "STORE_MOTION insert store at start of BB %d:\n",
+ fprintf (dump_file, "STORE_MOTION insert store at start of BB %d:\n",
bb->index);
- print_inline_rtx (gcse_file, insn, 6);
- fprintf (gcse_file, "\n");
+ print_inline_rtx (dump_file, insn, 6);
+ fprintf (dump_file, "\n");
}
}
/* This routine will insert a store on an edge. EXPR is the ldst entry for
- the memory reference, and E is the edge to insert it on. Returns non-zero
+ the memory reference, and E is the edge to insert it on. Returns nonzero
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 (expr->pattern, reg);
-
+ insn = gen_move_insn (copy_rtx (expr->pattern), reg);
+
/* If we are inserting this expression on ALL predecessor edges of a BB,
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);
}
- insert_insn_start_bb (insn, bb);
- 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);
+ insert_insn_start_basic_block (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);
-
- if (gcse_file)
+
+ if (dump_file)
{
- fprintf (gcse_file, "STORE_MOTION insert insn on edge (%d, %d):\n",
+ fprintf (dump_file, "STORE_MOTION insert insn on edge (%d, %d):\n",
e->src->index, e->dest->index);
- print_inline_rtx (gcse_file, insn, 6);
- fprintf (gcse_file, "\n");
+ print_inline_rtx (dump_file, insn, 6);
+ fprintf (dump_file, "\n");
}
-
+
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 = XNEWVEC (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 (dump_file)
+ fprintf (dump_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;
-
- insn = gen_move_insn (reg, SET_SRC (PATTERN (del)));
+ rtx insn, mem, note, set, ptr;
+
+ mem = smexpr->pattern;
+ insn = gen_move_insn (reg, SET_SRC (single_set (del)));
+
+ 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. */
+ REG_NOTES (insn) = REG_NOTES (del);
+
+ /* Emit the insn AFTER all the notes are transferred.
+ This is cheaper since we avoid df rescanning for the note change. */
insn = emit_insn_after (insn, del);
-
- if (gcse_file)
+
+ if (dump_file)
{
- fprintf (gcse_file,
+ fprintf (dump_file,
"STORE_MOTION delete insn in BB %d:\n ", bb->index);
- print_inline_rtx (gcse_file, del, 6);
- fprintf (gcse_file, "\nSTORE MOTION replaced with insn:\n ");
- print_inline_rtx (gcse_file, insn, 6);
- fprintf (gcse_file, "\n");
+ print_inline_rtx (dump_file, del, 6);
+ fprintf (dump_file, "\nSTORE MOTION replaced with insn:\n ");
+ print_inline_rtx (dump_file, insn, 6);
+ fprintf (dump_file, "\n");
}
-
+
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 (dump_file)
+ fprintf (dump_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;
if (expr->reaching_reg == NULL_RTX)
- expr->reaching_reg = gen_reg_rtx (GET_MODE (expr->pattern));
-
+ expr->reaching_reg = gen_reg_rtx_and_attrs (expr->pattern);
- /* If there is more than 1 store, the earlier ones will be dead,
- but it doesn't hurt to replace them here. */
reg = expr->reaching_reg;
-
+
for (i = AVAIL_STORE_LIST (expr); i; i = XEXP (i, 1))
{
del = XEXP (i, 0);
if (BLOCK_FOR_INSN (del) == bb)
{
- /* We know there is only one since we deleted redundant
+ /* 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 ()
+static void
+free_store_memory (void)
{
free_ldst_mems ();
-
+
if (ae_gen)
sbitmap_vector_free (ae_gen);
if (ae_kill)
sbitmap_vector_free (pre_delete_map);
if (reg_set_in_block)
sbitmap_vector_free (reg_set_in_block);
-
+
ae_gen = ae_kill = transp = st_antloc = NULL;
pre_insert_map = pre_delete_map = reg_set_in_block = NULL;
}
other way by looking at the flowgraph in reverse. */
static void
-store_motion ()
+store_motion (void)
{
+ basic_block bb;
int x;
struct ls_expr * ptr;
int update_flow = 0;
- if (gcse_file)
+ if (dump_file)
{
- fprintf (gcse_file, "before store motion\n");
- print_rtl (gcse_file, get_insns ());
+ fprintf (dump_file, "before store motion\n");
+ print_rtl (dump_file, get_insns ());
}
-
init_alias_analysis ();
- /* Find all the stores that are live to the end of their block. */
+ /* Find all the available and anticipatable stores. */
num_stores = compute_store_table ();
if (num_stores == 0)
{
+ htab_delete (pre_ldst_table);
+ pre_ldst_table = NULL;
sbitmap_vector_free (reg_set_in_block);
end_alias_analysis ();
return;
}
- /* Now compute whats actually available to move. */
- add_noreturn_fake_exit_edges ();
+ /* 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,
+ edge_list = pre_edge_rev_lcm (num_stores, transp, ae_gen,
+ st_antloc, ae_kill, &pre_insert_map,
&pre_delete_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))
{
- for (x = 0; x < n_basic_blocks; x++)
- if (TEST_BIT (pre_delete_map[x], ptr->index))
- delete_store (ptr, BASIC_BLOCK (x));
+ /* 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 (dump_file != NULL)
+ fprintf (dump_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);
for (x = 0; x < NUM_EDGES (edge_list); x++)
if (TEST_BIT (pre_insert_map[x], ptr->index))
free_store_memory ();
free_edge_list (edge_list);
- remove_fake_edges ();
+ remove_fake_exit_edges ();
+ end_alias_analysis ();
+}
+
+\f
+/* Entry point for jump bypassing optimization pass. */
+
+static int
+bypass_jumps (void)
+{
+ int changed;
+
+ /* We do not construct an accurate cfg in functions which call
+ setjmp, so just punt to be safe. */
+ if (cfun->calls_setjmp)
+ return 0;
+
+ /* Identify the basic block information for this function, including
+ successors and predecessors. */
+ max_gcse_regno = max_reg_num ();
+
+ if (dump_file)
+ dump_flow_info (dump_file, dump_flags);
+
+ /* Return if there's nothing to do, or it is too expensive. */
+ if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
+ || is_too_expensive (_ ("jump bypassing disabled")))
+ return 0;
+
+ gcc_obstack_init (&gcse_obstack);
+ bytes_used = 0;
+
+ /* We need alias. */
+ init_alias_analysis ();
+
+ /* Record where pseudo-registers are set. This data is kept accurate
+ during each pass. ??? We could also record hard-reg information here
+ [since it's unchanging], however it is currently done during hash table
+ computation.
+
+ It may be tempting to compute MEM set information here too, but MEM sets
+ will be subject to code motion one day and thus we need to compute
+ information about memory sets when we build the hash tables. */
+
+ alloc_reg_set_mem (max_gcse_regno);
+ compute_sets ();
+
+ max_gcse_regno = max_reg_num ();
+ alloc_gcse_mem ();
+ changed = one_cprop_pass (MAX_GCSE_PASSES + 2, true, true);
+ free_gcse_mem ();
+
+ if (dump_file)
+ {
+ fprintf (dump_file, "BYPASS of %s: %d basic blocks, ",
+ current_function_name (), n_basic_blocks);
+ fprintf (dump_file, "%d bytes\n\n", bytes_used);
+ }
+
+ obstack_free (&gcse_obstack, NULL);
+ free_reg_set_mem ();
+
+ /* We are finished with alias. */
end_alias_analysis ();
+
+ 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)
+ {
+ warning (OPT_Wdisabled_optimization,
+ "%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)
+ {
+ warning (OPT_Wdisabled_optimization,
+ "%s: %d basic blocks and %d registers",
+ pass, n_basic_blocks, max_reg_num ());
+
+ return true;
+ }
+
+ return false;
+}
+\f
+static bool
+gate_handle_jump_bypass (void)
+{
+ return optimize > 0 && flag_gcse
+ && dbg_cnt (jump_bypass);
+}
+
+/* Perform jump bypassing and control flow optimizations. */
+static unsigned int
+rest_of_handle_jump_bypass (void)
+{
+ delete_unreachable_blocks ();
+ if (bypass_jumps ())
+ {
+ delete_trivially_dead_insns (get_insns (), max_reg_num ());
+ rebuild_jump_labels (get_insns ());
+ cleanup_cfg (0);
+ }
+ return 0;
+}
+
+struct rtl_opt_pass pass_jump_bypass =
+{
+ {
+ RTL_PASS,
+ "bypass", /* name */
+ gate_handle_jump_bypass, /* gate */
+ rest_of_handle_jump_bypass, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_BYPASS, /* tv_id */
+ 0, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_dump_func |
+ TODO_ggc_collect | TODO_verify_flow /* todo_flags_finish */
+ }
+};
+
+
+static bool
+gate_handle_gcse (void)
+{
+ return optimize > 0 && flag_gcse
+ && dbg_cnt (gcse);
+}
+
+
+static unsigned int
+rest_of_handle_gcse (void)
+{
+ int save_csb, save_cfj;
+ int tem2 = 0, tem;
+ tem = gcse_main (get_insns ());
+ delete_trivially_dead_insns (get_insns (), max_reg_num ());
+ rebuild_jump_labels (get_insns ());
+ save_csb = flag_cse_skip_blocks;
+ save_cfj = flag_cse_follow_jumps;
+ flag_cse_skip_blocks = flag_cse_follow_jumps = 0;
+
+ /* If -fexpensive-optimizations, re-run CSE to clean up things done
+ by gcse. */
+ if (flag_expensive_optimizations)
+ {
+ timevar_push (TV_CSE);
+ tem2 = cse_main (get_insns (), max_reg_num ());
+ df_finish_pass (false);
+ purge_all_dead_edges ();
+ delete_trivially_dead_insns (get_insns (), max_reg_num ());
+ timevar_pop (TV_CSE);
+ cse_not_expected = !flag_rerun_cse_after_loop;
+ }
+
+ /* If gcse or cse altered any jumps, rerun jump optimizations to clean
+ things up. */
+ if (tem || tem2 == 2)
+ {
+ timevar_push (TV_JUMP);
+ rebuild_jump_labels (get_insns ());
+ cleanup_cfg (0);
+ timevar_pop (TV_JUMP);
+ }
+ else if (tem2 == 1)
+ cleanup_cfg (0);
+
+ flag_cse_skip_blocks = save_csb;
+ flag_cse_follow_jumps = save_cfj;
+ return 0;
}
+
+struct rtl_opt_pass pass_gcse =
+{
+ {
+ RTL_PASS,
+ "gcse1", /* name */
+ gate_handle_gcse, /* gate */
+ rest_of_handle_gcse, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_GCSE, /* tv_id */
+ 0, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_df_finish | TODO_verify_rtl_sharing |
+ TODO_dump_func |
+ TODO_verify_flow | TODO_ggc_collect /* todo_flags_finish */
+ }
+};
+
+
+#include "gt-gcse.h"
projects / msp430-gcc.git / blobdiff