/* Global common subexpression elimination/Partial redundancy elimination
and global constant/copy propagation for GNU compiler.
- Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004
- 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
- a store to the same address as a load does not kill the load if the
source of the store is also the destination of the load. Handling this
allows more load motion, particularly out of loops.
- - ability to realloc sbitmap vectors would allow one initial computation
- of reg_set_in_block with only subsequent additions, rather than
- recomputing it for each pass
*/
#include "intl.h"
#include "obstack.h"
#include "timevar.h"
+#include "tree-pass.h"
+#include "hashtab.h"
+#include "df.h"
+#include "dbgcnt.h"
+#include "target.h"
/* Propagate flow information through back edges and thus enable PRE's
moving loop invariant calculations out of loops.
We perform the following steps:
- 1) Compute basic block information.
+ 1) Compute table of places where registers are set.
- 2) Compute table of places where registers are set.
+ 2) Perform copy/constant propagation.
- 3) Perform copy/constant propagation.
-
- 4) Perform global cse using lazy code motion if not optimizing
+ 3) 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.
+ 4) Perform another pass of copy/constant propagation. Try to bypass
+ conditional jumps if the condition can be computed from a value of
+ an incoming edge.
+
+ 5) Perform store motion.
Two passes of copy/constant propagation are done because the first one
enables more GCSE and the second one helps to clean up the copies that
(set (pseudo-reg) (expression)).
Function want_to_gcse_p says what these are.
+ In addition, expressions in REG_EQUAL notes are candidates for GXSE-ing.
+ This allows PRE to hoist expressions that are expressed in multiple insns,
+ such as comprex address calculations (e.g. for PIC code, or loads with a
+ high part and as lowe part).
+
PRE handles moving invariant expressions out of loops (by treating them as
partially redundant).
It was found doing copy propagation between each pass enables further
substitutions.
+ This study was done before expressions in REG_EQUAL notes were added as
+ candidate expressions for optimization, and before the GIMPLE optimizers
+ were added. Probably, multiple passes is even less efficient now than
+ at the time when the study was conducted.
+
PRE is quite expensive in complicated functions because the DFA can take
- a while 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 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;
+/* Set to non-zero if CSE should run after all GCSE optimizations are done. */
+int flag_rerun_cse_after_global_opts;
/* An obstack for our working variables. */
static struct obstack gcse_obstack;
[one could build a mapping table without holes afterwards though].
Someday I'll perform the computation and figure it out. */
-struct hash_table
+struct hash_table_d
{
/* The table itself.
This is an array of `expr_hash_table_size' elements. */
};
/* Expression hash table. */
-static struct hash_table expr_hash_table;
+static struct hash_table_d expr_hash_table;
/* Copy propagation hash table. */
-static struct hash_table set_hash_table;
-
-/* Mapping of uids to cuids.
- Only real insns get cuids. */
-static int *uid_cuid;
-
-/* Highest UID in UID_CUID. */
-static int max_uid;
-
-/* Get the cuid of an insn. */
-#ifdef ENABLE_CHECKING
-#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;
-
-/* Table of registers that are modified.
-
- For each register, each element is a list of places where the pseudo-reg
- is set.
-
- For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
- requires knowledge of which blocks kill which regs [and thus could use
- a bitmap instead of the lists `reg_set_table' uses].
-
- `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
- num-regs) [however perhaps it may be useful to keep the data as is]. One
- advantage of recording things this way is that `reg_set_table' is fairly
- sparse with respect to pseudo regs but for hard regs could be fairly dense
- [relatively speaking]. And recording sets of pseudo-regs in lists speeds
- up functions like compute_transp since in the case of pseudo-regs we only
- need to iterate over the number of times a pseudo-reg is set, not over the
- number of basic blocks [clearly there is a bit of a slow down in the cases
- where a pseudo is set more than once in a block, however it is believed
- that the net effect is to speed things up]. This isn't done for hard-regs
- because recording call-clobbered hard-regs in `reg_set_table' at each
- function call can consume a fair bit of memory, and iterating over
- hard-regs stored this way in compute_transp will be more expensive. */
-
-typedef struct reg_set
-{
- /* The next setting of this register. */
- struct reg_set *next;
- /* The insn where it was set. */
- rtx insn;
-} reg_set;
-
-static reg_set **reg_set_table;
-
-/* Size of `reg_set_table'.
- The table starts out at max_gcse_regno + slop, and is enlarged as
- necessary. */
-static int reg_set_table_size;
-
-/* Amount to grow `reg_set_table' by when it's full. */
-#define REG_SET_TABLE_SLOP 100
+static struct hash_table_d set_hash_table;
/* This is a list of expressions which are MEMs and will be used by load
or store motion.
/* 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 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. */
-static sbitmap *reg_set_in_block;
-
/* Array, indexed by basic block number for a list of insns which modify
memory within that block. */
static rtx * modify_mem_list;
/* This array parallels modify_mem_list, but is kept canonicalized. */
static rtx * canon_modify_mem_list;
-static 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. */
static int gcse_create_count;
/* Number of local constants propagated. */
static int local_const_prop_count;
-/* Number of local copys propagated. */
+/* 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 copys propagated. */
+/* Number of global copies propagated. */
static int global_copy_prop_count;
\f
/* For available exprs */
-static sbitmap *ae_kill, *ae_gen;
-
-/* Objects of this type are passed around by the null-pointer check
- removal routines. */
-struct null_pointer_info
-{
- /* The basic block being processed. */
- basic_block current_block;
- /* The first register to be handled in this pass. */
- unsigned int min_reg;
- /* One greater than the last register to be handled in this pass. */
- unsigned int max_reg;
- sbitmap *nonnull_local;
- sbitmap *nonnull_killed;
-};
+static sbitmap *ae_kill;
\f
static void compute_can_copy (void);
static void *gmalloc (size_t) ATTRIBUTE_MALLOC;
static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC;
-static void *grealloc (void *, size_t);
static void *gcse_alloc (unsigned long);
-static void alloc_gcse_mem (rtx);
+static void 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 replace_one_set (int, rtx, rtx);
-static void record_set_info (rtx, rtx, void *);
-static void compute_sets (rtx);
-static void hash_scan_insn (rtx, struct hash_table *, int);
-static void hash_scan_set (rtx, rtx, struct hash_table *);
-static void hash_scan_clobber (rtx, rtx, struct hash_table *);
-static void hash_scan_call (rtx, rtx, struct hash_table *);
+static void hash_scan_insn (rtx, struct hash_table_d *);
+static void hash_scan_set (rtx, rtx, struct hash_table_d *);
+static void hash_scan_clobber (rtx, rtx, struct hash_table_d *);
+static void hash_scan_call (rtx, rtx, struct hash_table_d *);
static int want_to_gcse_p (rtx);
-static bool can_assign_to_reg_p (rtx);
-static bool gcse_constant_p (rtx);
-static int oprs_unchanged_p (rtx, rtx, int);
-static int oprs_anticipatable_p (rtx, rtx);
-static int oprs_available_p (rtx, rtx);
+static 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 (rtx, enum machine_mode, int *, int);
+ struct hash_table_d *);
+static void insert_set_in_table (rtx, rtx, struct hash_table_d *);
+static unsigned int hash_expr (const_rtx, enum machine_mode, int *, int);
static unsigned int hash_set (int, int);
-static int expr_equiv_p (rtx, rtx);
+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, 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 void record_last_set_info (rtx, const_rtx, void *);
+static void compute_hash_table (struct hash_table_d *);
+static void alloc_hash_table (int, struct hash_table_d *, int);
+static void free_hash_table (struct hash_table_d *);
+static void compute_hash_table_work (struct hash_table_d *);
+static void dump_hash_table (FILE *, const char *, struct hash_table_d *);
+static struct expr *lookup_set (unsigned int, struct hash_table_d *);
static struct expr *next_set (unsigned int, struct expr *);
static void reset_opr_set_tables (void);
-static int oprs_not_set_p (rtx, rtx);
+static 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 (rtx, int, sbitmap *, int);
+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 *);
+ struct hash_table_d *);
static void compute_cprop_data (void);
static void find_used_regs (rtx *, void *);
static int try_replace_reg (rtx, rtx, rtx);
static struct expr *find_avail_set (int, rtx);
static int cprop_jump (basic_block, rtx, rtx, rtx, rtx);
-static void mems_conflict_for_gcse_p (rtx, rtx, void *);
-static int load_killed_in_block_p (basic_block, int, rtx, int);
-static void canon_list_insert (rtx, rtx, void *);
-static int cprop_insn (rtx, int);
-static int cprop (int);
+static void 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);
static void find_implicit_sets (void);
-static int one_cprop_pass (int, int, int);
-static bool constprop_register (rtx, rtx, rtx, int);
+static int one_cprop_pass (void);
+static bool constprop_register (rtx, rtx, rtx);
static struct expr *find_bypass_set (int, int);
-static bool reg_killed_on_edge (rtx, edge);
+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 compute_pre_data (void);
static int pre_expr_reaches_here_p (basic_block, struct expr *,
basic_block);
-static void insert_insn_end_bb (struct expr *, basic_block, int);
+static void 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 int one_pre_gcse_pass (void);
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 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 void free_ldst_mems (void);
static void print_ldst_list (FILE *);
static struct ls_expr * find_rtx_in_ldst (rtx);
-static int enumerate_ldsts (void);
static inline struct ls_expr * first_ls_expr (void);
static inline struct ls_expr * next_ls_expr (struct ls_expr *);
-static int simple_mem (rtx);
+static 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, rtx, void *);
-static void reg_clear_last_set (rtx, rtx, void *);
-static bool store_ops_ok (rtx, int *);
-static rtx extract_mentioned_regs (rtx);
-static rtx extract_mentioned_regs_helper (rtx, rtx);
-static void find_moveable_store (rtx, int *, int *);
-static int compute_store_table (void);
-static bool load_kills_store (rtx, rtx, int);
-static bool find_loads (rtx, rtx, int);
-static bool store_killed_in_insn (rtx, rtx, rtx, int);
-static bool store_killed_after (rtx, rtx, rtx, basic_block, int *, rtx *);
-static bool store_killed_before (rtx, rtx, rtx, basic_block, int *);
-static void build_store_vectors (void);
-static void insert_insn_start_bb (rtx, basic_block);
-static int insert_store (struct ls_expr *, edge);
-static void remove_reachable_equiv_notes (basic_block, struct ls_expr *);
-static void replace_store_insn (rtx, rtx, basic_block, struct ls_expr *);
-static void delete_store (struct ls_expr *, basic_block);
-static void free_store_memory (void);
-static void store_motion (void);
static void free_insn_expr_list_list (rtx *);
static void clear_modify_mem_tables (void);
static void free_modify_mem_tables (void);
static rtx gcse_emit_move_after (rtx, rtx, rtx);
static void local_cprop_find_used_regs (rtx *, void *);
-static bool do_local_cprop (rtx, rtx, int, rtx*);
-static bool adjust_libcall_notes (rtx, rtx, rtx, rtx*);
-static void local_cprop_pass (int);
+static bool do_local_cprop (rtx, rtx);
+static int local_cprop_pass (void);
static bool is_too_expensive (const char *);
-\f
-
-/* Entry point for global common subexpression elimination.
- F is the first instruction in the function. Return nonzero if a
- change is mode. */
-
-int
-gcse_main (rtx f, FILE *file)
-{
- int changed, pass;
- /* Bytes used at start of pass. */
- int initial_bytes_used;
- /* Maximum number of bytes used by a pass. */
- int max_pass_bytes;
- /* Point to release obstack data from for each pass. */
- char *gcse_obstack_bottom;
-
- /* We do not construct an accurate cfg in functions which call
- setjmp, so just punt to be safe. */
- if (current_function_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);
-
- /* Return if there's nothing to do, or it is too expensive. */
- if (n_basic_blocks <= 1 || is_too_expensive (_("GCSE 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 (f);
-
- pass = 0;
- initial_bytes_used = bytes_used;
- max_pass_bytes = 0;
- gcse_obstack_bottom = gcse_alloc (1);
- changed = 1;
- while (changed && pass < MAX_GCSE_PASSES)
- {
- changed = 0;
- if (file)
- fprintf (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. */
- bytes_used = initial_bytes_used;
-
- /* Each pass may create new registers, so recalculate each time. */
- max_gcse_regno = max_reg_num ();
-
- alloc_gcse_mem (f);
-
- /* Don't allow constant propagation to modify jumps
- during this pass. */
- timevar_push (TV_CPROP1);
- changed = one_cprop_pass (pass + 1, 0, 0);
- timevar_pop (TV_CPROP1);
-
- if (optimize_size)
- /* Do nothing. */ ;
- else
- {
- timevar_push (TV_PRE);
- changed |= one_pre_gcse_pass (pass + 1);
- /* We may have just created new basic blocks. Release and
- recompute various things which are sized on the number of
- basic blocks. */
- if (changed)
- {
- free_modify_mem_tables ();
- modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
- canon_modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
- }
- free_reg_set_mem ();
- alloc_reg_set_mem (max_reg_num ());
- compute_sets (f);
- run_jump_opt_after_gcse = 1;
- timevar_pop (TV_PRE);
- }
-
- if (max_pass_bytes < bytes_used)
- max_pass_bytes = bytes_used;
-
- /* Free up memory, then reallocate for code hoisting. We can
- not re-use the existing allocated memory because the tables
- will not have info for the insns or registers created by
- partial redundancy elimination. */
- free_gcse_mem ();
-
- /* 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 don't run the partial redundancy algorithms). */
- if (optimize_size)
- {
- timevar_push (TV_HOIST);
- max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (f);
- changed |= one_code_hoisting_pass ();
- free_gcse_mem ();
-
- if (max_pass_bytes < bytes_used)
- max_pass_bytes = bytes_used;
- timevar_pop (TV_HOIST);
- }
-
- if (file)
- {
- fprintf (file, "\n");
- fflush (file);
- }
-
- obstack_free (&gcse_obstack, gcse_obstack_bottom);
- pass++;
- }
-
- /* 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. */
- timevar_push (TV_CPROP2);
- one_cprop_pass (pass + 1, 1, 0);
- timevar_pop (TV_CPROP2);
- free_gcse_mem ();
-
- if (file)
- {
- fprintf (file, "GCSE of %s: %d basic blocks, ",
- current_function_name (), n_basic_blocks);
- fprintf (file, "%d pass%s, %d bytes\n\n",
- pass, pass > 1 ? "es" : "", max_pass_bytes);
- }
- obstack_free (&gcse_obstack, NULL);
- free_reg_set_mem ();
+#define GNEW(T) ((T *) gmalloc (sizeof (T)))
+#define GCNEW(T) ((T *) gcalloc (1, sizeof (T)))
- /* We are finished with alias. */
- end_alias_analysis ();
- allocate_reg_info (max_reg_num (), FALSE, FALSE);
+#define GNEWVEC(T, N) ((T *) gmalloc (sizeof (T) * (N)))
+#define GCNEWVEC(T, N) ((T *) gcalloc ((N), sizeof (T)))
- if (!optimize_size && flag_gcse_sm)
- {
- timevar_push (TV_LSM);
- store_motion ();
- timevar_pop (TV_LSM);
- }
+#define GNEWVAR(T, S) ((T *) gmalloc ((S)))
+#define GCNEWVAR(T, S) ((T *) gcalloc (1, (S)))
- /* Record where pseudo-registers are set. */
- return run_jump_opt_after_gcse;
-}
+#define GOBNEW(T) ((T *) gcse_alloc (sizeof (T)))
+#define GOBNEWVAR(T, S) ((T *) gcse_alloc ((S)))
\f
/* Misc. utilities. */
return can_copy[mode] != 0;
}
+
\f
/* Cover function to xmalloc to record bytes allocated. */
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 void *
-grealloc (void *ptr, size_t size)
-{
- return xrealloc (ptr, size);
-}
-
/* Cover function to obstack_alloc. */
static void *
return obstack_alloc (&gcse_obstack, size);
}
-/* Allocate memory for the cuid mapping array,
- and reg/memory set tracking tables.
-
+/* Allocate memory for the reg/memory set tracking tables.
This is called at the start of each pass. */
static void
-alloc_gcse_mem (rtx f)
+alloc_gcse_mem (void)
{
- int i;
- 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. */
-
- max_uid = get_max_uid ();
- uid_cuid = gcalloc (max_uid + 1, sizeof (int));
- 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. */
-
- max_cuid = i;
- cuid_insn = gcalloc (max_cuid + 1, sizeof (rtx));
- for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- CUID_INSN (i++) = insn;
-
/* Allocate vars to track sets of regs. */
- reg_set_bitmap = BITMAP_XMALLOC ();
+ reg_set_bitmap = BITMAP_ALLOC (NULL);
- /* Allocate vars to track sets of regs, memory per block. */
- reg_set_in_block = sbitmap_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 = gcalloc (last_basic_block, sizeof (rtx));
- canon_modify_mem_list = gcalloc (last_basic_block, 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 (void)
{
- free (uid_cuid);
- free (cuid_insn);
-
- BITMAP_XFREE (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);
+ BITMAP_FREE (modify_mem_list_set);
+ BITMAP_FREE (blocks_with_calls);
}
\f
/* Compute the local properties of each recorded expression.
static void
compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc,
- struct hash_table *table)
+ struct hash_table_d *table)
{
unsigned int i;
}
}
\f
-/* Register set information.
-
- `reg_set_table' records where each register is set or otherwise
- modified. */
-
-static struct obstack reg_set_obstack;
-
-static void
-alloc_reg_set_mem (int n_regs)
-{
- reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
- reg_set_table = gcalloc (reg_set_table_size, sizeof (struct reg_set *));
-
- gcc_obstack_init (®_set_obstack);
-}
-
-static void
-free_reg_set_mem (void)
-{
- free (reg_set_table);
- obstack_free (®_set_obstack, NULL);
-}
-
-/* An OLD_INSN that used to set REGNO was replaced by NEW_INSN.
- Update the corresponding `reg_set_table' entry accordingly.
- We assume that NEW_INSN is not already recorded in reg_set_table[regno]. */
-
-static void
-replace_one_set (int regno, rtx old_insn, rtx new_insn)
-{
- struct reg_set *reg_info;
- if (regno >= reg_set_table_size)
- return;
- for (reg_info = reg_set_table[regno]; reg_info; reg_info = reg_info->next)
- if (reg_info->insn == old_insn)
- {
- reg_info->insn = new_insn;
- break;
- }
-}
-
-/* Record REGNO in the reg_set table. */
-
-static void
-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;
-
- /* If the table isn't big enough, enlarge it. */
- if (regno >= reg_set_table_size)
- {
- int new_size = regno + REG_SET_TABLE_SLOP;
-
- reg_set_table = grealloc (reg_set_table,
- new_size * sizeof (struct reg_set *));
- memset (reg_set_table + reg_set_table_size, 0,
- (new_size - reg_set_table_size) * sizeof (struct reg_set *));
- reg_set_table_size = new_size;
- }
-
- new_reg_info = obstack_alloc (®_set_obstack, sizeof (struct reg_set));
- bytes_used += sizeof (struct reg_set);
- new_reg_info->insn = insn;
- new_reg_info->next = reg_set_table[regno];
- reg_set_table[regno] = new_reg_info;
-}
-
-/* Called from compute_sets via note_stores to handle one SET or CLOBBER in
- an insn. The DATA is really the instruction in which the SET is
- occurring. */
-
-static void
-record_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED, void *data)
-{
- rtx record_set_insn = (rtx) data;
-
- 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 documentation. */
-
-static void
-compute_sets (rtx f)
-{
- rtx insn;
-
- for (insn = f; insn != 0; insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- note_stores (PATTERN (insn), record_set_info, insn);
-}
-\f
/* Hash table support. */
struct reg_avail_info
static int
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 REG:
case SUBREG:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case CALL:
return 0;
default:
- return can_assign_to_reg_p (x);
+ return can_assign_to_reg_without_clobbers_p (x);
}
}
-/* Used internally by can_assign_to_reg_p. */
+/* Used internally by can_assign_to_reg_without_clobbers_p. */
static GTY(()) rtx test_insn;
-/* Return true if we can assign X to a pseudo register. */
+/* Return true if we can assign X to a pseudo register such that the
+ resulting insn does not result in clobbering a hard register as a
+ side-effect.
-static bool
-can_assign_to_reg_p (rtx x)
+ Additionally, if the target requires it, check that the resulting insn
+ can be copied. If it cannot, this means that X is special and probably
+ has hidden side-effects we don't want to mess with.
+
+ This function is typically used by code motion passes, to verify
+ that it is safe to insert an insn without worrying about clobbering
+ maybe live hard regs. */
+
+bool
+can_assign_to_reg_without_clobbers_p (rtx x)
{
int num_clobbers = 0;
int icode;
valid. */
PUT_MODE (SET_DEST (PATTERN (test_insn)), GET_MODE (x));
SET_SRC (PATTERN (test_insn)) = x;
- return ((icode = recog (PATTERN (test_insn), test_insn, &num_clobbers)) >= 0
- && (num_clobbers == 0 || ! added_clobbers_hard_reg_p (icode)));
+
+ icode = recog (PATTERN (test_insn), test_insn, &num_clobbers);
+ if (icode < 0)
+ return false;
+
+ if (num_clobbers > 0 && added_clobbers_hard_reg_p (icode))
+ return false;
+
+ if (targetm.cannot_copy_insn_p && targetm.cannot_copy_insn_p (test_insn))
+ return false;
+
+ return true;
}
/* Return nonzero if the operands of expression X are unchanged from the
or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
static int
-oprs_unchanged_p (rtx x, rtx 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)
- return info->last_set < INSN_CUID (insn);
+ return info->last_set < DF_INSN_LUID (insn);
else
- return info->first_set >= INSN_CUID (insn);
+ return info->first_set >= DF_INSN_LUID (insn);
}
case MEM:
- if (load_killed_in_block_p (current_bb, INSN_CUID (insn),
+ if (load_killed_in_block_p (current_bb, DF_INSN_LUID (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 (rtx dest, rtx setter 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);
}
/* Return nonzero if the expression in X (a memory reference) is killed
- in block BB before or after the insn with the CUID in UID_LIMIT.
+ in block BB before or after the insn with the LUID in UID_LIMIT.
AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
before UID_LIMIT.
AVAIL_P to 0. */
static int
-load_killed_in_block_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;
/* Ignore entries in the list that do not apply. */
if ((avail_p
- && INSN_CUID (XEXP (list_entry, 0)) < uid_limit)
+ && DF_INSN_LUID (XEXP (list_entry, 0)) < uid_limit)
|| (! avail_p
- && INSN_CUID (XEXP (list_entry, 0)) > uid_limit))
+ && DF_INSN_LUID (XEXP (list_entry, 0)) > uid_limit))
{
list_entry = XEXP (list_entry, 1);
continue;
the start of INSN's basic block up to but not including INSN. */
static int
-oprs_anticipatable_p (rtx x, rtx insn)
+oprs_anticipatable_p (const_rtx x, const_rtx insn)
{
return oprs_unchanged_p (x, insn, 0);
}
INSN to the end of INSN's basic block. */
static int
-oprs_available_p (rtx x, rtx insn)
+oprs_available_p (const_rtx x, const_rtx insn)
{
return oprs_unchanged_p (x, insn, 1);
}
the current size of the hash table to be probed. */
static unsigned int
-hash_expr (rtx x, enum machine_mode mode, int *do_not_record_p,
+hash_expr (const_rtx x, enum machine_mode mode, int *do_not_record_p,
int hash_table_size)
{
unsigned int hash;
/* Return nonzero if exp1 is equivalent to exp2. */
static int
-expr_equiv_p (rtx x, rtx y)
+expr_equiv_p (const_rtx x, const_rtx y)
{
return exp_equiv_p (x, y, 0, true);
}
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 avail_p, struct hash_table_d *table)
{
int found, do_not_record_p;
unsigned int hash;
struct expr *cur_expr, *last_expr = NULL;
struct occr *antic_occr, *avail_occr;
- struct occr *last_occr = NULL;
hash = hash_expr (x, mode, &do_not_record_p, table->size);
if (! found)
{
- cur_expr = gcse_alloc (sizeof (struct expr));
+ 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. */
{
antic_occr = cur_expr->antic_occr;
- /* Search for another occurrence in the same basic block. */
- while (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
- {
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = antic_occr;
- antic_occr = antic_occr->next;
- }
+ if (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
+ antic_occr = NULL;
if (antic_occr)
/* Found another instance of the expression in the same basic block.
else
{
/* First occurrence of this expression in this basic block. */
- antic_occr = gcse_alloc (sizeof (struct occr));
+ antic_occr = GOBNEW (struct occr);
bytes_used += sizeof (struct occr);
- /* First occurrence of this expression in any block? */
- if (cur_expr->antic_occr == NULL)
- cur_expr->antic_occr = antic_occr;
- else
- last_occr->next = antic_occr;
-
antic_occr->insn = insn;
- antic_occr->next = NULL;
+ antic_occr->next = cur_expr->antic_occr;
antic_occr->deleted_p = 0;
+ cur_expr->antic_occr = antic_occr;
}
}
{
avail_occr = cur_expr->avail_occr;
- /* Search for another occurrence in the same basic block. */
- while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))
+ if (avail_occr && BLOCK_NUM (avail_occr->insn) == BLOCK_NUM (insn))
{
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = avail_occr;
- avail_occr = avail_occr->next;
+ /* Found another instance of the expression in the same basic block.
+ Prefer this occurrence to the currently recorded one. We want
+ the last one in the block and the block is scanned from start
+ to end. */
+ avail_occr->insn = insn;
}
-
- if (avail_occr)
- /* Found another instance of the expression in the same basic block.
- Prefer this occurrence to the currently recorded one. We want
- the last one in the block and the block is scanned from start
- to end. */
- avail_occr->insn = insn;
else
{
/* First occurrence of this expression in this basic block. */
- avail_occr = gcse_alloc (sizeof (struct occr));
+ avail_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 = avail_occr;
- else
- last_occr->next = avail_occr;
-
avail_occr->insn = insn;
- avail_occr->next = NULL;
+ avail_occr->next = cur_expr->avail_occr;
avail_occr->deleted_p = 0;
+ cur_expr->avail_occr = avail_occr;
}
}
}
basic block. */
static void
-insert_set_in_table (rtx x, rtx insn, struct hash_table *table)
+insert_set_in_table (rtx x, rtx insn, struct hash_table_d *table)
{
int found;
unsigned int hash;
struct expr *cur_expr, *last_expr = NULL;
- struct occr *cur_occr, *last_occr = NULL;
+ struct occr *cur_occr;
gcc_assert (GET_CODE (x) == SET && REG_P (SET_DEST (x)));
if (! found)
{
- cur_expr = gcse_alloc (sizeof (struct expr));
+ 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. */
/* 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 = 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->next = cur_expr->avail_occr;
cur_occr->deleted_p = 0;
+ cur_expr->avail_occr = cur_occr;
}
}
the purposes of GCSE's constant propagation. */
static bool
-gcse_constant_p (rtx x)
+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)
+ && CONST_INT_P (XEXP (x, 0))
+ && CONST_INT_P (XEXP (x, 1)))
return true;
/* Consider a COMPARE of the same registers is a constant
&& ! FLOAT_MODE_P (GET_MODE (XEXP (x, 1))))
return true;
- return CONSTANT_P (x);
+ /* Since X might be inserted more than once we have to take care that it
+ is sharable. */
+ return CONSTANT_P (x) && (GET_CODE (x) != CONST || shared_const_p (x));
}
/* Scan pattern PAT of INSN and add an entry to the hash TABLE (set or
expression one). */
static void
-hash_scan_set (rtx pat, rtx insn, struct hash_table *table)
+hash_scan_set (rtx pat, rtx insn, struct hash_table_d *table)
{
rtx src = SET_SRC (pat);
rtx dest = SET_DEST (pat);
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 (table->set_p && (note = find_reg_equal_equiv_note (insn)) != 0
- && gcse_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. */
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))))
+ && (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
modified. Here we want to search from INSN+1 on, but
oprs_available_p searches from INSN on. */
&& (insn == BB_END (BLOCK_FOR_INSN (insn))
- || ((tmp = next_nonnote_insn (insn)) != NULL_RTX
- && oprs_available_p (pat, tmp))))
+ || (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
static void
hash_scan_clobber (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
- struct hash_table *table ATTRIBUTE_UNUSED)
+ struct hash_table_d *table ATTRIBUTE_UNUSED)
{
/* Currently nothing to do. */
}
static void
hash_scan_call (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
- struct hash_table *table ATTRIBUTE_UNUSED)
+ struct hash_table_d *table ATTRIBUTE_UNUSED)
{
/* Currently nothing to do. */
}
are also in the PARALLEL. Later.
If SET_P is nonzero, 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. */
+ otherwise it is for the expression hash table. */
static void
-hash_scan_insn (rtx insn, struct hash_table *table, int in_libcall_block)
+hash_scan_insn (rtx insn, struct hash_table_d *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. */
}
static void
-dump_hash_table (FILE *file, const char *name, struct hash_table *table)
+dump_hash_table (FILE *file, const char *name, struct hash_table_d *table)
{
int i;
/* Flattened out table, so it's printed in proper order. */
unsigned int *hash_val;
struct expr *expr;
- flat_table = xcalloc (table->n_elems, sizeof (struct expr *));
- hash_val = xmalloc (table->n_elems * sizeof (unsigned int));
+ flat_table = XCNEWVEC (struct expr *, table->n_elems);
+ hash_val = XNEWVEC (unsigned int, table->n_elems);
for (i = 0; i < (int) table->size; i++)
for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
is set and is used to compute "availability".
last_bb records the block for which first_set and last_set are
- valid, as a quick test to invalidate them.
-
- reg_set_in_block records whether the register is set in the block
- and is used to compute "transparency". */
+ valid, as a quick test to invalidate them. */
static void
record_last_reg_set_info (rtx insn, int regno)
{
struct reg_avail_info *info = ®_avail_info[regno];
- int cuid = INSN_CUID (insn);
+ int luid = DF_INSN_LUID (insn);
- info->last_set = cuid;
+ info->last_set = luid;
if (info->last_bb != current_bb)
{
info->last_bb = current_bb;
- info->first_set = cuid;
- SET_BIT (reg_set_in_block[current_bb->index], regno);
+ info->first_set = luid;
}
}
taken off pairwise. */
static void
-canon_list_insert (rtx dest ATTRIBUTE_UNUSED, rtx unused1 ATTRIBUTE_UNUSED,
+canon_list_insert (rtx dest ATTRIBUTE_UNUSED, const_rtx unused1 ATTRIBUTE_UNUSED,
void * v_insn)
{
rtx dest_addr, insn;
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);
alloc_EXPR_LIST (VOIDmode, dest_addr, canon_modify_mem_list[bb]);
canon_modify_mem_list[bb] =
alloc_EXPR_LIST (VOIDmode, dest, canon_modify_mem_list[bb]);
- bitmap_set_bit (canon_modify_mem_list_set, bb);
}
/* Record memory modification information for INSN. We do not actually care
need to insert a pair of items, as canon_list_insert does. */
canon_modify_mem_list[bb] =
alloc_INSN_LIST (insn, canon_modify_mem_list[bb]);
- bitmap_set_bit (canon_modify_mem_list_set, bb);
+ bitmap_set_bit (blocks_with_calls, bb);
}
else
note_stores (PATTERN (insn), canon_list_insert, (void*) insn);
the SET is taking place. */
static void
-record_last_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED, void *data)
+record_last_set_info (rtx dest, const_rtx setter ATTRIBUTE_UNUSED, void *data)
{
rtx last_set_insn = (rtx) data;
TABLE is the table computed. */
static void
-compute_hash_table_work (struct hash_table *table)
+compute_hash_table_work (struct hash_table_d *table)
{
- unsigned int i;
-
- /* While we compute the hash table we also compute a bit array of which
- 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, last_basic_block);
+ int i;
/* re-Cache any INSN_LIST nodes we have allocated. */
clear_modify_mem_tables ();
/* Some working arrays used to track first and last set in each block. */
- reg_avail_info = gmalloc (max_gcse_regno * sizeof (struct reg_avail_info));
+ reg_avail_info = GNEWVEC (struct reg_avail_info, max_reg_num ());
- for (i = 0; i < max_gcse_regno; ++i)
+ for (i = 0; i < max_reg_num (); ++i)
reg_avail_info[i].last_bb = NULL;
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 = BB_HEAD (current_bb);
- insn && insn != NEXT_INSN (BB_END (current_bb));
- insn = NEXT_INSN (insn))
+ determine when registers and memory are first and last set. */
+ FOR_BB_INSNS (current_bb, insn)
{
if (! INSN_P (insn))
continue;
BB_HEAD (current_bb), table);
/* The next pass builds the hash table. */
-
- for (insn = BB_HEAD (current_bb), in_libcall_block = 0;
- insn && insn != NEXT_INSN (BB_END (current_bb));
- insn = NEXT_INSN (insn))
+ 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 (table->set_p && find_reg_note (insn, REG_RETVAL, NULL_RTX))
- in_libcall_block = 0;
- hash_scan_insn (insn, table, in_libcall_block);
- if (!table->set_p && find_reg_note (insn, REG_RETVAL, NULL_RTX))
- in_libcall_block = 0;
- }
+ hash_scan_insn (insn, table);
}
free (reg_avail_info);
be created. */
static void
-alloc_hash_table (int n_insns, struct hash_table *table, int set_p)
+alloc_hash_table (int n_insns, struct hash_table_d *table, int set_p)
{
int n;
??? Later take some measurements. */
table->size |= 1;
n = table->size * sizeof (struct expr *);
- table->table = gmalloc (n);
+ table->table = GNEWVAR (struct expr *, n);
table->set_p = set_p;
}
/* Free things allocated by alloc_hash_table. */
static void
-free_hash_table (struct hash_table *table)
+free_hash_table (struct hash_table_d *table)
{
free (table->table);
}
expression hash table. */
static void
-compute_hash_table (struct hash_table *table)
+compute_hash_table (struct hash_table_d *table)
{
/* Initialize count of number of entries in hash table. */
table->n_elems = 0;
table entry, or NULL if not found. */
static struct expr *
-lookup_set (unsigned int regno, struct hash_table *table)
+lookup_set (unsigned int regno, struct hash_table_d *table)
{
unsigned int hash = hash_set (regno, table->size);
struct expr *expr;
EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set, 0, i, bi)
{
free_INSN_LIST_list (modify_mem_list + i);
- }
- bitmap_clear (modify_mem_list_set);
-
- EXECUTE_IF_SET_IN_BITMAP (canon_modify_mem_list_set, 0, i, bi)
- {
free_insn_expr_list_list (canon_modify_mem_list + i);
}
- bitmap_clear (canon_modify_mem_list_set);
+ bitmap_clear (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 (void)
INSN's basic block. */
static int
-oprs_not_set_p (rtx x, rtx 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:
case MEM:
if (load_killed_in_block_p (BLOCK_FOR_INSN (insn),
- INSN_CUID (insn), x, 0))
+ DF_INSN_LUID (insn), x, 0))
return 0;
else
return oprs_not_set_p (XEXP (x, 0), insn);
static void
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);
}
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);
bit in BMAP. */
static void
-compute_transp (rtx x, int indx, sbitmap *bmap, int set_p)
+compute_transp (const_rtx x, int indx, sbitmap *bmap, int set_p)
{
int i, j;
- basic_block bb;
enum rtx_code code;
- reg_set *r;
const char *fmt;
/* repeat is used to turn tail-recursion into iteration since GCC
case REG:
if (set_p)
{
- if (REGNO (x) < FIRST_PSEUDO_REGISTER)
- {
- FOR_EACH_BB (bb)
- if (TEST_BIT (reg_set_in_block[bb->index], REGNO (x)))
- SET_BIT (bmap[bb->index], indx);
- }
- else
- {
- for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
- SET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
- }
+ df_ref def;
+ for (def = DF_REG_DEF_CHAIN (REGNO (x));
+ def;
+ def = DF_REF_NEXT_REG (def))
+ SET_BIT (bmap[DF_REF_BB (def)->index], indx);
}
else
{
- 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[BLOCK_NUM (r->insn)], indx);
- }
+ df_ref def;
+ for (def = DF_REG_DEF_CHAIN (REGNO (x));
+ def;
+ def = DF_REF_NEXT_REG (def))
+ RESET_BIT (bmap[DF_REF_BB (def)->index], indx);
}
return;
case MEM:
- FOR_EACH_BB (bb)
+ if (! MEM_READONLY_P (x))
{
- rtx list_entry = canon_modify_mem_list[bb->index];
+ bitmap_iterator bi;
+ unsigned bb_index;
- while (list_entry)
+ /* 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)
{
- rtx dest, dest_addr;
+ if (set_p)
+ SET_BIT (bmap[bb_index], indx);
+ else
+ RESET_BIT (bmap[bb_index], indx);
+ }
- if (CALL_P (XEXP (list_entry, 0)))
- {
- if (set_p)
- SET_BIT (bmap[bb->index], indx);
- else
- RESET_BIT (bmap[bb->index], indx);
- break;
- }
- /* LIST_ENTRY must be an INSN of some kind that sets memory.
- Examine each hunk of memory that is modified. */
+ /* Now iterate over the blocks which have memory modifications
+ but which do not have any calls. */
+ EXECUTE_IF_AND_COMPL_IN_BITMAP (modify_mem_list_set,
+ blocks_with_calls,
+ 0, bb_index, bi)
+ {
+ rtx list_entry = canon_modify_mem_list[bb_index];
- dest = XEXP (list_entry, 0);
- list_entry = XEXP (list_entry, 1);
- dest_addr = XEXP (list_entry, 0);
+ while (list_entry)
+ {
+ rtx dest, dest_addr;
- if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
- x, rtx_addr_varies_p))
- {
- if (set_p)
- SET_BIT (bmap[bb->index], indx);
- else
- RESET_BIT (bmap[bb->index], indx);
- break;
- }
- list_entry = XEXP (list_entry, 1);
- }
+ /* LIST_ENTRY must be an INSN of some kind that sets memory.
+ Examine each hunk of memory that is modified. */
+
+ dest = XEXP (list_entry, 0);
+ list_entry = XEXP (list_entry, 1);
+ dest_addr = XEXP (list_entry, 0);
+
+ if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
+ x, 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);
+ }
+ }
}
x = XEXP (x, 0);
case CONST:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case SYMBOL_REF:
case LABEL_REF:
int success = 0;
rtx set = single_set (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;
validate_change (insn, &SET_SRC (set), src, 0);
}
- /* If there is already a NOTE, update the expression in it with our
- replacement. */
- if (note != 0)
- XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
-
+ /* If 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
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 (XEXP (set, 0)) != ZERO_EXTRACT
- && GET_CODE (XEXP (set, 0)) != SIGN_EXTRACT)
+ && 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));
}
We don't allow that. Remove that note. This code ought
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;
static int
cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
{
- rtx new, set_src, note_src;
+ rtx new_rtx, set_src, note_src;
rtx set = pc_set (jump);
rtx note = find_reg_equal_equiv_note (jump);
else
setcc = NULL_RTX;
- new = simplify_replace_rtx (set_src, from, src);
+ 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, SET_SRC (set)))
+ 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)
+ if (new_rtx == pc_rtx)
delete_insn (jump);
else
{
/* Ensure the value computed inside the jump insn to be equivalent
to one computed by setcc. */
- if (setcc && modified_in_p (new, setcc))
+ if (setcc && modified_in_p (new_rtx, setcc))
return 0;
- if (! validate_change (jump, &SET_SRC (set), new, 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
we need to attach a note to the branch itself to make this
optimization work. */
- if (!rtx_equal_p (new, note_src))
- set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new));
+ if (!rtx_equal_p (new_rtx, note_src))
+ set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new_rtx));
return 0;
}
/* Remove REG_EQUAL note after simplification. */
if (note_src)
remove_note (jump, note);
-
- /* 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 (jump);
}
#ifdef HAVE_cc0
delete_insn (setcc);
#endif
- run_jump_opt_after_gcse = 1;
-
global_const_prop_count++;
- if (gcse_file != NULL)
+ if (dump_file != NULL)
{
- fprintf (gcse_file,
+ fprintf (dump_file,
"GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
REGNO (from), INSN_UID (jump));
- print_rtl (gcse_file, src);
- fprintf (gcse_file, "\n");
+ 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;
}
static bool
-constprop_register (rtx insn, rtx from, rtx to, int alter_jumps)
+constprop_register (rtx insn, rtx from, rtx to)
{
rtx sset;
/* Check for reg or cc0 setting instructions followed by
conditional branch instructions first. */
- if (alter_jumps
- && (sset = single_set (insn)) != NULL
+ if ((sset = single_set (insn)) != NULL
&& NEXT_INSN (insn)
&& any_condjump_p (NEXT_INSN (insn)) && onlyjump_p (NEXT_INSN (insn)))
{
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))
+ else if (any_condjump_p (insn) && onlyjump_p (insn))
return cprop_jump (BLOCK_FOR_INSN (insn), NULL, insn, from, to);
return 0;
}
The result is nonzero if a change was made. */
static int
-cprop_insn (rtx insn, int alter_jumps)
+cprop_insn (rtx insn)
{
struct reg_use *reg_used;
int changed = 0;
rtx pat, src;
struct expr *set;
- /* Ignore registers created by GCSE.
- We do this because ... */
- if (regno >= max_gcse_regno)
- continue;
-
/* If the register has already been set in this block, there's
nothing we can do. */
if (! oprs_not_set_p (reg_used->reg_rtx, insn))
/* Constant propagation. */
if (gcse_constant_p (src))
{
- if (constprop_register (insn, reg_used->reg_rtx, src, alter_jumps))
+ if (constprop_register (insn, reg_used->reg_rtx, src))
{
changed = 1;
global_const_prop_count++;
- if (gcse_file != NULL)
+ if (dump_file != NULL)
{
- fprintf (gcse_file, "GLOBAL 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");
}
if (INSN_DELETED_P (insn))
return 1;
{
changed = 1;
global_copy_prop_count++;
- if (gcse_file != NULL)
+ if (dump_file != NULL)
{
- fprintf (gcse_file, "GLOBAL 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
find_used_regs (xptr, data);
}
-/* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
- their REG_EQUAL notes need updating. */
+/* Try to perform local const/copy propagation on X in INSN. */
static bool
-do_local_cprop (rtx x, rtx insn, int alter_jumps, rtx *libcall_sp)
+do_local_cprop (rtx x, rtx insn)
{
rtx newreg = NULL, newcnst = NULL;
rtx this_rtx = l->loc;
rtx note;
- /* Don't CSE non-constant values out of libcall blocks. */
- if (l->in_libcall && ! CONSTANT_P (this_rtx))
- continue;
-
if (gcse_constant_p (this_rtx))
newcnst = this_rtx;
if (REG_P (this_rtx) && REGNO (this_rtx) >= FIRST_PSEUDO_REGISTER
|| ! MEM_P (XEXP (note, 0))))
newreg = this_rtx;
}
- if (newcnst && constprop_register (insn, x, newcnst, alter_jumps))
+ if (newcnst && constprop_register (insn, x, newcnst))
{
- /* If we find a case where we can't fix the retval REG_EQUAL notes
- match the new register, we either have to abandon this replacement
- or fix delete_trivially_dead_insns to preserve the setting insn,
- or make it delete the REG_EUAQL note, and fix up all passes that
- require the REG_EQUAL note there. */
- bool adjusted;
-
- adjusted = adjust_libcall_notes (x, newcnst, insn, libcall_sp);
- gcc_assert (adjusted);
-
- if (gcse_file != NULL)
+ if (dump_file != NULL)
{
- fprintf (gcse_file, "LOCAL CONST-PROP: Replacing reg %d in ",
+ fprintf (dump_file, "LOCAL CONST-PROP: Replacing reg %d in ",
REGNO (x));
- fprintf (gcse_file, "insn %d with constant ",
+ fprintf (dump_file, "insn %d with constant ",
INSN_UID (insn));
- print_rtl (gcse_file, newcnst);
- fprintf (gcse_file, "\n");
+ 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))
{
- adjust_libcall_notes (x, newreg, insn, libcall_sp);
- if (gcse_file != NULL)
+ if (dump_file != NULL)
{
- fprintf (gcse_file,
+ fprintf (dump_file,
"LOCAL COPY-PROP: Replacing reg %d in insn %d",
REGNO (x), INSN_UID (insn));
- fprintf (gcse_file, " with reg %d\n", REGNO (newreg));
+ fprintf (dump_file, " with reg %d\n", REGNO (newreg));
}
local_copy_prop_count++;
return true;
return false;
}
-/* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
- their REG_EQUAL notes need updating to reflect that OLDREG has been
- replaced with NEWVAL in INSN. Return true if all substitutions could
- be made. */
-static bool
-adjust_libcall_notes (rtx oldreg, rtx newval, rtx insn, rtx *libcall_sp)
+/* Do local const/copy propagation (i.e. within each basic block). */
+
+static int
+local_cprop_pass (void)
{
- rtx end;
+ basic_block bb;
+ rtx insn;
+ struct reg_use *reg_used;
+ bool changed = false;
- while ((end = *libcall_sp++))
+ cselib_init (false);
+ FOR_EACH_BB (bb)
{
- rtx note = find_reg_equal_equiv_note (end);
-
- if (! note)
- continue;
-
- if (REG_P (newval))
+ FOR_BB_INSNS (bb, insn)
{
- if (reg_set_between_p (newval, PREV_INSN (insn), end))
+ if (INSN_P (insn))
{
+ rtx note = find_reg_equal_equiv_note (insn);
do
{
- note = find_reg_equal_equiv_note (end);
- if (! note)
- continue;
- if (reg_mentioned_p (newval, XEXP (note, 0)))
- return false;
- }
- while ((end = *libcall_sp++));
- return true;
- }
- }
- XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), oldreg, newval);
- insn = end;
- }
- return true;
-}
-
-#define MAX_NESTED_LIBCALLS 9
-
-static void
-local_cprop_pass (int alter_jumps)
-{
- rtx insn;
- struct reg_use *reg_used;
- rtx libcall_stack[MAX_NESTED_LIBCALLS + 1], *libcall_sp;
- bool changed = false;
-
- cselib_init (false);
- libcall_sp = &libcall_stack[MAX_NESTED_LIBCALLS];
- *libcall_sp = 0;
- for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
- {
- if (INSN_P (insn))
- {
- rtx note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
-
- if (note)
- {
- gcc_assert (libcall_sp != libcall_stack);
- *--libcall_sp = XEXP (note, 0);
- }
- note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
- if (note)
- libcall_sp++;
- 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,
- libcall_sp))
- {
- changed = true;
+ 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))
+ {
+ changed = true;
+ break;
+ }
+ }
+ if (INSN_DELETED_P (insn))
break;
- }
- if (INSN_DELETED_P (insn))
- break;
+ }
+ while (reg_use_count);
}
- while (reg_use_count);
+ cselib_process_insn (insn);
}
- cselib_process_insn (insn);
- }
- 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 (get_insns ());
- }
-}
-
-/* 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 (gcse_file != NULL)
- fprintf (gcse_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 (insn = BB_HEAD (bb);
- insn != NULL && insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (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);
- }
+ /* Forget everything at the end of a basic block. */
+ cselib_clear_table ();
}
- if (gcse_file != NULL)
- fprintf (gcse_file, "\n");
+ cselib_finish ();
return changed;
}
settle for the condition variable in the jump instruction being integral.
We prefer to be able to record the value of a user variable, rather than
the value of a temporary used in a condition. This could be solved by
- recording the value of *every* register scaned by canonicalize_condition,
+ recording the value of *every* register scanned by canonicalize_condition,
but this would require some code reorganization. */
rtx
it. COND is either an EQ or NE comparison. */
static bool
-implicit_set_cond_p (rtx cond)
+implicit_set_cond_p (const_rtx cond)
{
- enum machine_mode mode = GET_MODE (XEXP (cond, 0));
- rtx cst = XEXP (cond, 1);
+ 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. */
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. */
+ basic block.
+
+ FIXME: This would be more effective if critical edges are pre-split. As
+ it is now, we can't record implicit sets for blocks that have
+ critical successor edges. This results in missed optimizations
+ and in more (unnecessary) work in cfgcleanup.c:thread_jump(). */
static void
find_implicit_sets (void)
{
basic_block bb, dest;
unsigned int count;
- rtx cond, new;
+ rtx cond, new_rtx;
count = 0;
FOR_EACH_BB (bb)
dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
: FALLTHRU_EDGE (bb)->dest;
- if (dest && EDGE_COUNT (dest->preds) == 1
+ if (dest
+ /* Record nothing for a critical edge. */
+ && single_pred_p (dest)
&& dest != EXIT_BLOCK_PTR)
{
- new = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
+ new_rtx = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
XEXP (cond, 1));
- implicit_sets[dest->index] = new;
- if (gcse_file)
+ implicit_sets[dest->index] = new_rtx;
+ if (dump_file)
{
- fprintf(gcse_file, "Implicit set of reg %d in ",
+ fprintf(dump_file, "Implicit set of reg %d in ",
REGNO (XEXP (cond, 0)));
- fprintf(gcse_file, "basic block %d\n", dest->index);
+ fprintf(dump_file, "basic block %d\n", dest->index);
}
count++;
}
}
}
- if (gcse_file)
- fprintf (gcse_file, "Found %d implicit sets\n", 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, int cprop_jumps, int bypass_jumps)
-{
- int changed = 0;
-
- global_const_prop_count = local_const_prop_count = 0;
- global_copy_prop_count = local_copy_prop_count = 0;
-
- local_cprop_pass (cprop_jumps);
-
- /* Determine implicit sets. */
- implicit_sets = xcalloc (last_basic_block, sizeof (rtx));
- 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 (gcse_file)
- dump_hash_table (gcse_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 (gcse_file)
- {
- fprintf (gcse_file, "CPROP of %s, pass %d: %d bytes needed, ",
- current_function_name (), pass, bytes_used);
- fprintf (gcse_file, "%d local const props, %d local copy props\n\n",
- local_const_prop_count, local_copy_prop_count);
- fprintf (gcse_file, "%d global const props, %d global copy props\n\n",
- global_const_prop_count, global_copy_prop_count);
- }
- /* Global analysis may get into infinite loops for unreachable blocks. */
- if (changed && cprop_jumps)
- delete_unreachable_blocks ();
-
- return changed;
-}
-\f
/* Bypass conditional jumps. */
/* The value of last_basic_block at the beginning of the jump_bypass
valid prior to commit_edge_insertions. */
static bool
-reg_killed_on_edge (rtx reg, edge e)
+reg_killed_on_edge (const_rtx reg, const_edge e)
{
rtx insn;
unsigned int regno = REGNO (reg_used->reg_rtx);
basic_block dest, old_dest;
struct expr *set;
- rtx src, new;
-
- if (regno >= max_gcse_regno)
- continue;
+ rtx src, new_rtx;
set = find_bypass_set (regno, e->src->index);
SET_DEST (PATTERN (setcc)),
SET_SRC (PATTERN (setcc)));
- new = simplify_replace_rtx (src, reg_used->reg_rtx,
+ new_rtx = simplify_replace_rtx (src, reg_used->reg_rtx,
SET_SRC (set->expr));
/* Jump bypassing may have already placed instructions on
has instructions associated with it, as these insns won't
get executed if the incoming edge is redirected. */
- if (new == pc_rtx)
+ if (new_rtx == pc_rtx)
{
edest = FALLTHRU_EDGE (bb);
dest = edest->insns.r ? NULL : edest->dest;
}
- else if (GET_CODE (new) == LABEL_REF)
+ else if (GET_CODE (new_rtx) == LABEL_REF)
{
- edge_iterator ei2;
-
- dest = BLOCK_FOR_INSN (XEXP (new, 0));
+ dest = BLOCK_FOR_INSN (XEXP (new_rtx, 0));
/* Don't bypass edges containing instructions. */
- FOR_EACH_EDGE (edest, ei2, bb->succs)
- if (edest->dest == dest && edest->insns.r)
- {
- dest = NULL;
- break;
- }
+ edest = find_edge (bb, dest);
+ if (edest && edest->insns.r)
+ dest = NULL;
}
else
dest = NULL;
branch. We would end up emitting the instruction on "both"
edges. */
- if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc))))
- {
- edge e2;
- edge_iterator ei2;
-
- FOR_EACH_EDGE (e2, ei2, e->src->succs)
- if (e2->dest == dest)
- {
- dest = NULL;
- break;
- }
- }
+ if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc)))
+ && find_edge (e->src, dest))
+ dest = NULL;
old_dest = e->dest;
if (dest != NULL
insert_insn_on_edge (copy_insn (pat), e);
}
- if (gcse_file != NULL)
+ if (dump_file != NULL)
{
- fprintf (gcse_file, "JUMP-BYPASS: Proved reg %d "
+ fprintf (dump_file, "JUMP-BYPASS: Proved reg %d "
"in jump_insn %d equals constant ",
regno, INSN_UID (jump));
- print_rtl (gcse_file, SET_SRC (set->expr));
- fprintf (gcse_file, "\nBypass edge from %d->%d to %d\n",
+ 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;
EXIT_BLOCK_PTR, next_bb)
{
/* Check for more than one predecessor. */
- if (EDGE_COUNT (bb->preds) > 1)
+ if (!single_pred_p (bb))
{
setcc = NULL_RTX;
- for (insn = BB_HEAD (bb);
- insn != NULL && insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
+ FOR_BB_INSNS (bb, insn)
if (NONJUMP_INSN_P (insn))
{
if (setcc)
/* 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();
+ commit_edge_insertions ();
return changed;
}
/* Contains the edge_list returned by pre_edge_lcm. */
static struct edge_list *edge_list;
-/* Redundant insns. */
-static sbitmap pre_redundant_insns;
-
/* Allocate vars used for PRE analysis. */
static void
sbitmap_not (ae_kill[bb->index], ae_kill[bb->index]);
}
- edge_list = pre_edge_lcm (gcse_file, expr_hash_table.n_elems, 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;
pre_expr_reaches_here_p (basic_block occr_bb, struct expr *expr, basic_block bb)
{
int rval;
- char *visited = xcalloc (last_basic_block, 1);
+ char *visited = XCNEWVEC (char, last_basic_block);
rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
no sense for code hoisting. */
static void
-insert_insn_end_bb (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 (bb);
rtx new_insn;
if (JUMP_P (insn)
|| (NONJUMP_INSN_P (insn)
- && (EDGE_COUNT (bb->succs) > 1
- || EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL)))
+ && (!single_succ_p (bb)
+ || single_succ_edge (bb)->flags & EDGE_ABNORMAL)))
{
#ifdef HAVE_cc0
rtx note;
}
#endif
/* FIXME: What if something in cc0/jump uses value set in new insn? */
- new_insn = emit_insn_before_noloc (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 (CALL_P (insn)
- && (EDGE_COUNT (bb->succs) > 1 || EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL))
+ && (!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
|| NOTE_INSN_BASIC_BLOCK_P (insn))
insn = NEXT_INSN (insn);
- new_insn = emit_insn_before_noloc (pat, insn);
+ new_insn = emit_insn_before_noloc (pat, insn, bb);
}
else
- new_insn = emit_insn_after_noloc (pat, insn);
+ new_insn = emit_insn_after_noloc (pat, insn, bb);
while (1)
{
if (INSN_P (pat))
- {
- add_label_notes (PATTERN (pat), new_insn);
- note_stores (PATTERN (pat), record_set_info, pat);
- }
+ add_label_notes (PATTERN (pat), 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);
}
}
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))
{
now. */
if (eg->flags & EDGE_ABNORMAL)
- insert_insn_end_bb (index_map[j], bb, 0);
+ 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: 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);
}
int regno = REGNO (reg);
int indx = expr->bitmap_index;
rtx pat = PATTERN (insn);
- rtx set, new_insn;
+ rtx set, first_set, new_insn;
rtx old_reg;
int i;
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
- && expr_equiv_p (SET_SRC (x), expr->expr))
+ if (GET_CODE (x) == SET)
{
- set = x;
- break;
+ /* 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:
{
new_insn = gen_move_insn (old_reg, reg);
new_insn = emit_insn_after (new_insn, insn);
-
- /* Keep register set table up to date. */
- replace_one_set (REGNO (old_reg), insn, new_insn);
- 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. */
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);
}
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);
continue;
/* Don't handle this one if it's a redundant one. */
- if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
+ if (INSN_DELETED_P (insn))
continue;
/* Or if the expression doesn't reach the deleted one. */
static rtx
gcse_emit_move_after (rtx src, rtx dest, rtx insn)
{
- rtx new;
+ 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 = emit_insn_after (gen_move_insn (dest, src), insn);
+ new_rtx = emit_insn_after (gen_move_insn (dest, src), insn);
/* Note the equivalence for local CSE pass. */
- set2 = single_set (new);
+ set2 = single_set (new_rtx);
if (!set2 || !rtx_equal_p (SET_DEST (set2), dest))
- return new;
+ 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, REG_EQUAL, copy_insn_1 (eqv));
+ set_unique_reg_note (new_rtx, REG_EQUAL, copy_insn_1 (eqv));
- return new;
+ return new_rtx;
}
/* Delete redundant computations.
/* We only delete insns that have a single_set. */
if (TEST_BIT (pre_delete_map[bb->index], indx)
- && (set = single_set (insn)) != 0)
+ && (set = single_set (insn)) != 0
+ && dbg_cnt (pre_insn))
{
/* 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)));
+ expr->reaching_reg = 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;
- SET_BIT (pre_redundant_insns, INSN_CUID (insn));
changed = 1;
gcse_subst_count++;
- if (gcse_file)
+ 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));
}
}
/* Compute a mapping from expression number (`bitmap_index') to
hash table entry. */
- index_map = xcalloc (expr_hash_table.n_elems, sizeof (struct expr *));
+ 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. */
- pre_redundant_insns = sbitmap_alloc (max_cuid);
- sbitmap_zero (pre_redundant_insns);
-
/* Delete the redundant insns first so that
- we know what register to use for the new insns and for the other
ones with reaching expressions
- 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
}
free (index_map);
- sbitmap_free (pre_redundant_insns);
return changed;
}
Return nonzero if a change was made. */
static int
-one_pre_gcse_pass (int pass)
+one_pre_gcse_pass (void)
{
int changed = 0;
gcse_subst_count = 0;
gcse_create_count = 0;
- alloc_hash_table (max_cuid, &expr_hash_table, 0);
+ /* Return if there's nothing to do, or it is too expensive. */
+ if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
+ || is_too_expensive (_("PRE disabled")))
+ return 0;
+
+ /* We need alias. */
+ init_alias_analysis ();
+
+ bytes_used = 0;
+ gcc_obstack_init (&gcse_obstack);
+ alloc_gcse_mem ();
+
+ alloc_hash_table (get_max_uid (), &expr_hash_table, 0);
add_noreturn_fake_exit_edges ();
if (flag_gcse_lm)
compute_ld_motion_mems ();
compute_hash_table (&expr_hash_table);
trim_ld_motion_mems ();
- if (gcse_file)
- dump_hash_table (gcse_file, "Expression", &expr_hash_table);
+ if (dump_file)
+ dump_hash_table (dump_file, "Expression", &expr_hash_table);
if (expr_hash_table.n_elems > 0)
{
remove_fake_exit_edges ();
free_hash_table (&expr_hash_table);
- if (gcse_file)
+ free_gcse_mem ();
+ obstack_free (&gcse_obstack, NULL);
+
+ /* We are finished with alias. */
+ end_alias_analysis ();
+
+ 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, "PRE GCSE of %s, %d basic blocks, %d bytes needed, ",
+ current_function_name (), n_basic_blocks, 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 (rtx x, rtx insn)
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))++;
+
return;
}
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 (! CALL_P (BB_END (bb)))
the convergence. */
FOR_EACH_BB_REVERSE (bb)
{
- changed |= sbitmap_a_or_b_and_c_cg (hoist_vbein[bb->index], antloc[bb->index],
- hoist_vbeout[bb->index], transp[bb->index]);
if (bb->next_bb != EXIT_BLOCK_PTR)
- sbitmap_intersection_of_succs (hoist_vbeout[bb->index], hoist_vbein, bb->index);
+ 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. */
compute_transpout ();
compute_code_hoist_vbeinout ();
calculate_dominance_info (CDI_DOMINATORS);
- if (gcse_file)
- fprintf (gcse_file, "\n");
+ if (dump_file)
+ fprintf (dump_file, "\n");
}
/* Determine if the expression identified by EXPR_INDEX would
if (visited == NULL)
{
visited_allocated_locally = 1;
- visited = xcalloc (last_basic_block, 1);
+ visited = XCNEWVEC (char, last_basic_block);
}
FOR_EACH_EDGE (pred, ei, bb->preds)
\f
/* Actually perform code hoisting. */
-static void
+static int
hoist_code (void)
{
basic_block bb, dominated;
- basic_block *domby;
- unsigned int domby_len;
+ VEC (basic_block, heap) *domby;
unsigned int i,j;
struct expr **index_map;
struct expr *expr;
+ int changed = 0;
sbitmap_vector_zero (hoist_exprs, last_basic_block);
/* Compute a mapping from expression number (`bitmap_index') to
hash table entry. */
- index_map = xcalloc (expr_hash_table.n_elems, sizeof (struct expr *));
+ 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;
int found = 0;
int insn_inserted_p;
- domby_len = get_dominated_by (CDI_DOMINATORS, bb, &domby);
+ 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->index]->n_bits; i++)
/* We've found a potentially hoistable expression, now
we look at every block BB dominates to see if it
computes the expression. */
- for (j = 0; j < domby_len; j++)
+ for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
{
- dominated = domby[j];
/* Ignore self dominance. */
if (bb == dominated)
continue;
/* If we found nothing to hoist, then quit now. */
if (! found)
{
- free (domby);
- continue;
+ VEC_free (basic_block, heap, domby);
+ continue;
}
/* Loop over all the hoistable expressions. */
insn_inserted_p = 0;
/* These tests should be the same as the tests above. */
- if (TEST_BIT (hoist_vbeout[bb->index], 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 (j = 0; j < domby_len; j++)
+ for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
{
- dominated = domby[j];
/* Ignore self dominance. */
if (bb == dominated)
continue;
from the mode of the original destination pseudo. */
if (expr->reaching_reg == NULL)
expr->reaching_reg
- = gen_reg_rtx (GET_MODE (SET_DEST (set)));
+ = 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;
+ changed = 1;
+ gcse_subst_count++;
+
if (!insn_inserted_p)
{
- insert_insn_end_bb (index_map[i], bb, 0);
+ insert_insn_end_basic_block (index_map[i], bb, 0);
insn_inserted_p = 1;
}
}
}
}
}
- free (domby);
+ VEC_free (basic_block, heap, domby);
}
free (index_map);
+
+ return changed;
}
/* Top level routine to perform one code hoisting (aka unification) pass
{
int changed = 0;
- alloc_hash_table (max_cuid, &expr_hash_table, 0);
+ gcse_subst_count = 0;
+ gcse_create_count = 0;
+
+ /* 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;
+
+ /* We need alias. */
+ init_alias_analysis ();
+
+ bytes_used = 0;
+ gcc_obstack_init (&gcse_obstack);
+ alloc_gcse_mem ();
+
+ alloc_hash_table (get_max_uid (), &expr_hash_table, 0);
compute_hash_table (&expr_hash_table);
- if (gcse_file)
- dump_hash_table (gcse_file, "Code Hosting Expressions", &expr_hash_table);
+ if (dump_file)
+ dump_hash_table (dump_file, "Code Hosting Expressions", &expr_hash_table);
if (expr_hash_table.n_elems > 0)
{
alloc_code_hoist_mem (last_basic_block, expr_hash_table.n_elems);
compute_code_hoist_data ();
- hoist_code ();
+ changed = hoist_code ();
free_code_hoist_mem ();
}
free_hash_table (&expr_hash_table);
+ free_gcse_mem ();
+ obstack_free (&gcse_obstack, NULL);
+
+ /* We are finished with alias. */
+ end_alias_analysis ();
+
+ if (dump_file)
+ {
+ fprintf (dump_file, "HOIST of %s, %d basic blocks, %d bytes needed, ",
+ current_function_name (), n_basic_blocks, bytes_used);
+ fprintf (dump_file, "%d substs, %d insns created\n",
+ gcse_subst_count, gcse_create_count);
+ }
return changed;
}
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. */
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);
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- if (ptr->hash_index == hash && expr_equiv_p (ptr->pattern, x))
- return ptr;
+ e.pattern = x;
+ slot = htab_find_slot_with_hash (pre_ldst_table, &e, hash, INSERT);
+ if (*slot)
+ return (struct ls_expr *)*slot;
- ptr = xmalloc (sizeof (struct ls_expr));
+ ptr = XNEW (struct ls_expr);
ptr->next = pre_ldst_mems;
ptr->expr = NULL;
ptr->index = 0;
ptr->hash_index = hash;
pre_ldst_mems = ptr;
+ *slot = ptr;
return ptr;
}
static void
free_ldst_mems (void)
{
+ if (pre_ldst_table)
+ htab_delete (pre_ldst_table);
+ pre_ldst_table = NULL;
+
while (pre_ldst_mems)
{
struct ls_expr * tmp = pre_ldst_mems;
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);
static struct ls_expr *
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;
-}
-
-/* Assign each element of the list of mems a monotonically increasing value. */
-
-static int
-enumerate_ldsts (void)
-{
- struct ls_expr * ptr;
- int n = 0;
-
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- ptr->index = n++;
-
- return n;
+ 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;
}
/* Return first item in the list. */
ld_motion list, otherwise we let the usual aliasing take care of it. */
static int
-simple_mem (rtx x)
+simple_mem (const_rtx x)
{
if (! MEM_P (x))
return 0;
rtx insn;
pre_ldst_mems = NULL;
+ pre_ldst_table = htab_create (13, pre_ldst_expr_hash,
+ pre_ldst_expr_eq, NULL);
FOR_EACH_BB (bb)
{
- for (insn = BB_HEAD (bb);
- insn && insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
+ FOR_BB_INSNS (bb, insn)
{
if (INSN_P (insn))
{
&& 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))
+ && can_assign_to_reg_without_clobbers_p (src))
ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
else
ptr->invalid = 1;
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
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, copy_rtx (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);
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. */
+/* Return true if the graph is too expensive to optimize. PASS is the
+ optimization about to be performed. */
-#define ANTIC_STORE_LIST(x) ((x)->loads)
-#define AVAIL_STORE_LIST(x) ((x)->stores)
-#define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
+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.
-/* 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 int * regvec;
+ 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);
-/* And current insn, for the same routine. */
-static rtx compute_store_table_current_insn;
+ return true;
+ }
-/* Used in computing the reverse edge graph bit vectors. */
-static sbitmap * st_antloc;
+ /* 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 ());
-/* Global holding the number of store expressions we are dealing with. */
-static int num_stores;
+ return true;
+ }
-/* Checks to set if we need to mark a register set. Called from
- note_stores. */
+ return false;
+}
-static void
-reg_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED,
- void *data)
+\f
+/* Main function for the CPROP pass. */
+
+static int
+one_cprop_pass (void)
{
- sbitmap bb_reg = data;
+ int changed = 0;
- if (GET_CODE (dest) == SUBREG)
- dest = SUBREG_REG (dest);
+ /* Return if there's nothing to do, or it is too expensive. */
+ if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
+ || is_too_expensive (_ ("const/copy propagation disabled")))
+ return 0;
- if (REG_P (dest))
+ global_const_prop_count = local_const_prop_count = 0;
+ global_copy_prop_count = local_copy_prop_count = 0;
+
+ bytes_used = 0;
+ gcc_obstack_init (&gcse_obstack);
+ alloc_gcse_mem ();
+
+ /* Do a local const/copy propagation pass first. The global pass
+ only handles global opportunities.
+ If the local pass changes something, remove any unreachable blocks
+ because the CPROP global dataflow analysis may get into infinite
+ loops for CFGs with unreachable blocks.
+
+ FIXME: This local pass should not be necessary after CSE (but for
+ some reason it still is). It is also (proven) not necessary
+ to run the local pass right after FWPWOP.
+
+ FIXME: The global analysis would not get into infinite loops if it
+ would use the DF solver (via df_simple_dataflow) instead of
+ the solver implemented in this file. */
+ if (local_cprop_pass ())
{
- regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
- if (bb_reg)
- SET_BIT (bb_reg, REGNO (dest));
+ delete_unreachable_blocks ();
+ df_analyze ();
}
-}
-/* Clear any mark that says that this insn sets dest. Called from
- note_stores. */
+ /* Determine implicit sets. */
+ implicit_sets = XCNEWVEC (rtx, last_basic_block);
+ find_implicit_sets ();
-static void
-reg_clear_last_set (rtx dest, rtx setter ATTRIBUTE_UNUSED,
- void *data)
-{
- int *dead_vec = data;
+ alloc_hash_table (get_max_uid (), &set_hash_table, 1);
+ compute_hash_table (&set_hash_table);
- if (GET_CODE (dest) == SUBREG)
- dest = SUBREG_REG (dest);
+ /* Free implicit_sets before peak usage. */
+ free (implicit_sets);
+ implicit_sets = NULL;
- if (REG_P (dest) &&
- dead_vec[REGNO (dest)] == INSN_UID (compute_store_table_current_insn))
- dead_vec[REGNO (dest)] = 0;
-}
+ if (dump_file)
+ dump_hash_table (dump_file, "SET", &set_hash_table);
+ if (set_hash_table.n_elems > 0)
+ {
+ basic_block bb;
+ rtx insn;
-/* Return zero if some of the registers in list X are killed
- due to set of registers in bitmap REGS_SET. */
+ alloc_cprop_mem (last_basic_block, set_hash_table.n_elems);
+ compute_cprop_data ();
-static bool
-store_ops_ok (rtx x, int *regs_set)
-{
- 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;
- const char * fmt;
-
- /* Repeat is used to turn tail-recursion into iteration. */
- repeat:
-
- if (x == 0)
- return accum;
-
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- return alloc_EXPR_LIST (0, x, accum);
-
- case MEM:
- x = XEXP (x, 0);
- goto repeat;
-
- case PRE_DEC:
- case PRE_INC:
- case POST_DEC:
- case POST_INC:
- /* 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_VECTOR:
- case SYMBOL_REF:
- case LABEL_REF:
- case ADDR_VEC:
- case ADDR_DIFF_VEC:
- return accum;
-
- default:
- break;
- }
-
- i = GET_RTX_LENGTH (code) - 1;
- fmt = GET_RTX_FORMAT (code);
-
- for (; i >= 0; i--)
- {
- if (fmt[i] == 'e')
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
{
- rtx tem = XEXP (x, i);
+ /* Reset tables used to keep track of what's still valid [since
+ the start of the block]. */
+ reset_opr_set_tables ();
- /* If we are about to do the last recursive call
- needed at this level, change it into iteration. */
- if (i == 0)
- {
- x = tem;
- goto repeat;
- }
-
- accum = extract_mentioned_regs_helper (tem, accum);
+ FOR_BB_INSNS (bb, insn)
+ if (INSN_P (insn))
+ {
+ changed |= cprop_insn (insn);
+
+ /* Keep track of everything modified by this insn. */
+ /* ??? Need to be careful w.r.t. mods done to INSN.
+ Don't call mark_oprs_set if we turned the
+ insn into a NOTE. */
+ if (! NOTE_P (insn))
+ mark_oprs_set (insn);
+ }
}
- else if (fmt[i] == 'E')
- {
- int j;
- for (j = 0; j < XVECLEN (x, i); j++)
- accum = extract_mentioned_regs_helper (XVECEXP (x, i, j), accum);
- }
+ changed |= bypass_conditional_jumps ();
+ free_cprop_mem ();
}
- return accum;
-}
-
-/* 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 (rtx insn, int *regs_set_before, int *regs_set_after)
-{
- struct ls_expr * ptr;
- rtx dest, set, tmp;
- int check_anticipatable, check_available;
- basic_block bb = BLOCK_FOR_INSN (insn);
-
- set = single_set (insn);
- if (!set)
- return;
-
- dest = SET_DEST (set);
-
- if (! MEM_P (dest) || MEM_VOLATILE_P (dest)
- || GET_MODE (dest) == BLKmode)
- 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;
-
- /* Even if the destination cannot trap, the source may. In this case we'd
- need to handle updating the REG_EH_REGION note. */
- if (find_reg_note (insn, REG_EH_REGION, NULL_RTX))
- return;
-
- ptr = ldst_entry (dest);
- if (!ptr->pattern_regs)
- ptr->pattern_regs = extract_mentioned_regs (dest);
+ free_hash_table (&set_hash_table);
+ free_gcse_mem ();
+ obstack_free (&gcse_obstack, NULL);
- /* 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
+ if (dump_file)
{
- 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));
+ fprintf (dump_file, "CPROP of %s, %d basic blocks, %d bytes needed, ",
+ current_function_name (), n_basic_blocks, 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);
}
- /* 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));
+ return changed;
}
-/* Find available and anticipatable stores. */
-
-static int
-compute_store_table (void)
-{
- int ret;
- basic_block bb;
- unsigned regno;
- 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_vector_alloc (last_basic_block,
- max_gcse_regno);
- sbitmap_vector_zero (reg_set_in_block, last_basic_block);
- pre_ldst_mems = 0;
- last_set_in = xcalloc (max_gcse_regno, sizeof (int));
- already_set = xmalloc (sizeof (int) * max_gcse_regno);
-
- /* Find all the stores we care about. */
- FOR_EACH_BB (bb)
- {
- /* First compute the registers set in this block. */
- regvec = last_set_in;
-
- for (insn = BB_HEAD (bb);
- insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
- {
- if (! INSN_P (insn))
- continue;
-
- 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);
- 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 (insn = BB_HEAD (bb);
- insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
- {
- if (! INSN_P (insn))
- continue;
-
- if (CALL_P (insn))
- {
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; 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. */
- 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;
- free_ldst_entry (ptr);
- }
- else
- prev_next_ptr_ptr = &ptr->next;
- }
-
- ret = enumerate_ldsts ();
-
- if (gcse_file)
- {
- fprintf (gcse_file, "ST_avail and ST_antic (shown under loads..)\n");
- print_ldst_list (gcse_file);
- }
-
- free (last_set_in);
- free (already_set);
- return ret;
-}
+\f
+/* All the passes implemented in this file. Each pass has its
+ own gate and execute function, and at the end of the file a
+ pass definition for passes.c.
-/* 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. */
+ We do not construct an accurate cfg in functions which call
+ setjmp, so none of these passes runs if the function calls
+ setjmp.
+ FIXME: Should just handle setjmp via REG_SETJMP notes. */
static bool
-load_kills_store (rtx x, rtx store_pattern, int after)
+gate_rtl_cprop (void)
{
- if (after)
- return anti_dependence (x, store_pattern);
- else
- return true_dependence (store_pattern, GET_MODE (store_pattern), x,
- rtx_addr_varies_p);
+ return optimize > 0 && flag_gcse
+ && !cfun->calls_setjmp
+ && dbg_cnt (cprop);
}
-/* 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 bool
-find_loads (rtx x, rtx store_pattern, int after)
+static unsigned int
+execute_rtl_cprop (void)
{
- const char * fmt;
- int i, j;
- int ret = false;
-
- if (!x)
- return false;
-
- if (GET_CODE (x) == SET)
- x = SET_SRC (x);
-
- if (MEM_P (x))
- {
- 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, after);
- else if (fmt[i] == 'E')
- for (j = XVECLEN (x, i) - 1; j >= 0; j--)
- ret |= find_loads (XVECEXP (x, i, j), store_pattern, after);
- }
- return ret;
+ delete_unreachable_blocks ();
+ df_note_add_problem ();
+ df_set_flags (DF_LR_RUN_DCE);
+ df_analyze ();
+ flag_rerun_cse_after_global_opts |= one_cprop_pass ();
+ return 0;
}
-/* 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 it does. */
-
static bool
-store_killed_in_insn (rtx x, rtx x_regs, rtx insn, int after)
+gate_rtl_pre (void)
{
- rtx reg, base, note;
-
- 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. */
- if (! CONST_OR_PURE_CALL_P (insn) || pure_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;
- }
-
- if (GET_CODE (PATTERN (insn)) == SET)
- {
- rtx pat = PATTERN (insn);
- rtx dest = SET_DEST (pat);
-
- if (GET_CODE (dest) == SIGN_EXTRACT
- || 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 (SET_SRC (pat), x, after))
- return true;
- }
- 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);
+ return optimize > 0 && flag_gcse
+ && !cfun->calls_setjmp
+ && optimize_function_for_speed_p (cfun)
+ && dbg_cnt (pre);
}
-/* 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 bool
-store_killed_after (rtx x, rtx x_regs, rtx insn, basic_block bb,
- int *regs_set_after, rtx *fail_insn)
+static unsigned int
+execute_rtl_pre (void)
{
- rtx last = BB_END (bb), act;
-
- if (!store_ops_ok (x_regs, regs_set_after))
- {
- /* We do not know where it will happen. */
- if (fail_insn)
- *fail_insn = NULL_RTX;
- return true;
- }
-
- /* 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;
+ delete_unreachable_blocks ();
+ df_note_add_problem ();
+ df_analyze ();
+ flag_rerun_cse_after_global_opts |= one_pre_gcse_pass ();
+ return 0;
}
-/* 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 (rtx x, rtx x_regs, rtx insn, basic_block bb,
- int *regs_set_before)
+gate_rtl_hoist (void)
{
- rtx first = BB_HEAD (bb);
-
- if (!store_ops_ok (x_regs, regs_set_before))
- return true;
-
- for ( ; insn != PREV_INSN (first); insn = PREV_INSN (insn))
- if (store_killed_in_insn (x, x_regs, insn, true))
- return true;
-
- return false;
+ return optimize > 0 && flag_gcse
+ && !cfun->calls_setjmp
+ /* It does not make sense to run code hoisting unless we are optimizing
+ for code size -- it rarely makes programs faster, and can make then
+ bigger if we did PRE (when optimizing for space, we don't run PRE). */
+ && optimize_function_for_size_p (cfun)
+ && dbg_cnt (hoist);
}
-/* Fill in available, anticipatable, transparent and kill vectors in
- STORE_DATA, based on lists of available and anticipatable stores. */
-static void
-build_store_vectors (void)
+static unsigned int
+execute_rtl_hoist (void)
{
- basic_block bb;
- int *regs_set_in_block;
- rtx insn, st;
- struct ls_expr * ptr;
- unsigned regno;
+ delete_unreachable_blocks ();
+ df_note_add_problem ();
+ df_analyze ();
+ flag_rerun_cse_after_global_opts |= one_code_hoisting_pass ();
+ return 0;
+}
- /* Build the gen_vector. This is any store in the table which is not killed
- by aliasing later in its block. */
- ae_gen = sbitmap_vector_alloc (last_basic_block, num_stores);
- sbitmap_vector_zero (ae_gen, last_basic_block);
+struct rtl_opt_pass pass_rtl_cprop =
+{
+ {
+ RTL_PASS,
+ "cprop", /* name */
+ gate_rtl_cprop, /* gate */
+ execute_rtl_cprop, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_CPROP, /* tv_id */
+ PROP_cfglayout, /* 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 */
+ }
+};
- st_antloc = sbitmap_vector_alloc (last_basic_block, num_stores);
- sbitmap_vector_zero (st_antloc, last_basic_block);
+struct rtl_opt_pass pass_rtl_pre =
+{
+ {
+ RTL_PASS,
+ "pre", /* name */
+ gate_rtl_pre, /* gate */
+ execute_rtl_pre, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_PRE, /* tv_id */
+ PROP_cfglayout, /* 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 */
+ }
+};
- for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
- {
- for (st = AVAIL_STORE_LIST (ptr); st != NULL; st = XEXP (st, 1))
- {
- insn = XEXP (st, 0);
- bb = BLOCK_FOR_INSN (insn);
-
- /* 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))
- {
- 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, ptr);
- continue;
- }
- 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 = xmalloc (sizeof (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))
- {
- /* 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 (regs_set_in_block);
-
- if (gcse_file)
- {
- dump_sbitmap_vector (gcse_file, "st_antloc", "", st_antloc, last_basic_block);
- dump_sbitmap_vector (gcse_file, "st_kill", "", ae_kill, last_basic_block);
- dump_sbitmap_vector (gcse_file, "Transpt", "", transp, last_basic_block);
- dump_sbitmap_vector (gcse_file, "st_avloc", "", ae_gen, last_basic_block);
- }
-}
-
-/* Insert an instruction at the beginning of a basic block, and update
- the BB_HEAD if needed. */
-
-static void
-insert_insn_start_bb (rtx insn, basic_block bb)
-{
- /* Insert at start of successor block. */
- rtx prev = PREV_INSN (BB_HEAD (bb));
- rtx before = BB_HEAD (bb);
- while (before != 0)
- {
- if (! LABEL_P (before)
- && (! NOTE_P (before)
- || NOTE_LINE_NUMBER (before) != NOTE_INSN_BASIC_BLOCK))
- break;
- prev = before;
- if (prev == BB_END (bb))
- break;
- before = NEXT_INSN (before);
- }
-
- insn = emit_insn_after_noloc (insn, prev);
-
- if (gcse_file)
- {
- fprintf (gcse_file, "STORE_MOTION insert store at start of BB %d:\n",
- bb->index);
- print_inline_rtx (gcse_file, insn, 6);
- fprintf (gcse_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 nonzero
- if an edge insertion was performed. */
-
-static int
-insert_store (struct ls_expr * expr, edge e)
-{
- rtx reg, insn;
- basic_block bb;
- edge tmp;
- edge_iterator ei;
-
- /* We did all the deleted before this insert, so if we didn't delete a
- store, then we haven't set the reaching reg yet either. */
- if (expr->reaching_reg == NULL_RTX)
- return 0;
-
- if (e->flags & EDGE_FAKE)
- return 0;
-
- reg = expr->reaching_reg;
- insn = gen_move_insn (copy_rtx (expr->pattern), reg);
-
- /* 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_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_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 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)
- {
- fprintf (gcse_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");
- }
-
- return 1;
-}
-
-/* Remove any REG_EQUAL or REG_EQUIV notes containing a reference to the
- memory location in SMEXPR set in basic block BB.
-
- This could be rather expensive. */
-
-static void
-remove_reachable_equiv_notes (basic_block bb, struct ls_expr *smexpr)
-{
- edge_iterator *stack, ei;
- int sp;
- edge act;
- sbitmap visited = sbitmap_alloc (last_basic_block);
- rtx last, insn, note;
- rtx mem = smexpr->pattern;
-
- stack = xmalloc (sizeof (edge_iterator) * n_basic_blocks);
- sp = 0;
- ei = ei_start (bb->succs);
-
- sbitmap_zero (visited);
-
- act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
- while (1)
- {
- if (!act)
- {
- if (!sp)
- {
- free (stack);
- sbitmap_free (visited);
- return;
- }
- act = ei_edge (stack[--sp]);
- }
- bb = act->dest;
-
- if (bb == EXIT_BLOCK_PTR
- || TEST_BIT (visited, bb->index))
- {
- if (!ei_end_p (ei))
- ei_next (&ei);
- act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
- continue;
- }
- SET_BIT (visited, bb->index);
-
- if (TEST_BIT (st_antloc[bb->index], smexpr->index))
- {
- for (last = ANTIC_STORE_LIST (smexpr);
- BLOCK_FOR_INSN (XEXP (last, 0)) != bb;
- last = XEXP (last, 1))
- continue;
- last = XEXP (last, 0);
- }
- else
- last = NEXT_INSN (BB_END (bb));
-
- for (insn = BB_HEAD (bb); insn != last; insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- {
- note = find_reg_equal_equiv_note (insn);
- if (!note || !expr_equiv_p (XEXP (note, 0), mem))
- continue;
-
- if (gcse_file)
- fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
- INSN_UID (insn));
- remove_note (insn, note);
- }
-
- if (!ei_end_p (ei))
- ei_next (&ei);
- act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
-
- if (EDGE_COUNT (bb->succs) > 0)
- {
- if (act)
- stack[sp++] = ei;
- ei = ei_start (bb->succs);
- act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
- }
- }
-}
-
-/* This routine will replace a store with a SET to a specified register. */
-
-static void
-replace_store_insn (rtx reg, rtx del, basic_block bb, struct ls_expr *smexpr)
-{
- rtx insn, mem, note, set, ptr, pair;
-
- mem = smexpr->pattern;
- insn = gen_move_insn (reg, SET_SRC (single_set (del)));
- insn = emit_insn_after (insn, del);
-
- if (gcse_file)
- {
- fprintf (gcse_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");
- }
-
- for (ptr = ANTIC_STORE_LIST (smexpr); ptr; ptr = XEXP (ptr, 1))
- if (XEXP (ptr, 0) == del)
- {
- XEXP (ptr, 0) = insn;
- break;
- }
-
- /* Move the notes from the deleted insn to its replacement, and patch
- up the LIBCALL notes. */
- REG_NOTES (insn) = REG_NOTES (del);
-
- note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
- if (note)
- {
- pair = XEXP (note, 0);
- note = find_reg_note (pair, REG_LIBCALL, NULL_RTX);
- XEXP (note, 0) = insn;
- }
- note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
- if (note)
- {
- pair = XEXP (note, 0);
- note = find_reg_note (pair, REG_RETVAL, NULL_RTX);
- XEXP (note, 0) = insn;
- }
-
- delete_insn (del);
-
- /* Now we must handle REG_EQUAL notes whose contents is equal to the mem;
- they are no longer accurate provided that they are reached by this
- definition, so drop them. */
- for (; insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- {
- set = single_set (insn);
- if (!set)
- continue;
- if (expr_equiv_p (SET_DEST (set), mem))
- return;
- note = find_reg_equal_equiv_note (insn);
- if (!note || !expr_equiv_p (XEXP (note, 0), mem))
- continue;
-
- if (gcse_file)
- fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
- INSN_UID (insn));
- remove_note (insn, note);
- }
- remove_reachable_equiv_notes (bb, smexpr);
-}
-
-
-/* Delete a store, but copy the value that would have been stored into
- the reaching_reg for later storing. */
-
-static void
-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));
-
- 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
- ones during the available computation. */
- replace_store_insn (reg, del, bb, expr);
- break;
- }
- }
-}
-
-/* Free memory used by store motion. */
-
-static void
-free_store_memory (void)
-{
- free_ldst_mems ();
-
- if (ae_gen)
- sbitmap_vector_free (ae_gen);
- if (ae_kill)
- sbitmap_vector_free (ae_kill);
- if (transp)
- sbitmap_vector_free (transp);
- if (st_antloc)
- sbitmap_vector_free (st_antloc);
- if (pre_insert_map)
- sbitmap_vector_free (pre_insert_map);
- if (pre_delete_map)
- 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;
-}
-
-/* Perform store motion. Much like gcse, except we move expressions the
- other way by looking at the flowgraph in reverse. */
-
-static void
-store_motion (void)
-{
- basic_block bb;
- int x;
- struct ls_expr * ptr;
- int update_flow = 0;
-
- if (gcse_file)
- {
- fprintf (gcse_file, "before store motion\n");
- print_rtl (gcse_file, get_insns ());
- }
-
- init_alias_analysis ();
-
- /* Find all the available and anticipatable stores. */
- num_stores = compute_store_table ();
- if (num_stores == 0)
- {
- sbitmap_vector_free (reg_set_in_block);
- end_alias_analysis ();
- return;
- }
-
- /* 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,
- &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))
- {
- /* If any of the edges we have above are abnormal, we can't move this
- store. */
- for (x = NUM_EDGES (edge_list) - 1; x >= 0; x--)
- if (TEST_BIT (pre_insert_map[x], ptr->index)
- && (INDEX_EDGE (edge_list, x)->flags & EDGE_ABNORMAL))
- break;
-
- if (x >= 0)
- {
- if (gcse_file != NULL)
- fprintf (gcse_file,
- "Can't replace store %d: abnormal edge from %d to %d\n",
- ptr->index, INDEX_EDGE (edge_list, x)->src->index,
- INDEX_EDGE (edge_list, x)->dest->index);
- continue;
- }
-
- /* Now we want to insert the new stores which are going to be needed. */
-
- FOR_EACH_BB (bb)
- if (TEST_BIT (pre_delete_map[bb->index], ptr->index))
- delete_store (ptr, bb);
-
- for (x = 0; x < NUM_EDGES (edge_list); x++)
- if (TEST_BIT (pre_insert_map[x], ptr->index))
- update_flow |= insert_store (ptr, INDEX_EDGE (edge_list, x));
- }
-
- if (update_flow)
- commit_edge_insertions ();
-
- free_store_memory ();
- free_edge_list (edge_list);
- remove_fake_exit_edges ();
- end_alias_analysis ();
-}
-
-\f
-/* Entry point for jump bypassing optimization pass. */
-
-int
-bypass_jumps (FILE *file)
-{
- int changed;
-
- /* We do not construct an accurate cfg in functions which call
- setjmp, so just punt to be safe. */
- if (current_function_calls_setjmp)
- return 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);
-
- /* Return if there's nothing to do, or it is too expensive. */
- if (n_basic_blocks <= 1 || is_too_expensive (_ ("jump bypassing disabled")))
- return 0;
-
- 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 (get_insns ());
-
- max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (get_insns ());
- changed = one_cprop_pass (MAX_GCSE_PASSES + 2, 1, 1);
- free_gcse_mem ();
-
- if (file)
- {
- fprintf (file, "BYPASS of %s: %d basic blocks, ",
- current_function_name (), n_basic_blocks);
- fprintf (file, "%d bytes\n\n", bytes_used);
- }
-
- 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);
-
- return changed;
-}
-
-/* Return true if the graph is too expensive to optimize. PASS is the
- optimization about to be performed. */
-
-static bool
-is_too_expensive (const char *pass)
-{
- /* Trying to perform global optimizations on flow graphs which have
- a high connectivity will take a long time and is unlikely to be
- particularly useful.
-
- In normal circumstances a cfg should have about twice as many
- edges as blocks. But we do not want to punish small functions
- which have a couple switch statements. Rather than simply
- threshold the number of blocks, uses something with a more
- graceful degradation. */
- if (n_edges > 20000 + n_basic_blocks * 4)
- {
- if (warn_disabled_optimization)
- warning ("%s: %d basic blocks and %d edges/basic block",
- pass, n_basic_blocks, n_edges / n_basic_blocks);
-
- return true;
- }
-
- /* If allocating memory for the cprop bitmap would take up too much
- storage it's better just to disable the optimization. */
- if ((n_basic_blocks
- * SBITMAP_SET_SIZE (max_reg_num ())
- * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
- {
- if (warn_disabled_optimization)
- warning ("%s: %d basic blocks and %d registers",
- pass, n_basic_blocks, max_reg_num ());
-
- return true;
- }
-
- return false;
-}
+struct rtl_opt_pass pass_rtl_hoist =
+{
+ {
+ RTL_PASS,
+ "hoist", /* name */
+ gate_rtl_hoist, /* gate */
+ execute_rtl_hoist, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_HOIST, /* tv_id */
+ PROP_cfglayout, /* 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"
OSDN Git Service
