/* Global common subexpression elimination/Partial redundancy elimination
and global constant/copy propagation for GNU compiler.
- Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003
- Free Software Foundation, Inc.
+ Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005,
+ 2006, 2007, 2008, 2009, 2010 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 "toplev.h"
#include "rtl.h"
+#include "tree.h"
#include "tm_p.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "ggc.h"
#include "params.h"
#include "cselib.h"
-
+#include "intl.h"
#include "obstack.h"
-
-/* Propagate flow information through back edges and thus enable PRE's
- moving loop invariant calculations out of loops.
-
- Originally this tended to create worse overall code, but several
- improvements during the development of PRE seem to have made following
- back edges generally a win.
-
- Note much of the loop invariant code motion done here would normally
- be done by loop.c, which has more heuristics for when to move invariants
- out of loops. At some point we might need to move some of those
- heuristics into gcse.c. */
+#include "timevar.h"
+#include "tree-pass.h"
+#include "hashtab.h"
+#include "df.h"
+#include "dbgcnt.h"
+#include "target.h"
/* We support GCSE via Partial Redundancy Elimination. PRE optimizations
- are a superset of those done by GCSE.
+ are a superset of those done by classic GCSE.
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.
+ 3) Perform global cse using lazy code motion if not optimizing
+ for size, or code hoisting if we are.
- 4) Perform global cse.
-
- 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.
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 GCSE-ing.
+ This allows PRE to hoist expressions that are expressed in multiple insns,
+ such as complex address calculations (e.g. for PIC code, or loads with a
+ high part and a low part).
+
PRE handles moving invariant expressions out of loops (by treating them as
partially redundant).
- Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
- assignment) based GVN (global value numbering). L. T. Simpson's paper
- (Rice University) on value numbering is a useful reference for this.
-
**********************
We used to support multiple passes but there are diminishing returns in
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
- awhile to converge. Hence we only perform one pass. The parameter max-gcse-passes can
- be modified if one wants to experiment.
+ a while to converge. Hence we only perform one pass.
**********************
argue it is not. The number of iterations for the algorithm to converge
is typically 2-4 so I don't view it as that expensive (relatively speaking).
- PRE GCSE depends heavily on the second CSE pass to clean up the copies
+ PRE GCSE depends heavily on the second CPROP pass to clean up the copies
we create. To make an expression reach the place where it's redundant,
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) (INSN_UID (INSN) > max_uid ? (abort (), 0) : 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.
Load motion tracks MEMs which aren't killed by
- anything except itself. (ie, loads and stores to a single location).
+ anything except itself. (i.e., loads and stores to a single location).
We can then allow movement of these MEM refs with a little special
allowance. (all stores copy the same value to the reaching reg used
for the loads). This means all values used to store into memory must have
struct ls_expr * next; /* Next in the list. */
int invalid; /* Invalid for some reason. */
int index; /* If it maps to a bitmap index. */
- int hash_index; /* Index when in a hash table. */
+ unsigned int hash_index; /* Index when in a hash table. */
rtx reaching_reg; /* Register to use when re-writing. */
};
/* Head of the list of load/store memory refs. */
static struct ls_expr * pre_ldst_mems = NULL;
+/* Hashtable for the load/store memory refs. */
+static htab_t pre_ldst_table = NULL;
+
/* Bitmap containing one bit for each register in the program.
Used when performing GCSE to track which registers have been set since
the start of the basic block. */
static regset reg_set_bitmap;
-/* For each block, a bitmap of registers set in the block.
- This is used by expr_killed_p and compute_transp.
- 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;
-bitmap modify_mem_list_set;
+static bitmap modify_mem_list_set;
/* This array parallels modify_mem_list, but is kept canonicalized. */
static rtx * canon_modify_mem_list;
-bitmap canon_modify_mem_list_set;
+
+/* Bitmap indexed by block numbers to record which blocks contain
+ function calls. */
+static bitmap blocks_with_calls;
+
/* Various variables for statistics gathering. */
/* Memory used in a pass.
static int gcse_subst_count;
/* Number of copy instructions created. */
static int gcse_create_count;
-/* Number of constants propagated. */
-static int const_prop_count;
-/* Number of copys propagated. */
-static int copy_prop_count;
+/* Number of local constants propagated. */
+static int local_const_prop_count;
+/* Number of local copies propagated. */
+static int local_copy_prop_count;
+/* Number of global constants propagated. */
+static int global_const_prop_count;
+/* Number of global copies propagated. */
+static int global_copy_prop_count;
\f
-/* These variables are used by classic GCSE.
- Normally they'd be defined a bit later, but `rd_gen' needs to
- be declared sooner. */
-
-/* Each block has a bitmap of each type.
- The length of each blocks bitmap is:
-
- max_cuid - for reaching definitions
- n_exprs - for available expressions
-
- Thus we view the bitmaps as 2 dimensional arrays. i.e.
- rd_kill[block_num][cuid_num]
- ae_kill[block_num][expr_num] */
-
-/* For reaching defs */
-static sbitmap *rd_kill, *rd_gen, *reaching_defs, *rd_out;
-
-/* for available exprs */
-static sbitmap *ae_kill, *ae_gen, *ae_in, *ae_out;
-
-/* Objects of this type are passed around by the null-pointer check
- removal routines. */
-struct null_pointer_info
-{
- /* The basic block being processed. */
- basic_block current_block;
- /* The first register to be handled in this pass. */
- unsigned int min_reg;
- /* One greater than the last register to be handled in this pass. */
- unsigned int max_reg;
- sbitmap *nonnull_local;
- sbitmap *nonnull_killed;
-};
+/* For available exprs */
+static sbitmap *ae_kill;
\f
static void compute_can_copy (void);
-static void *gmalloc (unsigned int);
-static void *grealloc (void *, unsigned int);
+static void *gmalloc (size_t) ATTRIBUTE_MALLOC;
+static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC;
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 int get_bitmap_width (int, int, int);
-static void record_one_set (int, rtx);
-static void record_set_info (rtx, rtx, void *);
-static void compute_sets (rtx);
-static void hash_scan_insn (rtx, struct hash_table *, int);
-static void hash_scan_set (rtx, rtx, struct hash_table *);
-static void hash_scan_clobber (rtx, rtx, struct hash_table *);
-static void hash_scan_call (rtx, rtx, struct hash_table *);
+static 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 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);
-static unsigned int hash_expr_1 (rtx, enum machine_mode, int *);
-static unsigned int hash_string_1 (const char *);
+ 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_expr (rtx, 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 (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 void alloc_rd_mem (int, int);
-static void free_rd_mem (void);
-static void handle_rd_kill_set (rtx, int, basic_block);
-static void compute_kill_rd (void);
-static void compute_rd (void);
-static void alloc_avail_expr_mem (int, int);
-static void free_avail_expr_mem (void);
-static void compute_ae_gen (struct hash_table *);
-static int expr_killed_p (rtx, basic_block);
-static void compute_ae_kill (sbitmap *, sbitmap *, struct hash_table *);
-static int expr_reaches_here_p (struct occr *, struct expr *, basic_block,
- int);
-static rtx computing_insn (struct expr *, rtx);
-static int def_reaches_here_p (rtx, rtx);
-static int can_disregard_other_sets (struct reg_set **, rtx, int);
-static int handle_avail_expr (rtx, struct expr *);
-static int classic_gcse (void);
-static int one_classic_gcse_pass (int);
-static void invalidate_nonnull_info (rtx, rtx, void *);
-static int delete_null_pointer_checks_1 (unsigned int *, sbitmap *, sbitmap *,
- struct null_pointer_info *);
static rtx process_insert_insn (struct expr *);
static int pre_edge_insert (struct edge_list *, struct expr **);
-static int expr_reaches_here_p_work (struct occr *, struct expr *,
- basic_block, int, char *);
static int pre_expr_reaches_here_p_work (basic_block, struct expr *,
basic_block, char *);
static struct ls_expr * ldst_entry (rtx);
static void free_ldst_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 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 replace_store_insn (rtx, rtx, basic_block);
-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);
-\f
-/* Entry point for global common subexpression elimination.
- F is the first instruction in the function. */
-
-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. */
- if (n_basic_blocks <= 1)
- return 0;
-
- /* Trying to perform global optimizations on flow graphs which have
- a high connectivity will take a long time and is unlikely to be
- particularly useful.
-
- In normal circumstances a cfg should have about twice as many edges
- as blocks. But we do not want to punish small functions which have
- a couple switch statements. So we require a relatively large number
- of basic blocks and the ratio of edges to blocks to be high. */
- if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
- {
- if (warn_disabled_optimization)
- warning ("GCSE disabled: %d > 1000 basic blocks and %d >= 20 edges/basic block",
- n_basic_blocks, n_edges / n_basic_blocks);
- return 0;
- }
-
- /* If allocating memory for the cprop bitmap would take up too much
- storage it's better just to disable the optimization. */
- if ((n_basic_blocks
- * SBITMAP_SET_SIZE (max_gcse_regno)
- * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
- {
- if (warn_disabled_optimization)
- warning ("GCSE disabled: %d basic blocks and %d registers",
- n_basic_blocks, max_gcse_regno);
-
- return 0;
- }
-
- 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. */
- changed = one_cprop_pass (pass + 1, 0, 0);
-
- if (optimize_size)
- changed |= one_classic_gcse_pass (pass + 1);
- else
- {
- 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 = gmalloc (last_basic_block * sizeof (rtx));
- canon_modify_mem_list
- = gmalloc (last_basic_block * sizeof (rtx));
- memset (modify_mem_list, 0, last_basic_block * sizeof (rtx));
- memset (canon_modify_mem_list, 0, 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;
- }
-
- 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 optimizing
- for code size -- it rarely makes programs faster, and can make
- them bigger if we did partial redundancy elimination (when optimizing
- for space, we use a classic gcse algorithm instead of partial
- redundancy algorithms). */
- if (optimize_size)
- {
- max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (f);
- changed |= one_code_hoisting_pass ();
- free_gcse_mem ();
+static bool do_local_cprop (rtx, rtx);
+static int local_cprop_pass (void);
+static bool is_too_expensive (const char *);
- if (max_pass_bytes < bytes_used)
- max_pass_bytes = bytes_used;
- }
-
- 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. */
- one_cprop_pass (pass + 1, 1, 0);
- 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);
- }
+#define GNEW(T) ((T *) gmalloc (sizeof (T)))
+#define GCNEW(T) ((T *) gcalloc (1, sizeof (T)))
- 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);
+#define GNEWVEC(T, N) ((T *) gmalloc (sizeof (T) * (N)))
+#define GCNEWVEC(T, N) ((T *) gcalloc ((N), sizeof (T)))
- if (!optimize_size && flag_gcse_sm)
- store_motion ();
+#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. */
static void *
-gmalloc (unsigned int size)
+gmalloc (size_t size)
{
bytes_used += size;
return xmalloc (size);
}
-/* 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. */
+/* Cover function to xcalloc to record bytes allocated. */
static void *
-grealloc (void *ptr, unsigned int size)
+gcalloc (size_t nelem, size_t elsize)
{
- return xrealloc (ptr, size);
+ bytes_used += nelem * elsize;
+ return xcalloc (nelem, elsize);
}
/* Cover function to obstack_alloc. */
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, n;
- 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 ();
- n = (max_uid + 1) * sizeof (int);
- uid_cuid = gmalloc (n);
- memset (uid_cuid, 0, n);
- for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
- {
- if (INSN_P (insn))
- uid_cuid[INSN_UID (insn)] = i++;
- else
- uid_cuid[INSN_UID (insn)] = i;
- }
-
- /* Create a table mapping cuids to insns. */
-
- max_cuid = i;
- n = (max_cuid + 1) * sizeof (rtx);
- cuid_insn = gmalloc (n);
- memset (cuid_insn, 0, n);
- for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- CUID_INSN (i++) = insn;
-
/* Allocate vars to track sets of regs. */
- reg_set_bitmap = BITMAP_XMALLOC ();
+ reg_set_bitmap = BITMAP_ALLOC (NULL);
- /* Allocate vars to track sets of regs, memory per block. */
- reg_set_in_block = sbitmap_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 = gmalloc (last_basic_block * sizeof (rtx));
- canon_modify_mem_list = gmalloc (last_basic_block * sizeof (rtx));
- memset (modify_mem_list, 0, last_basic_block * sizeof (rtx));
- memset (canon_modify_mem_list, 0, last_basic_block * sizeof (rtx));
- modify_mem_list_set = BITMAP_XMALLOC ();
- canon_modify_mem_list_set = BITMAP_XMALLOC ();
+ modify_mem_list = 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);
-}
-
-/* Many of the global optimization algorithms work by solving dataflow
- equations for various expressions. Initially, some local value is
- computed for each expression in each block. Then, the values across the
- various blocks are combined (by following flow graph edges) to arrive at
- global values. Conceptually, each set of equations is independent. We
- may therefore solve all the equations in parallel, solve them one at a
- time, or pick any intermediate approach.
-
- When you're going to need N two-dimensional bitmaps, each X (say, the
- number of blocks) by Y (say, the number of expressions), call this
- function. It's not important what X and Y represent; only that Y
- correspond to the things that can be done in parallel. This function will
- return an appropriate chunking factor C; you should solve C sets of
- equations in parallel. By going through this function, we can easily
- trade space against time; by solving fewer equations in parallel we use
- less space. */
-
-static int
-get_bitmap_width (int n, int x, int y)
-{
- /* It's not really worth figuring out *exactly* how much memory will
- be used by a particular choice. The important thing is to get
- something approximately right. */
- size_t max_bitmap_memory = 10 * 1024 * 1024;
-
- /* The number of bytes we'd use for a single column of minimum
- width. */
- size_t column_size = n * x * sizeof (SBITMAP_ELT_TYPE);
-
- /* Often, it's reasonable just to solve all the equations in
- parallel. */
- if (column_size * SBITMAP_SET_SIZE (y) <= max_bitmap_memory)
- return y;
-
- /* Otherwise, pick the largest width we can, without going over the
- limit. */
- return SBITMAP_ELT_BITS * ((max_bitmap_memory + column_size - 1)
- / column_size);
+ BITMAP_FREE (modify_mem_list_set);
+ BITMAP_FREE (blocks_with_calls);
}
\f
/* Compute the local properties of each recorded expression.
ABSALTERED. */
static void
-compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc, struct hash_table *table)
+compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc,
+ struct hash_table_d *table)
{
unsigned int i;
if (antloc)
for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
{
- SET_BIT (antloc[BLOCK_NUM (occr->insn)], indx);
+ SET_BIT (antloc[BLOCK_FOR_INSN (occr->insn)->index], indx);
/* While we're scanning the table, this is a good place to
initialize this. */
if (comp)
for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
{
- SET_BIT (comp[BLOCK_NUM (occr->insn)], indx);
+ SET_BIT (comp[BLOCK_FOR_INSN (occr->insn)->index], indx);
/* While we're scanning the table, this is a good place to
initialize this. */
}
}
\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)
-{
- unsigned int n;
-
- reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
- n = reg_set_table_size * sizeof (struct reg_set *);
- reg_set_table = gmalloc (n);
- memset (reg_set_table, 0, n);
-
- gcc_obstack_init (®_set_obstack);
-}
-
-static void
-free_reg_set_mem (void)
-{
- free (reg_set_table);
- obstack_free (®_set_obstack, NULL);
-}
-
-/* Record REGNO in the reg_set table. */
-
-static void
-record_one_set (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 (GET_CODE (dest) == REG && 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
/* See whether X, the source of a set, is something we want to consider for
GCSE. */
-static GTY(()) rtx test_insn;
static int
want_to_gcse_p (rtx x)
{
- int num_clobbers = 0;
- int icode;
+#ifdef STACK_REGS
+ /* On register stack architectures, don't GCSE constants from the
+ constant pool, as the benefits are often swamped by the overhead
+ of shuffling the register stack between basic blocks. */
+ if (IS_STACK_MODE (GET_MODE (x)))
+ x = avoid_constant_pool_reference (x);
+#endif
switch (GET_CODE (x))
{
case SUBREG:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case CALL:
- case CONSTANT_P_RTX:
return 0;
default:
- break;
+ return can_assign_to_reg_without_clobbers_p (x);
}
+}
+
+/* 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 such that the
+ resulting insn does not result in clobbering a hard register as a
+ side-effect.
+
+ 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;
/* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
if (general_operand (x, GET_MODE (x)))
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);
/* If DEST is not a MEM, then it will not conflict with the load. Note
that function calls are assumed to clobber memory, but are handled
elsewhere. */
- if (GET_CODE (dest) != MEM)
+ if (! MEM_P (dest))
return;
/* If we are setting a MEM in our list of specially recognized MEMs,
}
/* 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;
/* If SETTER is a call everything is clobbered. Note that calls
to pure functions are never put on the list, so we need not
worry about them. */
- if (GET_CODE (setter) == CALL_INSN)
+ if (CALL_P (setter))
return 1;
/* SETTER must be an INSN of some kind that sets memory. Call
the start of INSN's basic block up to but not including INSN. */
static int
-oprs_anticipatable_p (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);
}
MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
indicating if a volatile operand is found or if the expression contains
- something we don't want to insert in the table.
-
- ??? One might want to merge this with canon_hash. Later. */
+ something we don't want to insert in the table. HASH_TABLE_SIZE is
+ the current size of the hash table to be probed. */
static unsigned int
-hash_expr (rtx x, enum machine_mode mode, int *do_not_record_p, int hash_table_size)
+hash_expr (const_rtx x, enum machine_mode mode, int *do_not_record_p,
+ int hash_table_size)
{
unsigned int hash;
*do_not_record_p = 0;
- hash = hash_expr_1 (x, mode, do_not_record_p);
+ hash = hash_rtx (x, mode, do_not_record_p,
+ NULL, /*have_reg_qty=*/false);
return hash % hash_table_size;
}
-/* Hash a string. Just add its bytes up. */
-
-static inline unsigned
-hash_string_1 (const char *ps)
-{
- unsigned hash = 0;
- const unsigned char *p = (const unsigned char *) ps;
-
- if (p)
- while (*p)
- hash += *p++;
+/* Hash a set of register REGNO.
- return hash;
-}
+ Sets are hashed on the register that is set. This simplifies the PRE copy
+ propagation code.
-/* Subroutine of hash_expr to do the actual work. */
+ ??? May need to make things more elaborate. Later, as necessary. */
static unsigned int
-hash_expr_1 (rtx x, enum machine_mode mode, int *do_not_record_p)
+hash_set (int regno, int hash_table_size)
{
- int i, j;
- unsigned hash = 0;
- enum rtx_code code;
- const char *fmt;
-
- /* Used to turn recursion into iteration. We can't rely on GCC's
- tail-recursion elimination since we need to keep accumulating values
- in HASH. */
+ unsigned int hash;
- if (x == 0)
- return hash;
+ hash = regno;
+ return hash % hash_table_size;
+}
- repeat:
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- hash += ((unsigned int) REG << 7) + REGNO (x);
- return hash;
+/* Return nonzero if exp1 is equivalent to exp2. */
- case CONST_INT:
- hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
- + (unsigned int) INTVAL (x));
- return hash;
+static int
+expr_equiv_p (const_rtx x, const_rtx y)
+{
+ return exp_equiv_p (x, y, 0, true);
+}
- case CONST_DOUBLE:
- /* This is like the general case, except that it only counts
- the integers representing the constant. */
- hash += (unsigned int) code + (unsigned int) GET_MODE (x);
- if (GET_MODE (x) != VOIDmode)
- for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
- hash += (unsigned int) XWINT (x, i);
- else
- hash += ((unsigned int) CONST_DOUBLE_LOW (x)
- + (unsigned int) CONST_DOUBLE_HIGH (x));
- return hash;
+/* Insert expression X in INSN in the hash TABLE.
+ If it is already present, record it as the last occurrence in INSN's
+ basic block.
- case CONST_VECTOR:
- {
- int units;
- rtx elt;
+ MODE is the mode of the value X is being stored into.
+ It is only used if X is a CONST_INT.
- units = CONST_VECTOR_NUNITS (x);
+ ANTIC_P is nonzero if X is an anticipatable expression.
+ AVAIL_P is nonzero if X is an available expression. */
- for (i = 0; i < units; ++i)
- {
- elt = CONST_VECTOR_ELT (x, i);
- hash += hash_expr_1 (elt, GET_MODE (elt), do_not_record_p);
- }
+static void
+insert_expr_in_table (rtx x, enum machine_mode mode, rtx insn, int antic_p,
+ 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;
- return hash;
- }
+ hash = hash_expr (x, mode, &do_not_record_p, table->size);
- /* Assume there is only one rtx object for any given label. */
- case LABEL_REF:
- /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
- differences and differences between each stage's debugging dumps. */
- hash += (((unsigned int) LABEL_REF << 7)
- + CODE_LABEL_NUMBER (XEXP (x, 0)));
- return hash;
+ /* Do not insert expression in table if it contains volatile operands,
+ or if hash_expr determines the expression is something we don't want
+ to or can't handle. */
+ if (do_not_record_p)
+ return;
- case SYMBOL_REF:
- {
- /* Don't hash on the symbol's address to avoid bootstrap differences.
- Different hash values may cause expressions to be recorded in
- different orders and thus different registers to be used in the
- final assembler. This also avoids differences in the dump files
- between various stages. */
- unsigned int h = 0;
- const unsigned char *p = (const unsigned char *) XSTR (x, 0);
-
- while (*p)
- h += (h << 7) + *p++; /* ??? revisit */
-
- hash += ((unsigned int) SYMBOL_REF << 7) + h;
- return hash;
- }
-
- case MEM:
- if (MEM_VOLATILE_P (x))
- {
- *do_not_record_p = 1;
- return 0;
- }
-
- hash += (unsigned int) MEM;
- /* We used alias set for hashing, but this is not good, since the alias
- set may differ in -fprofile-arcs and -fbranch-probabilities compilation
- causing the profiles to fail to match. */
- x = XEXP (x, 0);
- goto repeat;
-
- case PRE_DEC:
- case PRE_INC:
- case POST_DEC:
- case POST_INC:
- case PC:
- case CC0:
- case CALL:
- case UNSPEC_VOLATILE:
- *do_not_record_p = 1;
- return 0;
-
- case ASM_OPERANDS:
- if (MEM_VOLATILE_P (x))
- {
- *do_not_record_p = 1;
- return 0;
- }
- else
- {
- /* We don't want to take the filename and line into account. */
- hash += (unsigned) code + (unsigned) GET_MODE (x)
- + hash_string_1 (ASM_OPERANDS_TEMPLATE (x))
- + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
- + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
-
- if (ASM_OPERANDS_INPUT_LENGTH (x))
- {
- for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
- {
- hash += (hash_expr_1 (ASM_OPERANDS_INPUT (x, i),
- GET_MODE (ASM_OPERANDS_INPUT (x, i)),
- do_not_record_p)
- + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
- (x, i)));
- }
-
- hash += hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
- x = ASM_OPERANDS_INPUT (x, 0);
- mode = GET_MODE (x);
- goto repeat;
- }
- return hash;
- }
-
- default:
- break;
- }
-
- hash += (unsigned) code + (unsigned) GET_MODE (x);
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- /* If we are about to do the last recursive call
- needed at this level, change it into iteration.
- This function is called enough to be worth it. */
- if (i == 0)
- {
- x = XEXP (x, i);
- goto repeat;
- }
-
- hash += hash_expr_1 (XEXP (x, i), 0, do_not_record_p);
- if (*do_not_record_p)
- return 0;
- }
-
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- {
- hash += hash_expr_1 (XVECEXP (x, i, j), 0, do_not_record_p);
- if (*do_not_record_p)
- return 0;
- }
-
- else if (fmt[i] == 's')
- hash += hash_string_1 (XSTR (x, i));
- else if (fmt[i] == 'i')
- hash += (unsigned int) XINT (x, i);
- else
- abort ();
- }
-
- return hash;
-}
-
-/* Hash a set of register REGNO.
-
- Sets are hashed on the register that is set. This simplifies the PRE copy
- propagation code.
-
- ??? May need to make things more elaborate. Later, as necessary. */
-
-static unsigned int
-hash_set (int regno, int hash_table_size)
-{
- unsigned int hash;
-
- hash = regno;
- return hash % hash_table_size;
-}
-
-/* Return nonzero if exp1 is equivalent to exp2.
- ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
-
-static int
-expr_equiv_p (rtx x, rtx y)
-{
- int i, j;
- enum rtx_code code;
- const char *fmt;
-
- if (x == y)
- return 1;
-
- if (x == 0 || y == 0)
- return 0;
-
- code = GET_CODE (x);
- if (code != GET_CODE (y))
- return 0;
-
- /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
- if (GET_MODE (x) != GET_MODE (y))
- return 0;
-
- switch (code)
- {
- case PC:
- case CC0:
- case CONST_INT:
- return 0;
-
- case LABEL_REF:
- return XEXP (x, 0) == XEXP (y, 0);
-
- case SYMBOL_REF:
- return XSTR (x, 0) == XSTR (y, 0);
-
- case REG:
- return REGNO (x) == REGNO (y);
-
- case MEM:
- /* Can't merge two expressions in different alias sets, since we can
- decide that the expression is transparent in a block when it isn't,
- due to it being set with the different alias set. */
- if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
- return 0;
- break;
-
- /* For commutative operations, check both orders. */
- case PLUS:
- case MULT:
- case AND:
- case IOR:
- case XOR:
- case NE:
- case EQ:
- return ((expr_equiv_p (XEXP (x, 0), XEXP (y, 0))
- && expr_equiv_p (XEXP (x, 1), XEXP (y, 1)))
- || (expr_equiv_p (XEXP (x, 0), XEXP (y, 1))
- && expr_equiv_p (XEXP (x, 1), XEXP (y, 0))));
-
- case ASM_OPERANDS:
- /* We don't use the generic code below because we want to
- disregard filename and line numbers. */
-
- /* A volatile asm isn't equivalent to any other. */
- if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
- return 0;
-
- if (GET_MODE (x) != GET_MODE (y)
- || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
- || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
- ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
- || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
- || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
- return 0;
-
- if (ASM_OPERANDS_INPUT_LENGTH (x))
- {
- for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
- if (! expr_equiv_p (ASM_OPERANDS_INPUT (x, i),
- ASM_OPERANDS_INPUT (y, i))
- || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
- ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
- return 0;
- }
-
- return 1;
-
- default:
- break;
- }
-
- /* Compare the elements. If any pair of corresponding elements
- fail to match, return 0 for the whole thing. */
-
- fmt = GET_RTX_FORMAT (code);
- for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
- {
- switch (fmt[i])
- {
- case 'e':
- if (! expr_equiv_p (XEXP (x, i), XEXP (y, i)))
- return 0;
- break;
-
- case 'E':
- if (XVECLEN (x, i) != XVECLEN (y, i))
- return 0;
- for (j = 0; j < XVECLEN (x, i); j++)
- if (! expr_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j)))
- return 0;
- break;
-
- case 's':
- if (strcmp (XSTR (x, i), XSTR (y, i)))
- return 0;
- break;
-
- case 'i':
- if (XINT (x, i) != XINT (y, i))
- return 0;
- break;
-
- case 'w':
- if (XWINT (x, i) != XWINT (y, i))
- return 0;
- break;
-
- case '0':
- break;
-
- default:
- abort ();
- }
- }
-
- return 1;
-}
-
-/* Insert expression X in INSN in the hash TABLE.
- If it is already present, record it as the last occurrence in INSN's
- basic block.
-
- MODE is the mode of the value X is being stored into.
- It is only used if X is a CONST_INT.
-
- ANTIC_P is nonzero if X is an anticipatable expression.
- AVAIL_P is nonzero if X is an available expression. */
-
-static void
-insert_expr_in_table (rtx x, enum machine_mode mode, rtx insn, int antic_p,
- int avail_p, struct hash_table *table)
-{
- int found, do_not_record_p;
- unsigned int hash;
- struct expr *cur_expr, *last_expr = NULL;
- struct occr *antic_occr, *avail_occr;
- struct occr *last_occr = NULL;
-
- hash = hash_expr (x, mode, &do_not_record_p, table->size);
-
- /* Do not insert expression in table if it contains volatile operands,
- or if hash_expr determines the expression is something we don't want
- to or can't handle. */
- if (do_not_record_p)
- return;
-
- cur_expr = table->table[hash];
- found = 0;
+ cur_expr = table->table[hash];
+ found = 0;
while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, 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. */
{
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_FOR_INSN (antic_occr->insn) != BLOCK_FOR_INSN (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_FOR_INSN (avail_occr->insn) == BLOCK_FOR_INSN (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;
- if (GET_CODE (x) != SET
- || GET_CODE (SET_DEST (x)) != REG)
- abort ();
+ gcc_assert (GET_CODE (x) == SET && REG_P (SET_DEST (x)));
hash = hash_set (REGNO (SET_DEST (x)), table->size);
if (! found)
{
- cur_expr = 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_FOR_INSN (cur_occr->insn) == BLOCK_FOR_INSN (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
- if they are not floating point registers. */
+ if they are not floating point registers. */
if (GET_CODE(x) == COMPARE
- && GET_CODE (XEXP (x, 0)) == REG
- && GET_CODE (XEXP (x, 1)) == REG
+ && REG_P (XEXP (x, 0)) && REG_P (XEXP (x, 1))
&& REGNO (XEXP (x, 0)) == REGNO (XEXP (x, 1))
&& ! FLOAT_MODE_P (GET_MODE (XEXP (x, 0)))
&& ! FLOAT_MODE_P (GET_MODE (XEXP (x, 1))))
return true;
- if (GET_CODE (x) == CONSTANT_P_RTX)
- return false;
-
- 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);
if (GET_CODE (src) == CALL)
hash_scan_call (src, insn, table);
- else if (GET_CODE (dest) == REG)
+ else if (REG_P (dest))
{
unsigned int regno = REGNO (dest);
rtx tmp;
- /* 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. */
/* Don't GCSE something if we can't do a reg/reg copy. */
&& can_copy_p (GET_MODE (dest))
/* GCSE commonly inserts instruction after the insn. We can't
- do that easily for EH_REGION notes so disable GCSE on these
- for now. */
- && !find_reg_note (insn, REG_EH_REGION, NULL_RTX)
+ do that easily for EH edges so disable GCSE on these for now. */
+ /* ??? We can now easily create new EH landing pads at the
+ gimple level, for splitting edges; there's no reason we
+ can't do the same thing at the rtl level. */
+ && !can_throw_internal (insn)
/* Is SET_SRC something we want to gcse? */
&& want_to_gcse_p (src)
/* Don't CSE a nop. */
REG_EQUIV notes and if the argument slot is used somewhere
explicitly, it means address of parameter has been taken,
so we should not extend the lifetime of the pseudo. */
- && ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
- || GET_CODE (XEXP (note, 0)) != MEM))
+ && (note == NULL_RTX || ! MEM_P (XEXP (note, 0))))
{
/* An expression is not anticipatable if its operands are
modified before this insn or if this is not the only SET in
- this insn. */
- int antic_p = oprs_anticipatable_p (src, insn) && single_set (insn);
+ this insn. The latter condition does not have to mean that
+ SRC itself is not anticipatable, but we just will not be
+ able to handle code motion of insns with multiple sets. */
+ int antic_p = oprs_anticipatable_p (src, insn)
+ && !multiple_sets (insn);
/* An expression is not available if its operands are
subsequently modified, including this insn. It's also not
available if this is a branch, because we can't insert
/* Record sets for constant/copy propagation. */
else if (table->set_p
&& regno >= FIRST_PSEUDO_REGISTER
- && ((GET_CODE (src) == REG
+ && ((REG_P (src)
&& REGNO (src) >= FIRST_PSEUDO_REGISTER
&& can_copy_p (GET_MODE (dest))
&& REGNO (src) != regno)
/* A copy is not available if its src or dest is subsequently
modified. Here we want to search from INSN+1 on, but
oprs_available_p searches from INSN on. */
- && (insn == BLOCK_END (BLOCK_NUM (insn))
- || ((tmp = next_nonnote_insn (insn)) != NULL_RTX
- && oprs_available_p (pat, tmp))))
+ && (insn == BB_END (BLOCK_FOR_INSN (insn))
+ || (tmp = next_nonnote_insn (insn)) == NULL_RTX
+ || BLOCK_FOR_INSN (tmp) != BLOCK_FOR_INSN (insn)
+ || oprs_available_p (pat, tmp)))
insert_set_in_table (pat, insn, table);
}
+ /* In case of store we want to consider the memory value as available in
+ the REG stored in that memory. This makes it possible to remove
+ redundant loads from due to stores to the same location. */
+ else if (flag_gcse_las && REG_P (src) && MEM_P (dest))
+ {
+ unsigned int regno = REGNO (src);
+
+ /* Do not do this for constant/copy propagation. */
+ if (! table->set_p
+ /* Only record sets of pseudo-regs in the hash table. */
+ && regno >= FIRST_PSEUDO_REGISTER
+ /* Don't GCSE something if we can't do a reg/reg copy. */
+ && can_copy_p (GET_MODE (src))
+ /* GCSE commonly inserts instruction after the insn. We can't
+ do that easily for EH edges so disable GCSE on these for now. */
+ && !can_throw_internal (insn)
+ /* Is SET_DEST something we want to gcse? */
+ && want_to_gcse_p (dest)
+ /* Don't CSE a nop. */
+ && ! set_noop_p (pat)
+ /* Don't GCSE if it has attached REG_EQUIV note.
+ At this point this only function parameters should have
+ REG_EQUIV notes and if the argument slot is used somewhere
+ explicitly, it means address of parameter has been taken,
+ so we should not extend the lifetime of the pseudo. */
+ && ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
+ || ! MEM_P (XEXP (note, 0))))
+ {
+ /* Stores are never anticipatable. */
+ int antic_p = 0;
+ /* An expression is not available if its operands are
+ subsequently modified, including this insn. It's also not
+ available if this is a branch, because we can't insert
+ a set after the branch. */
+ int avail_p = oprs_available_p (dest, insn)
+ && ! JUMP_P (insn);
+
+ /* Record the memory expression (DEST) in the hash table. */
+ insert_expr_in_table (dest, GET_MODE (dest), insn,
+ antic_p, avail_p, table);
+ }
+ }
}
static void
hash_scan_clobber (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);
that function calls are assumed to clobber memory, but are handled
elsewhere. */
- if (GET_CODE (dest) != MEM)
+ if (! MEM_P (dest))
return;
dest_addr = get_addr (XEXP (dest, 0));
dest_addr = canon_rtx (dest_addr);
insn = (rtx) v_insn;
- bb = BLOCK_NUM (insn);
+ bb = BLOCK_FOR_INSN (insn)->index;
canon_modify_mem_list[bb] =
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
static void
record_last_mem_set_info (rtx insn)
{
- int bb = BLOCK_NUM (insn);
+ int bb = BLOCK_FOR_INSN (insn)->index;
/* load_killed_in_block_p will handle the case of calls clobbering
everything. */
modify_mem_list[bb] = alloc_INSN_LIST (insn, modify_mem_list[bb]);
bitmap_set_bit (modify_mem_list_set, bb);
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
/* Note that traversals of this loop (other than for free-ing)
will break after encountering a CALL_INSN. So, there's no
need to insert a pair of items, as canon_list_insert does. */
canon_modify_mem_list[bb] =
alloc_INSN_LIST (insn, canon_modify_mem_list[bb]);
- bitmap_set_bit (canon_modify_mem_list_set, bb);
+ bitmap_set_bit (blocks_with_calls, bb);
}
else
note_stores (PATTERN (insn), canon_list_insert, (void*) insn);
the SET is taking place. */
static void
-record_last_set_info (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;
if (GET_CODE (dest) == SUBREG)
dest = SUBREG_REG (dest);
- if (GET_CODE (dest) == REG)
+ if (REG_P (dest))
record_last_reg_set_info (last_set_insn, REGNO (dest));
- else if (GET_CODE (dest) == MEM
+ else if (MEM_P (dest)
/* Ignore pushes, they clobber nothing. */
&& ! push_operand (dest, GET_MODE (dest)))
record_last_mem_set_info (last_set_insn);
TABLE is the table computed. */
static void
-compute_hash_table_work (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 = current_bb->head;
- insn && insn != NEXT_INSN (current_bb->end);
- 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;
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
- bool clobbers_all = false;
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
-
for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
record_last_reg_set_info (insn, regno);
mark_call (insn);
if (table->set_p
&& implicit_sets[current_bb->index] != NULL_RTX)
hash_scan_set (implicit_sets[current_bb->index],
- current_bb->head, table);
+ BB_HEAD (current_bb), table);
/* The next pass builds the hash table. */
-
- for (insn = current_bb->head, in_libcall_block = 0;
- insn && insn != NEXT_INSN (current_bb->end);
- 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);
}
/* Allocate space for the set/expr hash TABLE.
- N_INSNS is the number of instructions in the function.
It is used to determine the number of buckets to use.
SET_P determines whether set or expression table will
be created. */
static void
-alloc_hash_table (int n_insns, struct hash_table *table, int set_p)
+alloc_hash_table (struct hash_table_d *table, int set_p)
{
int n;
- table->size = n_insns / 4;
+ n = get_max_insn_count ();
+
+ table->size = n / 4;
if (table->size < 11)
table->size = 11;
??? 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;
\f
/* Expression tracking support. */
-/* Lookup pattern PAT in the expression TABLE.
- The result is a pointer to the table entry, or NULL if not found. */
-
-static struct expr *
-lookup_expr (rtx pat, struct hash_table *table)
-{
- int do_not_record_p;
- unsigned int hash = hash_expr (pat, GET_MODE (pat), &do_not_record_p,
- table->size);
- struct expr *expr;
-
- if (do_not_record_p)
- return NULL;
-
- expr = table->table[hash];
-
- while (expr && ! expr_equiv_p (expr->expr, pat))
- expr = expr->next_same_hash;
-
- return expr;
-}
-
/* Lookup REGNO in the set TABLE. The result is a pointer to the
table entry, or NULL if not found. */
static struct expr *
-lookup_set (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;
static void
clear_modify_mem_tables (void)
{
- int i;
+ unsigned i;
+ bitmap_iterator bi;
- EXECUTE_IF_SET_IN_BITMAP
- (modify_mem_list_set, 0, i, free_INSN_LIST_list (modify_mem_list + i));
+ EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set, 0, i, bi)
+ {
+ free_INSN_LIST_list (modify_mem_list + i);
+ free_insn_expr_list_list (canon_modify_mem_list + i);
+ }
bitmap_clear (modify_mem_list_set);
-
- EXECUTE_IF_SET_IN_BITMAP
- (canon_modify_mem_list_set, 0, i,
- free_insn_expr_list_list (canon_modify_mem_list + i));
- bitmap_clear (canon_modify_mem_list_set);
+ bitmap_clear (blocks_with_calls);
}
-/* Release memory used by modify_mem_list_set and canon_modify_mem_list_set. */
+/* Release memory used by modify_mem_list_set. */
static void
free_modify_mem_tables (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);
- if (GET_CODE (dest) == REG)
+ if (REG_P (dest))
SET_REGNO_REG_SET (reg_set_bitmap, REGNO (dest));
- else if (GET_CODE (dest) == MEM)
+ else if (MEM_P (dest))
record_last_mem_set_info (insn);
if (GET_CODE (SET_SRC (pat)) == CALL)
while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
clob = XEXP (clob, 0);
- if (GET_CODE (clob) == REG)
+ if (REG_P (clob))
SET_REGNO_REG_SET (reg_set_bitmap, REGNO (clob));
else
record_last_mem_set_info (insn);
}
\f
-/* Classic GCSE reaching definition support. */
+/* Compute copy/constant propagation working variables. */
-/* Allocate reaching def variables. */
+/* Local properties of assignments. */
+static sbitmap *cprop_pavloc;
+static sbitmap *cprop_absaltered;
-static void
-alloc_rd_mem (int n_blocks, int n_insns)
-{
- rd_kill = sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_kill, n_blocks);
+/* Global properties of assignments (computed from the local properties). */
+static sbitmap *cprop_avin;
+static sbitmap *cprop_avout;
- rd_gen = sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_gen, n_blocks);
+/* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
+ basic blocks. N_SETS is the number of sets. */
- reaching_defs = sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (reaching_defs, n_blocks);
+static void
+alloc_cprop_mem (int n_blocks, int n_sets)
+{
+ cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
+ cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
- rd_out = sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_out, n_blocks);
+ cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
+ cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
}
-/* Free reaching def variables. */
+/* Free vars used by copy/const propagation. */
static void
-free_rd_mem (void)
+free_cprop_mem (void)
{
- sbitmap_vector_free (rd_kill);
- sbitmap_vector_free (rd_gen);
- sbitmap_vector_free (reaching_defs);
- sbitmap_vector_free (rd_out);
+ sbitmap_vector_free (cprop_pavloc);
+ sbitmap_vector_free (cprop_absaltered);
+ sbitmap_vector_free (cprop_avin);
+ sbitmap_vector_free (cprop_avout);
}
-/* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
+/* For each block, compute whether X is transparent. X is either an
+ expression or an assignment [though we don't care which, for this context
+ an assignment is treated as an expression]. For each block where an
+ element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
+ bit in BMAP. */
static void
-handle_rd_kill_set (rtx insn, int regno, basic_block bb)
+compute_transp (const_rtx x, int indx, sbitmap *bmap, int set_p)
{
- struct reg_set *this_reg;
-
- for (this_reg = reg_set_table[regno]; this_reg; this_reg = this_reg ->next)
- if (BLOCK_NUM (this_reg->insn) != BLOCK_NUM (insn))
- SET_BIT (rd_kill[bb->index], INSN_CUID (this_reg->insn));
-}
+ int i, j;
+ enum rtx_code code;
+ const char *fmt;
-/* Compute the set of kill's for reaching definitions. */
+ /* repeat is used to turn tail-recursion into iteration since GCC
+ can't do it when there's no return value. */
+ repeat:
-static void
-compute_kill_rd (void)
-{
- int cuid;
- unsigned int regno;
- int i;
- basic_block bb;
+ if (x == 0)
+ return;
- /* For each block
- For each set bit in `gen' of the block (i.e each insn which
- generates a definition in the block)
- Call the reg set by the insn corresponding to that bit regx
- Look at the linked list starting at reg_set_table[regx]
- For each setting of regx in the linked list, which is not in
- this block
- Set the bit in `kill' corresponding to that insn. */
- FOR_EACH_BB (bb)
- for (cuid = 0; cuid < max_cuid; cuid++)
- if (TEST_BIT (rd_gen[bb->index], cuid))
+ code = GET_CODE (x);
+ switch (code)
+ {
+ case REG:
+ if (set_p)
{
- rtx insn = CUID_INSN (cuid);
- rtx pat = PATTERN (insn);
-
- if (GET_CODE (insn) == CALL_INSN)
- {
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- handle_rd_kill_set (insn, regno, bb);
- }
-
- if (GET_CODE (pat) == PARALLEL)
- {
- for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
- {
- enum rtx_code code = GET_CODE (XVECEXP (pat, 0, i));
-
- if ((code == SET || code == CLOBBER)
- && GET_CODE (XEXP (XVECEXP (pat, 0, i), 0)) == REG)
- handle_rd_kill_set (insn,
- REGNO (XEXP (XVECEXP (pat, 0, i), 0)),
- bb);
- }
- }
- else if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == REG)
- /* Each setting of this register outside of this block
- must be marked in the set of kills in this block. */
- handle_rd_kill_set (insn, REGNO (SET_DEST (pat)), bb);
+ 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);
}
-}
-
-/* Compute the reaching definitions as in
- Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
- Chapter 10. It is the same algorithm as used for computing available
- expressions but applied to the gens and kills of reaching definitions. */
-
-static void
-compute_rd (void)
-{
- int changed, passes;
- basic_block bb;
-
- FOR_EACH_BB (bb)
- sbitmap_copy (rd_out[bb->index] /*dst*/, rd_gen[bb->index] /*src*/);
-
- passes = 0;
- changed = 1;
- while (changed)
- {
- changed = 0;
- FOR_EACH_BB (bb)
+ else
{
- sbitmap_union_of_preds (reaching_defs[bb->index], rd_out, bb->index);
- changed |= sbitmap_union_of_diff_cg (rd_out[bb->index], rd_gen[bb->index],
- reaching_defs[bb->index], rd_kill[bb->index]);
+ 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);
}
- passes++;
- }
-
- if (gcse_file)
- fprintf (gcse_file, "reaching def computation: %d passes\n", passes);
-}
-\f
-/* Classic GCSE available expression support. */
-
-/* Allocate memory for available expression computation. */
-
-static void
-alloc_avail_expr_mem (int n_blocks, int n_exprs)
-{
- ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_kill, n_blocks);
-
- ae_gen = sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_gen, n_blocks);
-
- ae_in = sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_in, n_blocks);
- ae_out = sbitmap_vector_alloc (n_blocks, n_exprs);
- sbitmap_vector_zero (ae_out, n_blocks);
-}
-
-static void
-free_avail_expr_mem (void)
-{
- sbitmap_vector_free (ae_kill);
- sbitmap_vector_free (ae_gen);
- sbitmap_vector_free (ae_in);
- sbitmap_vector_free (ae_out);
-}
+ return;
-/* Compute the set of available expressions generated in each basic block. */
+ case MEM:
+ if (! MEM_READONLY_P (x))
+ {
+ bitmap_iterator bi;
+ unsigned bb_index;
-static void
-compute_ae_gen (struct hash_table *expr_hash_table)
-{
- unsigned int i;
- struct expr *expr;
- struct occr *occr;
+ /* First handle all the blocks with calls. We don't need to
+ do any list walking for them. */
+ EXECUTE_IF_SET_IN_BITMAP (blocks_with_calls, 0, bb_index, bi)
+ {
+ if (set_p)
+ SET_BIT (bmap[bb_index], indx);
+ else
+ RESET_BIT (bmap[bb_index], indx);
+ }
- /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
- This is all we have to do because an expression is not recorded if it
- is not available, and the only expressions we want to work with are the
- ones that are recorded. */
- for (i = 0; i < expr_hash_table->size; i++)
- for (expr = expr_hash_table->table[i]; expr != 0; expr = expr->next_same_hash)
- for (occr = expr->avail_occr; occr != 0; occr = occr->next)
- SET_BIT (ae_gen[BLOCK_NUM (occr->insn)], expr->bitmap_index);
-}
+ /* Now iterate over the blocks which have memory modifications
+ but which do not have any calls. */
+ EXECUTE_IF_AND_COMPL_IN_BITMAP (modify_mem_list_set,
+ blocks_with_calls,
+ 0, bb_index, bi)
+ {
+ rtx list_entry = canon_modify_mem_list[bb_index];
-/* Return nonzero if expression X is killed in BB. */
+ while (list_entry)
+ {
+ rtx dest, dest_addr;
-static int
-expr_killed_p (rtx x, basic_block bb)
-{
- int i, j;
- enum rtx_code code;
- const char *fmt;
+ /* LIST_ENTRY must be an INSN of some kind that sets memory.
+ Examine each hunk of memory that is modified. */
- if (x == 0)
- return 1;
+ dest = XEXP (list_entry, 0);
+ list_entry = XEXP (list_entry, 1);
+ dest_addr = XEXP (list_entry, 0);
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- return TEST_BIT (reg_set_in_block[bb->index], REGNO (x));
+ if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
+ x, NULL_RTX, rtx_addr_varies_p))
+ {
+ if (set_p)
+ SET_BIT (bmap[bb_index], indx);
+ else
+ RESET_BIT (bmap[bb_index], indx);
+ break;
+ }
+ list_entry = XEXP (list_entry, 1);
+ }
+ }
+ }
- case MEM:
- if (load_killed_in_block_p (bb, get_max_uid () + 1, x, 0))
- return 1;
- else
- return expr_killed_p (XEXP (x, 0), bb);
+ x = XEXP (x, 0);
+ goto repeat;
case PC:
case CC0: /*FIXME*/
case CONST:
case CONST_INT:
case CONST_DOUBLE:
+ case CONST_FIXED:
case CONST_VECTOR:
case SYMBOL_REF:
case LABEL_REF:
case ADDR_VEC:
case ADDR_DIFF_VEC:
- return 0;
+ return;
default:
break;
needed at this level, change it into iteration.
This function is called enough to be worth it. */
if (i == 0)
- return expr_killed_p (XEXP (x, i), bb);
- else if (expr_killed_p (XEXP (x, i), bb))
- return 1;
+ {
+ x = XEXP (x, i);
+ goto repeat;
+ }
+
+ compute_transp (XEXP (x, i), indx, bmap, set_p);
}
else if (fmt[i] == 'E')
for (j = 0; j < XVECLEN (x, i); j++)
- if (expr_killed_p (XVECEXP (x, i, j), bb))
- return 1;
+ compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
}
-
- return 0;
}
-/* Compute the set of available expressions killed in each basic block. */
+/* Top level routine to do the dataflow analysis needed by copy/const
+ propagation. */
static void
-compute_ae_kill (sbitmap *ae_gen, sbitmap *ae_kill,
- struct hash_table *expr_hash_table)
+compute_cprop_data (void)
{
- basic_block bb;
- unsigned int i;
- struct expr *expr;
-
- FOR_EACH_BB (bb)
- for (i = 0; i < expr_hash_table->size; i++)
- for (expr = expr_hash_table->table[i]; expr; expr = expr->next_same_hash)
- {
- /* Skip EXPR if generated in this block. */
- if (TEST_BIT (ae_gen[bb->index], expr->bitmap_index))
- continue;
-
- if (expr_killed_p (expr->expr, bb))
- SET_BIT (ae_kill[bb->index], expr->bitmap_index);
- }
+ compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, &set_hash_table);
+ compute_available (cprop_pavloc, cprop_absaltered,
+ cprop_avout, cprop_avin);
}
\f
-/* Actually perform the Classic GCSE optimizations. */
-
-/* Return nonzero if occurrence OCCR of expression EXPR reaches block BB.
+/* Copy/constant propagation. */
- CHECK_SELF_LOOP is nonzero if we should consider a block reaching itself
- as a positive reach. We want to do this when there are two computations
- of the expression in the block.
+/* Maximum number of register uses in an insn that we handle. */
+#define MAX_USES 8
- VISITED is a pointer to a working buffer for tracking which BB's have
- been visited. It is NULL for the top-level call.
+/* Table of uses found in an insn.
+ Allocated statically to avoid alloc/free complexity and overhead. */
+static struct reg_use reg_use_table[MAX_USES];
- We treat reaching expressions that go through blocks containing the same
- reaching expression as "not reaching". E.g. if EXPR is generated in blocks
- 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
- 2 as not reaching. The intent is to improve the probability of finding
- only one reaching expression and to reduce register lifetimes by picking
- the closest such expression. */
+/* Index into `reg_use_table' while building it. */
+static int reg_use_count;
-static int
-expr_reaches_here_p_work (struct occr *occr, struct expr *expr,
- basic_block bb, int check_self_loop, char *visited)
-{
- edge pred;
+/* Set up a list of register numbers used in INSN. The found uses are stored
+ in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
+ and contains the number of uses in the table upon exit.
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
- {
- basic_block pred_bb = pred->src;
+ ??? If a register appears multiple times we will record it multiple times.
+ This doesn't hurt anything but it will slow things down. */
- if (visited[pred_bb->index])
- /* This predecessor has already been visited. Nothing to do. */
- ;
- else if (pred_bb == bb)
- {
- /* BB loops on itself. */
- if (check_self_loop
- && TEST_BIT (ae_gen[pred_bb->index], expr->bitmap_index)
- && BLOCK_NUM (occr->insn) == pred_bb->index)
- return 1;
+static void
+find_used_regs (rtx *xptr, void *data ATTRIBUTE_UNUSED)
+{
+ int i, j;
+ enum rtx_code code;
+ const char *fmt;
+ rtx x = *xptr;
- visited[pred_bb->index] = 1;
- }
+ /* repeat is used to turn tail-recursion into iteration since GCC
+ can't do it when there's no return value. */
+ repeat:
+ if (x == 0)
+ return;
- /* Ignore this predecessor if it kills the expression. */
- else if (TEST_BIT (ae_kill[pred_bb->index], expr->bitmap_index))
- visited[pred_bb->index] = 1;
+ code = GET_CODE (x);
+ if (REG_P (x))
+ {
+ if (reg_use_count == MAX_USES)
+ return;
- /* Does this predecessor generate this expression? */
- else if (TEST_BIT (ae_gen[pred_bb->index], expr->bitmap_index))
- {
- /* Is this the occurrence we're looking for?
- Note that there's only one generating occurrence per block
- so we just need to check the block number. */
- if (BLOCK_NUM (occr->insn) == pred_bb->index)
- return 1;
+ reg_use_table[reg_use_count].reg_rtx = x;
+ reg_use_count++;
+ }
- visited[pred_bb->index] = 1;
- }
+ /* Recursively scan the operands of this expression. */
- /* Neither gen nor kill. */
- else
+ for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
+ {
+ if (fmt[i] == 'e')
{
- visited[pred_bb->index] = 1;
- if (expr_reaches_here_p_work (occr, expr, pred_bb, check_self_loop,
- visited))
+ /* If we are about to do the last recursive call
+ needed at this level, change it into iteration.
+ This function is called enough to be worth it. */
+ if (i == 0)
+ {
+ x = XEXP (x, 0);
+ goto repeat;
+ }
- return 1;
+ find_used_regs (&XEXP (x, i), data);
}
+ else if (fmt[i] == 'E')
+ for (j = 0; j < XVECLEN (x, i); j++)
+ find_used_regs (&XVECEXP (x, i, j), data);
}
-
- /* All paths have been checked. */
- return 0;
}
-/* This wrapper for expr_reaches_here_p_work() is to ensure that any
- memory allocated for that function is returned. */
+/* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
+ Returns nonzero is successful. */
static int
-expr_reaches_here_p (struct occr *occr, struct expr *expr, basic_block bb,
- int check_self_loop)
+try_replace_reg (rtx from, rtx to, rtx insn)
{
- int rval;
- char *visited = xcalloc (last_basic_block, 1);
-
- rval = expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited);
-
- free (visited);
- return rval;
-}
-
-/* Return the instruction that computes EXPR that reaches INSN's basic block.
- If there is more than one such instruction, return NULL.
+ rtx note = find_reg_equal_equiv_note (insn);
+ rtx src = 0;
+ int success = 0;
+ rtx set = single_set (insn);
- Called only by handle_avail_expr. */
+ /* 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);
-static rtx
-computing_insn (struct expr *expr, rtx insn)
-{
- basic_block bb = BLOCK_FOR_INSN (insn);
+ validate_replace_src_group (from, to, insn);
+ if (num_changes_pending () && apply_change_group ())
+ success = 1;
- if (expr->avail_occr->next == NULL)
+ /* Try to simplify SET_SRC if we have substituted a constant. */
+ if (success && set && CONSTANT_P (to))
{
- if (BLOCK_FOR_INSN (expr->avail_occr->insn) == bb)
- /* The available expression is actually itself
- (i.e. a loop in the flow graph) so do nothing. */
- return NULL;
-
- /* (FIXME) Case that we found a pattern that was created by
- a substitution that took place. */
- return expr->avail_occr->insn;
+ src = simplify_rtx (SET_SRC (set));
+
+ if (src)
+ validate_change (insn, &SET_SRC (set), src, 0);
}
- else
+
+ /* 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, to));
+ if (!success && set && reg_mentioned_p (from, SET_SRC (set)))
{
- /* Pattern is computed more than once.
- Search backwards from this insn to see how many of these
- computations actually reach this insn. */
- struct occr *occr;
- rtx insn_computes_expr = NULL;
- int can_reach = 0;
-
- for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
- {
- if (BLOCK_FOR_INSN (occr->insn) == bb)
- {
- /* The expression is generated in this block.
- The only time we care about this is when the expression
- is generated later in the block [and thus there's a loop].
- We let the normal cse pass handle the other cases. */
- if (INSN_CUID (insn) < INSN_CUID (occr->insn)
- && expr_reaches_here_p (occr, expr, bb, 1))
- {
- can_reach++;
- if (can_reach > 1)
- return NULL;
+ /* If above failed and this is a single set, try to simplify the source of
+ the set given our substitution. We could perhaps try this for multiple
+ SETs, but it probably won't buy us anything. */
+ src = simplify_replace_rtx (SET_SRC (set), from, to);
- insn_computes_expr = occr->insn;
- }
- }
- else if (expr_reaches_here_p (occr, expr, bb, 0))
- {
- can_reach++;
- if (can_reach > 1)
- return NULL;
+ if (!rtx_equal_p (src, SET_SRC (set))
+ && validate_change (insn, &SET_SRC (set), src, 0))
+ success = 1;
- insn_computes_expr = occr->insn;
- }
- }
+ /* If we've failed to do replacement, have a single SET, don't already
+ have a note, and have no special SET, add a REG_EQUAL note to not
+ lose information. */
+ if (!success && note == 0 && set != 0
+ && GET_CODE (SET_DEST (set)) != ZERO_EXTRACT
+ && GET_CODE (SET_DEST (set)) != STRICT_LOW_PART)
+ note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
+ }
- if (insn_computes_expr == NULL)
- abort ();
+ /* REG_EQUAL may get simplified into register.
+ We don't allow that. Remove that note. This code ought
+ not to happen, because previous code ought to synthesize
+ reg-reg move, but be on the safe side. */
+ if (note && REG_NOTE_KIND (note) == REG_EQUAL && REG_P (XEXP (note, 0)))
+ remove_note (insn, note);
- return insn_computes_expr;
- }
+ return success;
}
-/* Return nonzero if the definition in DEF_INSN can reach INSN.
- Only called by can_disregard_other_sets. */
+/* Find a set of REGNOs that are available on entry to INSN's block. Returns
+ NULL no such set is found. */
-static int
-def_reaches_here_p (rtx insn, rtx def_insn)
+static struct expr *
+find_avail_set (int regno, rtx insn)
{
- rtx reg;
+ /* SET1 contains the last set found that can be returned to the caller for
+ use in a substitution. */
+ struct expr *set1 = 0;
- if (TEST_BIT (reaching_defs[BLOCK_NUM (insn)], INSN_CUID (def_insn)))
- return 1;
+ /* Loops are not possible here. To get a loop we would need two sets
+ available at the start of the block containing INSN. i.e. we would
+ need two sets like this available at the start of the block:
+
+ (set (reg X) (reg Y))
+ (set (reg Y) (reg X))
- if (BLOCK_NUM (insn) == BLOCK_NUM (def_insn))
+ This can not happen since the set of (reg Y) would have killed the
+ set of (reg X) making it unavailable at the start of this block. */
+ while (1)
{
- if (INSN_CUID (def_insn) < INSN_CUID (insn))
- {
- if (GET_CODE (PATTERN (def_insn)) == PARALLEL)
- return 1;
- else if (GET_CODE (PATTERN (def_insn)) == CLOBBER)
- reg = XEXP (PATTERN (def_insn), 0);
- else if (GET_CODE (PATTERN (def_insn)) == SET)
- reg = SET_DEST (PATTERN (def_insn));
- else
- abort ();
+ rtx src;
+ struct expr *set = lookup_set (regno, &set_hash_table);
- return ! reg_set_between_p (reg, NEXT_INSN (def_insn), insn);
+ /* Find a set that is available at the start of the block
+ which contains INSN. */
+ while (set)
+ {
+ if (TEST_BIT (cprop_avin[BLOCK_FOR_INSN (insn)->index],
+ set->bitmap_index))
+ break;
+ set = next_set (regno, set);
}
- else
- return 0;
- }
- return 0;
-}
+ /* If no available set was found we've reached the end of the
+ (possibly empty) copy chain. */
+ if (set == 0)
+ break;
-/* Return nonzero if *ADDR_THIS_REG can only have one value at INSN. The
- value returned is the number of definitions that reach INSN. Returning a
- value of zero means that [maybe] more than one definition reaches INSN and
- the caller can't perform whatever optimization it is trying. i.e. it is
- always safe to return zero. */
+ gcc_assert (GET_CODE (set->expr) == SET);
-static int
-can_disregard_other_sets (struct reg_set **addr_this_reg, rtx insn, int for_combine)
-{
- int number_of_reaching_defs = 0;
- struct reg_set *this_reg;
+ src = SET_SRC (set->expr);
- for (this_reg = *addr_this_reg; this_reg != 0; this_reg = this_reg->next)
- if (def_reaches_here_p (insn, this_reg->insn))
- {
- number_of_reaching_defs++;
- /* Ignore parallels for now. */
- if (GET_CODE (PATTERN (this_reg->insn)) == PARALLEL)
- return 0;
+ /* We know the set is available.
+ Now check that SRC is ANTLOC (i.e. none of the source operands
+ have changed since the start of the block).
- if (!for_combine
- && (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER
- || ! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
- SET_SRC (PATTERN (insn)))))
- /* A setting of the reg to a different value reaches INSN. */
- return 0;
+ If the source operand changed, we may still use it for the next
+ iteration of this loop, but we may not use it for substitutions. */
- if (number_of_reaching_defs > 1)
- {
- /* If in this setting the value the register is being set to is
- equal to the previous value the register was set to and this
- setting reaches the insn we are trying to do the substitution
- on then we are ok. */
- if (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER)
- return 0;
- else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
- SET_SRC (PATTERN (insn))))
- return 0;
- }
+ if (gcse_constant_p (src) || oprs_not_set_p (src, insn))
+ set1 = set;
- *addr_this_reg = this_reg;
- }
+ /* If the source of the set is anything except a register, then
+ we have reached the end of the copy chain. */
+ if (! REG_P (src))
+ break;
- return number_of_reaching_defs;
-}
+ /* Follow the copy chain, i.e. start another iteration of the loop
+ and see if we have an available copy into SRC. */
+ regno = REGNO (src);
+ }
-/* Expression computed by insn is available and the substitution is legal,
- so try to perform the substitution.
+ /* SET1 holds the last set that was available and anticipatable at
+ INSN. */
+ return set1;
+}
- The result is nonzero if any changes were made. */
+/* Subroutine of cprop_insn that tries to propagate constants into
+ JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
+ it is the instruction that immediately precedes JUMP, and must be a
+ single SET of a register. FROM is what we will try to replace,
+ SRC is the constant we will try to substitute for it. Returns nonzero
+ if a change was made. */
static int
-handle_avail_expr (rtx insn, struct expr *expr)
+cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
{
- rtx pat, insn_computes_expr, expr_set;
- rtx to;
- struct reg_set *this_reg;
- int found_setting, use_src;
- int changed = 0;
-
- /* We only handle the case where one computation of the expression
- reaches this instruction. */
- insn_computes_expr = computing_insn (expr, insn);
- if (insn_computes_expr == NULL)
- return 0;
- expr_set = single_set (insn_computes_expr);
- if (!expr_set)
- abort ();
-
- found_setting = 0;
- use_src = 0;
+ rtx new_rtx, set_src, note_src;
+ rtx set = pc_set (jump);
+ rtx note = find_reg_equal_equiv_note (jump);
- /* At this point we know only one computation of EXPR outside of this
- block reaches this insn. Now try to find a register that the
- expression is computed into. */
- if (GET_CODE (SET_SRC (expr_set)) == REG)
+ if (note)
{
- /* This is the case when the available expression that reaches
- here has already been handled as an available expression. */
- unsigned int regnum_for_replacing
- = REGNO (SET_SRC (expr_set));
-
- /* If the register was created by GCSE we can't use `reg_set_table',
- however we know it's set only once. */
- if (regnum_for_replacing >= max_gcse_regno
- /* If the register the expression is computed into is set only once,
- or only one set reaches this insn, we can use it. */
- || (((this_reg = reg_set_table[regnum_for_replacing]),
- this_reg->next == NULL)
- || can_disregard_other_sets (&this_reg, insn, 0)))
- {
- use_src = 1;
- found_setting = 1;
- }
+ note_src = XEXP (note, 0);
+ if (GET_CODE (note_src) == EXPR_LIST)
+ note_src = NULL_RTX;
}
+ else note_src = NULL_RTX;
- if (!found_setting)
- {
- unsigned int regnum_for_replacing
- = REGNO (SET_DEST (expr_set));
-
- /* This shouldn't happen. */
- if (regnum_for_replacing >= max_gcse_regno)
- abort ();
-
- this_reg = reg_set_table[regnum_for_replacing];
+ /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
+ set_src = note_src ? note_src : SET_SRC (set);
- /* If the register the expression is computed into is set only once,
- or only one set reaches this insn, use it. */
- if (this_reg->next == NULL
- || can_disregard_other_sets (&this_reg, insn, 0))
- found_setting = 1;
+ /* First substitute the SETCC condition into the JUMP instruction,
+ then substitute that given values into this expanded JUMP. */
+ if (setcc != NULL_RTX
+ && !modified_between_p (from, setcc, jump)
+ && !modified_between_p (src, setcc, jump))
+ {
+ rtx setcc_src;
+ rtx setcc_set = single_set (setcc);
+ rtx setcc_note = find_reg_equal_equiv_note (setcc);
+ setcc_src = (setcc_note && GET_CODE (XEXP (setcc_note, 0)) != EXPR_LIST)
+ ? XEXP (setcc_note, 0) : SET_SRC (setcc_set);
+ set_src = simplify_replace_rtx (set_src, SET_DEST (setcc_set),
+ setcc_src);
}
+ else
+ setcc = NULL_RTX;
- if (found_setting)
- {
- pat = PATTERN (insn);
- if (use_src)
- to = SET_SRC (expr_set);
- else
- to = SET_DEST (expr_set);
- changed = validate_change (insn, &SET_SRC (pat), to, 0);
+ new_rtx = simplify_replace_rtx (set_src, from, src);
- /* We should be able to ignore the return code from validate_change but
- to play it safe we check. */
- if (changed)
- {
- gcse_subst_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file, "GCSE: Replacing the source in insn %d with",
- INSN_UID (insn));
- fprintf (gcse_file, " reg %d %s insn %d\n",
- REGNO (to), use_src ? "from" : "set in",
- INSN_UID (insn_computes_expr));
- }
- }
- }
+ /* If no simplification can be made, then try the next register. */
+ if (rtx_equal_p (new_rtx, SET_SRC (set)))
+ return 0;
- /* The register that the expr is computed into is set more than once. */
- else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
+ /* If this is now a no-op delete it, otherwise this must be a valid insn. */
+ if (new_rtx == pc_rtx)
+ delete_insn (jump);
+ else
{
- /* Insert an insn after insnx that copies the reg set in insnx
- into a new pseudo register call this new register REGN.
- From insnb until end of basic block or until REGB is set
- replace all uses of REGB with REGN. */
- rtx new_insn;
-
- to = gen_reg_rtx (GET_MODE (SET_DEST (expr_set)));
-
- /* Generate the new insn. */
- /* ??? If the change fails, we return 0, even though we created
- an insn. I think this is ok. */
- new_insn
- = emit_insn_after (gen_rtx_SET (VOIDmode, to,
- SET_DEST (expr_set)),
- insn_computes_expr);
-
- /* Keep register set table up to date. */
- record_one_set (REGNO (to), new_insn);
-
- gcse_create_count++;
- if (gcse_file != NULL)
+ /* Ensure the value computed inside the jump insn to be equivalent
+ to one computed by setcc. */
+ if (setcc && modified_in_p (new_rtx, setcc))
+ return 0;
+ if (! validate_unshare_change (jump, &SET_SRC (set), new_rtx, 0))
{
- fprintf (gcse_file, "GCSE: Creating insn %d to copy value of reg %d",
- INSN_UID (NEXT_INSN (insn_computes_expr)),
- REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr)))));
- fprintf (gcse_file, ", computed in insn %d,\n",
- INSN_UID (insn_computes_expr));
- fprintf (gcse_file, " into newly allocated reg %d\n",
- REGNO (to));
+ /* When (some) constants are not valid in a comparison, and there
+ are two registers to be replaced by constants before the entire
+ comparison can be folded into a constant, we need to keep
+ intermediate information in REG_EQUAL notes. For targets with
+ separate compare insns, such notes are added by try_replace_reg.
+ When we have a combined compare-and-branch instruction, however,
+ we need to attach a note to the branch itself to make this
+ optimization work. */
+
+ if (!rtx_equal_p (new_rtx, note_src))
+ set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new_rtx));
+ return 0;
}
- pat = PATTERN (insn);
+ /* Remove REG_EQUAL note after simplification. */
+ if (note_src)
+ remove_note (jump, note);
+ }
- /* Do register replacement for INSN. */
- changed = validate_change (insn, &SET_SRC (pat),
- SET_DEST (PATTERN
- (NEXT_INSN (insn_computes_expr))),
- 0);
+#ifdef HAVE_cc0
+ /* Delete the cc0 setter. */
+ if (setcc != NULL && CC0_P (SET_DEST (single_set (setcc))))
+ delete_insn (setcc);
+#endif
- /* We should be able to ignore the return code from validate_change but
- to play it safe we check. */
- if (changed)
- {
- gcse_subst_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file,
- "GCSE: Replacing the source in insn %d with reg %d ",
- INSN_UID (insn),
- REGNO (SET_DEST (PATTERN (NEXT_INSN
- (insn_computes_expr)))));
- fprintf (gcse_file, "set in insn %d\n",
- INSN_UID (insn_computes_expr));
- }
- }
+ global_const_prop_count++;
+ if (dump_file != NULL)
+ {
+ fprintf (dump_file,
+ "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
+ REGNO (from), INSN_UID (jump));
+ print_rtl (dump_file, src);
+ fprintf (dump_file, "\n");
}
+ purge_dead_edges (bb);
- return changed;
-}
+ /* 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;
-/* Perform classic GCSE. This is called by one_classic_gcse_pass after all
- the dataflow analysis has been done.
+ 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);
+ }
- The result is nonzero if a change was made. */
+ return 1;
+}
-static int
-classic_gcse (void)
+static bool
+constprop_register (rtx insn, rtx from, rtx to)
{
- int changed;
- rtx insn;
- basic_block bb;
-
- /* Note we start at block 1. */
-
- if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
- return 0;
+ rtx sset;
- changed = 0;
- FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
+ /* Check for reg or cc0 setting instructions followed by
+ conditional branch instructions first. */
+ if ((sset = single_set (insn)) != NULL
+ && NEXT_INSN (insn)
+ && any_condjump_p (NEXT_INSN (insn)) && onlyjump_p (NEXT_INSN (insn)))
{
- /* Reset tables used to keep track of what's still valid [since the
- start of the block]. */
- reset_opr_set_tables ();
+ rtx dest = SET_DEST (sset);
+ if ((REG_P (dest) || CC0_P (dest))
+ && cprop_jump (BLOCK_FOR_INSN (insn), insn, NEXT_INSN (insn), from, to))
+ return 1;
+ }
- for (insn = bb->head;
- insn != NULL && insn != NEXT_INSN (bb->end);
- insn = NEXT_INSN (insn))
- {
- /* Is insn of form (set (pseudo-reg) ...)? */
- if (GET_CODE (insn) == INSN
- && GET_CODE (PATTERN (insn)) == SET
- && GET_CODE (SET_DEST (PATTERN (insn))) == REG
- && REGNO (SET_DEST (PATTERN (insn))) >= FIRST_PSEUDO_REGISTER)
- {
- rtx pat = PATTERN (insn);
- rtx src = SET_SRC (pat);
- struct expr *expr;
-
- if (want_to_gcse_p (src)
- /* Is the expression recorded? */
- && ((expr = lookup_expr (src, &expr_hash_table)) != NULL)
- /* Is the expression available [at the start of the
- block]? */
- && TEST_BIT (ae_in[bb->index], expr->bitmap_index)
- /* Are the operands unchanged since the start of the
- block? */
- && oprs_not_set_p (src, insn))
- changed |= handle_avail_expr (insn, expr);
- }
+ /* Handle normal insns next. */
+ if (NONJUMP_INSN_P (insn)
+ && try_replace_reg (from, to, insn))
+ return 1;
- /* Keep track of everything modified by this insn. */
- /* ??? Need to be careful w.r.t. mods done to INSN. */
- if (INSN_P (insn))
- mark_oprs_set (insn);
- }
- }
+ /* Try to propagate a CONST_INT into a conditional jump.
+ We're pretty specific about what we will handle in this
+ code, we can extend this as necessary over time.
- return changed;
+ Right now the insn in question must look like
+ (set (pc) (if_then_else ...)) */
+ else if (any_condjump_p (insn) && onlyjump_p (insn))
+ return cprop_jump (BLOCK_FOR_INSN (insn), NULL, insn, from, to);
+ return 0;
}
-/* Top level routine to perform one classic GCSE pass.
-
- Return nonzero if a change was made. */
+/* Perform constant and copy propagation on INSN.
+ The result is nonzero if a change was made. */
static int
-one_classic_gcse_pass (int pass)
+cprop_insn (rtx insn)
{
+ struct reg_use *reg_used;
int changed = 0;
+ rtx note;
- gcse_subst_count = 0;
- gcse_create_count = 0;
+ if (!INSN_P (insn))
+ return 0;
- alloc_hash_table (max_cuid, &expr_hash_table, 0);
- alloc_rd_mem (last_basic_block, max_cuid);
- compute_hash_table (&expr_hash_table);
- if (gcse_file)
- dump_hash_table (gcse_file, "Expression", &expr_hash_table);
+ reg_use_count = 0;
+ note_uses (&PATTERN (insn), find_used_regs, NULL);
- if (expr_hash_table.n_elems > 0)
- {
- compute_kill_rd ();
- compute_rd ();
- alloc_avail_expr_mem (last_basic_block, expr_hash_table.n_elems);
- compute_ae_gen (&expr_hash_table);
- compute_ae_kill (ae_gen, ae_kill, &expr_hash_table);
- compute_available (ae_gen, ae_kill, ae_out, ae_in);
- changed = classic_gcse ();
- free_avail_expr_mem ();
- }
+ note = find_reg_equal_equiv_note (insn);
- free_rd_mem ();
- free_hash_table (&expr_hash_table);
+ /* We may win even when propagating constants into notes. */
+ if (note)
+ find_used_regs (&XEXP (note, 0), NULL);
- if (gcse_file)
+ for (reg_used = ®_use_table[0]; reg_use_count > 0;
+ reg_used++, reg_use_count--)
{
- fprintf (gcse_file, "\n");
- fprintf (gcse_file, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
- current_function_name, pass, bytes_used, gcse_subst_count);
- fprintf (gcse_file, "%d insns created\n", gcse_create_count);
- }
-
- return changed;
-}
-\f
-/* Compute copy/constant propagation working variables. */
+ unsigned int regno = REGNO (reg_used->reg_rtx);
+ rtx pat, src;
+ struct expr *set;
-/* Local properties of assignments. */
-static sbitmap *cprop_pavloc;
-static sbitmap *cprop_absaltered;
+ /* 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))
+ continue;
-/* Global properties of assignments (computed from the local properties). */
-static sbitmap *cprop_avin;
-static sbitmap *cprop_avout;
+ /* Find an assignment that sets reg_used and is available
+ at the start of the block. */
+ set = find_avail_set (regno, insn);
+ if (! set)
+ continue;
-/* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
- basic blocks. N_SETS is the number of sets. */
+ pat = set->expr;
+ /* ??? We might be able to handle PARALLELs. Later. */
+ gcc_assert (GET_CODE (pat) == SET);
-static void
-alloc_cprop_mem (int n_blocks, int n_sets)
-{
- cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
- cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
+ src = SET_SRC (pat);
- cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
- cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
-}
+ /* Constant propagation. */
+ if (gcse_constant_p (src))
+ {
+ if (constprop_register (insn, reg_used->reg_rtx, src))
+ {
+ changed = 1;
+ global_const_prop_count++;
+ if (dump_file != NULL)
+ {
+ 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;
+ }
+ }
+ else if (REG_P (src)
+ && REGNO (src) >= FIRST_PSEUDO_REGISTER
+ && REGNO (src) != regno)
+ {
+ if (try_replace_reg (reg_used->reg_rtx, src, insn))
+ {
+ changed = 1;
+ global_copy_prop_count++;
+ if (dump_file != NULL)
+ {
+ fprintf (dump_file, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
+ regno, INSN_UID (insn));
+ fprintf (dump_file, " with reg %d\n", REGNO (src));
+ }
-/* Free vars used by copy/const propagation. */
+ /* The original insn setting reg_used may or may not now be
+ deletable. We leave the deletion to flow. */
+ /* FIXME: If it turns out that the insn isn't deletable,
+ then we may have unnecessarily extended register lifetimes
+ and made things worse. */
+ }
+ }
+ }
-static void
-free_cprop_mem (void)
-{
- sbitmap_vector_free (cprop_pavloc);
- sbitmap_vector_free (cprop_absaltered);
- sbitmap_vector_free (cprop_avin);
- sbitmap_vector_free (cprop_avout);
+ if (changed && DEBUG_INSN_P (insn))
+ return 0;
+
+ return changed;
}
-/* For each block, compute whether X is transparent. X is either an
- expression or an assignment [though we don't care which, for this context
- an assignment is treated as an expression]. For each block where an
- element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
- bit in BMAP. */
+/* Like find_used_regs, but avoid recording uses that appear in
+ input-output contexts such as zero_extract or pre_dec. This
+ restricts the cases we consider to those for which local cprop
+ can legitimately make replacements. */
static void
-compute_transp (rtx x, int indx, sbitmap *bmap, int set_p)
+local_cprop_find_used_regs (rtx *xptr, void *data)
{
- 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
- can't do it when there's no return value. */
- repeat:
+ rtx x = *xptr;
if (x == 0)
return;
- code = GET_CODE (x);
- switch (code)
+ switch (GET_CODE (x))
{
- case REG:
- if (set_p)
+ case ZERO_EXTRACT:
+ case SIGN_EXTRACT:
+ case STRICT_LOW_PART:
+ return;
+
+ case PRE_DEC:
+ case PRE_INC:
+ case POST_DEC:
+ case POST_INC:
+ case PRE_MODIFY:
+ case POST_MODIFY:
+ /* Can only legitimately appear this early in the context of
+ stack pushes for function arguments, but handle all of the
+ codes nonetheless. */
+ return;
+
+ case SUBREG:
+ /* Setting a subreg of a register larger than word_mode leaves
+ the non-written words unchanged. */
+ if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) > BITS_PER_WORD)
+ return;
+ break;
+
+ default:
+ break;
+ }
+
+ find_used_regs (xptr, data);
+}
+
+/* Try to perform local const/copy propagation on X in INSN. */
+
+static bool
+do_local_cprop (rtx x, rtx insn)
+{
+ rtx newreg = NULL, newcnst = NULL;
+
+ /* Rule out USE instructions and ASM statements as we don't want to
+ change the hard registers mentioned. */
+ if (REG_P (x)
+ && (REGNO (x) >= FIRST_PSEUDO_REGISTER
+ || (GET_CODE (PATTERN (insn)) != USE
+ && asm_noperands (PATTERN (insn)) < 0)))
+ {
+ cselib_val *val = cselib_lookup (x, GET_MODE (x), 0);
+ struct elt_loc_list *l;
+
+ if (!val)
+ return false;
+ for (l = val->locs; l; l = l->next)
{
- 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);
- }
+ rtx this_rtx = l->loc;
+ rtx note;
+
+ if (gcse_constant_p (this_rtx))
+ newcnst = this_rtx;
+ if (REG_P (this_rtx) && REGNO (this_rtx) >= FIRST_PSEUDO_REGISTER
+ /* Don't copy propagate if it has attached REG_EQUIV note.
+ At this point this only function parameters should have
+ REG_EQUIV notes and if the argument slot is used somewhere
+ explicitly, it means address of parameter has been taken,
+ so we should not extend the lifetime of the pseudo. */
+ && (!(note = find_reg_note (l->setting_insn, REG_EQUIV, NULL_RTX))
+ || ! MEM_P (XEXP (note, 0))))
+ newreg = this_rtx;
}
- else
+ if (newcnst && constprop_register (insn, x, newcnst))
{
- if (REGNO (x) < FIRST_PSEUDO_REGISTER)
+ if (dump_file != NULL)
{
- FOR_EACH_BB (bb)
- if (TEST_BIT (reg_set_in_block[bb->index], REGNO (x)))
- RESET_BIT (bmap[bb->index], indx);
+ fprintf (dump_file, "LOCAL CONST-PROP: Replacing reg %d in ",
+ REGNO (x));
+ fprintf (dump_file, "insn %d with constant ",
+ INSN_UID (insn));
+ print_rtl (dump_file, newcnst);
+ fprintf (dump_file, "\n");
}
- else
+ local_const_prop_count++;
+ return true;
+ }
+ else if (newreg && newreg != x && try_replace_reg (x, newreg, insn))
+ {
+ if (dump_file != NULL)
{
- for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
- RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
+ fprintf (dump_file,
+ "LOCAL COPY-PROP: Replacing reg %d in insn %d",
+ REGNO (x), INSN_UID (insn));
+ fprintf (dump_file, " with reg %d\n", REGNO (newreg));
}
+ local_copy_prop_count++;
+ return true;
}
+ }
+ return false;
+}
- return;
+/* Do local const/copy propagation (i.e. within each basic block). */
- case MEM:
- FOR_EACH_BB (bb)
- {
- rtx list_entry = canon_modify_mem_list[bb->index];
+static int
+local_cprop_pass (void)
+{
+ basic_block bb;
+ rtx insn;
+ struct reg_use *reg_used;
+ bool changed = false;
- while (list_entry)
+ cselib_init (0);
+ FOR_EACH_BB (bb)
+ {
+ FOR_BB_INSNS (bb, insn)
+ {
+ if (INSN_P (insn))
{
- rtx dest, dest_addr;
-
- if (GET_CODE (XEXP (list_entry, 0)) == CALL_INSN)
- {
- if (set_p)
- SET_BIT (bmap[bb->index], indx);
- else
- RESET_BIT (bmap[bb->index], indx);
- break;
- }
- /* LIST_ENTRY must be an INSN of some kind that sets memory.
- Examine each hunk of memory that is modified. */
-
- dest = XEXP (list_entry, 0);
- list_entry = XEXP (list_entry, 1);
- dest_addr = XEXP (list_entry, 0);
-
- if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
- x, rtx_addr_varies_p))
+ rtx note = find_reg_equal_equiv_note (insn);
+ do
{
- if (set_p)
- SET_BIT (bmap[bb->index], indx);
- else
- RESET_BIT (bmap[bb->index], indx);
- break;
+ 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;
}
- list_entry = XEXP (list_entry, 1);
+ while (reg_use_count);
}
+ cselib_process_insn (insn);
}
- x = XEXP (x, 0);
- goto repeat;
-
- case PC:
- case CC0: /*FIXME*/
- case CONST:
- case CONST_INT:
- case CONST_DOUBLE:
- case CONST_VECTOR:
- case SYMBOL_REF:
- case LABEL_REF:
- case ADDR_VEC:
- case ADDR_DIFF_VEC:
- return;
-
- default:
- break;
+ /* Forget everything at the end of a basic block. */
+ cselib_clear_table ();
}
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- /* If we are about to do the last recursive call
- needed at this level, change it into iteration.
- This function is called enough to be worth it. */
- if (i == 0)
- {
- x = XEXP (x, i);
- goto repeat;
- }
+ cselib_finish ();
- compute_transp (XEXP (x, i), indx, bmap, set_p);
- }
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
- }
+ return changed;
}
-/* Top level routine to do the dataflow analysis needed by copy/const
- propagation. */
+/* Similar to get_condition, only the resulting condition must be
+ valid at JUMP, instead of at EARLIEST.
-static void
-compute_cprop_data (void)
+ This differs from noce_get_condition in ifcvt.c in that we prefer not to
+ settle for the condition variable in the jump instruction being integral.
+ We prefer to be able to record the value of a user variable, rather than
+ the value of a temporary used in a condition. This could be solved by
+ recording the value of *every* register scanned by canonicalize_condition,
+ but this would require some code reorganization. */
+
+rtx
+fis_get_condition (rtx jump)
{
- compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, &set_hash_table);
- compute_available (cprop_pavloc, cprop_absaltered,
- cprop_avout, cprop_avin);
+ return get_condition (jump, NULL, false, true);
}
-\f
-/* Copy/constant propagation. */
-/* Maximum number of register uses in an insn that we handle. */
-#define MAX_USES 8
+/* Check the comparison COND to see if we can safely form an implicit set from
+ it. COND is either an EQ or NE comparison. */
-/* Table of uses found in an insn.
- Allocated statically to avoid alloc/free complexity and overhead. */
-static struct reg_use reg_use_table[MAX_USES];
+static bool
+implicit_set_cond_p (const_rtx cond)
+{
+ const enum machine_mode mode = GET_MODE (XEXP (cond, 0));
+ const_rtx cst = XEXP (cond, 1);
-/* Index into `reg_use_table' while building it. */
-static int reg_use_count;
+ /* We can't perform this optimization if either operand might be or might
+ contain a signed zero. */
+ if (HONOR_SIGNED_ZEROS (mode))
+ {
+ /* It is sufficient to check if CST is or contains a zero. We must
+ handle float, complex, and vector. If any subpart is a zero, then
+ the optimization can't be performed. */
+ /* ??? The complex and vector checks are not implemented yet. We just
+ always return zero for them. */
+ if (GET_CODE (cst) == CONST_DOUBLE)
+ {
+ REAL_VALUE_TYPE d;
+ REAL_VALUE_FROM_CONST_DOUBLE (d, cst);
+ if (REAL_VALUES_EQUAL (d, dconst0))
+ return 0;
+ }
+ else
+ return 0;
+ }
-/* Set up a list of register numbers used in INSN. The found uses are stored
- in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
- and contains the number of uses in the table upon exit.
+ return gcse_constant_p (cst);
+}
- ??? If a register appears multiple times we will record it multiple times.
- This doesn't hurt anything but it will slow things down. */
+/* Find the implicit sets of a function. An "implicit set" is a constraint
+ on the value of a variable, implied by a conditional jump. For example,
+ following "if (x == 2)", the then branch may be optimized as though the
+ conditional performed an "explicit set", in this example, "x = 2". This
+ function records the set patterns that are implicit at the start of each
+ basic block.
+
+ 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_used_regs (rtx *xptr, void *data ATTRIBUTE_UNUSED)
+find_implicit_sets (void)
{
- int i, j;
- enum rtx_code code;
- const char *fmt;
- rtx x = *xptr;
-
- /* repeat is used to turn tail-recursion into iteration since GCC
- can't do it when there's no return value. */
- repeat:
- if (x == 0)
- return;
-
- code = GET_CODE (x);
- if (REG_P (x))
- {
- if (reg_use_count == MAX_USES)
- return;
+ basic_block bb, dest;
+ unsigned int count;
+ rtx cond, new_rtx;
- reg_use_table[reg_use_count].reg_rtx = x;
- reg_use_count++;
- }
+ count = 0;
+ FOR_EACH_BB (bb)
+ /* Check for more than one successor. */
+ if (EDGE_COUNT (bb->succs) > 1)
+ {
+ cond = fis_get_condition (BB_END (bb));
- /* Recursively scan the operands of this expression. */
+ if (cond
+ && (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
+ && REG_P (XEXP (cond, 0))
+ && REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
+ && implicit_set_cond_p (cond))
+ {
+ dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
+ : FALLTHRU_EDGE (bb)->dest;
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- /* If we are about to do the last recursive call
- needed at this level, change it into iteration.
- This function is called enough to be worth it. */
- if (i == 0)
- {
- x = XEXP (x, 0);
- goto repeat;
- }
+ if (dest
+ /* Record nothing for a critical edge. */
+ && single_pred_p (dest)
+ && dest != EXIT_BLOCK_PTR)
+ {
+ new_rtx = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
+ XEXP (cond, 1));
+ implicit_sets[dest->index] = new_rtx;
+ if (dump_file)
+ {
+ fprintf(dump_file, "Implicit set of reg %d in ",
+ REGNO (XEXP (cond, 0)));
+ fprintf(dump_file, "basic block %d\n", dest->index);
+ }
+ count++;
+ }
+ }
+ }
- find_used_regs (&XEXP (x, i), data);
- }
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- find_used_regs (&XVECEXP (x, i, j), data);
- }
+ if (dump_file)
+ fprintf (dump_file, "Found %d implicit sets\n", count);
}
-/* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
- Returns nonzero is successful. */
-
-static int
-try_replace_reg (rtx from, rtx to, rtx insn)
-{
- rtx note = find_reg_equal_equiv_note (insn);
- rtx src = 0;
- int success = 0;
- rtx set = single_set (insn);
-
- validate_replace_src_group (from, to, insn);
- if (num_changes_pending () && apply_change_group ())
- success = 1;
-
- /* Try to simplify SET_SRC if we have substituted a constant. */
- if (success && set && CONSTANT_P (to))
- {
- src = simplify_rtx (SET_SRC (set));
-
- if (src)
- validate_change (insn, &SET_SRC (set), src, 0);
- }
-
- if (!success && set && reg_mentioned_p (from, SET_SRC (set)))
- {
- /* If above failed and this is a single set, try to simplify the source of
- the set given our substitution. We could perhaps try this for multiple
- SETs, but it probably won't buy us anything. */
- src = simplify_replace_rtx (SET_SRC (set), from, to);
-
- if (!rtx_equal_p (src, SET_SRC (set))
- && validate_change (insn, &SET_SRC (set), src, 0))
- success = 1;
-
- /* If we've failed to do replacement, have a single SET, don't already
- have a note, and have no special SET, add a REG_EQUAL note to not
- lose information. */
- if (!success && note == 0 && set != 0
- && GET_CODE (XEXP (set, 0)) != ZERO_EXTRACT
- && GET_CODE (XEXP (set, 0)) != SIGN_EXTRACT)
- note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
- }
-
- /* If there is already a NOTE, update the expression in it with our
- replacement. */
- else if (note != 0)
- XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
+/* Bypass conditional jumps. */
- /* REG_EQUAL may get simplified into register.
- We don't allow that. Remove that note. This code ought
- not to happen, because previous code ought to synthesize
- reg-reg move, but be on the safe side. */
- if (note && REG_P (XEXP (note, 0)))
- remove_note (insn, note);
+/* The value of last_basic_block at the beginning of the jump_bypass
+ pass. The use of redirect_edge_and_branch_force may introduce new
+ basic blocks, but the data flow analysis is only valid for basic
+ block indices less than bypass_last_basic_block. */
- return success;
-}
+static int bypass_last_basic_block;
-/* Find a set of REGNOs that are available on entry to INSN's block. Returns
- NULL no such set is found. */
+/* Find a set of REGNO to a constant that is available at the end of basic
+ block BB. Returns NULL if no such set is found. Based heavily upon
+ find_avail_set. */
static struct expr *
-find_avail_set (int regno, rtx insn)
+find_bypass_set (int regno, int bb)
{
- /* SET1 contains the last set found that can be returned to the caller for
- use in a substitution. */
- struct expr *set1 = 0;
-
- /* Loops are not possible here. To get a loop we would need two sets
- available at the start of the block containing INSN. ie we would
- need two sets like this available at the start of the block:
-
- (set (reg X) (reg Y))
- (set (reg Y) (reg X))
+ struct expr *result = 0;
- This can not happen since the set of (reg Y) would have killed the
- set of (reg X) making it unavailable at the start of this block. */
- while (1)
+ for (;;)
{
rtx src;
struct expr *set = lookup_set (regno, &set_hash_table);
- /* Find a set that is available at the start of the block
- which contains INSN. */
while (set)
{
- if (TEST_BIT (cprop_avin[BLOCK_NUM (insn)], set->bitmap_index))
+ if (TEST_BIT (cprop_avout[bb], set->bitmap_index))
break;
set = next_set (regno, set);
}
- /* If no available set was found we've reached the end of the
- (possibly empty) copy chain. */
if (set == 0)
break;
- if (GET_CODE (set->expr) != SET)
- abort ();
+ gcc_assert (GET_CODE (set->expr) == SET);
src = SET_SRC (set->expr);
+ if (gcse_constant_p (src))
+ result = set;
- /* We know the set is available.
- Now check that SRC is ANTLOC (i.e. none of the source operands
- have changed since the start of the block).
-
- If the source operand changed, we may still use it for the next
- iteration of this loop, but we may not use it for substitutions. */
-
- if (gcse_constant_p (src) || oprs_not_set_p (src, insn))
- set1 = set;
-
- /* If the source of the set is anything except a register, then
- we have reached the end of the copy chain. */
- if (GET_CODE (src) != REG)
+ if (! REG_P (src))
break;
- /* Follow the copy chain, ie start another iteration of the loop
- and see if we have an available copy into SRC. */
regno = REGNO (src);
}
+ return result;
+}
- /* SET1 holds the last set that was available and anticipatable at
- INSN. */
- return set1;
+
+/* Subroutine of bypass_block that checks whether a pseudo is killed by
+ any of the instructions inserted on an edge. Jump bypassing places
+ condition code setters on CFG edges using insert_insn_on_edge. This
+ function is required to check that our data flow analysis is still
+ valid prior to commit_edge_insertions. */
+
+static bool
+reg_killed_on_edge (const_rtx reg, const_edge e)
+{
+ rtx insn;
+
+ for (insn = e->insns.r; insn; insn = NEXT_INSN (insn))
+ if (INSN_P (insn) && reg_set_p (reg, insn))
+ return true;
+
+ return false;
}
-/* Subroutine of cprop_insn that tries to propagate constants into
- JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
- it is the instruction that immediately precedes JUMP, and must be a
- single SET of a register. FROM is what we will try to replace,
- SRC is the constant we will try to substitute for it. Returns nonzero
- if a change was made. */
+/* Subroutine of bypass_conditional_jumps that attempts to bypass the given
+ basic block BB which has more than one predecessor. If not NULL, SETCC
+ is the first instruction of BB, which is immediately followed by JUMP_INSN
+ JUMP. Otherwise, SETCC is NULL, and JUMP is the first insn of BB.
+ Returns nonzero if a change was made.
+
+ During the jump bypassing pass, we may place copies of SETCC instructions
+ on CFG edges. The following routine must be careful to pay attention to
+ these inserted insns when performing its transformations. */
static int
-cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
+bypass_block (basic_block bb, rtx setcc, rtx jump)
{
- rtx new, set_src, note_src;
- rtx set = pc_set (jump);
- rtx note = find_reg_equal_equiv_note (jump);
+ rtx insn, note;
+ edge e, edest;
+ int i, change;
+ int may_be_loop_header;
+ unsigned removed_p;
+ edge_iterator ei;
+
+ insn = (setcc != NULL) ? setcc : jump;
+ /* Determine set of register uses in INSN. */
+ reg_use_count = 0;
+ note_uses (&PATTERN (insn), find_used_regs, NULL);
+ note = find_reg_equal_equiv_note (insn);
if (note)
- {
- note_src = XEXP (note, 0);
- if (GET_CODE (note_src) == EXPR_LIST)
- note_src = NULL_RTX;
- }
- else note_src = NULL_RTX;
+ find_used_regs (&XEXP (note, 0), NULL);
- /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
- set_src = note_src ? note_src : SET_SRC (set);
+ may_be_loop_header = false;
+ FOR_EACH_EDGE (e, ei, bb->preds)
+ if (e->flags & EDGE_DFS_BACK)
+ {
+ may_be_loop_header = true;
+ break;
+ }
- /* First substitute the SETCC condition into the JUMP instruction,
- then substitute that given values into this expanded JUMP. */
- if (setcc != NULL_RTX
- && !modified_between_p (from, setcc, jump)
- && !modified_between_p (src, setcc, jump))
+ change = 0;
+ for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
{
- rtx setcc_src;
- rtx setcc_set = single_set (setcc);
- rtx setcc_note = find_reg_equal_equiv_note (setcc);
- setcc_src = (setcc_note && GET_CODE (XEXP (setcc_note, 0)) != EXPR_LIST)
- ? XEXP (setcc_note, 0) : SET_SRC (setcc_set);
- set_src = simplify_replace_rtx (set_src, SET_DEST (setcc_set),
- setcc_src);
- }
- else
- setcc = NULL_RTX;
+ removed_p = 0;
- new = 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)))
- return 0;
+ if (e->flags & EDGE_COMPLEX)
+ {
+ ei_next (&ei);
+ continue;
+ }
- /* If this is now a no-op delete it, otherwise this must be a valid insn. */
- if (new == 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))
- return 0;
- if (! validate_change (jump, &SET_SRC (set), new, 0))
+ /* We can't redirect edges from new basic blocks. */
+ if (e->src->index >= bypass_last_basic_block)
{
- /* When (some) constants are not valid in a comparison, and there
- are two registers to be replaced by constants before the entire
- comparison can be folded into a constant, we need to keep
- intermediate information in REG_EQUAL notes. For targets with
- separate compare insns, such notes are added by try_replace_reg.
- When we have a combined compare-and-branch instruction, however,
- we need to attach a note to the branch itself to make this
- optimization work. */
+ ei_next (&ei);
+ continue;
+ }
- if (!rtx_equal_p (new, note_src))
- set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new));
- return 0;
+ /* The irreducible loops created by redirecting of edges entering the
+ loop from outside would decrease effectiveness of some of the following
+ optimizations, so prevent this. */
+ if (may_be_loop_header
+ && !(e->flags & EDGE_DFS_BACK))
+ {
+ ei_next (&ei);
+ continue;
}
- /* Remove REG_EQUAL note after simplification. */
- if (note_src)
- remove_note (jump, note);
+ for (i = 0; i < reg_use_count; i++)
+ {
+ struct reg_use *reg_used = ®_use_table[i];
+ unsigned int regno = REGNO (reg_used->reg_rtx);
+ basic_block dest, old_dest;
+ struct expr *set;
+ rtx src, new_rtx;
- /* If 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);
- }
+ set = find_bypass_set (regno, e->src->index);
-#ifdef HAVE_cc0
- /* Delete the cc0 setter. */
- if (setcc != NULL && CC0_P (SET_DEST (single_set (setcc))))
- delete_insn (setcc);
-#endif
+ if (! set)
+ continue;
- run_jump_opt_after_gcse = 1;
+ /* Check the data flow is valid after edge insertions. */
+ if (e->insns.r && reg_killed_on_edge (reg_used->reg_rtx, e))
+ continue;
- const_prop_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file,
- "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");
- }
- purge_dead_edges (bb);
+ src = SET_SRC (pc_set (jump));
- return 1;
-}
+ if (setcc != NULL)
+ src = simplify_replace_rtx (src,
+ SET_DEST (PATTERN (setcc)),
+ SET_SRC (PATTERN (setcc)));
-static bool
-constprop_register (rtx insn, rtx from, rtx to, int alter_jumps)
-{
- rtx sset;
+ new_rtx = simplify_replace_rtx (src, reg_used->reg_rtx,
+ SET_SRC (set->expr));
- /* Check for reg or cc0 setting instructions followed by
- conditional branch instructions first. */
- if (alter_jumps
- && (sset = single_set (insn)) != NULL
- && NEXT_INSN (insn)
- && any_condjump_p (NEXT_INSN (insn)) && onlyjump_p (NEXT_INSN (insn)))
- {
- rtx dest = SET_DEST (sset);
- if ((REG_P (dest) || CC0_P (dest))
- && cprop_jump (BLOCK_FOR_INSN (insn), insn, NEXT_INSN (insn), from, to))
- return 1;
- }
+ /* Jump bypassing may have already placed instructions on
+ edges of the CFG. We can't bypass an outgoing edge that
+ has instructions associated with it, as these insns won't
+ get executed if the incoming edge is redirected. */
- /* Handle normal insns next. */
- if (GET_CODE (insn) == INSN
- && try_replace_reg (from, to, insn))
- return 1;
+ if (new_rtx == pc_rtx)
+ {
+ edest = FALLTHRU_EDGE (bb);
+ dest = edest->insns.r ? NULL : edest->dest;
+ }
+ else if (GET_CODE (new_rtx) == LABEL_REF)
+ {
+ dest = BLOCK_FOR_INSN (XEXP (new_rtx, 0));
+ /* Don't bypass edges containing instructions. */
+ edest = find_edge (bb, dest);
+ if (edest && edest->insns.r)
+ dest = NULL;
+ }
+ else
+ dest = NULL;
- /* Try to propagate a CONST_INT into a conditional jump.
- We're pretty specific about what we will handle in this
- code, we can extend this as necessary over time.
+ /* Avoid unification of the edge with other edges from original
+ branch. We would end up emitting the instruction on "both"
+ edges. */
- Right now the insn in question must look like
- (set (pc) (if_then_else ...)) */
- else if (alter_jumps && any_condjump_p (insn) && onlyjump_p (insn))
- return cprop_jump (BLOCK_FOR_INSN (insn), NULL, insn, from, to);
- return 0;
+ if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc)))
+ && find_edge (e->src, dest))
+ dest = NULL;
+
+ old_dest = e->dest;
+ if (dest != NULL
+ && dest != old_dest
+ && dest != EXIT_BLOCK_PTR)
+ {
+ redirect_edge_and_branch_force (e, dest);
+
+ /* Copy the register setter to the redirected edge.
+ Don't copy CC0 setters, as CC0 is dead after jump. */
+ if (setcc)
+ {
+ rtx pat = PATTERN (setcc);
+ if (!CC0_P (SET_DEST (pat)))
+ insert_insn_on_edge (copy_insn (pat), e);
+ }
+
+ if (dump_file != NULL)
+ {
+ fprintf (dump_file, "JUMP-BYPASS: Proved reg %d "
+ "in jump_insn %d equals constant ",
+ regno, INSN_UID (jump));
+ print_rtl (dump_file, SET_SRC (set->expr));
+ fprintf (dump_file, "\nBypass edge from %d->%d to %d\n",
+ e->src->index, old_dest->index, dest->index);
+ }
+ change = 1;
+ removed_p = 1;
+ break;
+ }
+ }
+ if (!removed_p)
+ ei_next (&ei);
+ }
+ return change;
}
-/* Perform constant and copy propagation on INSN.
- The result is nonzero if a change was made. */
+/* Find basic blocks with more than one predecessor that only contain a
+ single conditional jump. If the result of the comparison is known at
+ compile-time from any incoming edge, redirect that edge to the
+ appropriate target. Returns nonzero if a change was made.
+
+ This function is now mis-named, because we also handle indirect jumps. */
static int
-cprop_insn (rtx insn, int alter_jumps)
+bypass_conditional_jumps (void)
{
- struct reg_use *reg_used;
- int changed = 0;
- rtx note;
+ basic_block bb;
+ int changed;
+ rtx setcc;
+ rtx insn;
+ rtx dest;
- if (!INSN_P (insn))
+ /* Note we start at block 1. */
+ if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
return 0;
- reg_use_count = 0;
- note_uses (&PATTERN (insn), find_used_regs, NULL);
-
- note = find_reg_equal_equiv_note (insn);
-
- /* We may win even when propagating constants into notes. */
- if (note)
- find_used_regs (&XEXP (note, 0), NULL);
+ bypass_last_basic_block = last_basic_block;
+ mark_dfs_back_edges ();
- for (reg_used = ®_use_table[0]; reg_use_count > 0;
- reg_used++, reg_use_count--)
+ changed = 0;
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb,
+ EXIT_BLOCK_PTR, next_bb)
{
- unsigned int regno = REGNO (reg_used->reg_rtx);
- 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))
- continue;
-
- /* Find an assignment that sets reg_used and is available
- at the start of the block. */
- set = find_avail_set (regno, insn);
- if (! set)
- continue;
-
- pat = set->expr;
- /* ??? We might be able to handle PARALLELs. Later. */
- if (GET_CODE (pat) != SET)
- abort ();
-
- src = SET_SRC (pat);
-
- /* Constant propagation. */
- if (gcse_constant_p (src))
- {
- if (constprop_register (insn, reg_used->reg_rtx, src, alter_jumps))
- {
- changed = 1;
- const_prop_count++;
- if (gcse_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");
- }
- if (INSN_DELETED_P (insn))
- return 1;
- }
- }
- else if (GET_CODE (src) == REG
- && REGNO (src) >= FIRST_PSEUDO_REGISTER
- && REGNO (src) != regno)
+ /* Check for more than one predecessor. */
+ if (!single_pred_p (bb))
{
- if (try_replace_reg (reg_used->reg_rtx, src, insn))
- {
- changed = 1;
- copy_prop_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
- regno, INSN_UID (insn));
- fprintf (gcse_file, " with reg %d\n", REGNO (src));
- }
+ setcc = NULL_RTX;
+ FOR_BB_INSNS (bb, insn)
+ if (DEBUG_INSN_P (insn))
+ continue;
+ else if (NONJUMP_INSN_P (insn))
+ {
+ if (setcc)
+ break;
+ if (GET_CODE (PATTERN (insn)) != SET)
+ break;
- /* The original insn setting reg_used may or may not now be
- deletable. We leave the deletion to flow. */
- /* FIXME: If it turns out that the insn isn't deletable,
- then we may have unnecessarily extended register lifetimes
- and made things worse. */
- }
+ dest = SET_DEST (PATTERN (insn));
+ if (REG_P (dest) || CC0_P (dest))
+ setcc = insn;
+ else
+ break;
+ }
+ else if (JUMP_P (insn))
+ {
+ if ((any_condjump_p (insn) || computed_jump_p (insn))
+ && onlyjump_p (insn))
+ changed |= bypass_block (bb, setcc, insn);
+ break;
+ }
+ else if (INSN_P (insn))
+ break;
}
}
+ /* If we bypassed any register setting insns, we inserted a
+ copy on the redirected edge. These need to be committed. */
+ if (changed)
+ commit_edge_insertions ();
+
return changed;
}
+\f
+/* Compute PRE+LCM working variables. */
-/* Like find_used_regs, but avoid recording uses that appear in
- input-output contexts such as zero_extract or pre_dec. This
- restricts the cases we consider to those for which local cprop
- can legitimately make replacements. */
+/* Local properties of expressions. */
+/* Nonzero for expressions that are transparent in the block. */
+static sbitmap *transp;
-static void
-local_cprop_find_used_regs (rtx *xptr, void *data)
-{
- rtx x = *xptr;
+/* Nonzero for expressions that are transparent at the end of the block.
+ This is only zero for expressions killed by abnormal critical edge
+ created by a calls. */
+static sbitmap *transpout;
- if (x == 0)
- return;
+/* Nonzero for expressions that are computed (available) in the block. */
+static sbitmap *comp;
- switch (GET_CODE (x))
- {
- case ZERO_EXTRACT:
- case SIGN_EXTRACT:
- case STRICT_LOW_PART:
- return;
+/* Nonzero for expressions that are locally anticipatable in the block. */
+static sbitmap *antloc;
- case PRE_DEC:
- case PRE_INC:
- case POST_DEC:
- case POST_INC:
- case PRE_MODIFY:
- case POST_MODIFY:
- /* Can only legitimately appear this early in the context of
- stack pushes for function arguments, but handle all of the
- codes nonetheless. */
- return;
-
- case SUBREG:
- /* Setting a subreg of a register larger than word_mode leaves
- the non-written words unchanged. */
- if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) > BITS_PER_WORD)
- return;
- break;
+/* Nonzero for expressions where this block is an optimal computation
+ point. */
+static sbitmap *pre_optimal;
- default:
- break;
- }
+/* Nonzero for expressions which are redundant in a particular block. */
+static sbitmap *pre_redundant;
- find_used_regs (xptr, data);
-}
+/* Nonzero for expressions which should be inserted on a specific edge. */
+static sbitmap *pre_insert_map;
-/* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
- their REG_EQUAL notes need updating. */
+/* Nonzero for expressions which should be deleted in a specific block. */
+static sbitmap *pre_delete_map;
-static bool
-do_local_cprop (rtx x, rtx insn, int alter_jumps, rtx *libcall_sp)
-{
- rtx newreg = NULL, newcnst = NULL;
+/* Contains the edge_list returned by pre_edge_lcm. */
+static struct edge_list *edge_list;
- /* Rule out USE instructions and ASM statements as we don't want to
- change the hard registers mentioned. */
- if (GET_CODE (x) == REG
- && (REGNO (x) >= FIRST_PSEUDO_REGISTER
- || (GET_CODE (PATTERN (insn)) != USE
- && asm_noperands (PATTERN (insn)) < 0)))
- {
- cselib_val *val = cselib_lookup (x, GET_MODE (x), 0);
- struct elt_loc_list *l;
+/* Allocate vars used for PRE analysis. */
- if (!val)
- return false;
- for (l = val->locs; l; l = l->next)
- {
- rtx this_rtx = l->loc;
- rtx note;
+static void
+alloc_pre_mem (int n_blocks, int n_exprs)
+{
+ transp = sbitmap_vector_alloc (n_blocks, n_exprs);
+ comp = sbitmap_vector_alloc (n_blocks, n_exprs);
+ antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
- if (l->in_libcall)
- continue;
+ pre_optimal = NULL;
+ pre_redundant = NULL;
+ pre_insert_map = NULL;
+ pre_delete_map = NULL;
+ ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
- if (gcse_constant_p (this_rtx))
- newcnst = this_rtx;
- if (REG_P (this_rtx) && REGNO (this_rtx) >= FIRST_PSEUDO_REGISTER
- /* Don't copy propagate if it has attached REG_EQUIV note.
- At this point this only function parameters should have
- REG_EQUIV notes and if the argument slot is used somewhere
- explicitly, it means address of parameter has been taken,
- so we should not extend the lifetime of the pseudo. */
- && (!(note = find_reg_note (l->setting_insn, REG_EQUIV, NULL_RTX))
- || GET_CODE (XEXP (note, 0)) != MEM))
- newreg = this_rtx;
- }
- if (newcnst && constprop_register (insn, x, newcnst, alter_jumps))
- {
- /* 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. */
- if (!adjust_libcall_notes (x, newcnst, insn, libcall_sp))
- abort ();
- if (gcse_file != NULL)
- {
- fprintf (gcse_file, "LOCAL CONST-PROP: Replacing reg %d in ",
- REGNO (x));
- fprintf (gcse_file, "insn %d with constant ",
- INSN_UID (insn));
- print_rtl (gcse_file, newcnst);
- fprintf (gcse_file, "\n");
- }
- 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)
- {
- fprintf (gcse_file,
- "LOCAL COPY-PROP: Replacing reg %d in insn %d",
- REGNO (x), INSN_UID (insn));
- fprintf (gcse_file, " with reg %d\n", REGNO (newreg));
- }
- copy_prop_count++;
- return true;
- }
- }
- return false;
+ /* pre_insert and pre_delete are allocated later. */
}
-/* 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)
+/* Free vars used for PRE analysis. */
+
+static void
+free_pre_mem (void)
{
- rtx end;
+ sbitmap_vector_free (transp);
+ sbitmap_vector_free (comp);
- while ((end = *libcall_sp++))
- {
- rtx note = find_reg_equal_equiv_note (end);
+ /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
- if (! note)
- continue;
+ if (pre_optimal)
+ sbitmap_vector_free (pre_optimal);
+ if (pre_redundant)
+ sbitmap_vector_free (pre_redundant);
+ if (pre_insert_map)
+ sbitmap_vector_free (pre_insert_map);
+ if (pre_delete_map)
+ sbitmap_vector_free (pre_delete_map);
- if (REG_P (newval))
- {
- if (reg_set_between_p (newval, PREV_INSN (insn), end))
- {
- 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) = replace_rtx (XEXP (note, 0), oldreg, newval);
- insn = end;
- }
- return true;
+ transp = comp = NULL;
+ pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
}
-#define MAX_NESTED_LIBCALLS 9
+/* Top level routine to do the dataflow analysis needed by PRE. */
static void
-local_cprop_pass (int alter_jumps)
+compute_pre_data (void)
{
- rtx insn;
- struct reg_use *reg_used;
- rtx libcall_stack[MAX_NESTED_LIBCALLS + 1], *libcall_sp;
- bool changed = false;
+ sbitmap trapping_expr;
+ basic_block bb;
+ unsigned int ui;
- cselib_init ();
- 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);
+ compute_local_properties (transp, comp, antloc, &expr_hash_table);
+ sbitmap_vector_zero (ae_kill, last_basic_block);
- if (note)
- {
- if (libcall_sp == libcall_stack)
- abort ();
- *--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;
- break;
- }
- if (INSN_DELETED_P (insn))
- break;
- }
- while (reg_use_count);
- }
- cselib_process_insn (insn);
- }
- cselib_finish ();
- /* Global analysis may get into infinite loops for unreachable blocks. */
- if (changed && alter_jumps)
+ /* Collect expressions which might trap. */
+ trapping_expr = sbitmap_alloc (expr_hash_table.n_elems);
+ sbitmap_zero (trapping_expr);
+ for (ui = 0; ui < expr_hash_table.size; ui++)
{
- delete_unreachable_blocks ();
- free_reg_set_mem ();
- alloc_reg_set_mem (max_reg_num ());
- compute_sets (get_insns ());
+ struct expr *e;
+ for (e = expr_hash_table.table[ui]; e != NULL; e = e->next_same_hash)
+ if (may_trap_p (e->expr))
+ SET_BIT (trapping_expr, e->bitmap_index);
}
-}
-
-/* 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;
+ /* Compute ae_kill for each basic block using:
- /* 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;
- }
+ ~(TRANSP | COMP)
+ */
- changed = 0;
- FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
+ FOR_EACH_BB (bb)
{
- /* Reset tables used to keep track of what's still valid [since the
- start of the block]. */
- reset_opr_set_tables ();
+ edge e;
+ edge_iterator ei;
- for (insn = bb->head;
- insn != NULL && insn != NEXT_INSN (bb->end);
- insn = NEXT_INSN (insn))
- if (INSN_P (insn))
+ /* If the current block is the destination of an abnormal edge, we
+ kill all trapping expressions because we won't be able to properly
+ place the instruction on the edge. So make them neither
+ anticipatable nor transparent. This is fairly conservative. */
+ FOR_EACH_EDGE (e, ei, bb->preds)
+ if (e->flags & EDGE_ABNORMAL)
{
- 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 (GET_CODE (insn) != NOTE)
- mark_oprs_set (insn);
+ sbitmap_difference (antloc[bb->index], antloc[bb->index], trapping_expr);
+ sbitmap_difference (transp[bb->index], transp[bb->index], trapping_expr);
+ break;
}
- }
- if (gcse_file != NULL)
- fprintf (gcse_file, "\n");
+ sbitmap_a_or_b (ae_kill[bb->index], transp[bb->index], comp[bb->index]);
+ sbitmap_not (ae_kill[bb->index], ae_kill[bb->index]);
+ }
- return changed;
+ edge_list = pre_edge_lcm (expr_hash_table.n_elems, transp, comp, antloc,
+ ae_kill, &pre_insert_map, &pre_delete_map);
+ sbitmap_vector_free (antloc);
+ antloc = NULL;
+ sbitmap_vector_free (ae_kill);
+ ae_kill = NULL;
+ sbitmap_free (trapping_expr);
}
+\f
+/* PRE utilities */
-/* Similar to get_condition, only the resulting condition must be
- valid at JUMP, instead of at EARLIEST.
-
- This differs from noce_get_condition in ifcvt.c in that we prefer not to
- settle for the condition variable in the jump instruction being integral.
- We prefer to be able to record the value of a user variable, rather than
- the value of a temporary used in a condition. This could be solved by
- recording the value of *every* register scaned by canonicalize_condition,
- but this would require some code reorganization. */
+/* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
+ block BB.
-rtx
-fis_get_condition (rtx jump)
-{
- rtx cond, set, tmp, insn, earliest;
- bool reverse;
-
- if (! any_condjump_p (jump))
- return NULL_RTX;
-
- set = pc_set (jump);
- cond = XEXP (SET_SRC (set), 0);
-
- /* If this branches to JUMP_LABEL when the condition is false,
- reverse the condition. */
- reverse = (GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
- && XEXP (XEXP (SET_SRC (set), 2), 0) == JUMP_LABEL (jump));
-
- /* Use canonicalize_condition to do the dirty work of manipulating
- MODE_CC values and COMPARE rtx codes. */
- tmp = canonicalize_condition (jump, cond, reverse, &earliest, NULL_RTX);
- if (!tmp)
- return NULL_RTX;
-
- /* Verify that the given condition is valid at JUMP by virtue of not
- having been modified since EARLIEST. */
- for (insn = earliest; insn != jump; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && modified_in_p (tmp, insn))
- break;
- if (insn == jump)
- return tmp;
-
- /* The condition was modified. See if we can get a partial result
- that doesn't follow all the reversals. Perhaps combine can fold
- them together later. */
- tmp = XEXP (tmp, 0);
- if (!REG_P (tmp) || GET_MODE_CLASS (GET_MODE (tmp)) != MODE_INT)
- return NULL_RTX;
- tmp = canonicalize_condition (jump, cond, reverse, &earliest, tmp);
- if (!tmp)
- return NULL_RTX;
-
- /* For sanity's sake, re-validate the new result. */
- for (insn = earliest; insn != jump; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && modified_in_p (tmp, insn))
- return NULL_RTX;
-
- return tmp;
-}
+ VISITED is a pointer to a working buffer for tracking which BB's have
+ been visited. It is NULL for the top-level call.
-/* Find the implicit sets of a function. An "implicit set" is a constraint
- on the value of a variable, implied by a conditional jump. For example,
- following "if (x == 2)", the then branch may be optimized as though the
- conditional performed an "explicit set", in this example, "x = 2". This
- function records the set patterns that are implicit at the start of each
- basic block. */
+ We treat reaching expressions that go through blocks containing the same
+ reaching expression as "not reaching". E.g. if EXPR is generated in blocks
+ 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
+ 2 as not reaching. The intent is to improve the probability of finding
+ only one reaching expression and to reduce register lifetimes by picking
+ the closest such expression. */
-static void
-find_implicit_sets (void)
+static int
+pre_expr_reaches_here_p_work (basic_block occr_bb, struct expr *expr, basic_block bb, char *visited)
{
- basic_block bb, dest;
- unsigned int count;
- rtx cond, new;
-
- count = 0;
- FOR_EACH_BB (bb)
- /* Check for more than one sucessor. */
- if (bb->succ && bb->succ->succ_next)
- {
- cond = fis_get_condition (bb->end);
+ edge pred;
+ edge_iterator ei;
- if (cond
- && (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
- && GET_CODE (XEXP (cond, 0)) == REG
- && REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
- && gcse_constant_p (XEXP (cond, 1)))
- {
- dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
- : FALLTHRU_EDGE (bb)->dest;
+ FOR_EACH_EDGE (pred, ei, bb->preds)
+ {
+ basic_block pred_bb = pred->src;
- if (dest && ! dest->pred->pred_next
- && dest != EXIT_BLOCK_PTR)
- {
- new = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
- XEXP (cond, 1));
- implicit_sets[dest->index] = new;
- if (gcse_file)
- {
- fprintf(gcse_file, "Implicit set of reg %d in ",
- REGNO (XEXP (cond, 0)));
- fprintf(gcse_file, "basic block %d\n", dest->index);
- }
- count++;
- }
- }
- }
+ if (pred->src == ENTRY_BLOCK_PTR
+ /* Has predecessor has already been visited? */
+ || visited[pred_bb->index])
+ ;/* Nothing to do. */
- if (gcse_file)
- fprintf (gcse_file, "Found %d implicit sets\n", count);
+ /* Does this predecessor generate this expression? */
+ else if (TEST_BIT (comp[pred_bb->index], expr->bitmap_index))
+ {
+ /* Is this the occurrence we're looking for?
+ Note that there's only one generating occurrence per block
+ so we just need to check the block number. */
+ if (occr_bb == pred_bb)
+ return 1;
+
+ visited[pred_bb->index] = 1;
+ }
+ /* Ignore this predecessor if it kills the expression. */
+ else if (! TEST_BIT (transp[pred_bb->index], expr->bitmap_index))
+ visited[pred_bb->index] = 1;
+
+ /* Neither gen nor kill. */
+ else
+ {
+ visited[pred_bb->index] = 1;
+ if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
+ return 1;
+ }
+ }
+
+ /* All paths have been checked. */
+ return 0;
}
-/* 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. */
+/* The wrapper for pre_expr_reaches_here_work that ensures that any
+ memory allocated for that function is returned. */
static int
-one_cprop_pass (int pass, int cprop_jumps, int bypass_jumps)
+pre_expr_reaches_here_p (basic_block occr_bb, struct expr *expr, basic_block bb)
{
- int changed = 0;
+ int rval;
+ char *visited = XCNEWVEC (char, last_basic_block);
+
+ rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
- const_prop_count = 0;
- copy_prop_count = 0;
+ free (visited);
+ return rval;
+}
+\f
- local_cprop_pass (cprop_jumps);
+/* Given an expr, generate RTL which we can insert at the end of a BB,
+ or on an edge. Set the block number of any insns generated to
+ the value of BB. */
- /* Determine implicit sets. */
- implicit_sets = xcalloc (last_basic_block, sizeof (rtx));
- find_implicit_sets ();
+static rtx
+process_insert_insn (struct expr *expr)
+{
+ rtx reg = expr->reaching_reg;
+ rtx exp = copy_rtx (expr->expr);
+ rtx pat;
- alloc_hash_table (max_cuid, &set_hash_table, 1);
- compute_hash_table (&set_hash_table);
+ start_sequence ();
- /* Free implicit_sets before peak usage. */
- free (implicit_sets);
- implicit_sets = NULL;
+ /* If the expression is something that's an operand, like a constant,
+ just copy it to a register. */
+ if (general_operand (exp, GET_MODE (reg)))
+ emit_move_insn (reg, exp);
- if (gcse_file)
- dump_hash_table (gcse_file, "SET", &set_hash_table);
- if (set_hash_table.n_elems > 0)
+ /* Otherwise, make a new insn to compute this expression and make sure the
+ insn will be recognized (this also adds any needed CLOBBERs). Copy the
+ expression to make sure we don't have any sharing issues. */
+ else
{
- 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 ();
+ rtx insn = emit_insn (gen_rtx_SET (VOIDmode, reg, exp));
+
+ if (insn_invalid_p (insn))
+ gcc_unreachable ();
}
- 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 const props, %d copy props\n\n",
- const_prop_count, copy_prop_count);
- }
- /* Global analysis may get into infinite loops for unreachable blocks. */
- if (changed && cprop_jumps)
- delete_unreachable_blocks ();
+ pat = get_insns ();
+ end_sequence ();
- return changed;
+ return pat;
}
-\f
-/* Bypass conditional jumps. */
-/* The value of last_basic_block at the beginning of the jump_bypass
- pass. The use of redirect_edge_and_branch_force may introduce new
- basic blocks, but the data flow analysis is only valid for basic
- block indices less than bypass_last_basic_block. */
+/* Add EXPR to the end of basic block BB.
-static int bypass_last_basic_block;
+ This is used by both the PRE and code hoisting.
-/* Find a set of REGNO to a constant that is available at the end of basic
- block BB. Returns NULL if no such set is found. Based heavily upon
- find_avail_set. */
+ For PRE, we want to verify that the expr is either transparent
+ or locally anticipatable in the target block. This check makes
+ no sense for code hoisting. */
-static struct expr *
-find_bypass_set (int regno, int bb)
+static void
+insert_insn_end_basic_block (struct expr *expr, basic_block bb, int pre)
{
- struct expr *result = 0;
+ rtx insn = BB_END (bb);
+ rtx new_insn;
+ rtx reg = expr->reaching_reg;
+ int regno = REGNO (reg);
+ rtx pat, pat_end;
- for (;;)
+ pat = process_insert_insn (expr);
+ gcc_assert (pat && INSN_P (pat));
+
+ pat_end = pat;
+ while (NEXT_INSN (pat_end) != NULL_RTX)
+ pat_end = NEXT_INSN (pat_end);
+
+ /* If the last insn is a jump, insert EXPR in front [taking care to
+ handle cc0, etc. properly]. Similarly we need to care trapping
+ instructions in presence of non-call exceptions. */
+
+ if (JUMP_P (insn)
+ || (NONJUMP_INSN_P (insn)
+ && (!single_succ_p (bb)
+ || single_succ_edge (bb)->flags & EDGE_ABNORMAL)))
{
- rtx src;
- struct expr *set = lookup_set (regno, &set_hash_table);
+#ifdef HAVE_cc0
+ rtx note;
+#endif
+ /* It should always be the case that we can put these instructions
+ anywhere in the basic block with performing PRE optimizations.
+ Check this. */
+ gcc_assert (!NONJUMP_INSN_P (insn) || !pre
+ || TEST_BIT (antloc[bb->index], expr->bitmap_index)
+ || TEST_BIT (transp[bb->index], expr->bitmap_index));
- while (set)
+ /* If this is a jump table, then we can't insert stuff here. Since
+ we know the previous real insn must be the tablejump, we insert
+ the new instruction just before the tablejump. */
+ if (GET_CODE (PATTERN (insn)) == ADDR_VEC
+ || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
+ insn = prev_real_insn (insn);
+
+#ifdef HAVE_cc0
+ /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
+ if cc0 isn't set. */
+ note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
+ if (note)
+ insn = XEXP (note, 0);
+ else
{
- if (TEST_BIT (cprop_avout[bb], set->bitmap_index))
- break;
- set = next_set (regno, set);
+ rtx maybe_cc0_setter = prev_nonnote_insn (insn);
+ if (maybe_cc0_setter
+ && INSN_P (maybe_cc0_setter)
+ && sets_cc0_p (PATTERN (maybe_cc0_setter)))
+ insn = maybe_cc0_setter;
}
+#endif
+ /* FIXME: What if something in cc0/jump uses value set in new insn? */
+ new_insn = emit_insn_before_noloc (pat, insn, bb);
+ }
- if (set == 0)
- break;
+ /* Likewise if the last insn is a call, as will happen in the presence
+ of exception handling. */
+ else if (CALL_P (insn)
+ && (!single_succ_p (bb)
+ || single_succ_edge (bb)->flags & EDGE_ABNORMAL))
+ {
+ /* Keeping in mind targets with small register classes and parameters
+ in registers, we search backward and place the instructions before
+ the first parameter is loaded. Do this for everyone for consistency
+ and a presumption that we'll get better code elsewhere as well.
- if (GET_CODE (set->expr) != SET)
- abort ();
+ It should always be the case that we can put these instructions
+ anywhere in the basic block with performing PRE optimizations.
+ Check this. */
- src = SET_SRC (set->expr);
- if (gcse_constant_p (src))
- result = set;
+ gcc_assert (!pre
+ || TEST_BIT (antloc[bb->index], expr->bitmap_index)
+ || TEST_BIT (transp[bb->index], expr->bitmap_index));
- if (GET_CODE (src) != REG)
- break;
+ /* Since different machines initialize their parameter registers
+ in different orders, assume nothing. Collect the set of all
+ parameter registers. */
+ insn = find_first_parameter_load (insn, BB_HEAD (bb));
- regno = REGNO (src);
- }
- return result;
-}
+ /* If we found all the parameter loads, then we want to insert
+ before the first parameter load.
+ If we did not find all the parameter loads, then we might have
+ stopped on the head of the block, which could be a CODE_LABEL.
+ If we inserted before the CODE_LABEL, then we would be putting
+ the insn in the wrong basic block. In that case, put the insn
+ after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
+ while (LABEL_P (insn)
+ || NOTE_INSN_BASIC_BLOCK_P (insn))
+ insn = NEXT_INSN (insn);
-/* Subroutine of bypass_block that checks whether a pseudo is killed by
- any of the instructions inserted on an edge. Jump bypassing places
- condition code setters on CFG edges using insert_insn_on_edge. This
- function is required to check that our data flow analysis is still
- valid prior to commit_edge_insertions. */
+ new_insn = emit_insn_before_noloc (pat, insn, bb);
+ }
+ else
+ new_insn = emit_insn_after_noloc (pat, insn, bb);
-static bool
-reg_killed_on_edge (rtx reg, edge e)
-{
- rtx insn;
+ while (1)
+ {
+ if (INSN_P (pat))
+ add_label_notes (PATTERN (pat), new_insn);
+ if (pat == pat_end)
+ break;
+ pat = NEXT_INSN (pat);
+ }
- for (insn = e->insns; insn; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && reg_set_p (reg, insn))
- return true;
+ gcse_create_count++;
- return false;
+ if (dump_file)
+ {
+ fprintf (dump_file, "PRE/HOIST: end of bb %d, insn %d, ",
+ bb->index, INSN_UID (new_insn));
+ fprintf (dump_file, "copying expression %d to reg %d\n",
+ expr->bitmap_index, regno);
+ }
}
-/* Subroutine of bypass_conditional_jumps that attempts to bypass the given
- basic block BB which has more than one predecessor. If not NULL, SETCC
- is the first instruction of BB, which is immediately followed by JUMP_INSN
- JUMP. Otherwise, SETCC is NULL, and JUMP is the first insn of BB.
- Returns nonzero if a change was made.
-
- During the jump bypassing pass, we may place copies of SETCC instructions
- on CFG edges. The following routine must be careful to pay attention to
- these inserted insns when performing its transformations. */
+/* Insert partially redundant expressions on edges in the CFG to make
+ the expressions fully redundant. */
static int
-bypass_block (basic_block bb, rtx setcc, rtx jump)
+pre_edge_insert (struct edge_list *edge_list, struct expr **index_map)
{
- rtx insn, note;
- edge e, enext, edest;
- int i, change;
- int may_be_loop_header;
-
- insn = (setcc != NULL) ? setcc : jump;
+ int e, i, j, num_edges, set_size, did_insert = 0;
+ sbitmap *inserted;
- /* Determine set of register uses in INSN. */
- reg_use_count = 0;
- note_uses (&PATTERN (insn), find_used_regs, NULL);
- note = find_reg_equal_equiv_note (insn);
- if (note)
- find_used_regs (&XEXP (note, 0), NULL);
+ /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
+ if it reaches any of the deleted expressions. */
- may_be_loop_header = false;
- for (e = bb->pred; e; e = e->pred_next)
- if (e->flags & EDGE_DFS_BACK)
- {
- may_be_loop_header = true;
- break;
- }
+ set_size = pre_insert_map[0]->size;
+ num_edges = NUM_EDGES (edge_list);
+ inserted = sbitmap_vector_alloc (num_edges, expr_hash_table.n_elems);
+ sbitmap_vector_zero (inserted, num_edges);
- change = 0;
- for (e = bb->pred; e; e = enext)
+ for (e = 0; e < num_edges; e++)
{
- enext = e->pred_next;
- if (e->flags & EDGE_COMPLEX)
- continue;
-
- /* We can't redirect edges from new basic blocks. */
- if (e->src->index >= bypass_last_basic_block)
- continue;
-
- /* The irreducible loops created by redirecting of edges entering the
- loop from outside would decrease effectiveness of some of the following
- optimizations, so prevent this. */
- if (may_be_loop_header
- && !(e->flags & EDGE_DFS_BACK))
- continue;
+ int indx;
+ basic_block bb = INDEX_EDGE_PRED_BB (edge_list, e);
- for (i = 0; i < reg_use_count; i++)
+ for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
{
- struct reg_use *reg_used = ®_use_table[i];
- unsigned int regno = REGNO (reg_used->reg_rtx);
- basic_block dest, old_dest;
- struct expr *set;
- rtx src, new;
+ SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
- if (regno >= max_gcse_regno)
- continue;
-
- set = find_bypass_set (regno, e->src->index);
-
- if (! set)
- continue;
-
- /* Check the data flow is valid after edge insertions. */
- if (e->insns && reg_killed_on_edge (reg_used->reg_rtx, e))
- continue;
-
- src = SET_SRC (pc_set (jump));
-
- if (setcc != NULL)
- src = simplify_replace_rtx (src,
- SET_DEST (PATTERN (setcc)),
- SET_SRC (PATTERN (setcc)));
-
- new = simplify_replace_rtx (src, reg_used->reg_rtx,
- SET_SRC (set->expr));
-
- /* Jump bypassing may have already placed instructions on
- edges of the CFG. We can't bypass an outgoing edge that
- has instructions associated with it, as these insns won't
- get executed if the incoming edge is redirected. */
-
- if (new == pc_rtx)
- {
- edest = FALLTHRU_EDGE (bb);
- dest = edest->insns ? NULL : edest->dest;
- }
- else if (GET_CODE (new) == LABEL_REF)
- {
- dest = BLOCK_FOR_INSN (XEXP (new, 0));
- /* Don't bypass edges containing instructions. */
- for (edest = bb->succ; edest; edest = edest->succ_next)
- if (edest->dest == dest && edest->insns)
- {
- dest = NULL;
- break;
- }
- }
- else
- dest = NULL;
-
- old_dest = e->dest;
- if (dest != NULL
- && dest != old_dest
- && dest != EXIT_BLOCK_PTR)
- {
- redirect_edge_and_branch_force (e, dest);
-
- /* Copy the register setter to the redirected edge.
- Don't copy CC0 setters, as CC0 is dead after jump. */
- if (setcc)
- {
- rtx pat = PATTERN (setcc);
- if (!CC0_P (SET_DEST (pat)))
- insert_insn_on_edge (copy_insn (pat), e);
- }
-
- if (gcse_file != NULL)
- {
- fprintf (gcse_file, "JUMP-BYPASS: Proved reg %d in jump_insn %d equals constant ",
- regno, INSN_UID (jump));
- print_rtl (gcse_file, SET_SRC (set->expr));
- fprintf (gcse_file, "\nBypass edge from %d->%d to %d\n",
- e->src->index, old_dest->index, dest->index);
- }
- change = 1;
- break;
- }
- }
- }
- return change;
-}
-
-/* Find basic blocks with more than one predecessor that only contain a
- single conditional jump. If the result of the comparison is known at
- compile-time from any incoming edge, redirect that edge to the
- appropriate target. Returns nonzero if a change was made.
-
- This function is now mis-named, because we also handle indirect jumps. */
-
-static int
-bypass_conditional_jumps (void)
-{
- basic_block bb;
- int changed;
- rtx setcc;
- rtx insn;
- rtx dest;
-
- /* Note we start at block 1. */
- if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
- return 0;
-
- bypass_last_basic_block = last_basic_block;
- mark_dfs_back_edges ();
-
- changed = 0;
- FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb,
- EXIT_BLOCK_PTR, next_bb)
- {
- /* Check for more than one predecessor. */
- if (bb->pred && bb->pred->pred_next)
- {
- setcc = NULL_RTX;
- for (insn = bb->head;
- insn != NULL && insn != NEXT_INSN (bb->end);
- insn = NEXT_INSN (insn))
- if (GET_CODE (insn) == INSN)
- {
- if (setcc)
- break;
- if (GET_CODE (PATTERN (insn)) != SET)
- break;
-
- dest = SET_DEST (PATTERN (insn));
- if (REG_P (dest) || CC0_P (dest))
- setcc = insn;
- else
- break;
- }
- else if (GET_CODE (insn) == JUMP_INSN)
- {
- if ((any_condjump_p (insn) || computed_jump_p (insn))
- && onlyjump_p (insn))
- changed |= bypass_block (bb, setcc, insn);
- break;
- }
- else if (INSN_P (insn))
- break;
- }
- }
-
- /* If we bypassed any register setting insns, we inserted a
- copy on the redirected edge. These need to be committed. */
- if (changed)
- commit_edge_insertions();
-
- return changed;
-}
-\f
-/* Compute PRE+LCM working variables. */
-
-/* Local properties of expressions. */
-/* Nonzero for expressions that are transparent in the block. */
-static sbitmap *transp;
-
-/* Nonzero for expressions that are transparent at the end of the block.
- This is only zero for expressions killed by abnormal critical edge
- created by a calls. */
-static sbitmap *transpout;
-
-/* Nonzero for expressions that are computed (available) in the block. */
-static sbitmap *comp;
-
-/* Nonzero for expressions that are locally anticipatable in the block. */
-static sbitmap *antloc;
-
-/* Nonzero for expressions where this block is an optimal computation
- point. */
-static sbitmap *pre_optimal;
-
-/* Nonzero for expressions which are redundant in a particular block. */
-static sbitmap *pre_redundant;
-
-/* Nonzero for expressions which should be inserted on a specific edge. */
-static sbitmap *pre_insert_map;
-
-/* Nonzero for expressions which should be deleted in a specific block. */
-static sbitmap *pre_delete_map;
-
-/* 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
-alloc_pre_mem (int n_blocks, int n_exprs)
-{
- transp = sbitmap_vector_alloc (n_blocks, n_exprs);
- comp = sbitmap_vector_alloc (n_blocks, n_exprs);
- antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
-
- pre_optimal = NULL;
- pre_redundant = NULL;
- pre_insert_map = NULL;
- pre_delete_map = NULL;
- ae_in = NULL;
- ae_out = NULL;
- ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
-
- /* pre_insert and pre_delete are allocated later. */
-}
-
-/* Free vars used for PRE analysis. */
-
-static void
-free_pre_mem (void)
-{
- sbitmap_vector_free (transp);
- sbitmap_vector_free (comp);
-
- /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
-
- if (pre_optimal)
- sbitmap_vector_free (pre_optimal);
- if (pre_redundant)
- sbitmap_vector_free (pre_redundant);
- if (pre_insert_map)
- sbitmap_vector_free (pre_insert_map);
- if (pre_delete_map)
- sbitmap_vector_free (pre_delete_map);
- if (ae_in)
- sbitmap_vector_free (ae_in);
- if (ae_out)
- sbitmap_vector_free (ae_out);
-
- transp = comp = NULL;
- pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
- ae_in = ae_out = NULL;
-}
-
-/* Top level routine to do the dataflow analysis needed by PRE. */
-
-static void
-compute_pre_data (void)
-{
- sbitmap trapping_expr;
- basic_block bb;
- unsigned int ui;
-
- compute_local_properties (transp, comp, antloc, &expr_hash_table);
- sbitmap_vector_zero (ae_kill, last_basic_block);
-
- /* Collect expressions which might trap. */
- trapping_expr = sbitmap_alloc (expr_hash_table.n_elems);
- sbitmap_zero (trapping_expr);
- for (ui = 0; ui < expr_hash_table.size; ui++)
- {
- struct expr *e;
- for (e = expr_hash_table.table[ui]; e != NULL; e = e->next_same_hash)
- if (may_trap_p (e->expr))
- SET_BIT (trapping_expr, e->bitmap_index);
- }
-
- /* Compute ae_kill for each basic block using:
-
- ~(TRANSP | COMP)
-
- This is significantly faster than compute_ae_kill. */
-
- FOR_EACH_BB (bb)
- {
- edge e;
-
- /* If the current block is the destination of an abnormal edge, we
- kill all trapping expressions because we won't be able to properly
- place the instruction on the edge. So make them neither
- anticipatable nor transparent. This is fairly conservative. */
- for (e = bb->pred; e ; e = e->pred_next)
- if (e->flags & EDGE_ABNORMAL)
- {
- sbitmap_difference (antloc[bb->index], antloc[bb->index], trapping_expr);
- sbitmap_difference (transp[bb->index], transp[bb->index], trapping_expr);
- break;
- }
-
- sbitmap_a_or_b (ae_kill[bb->index], transp[bb->index], comp[bb->index]);
- sbitmap_not (ae_kill[bb->index], ae_kill[bb->index]);
- }
-
- edge_list = pre_edge_lcm (gcse_file, expr_hash_table.n_elems, transp, comp, antloc,
- ae_kill, &pre_insert_map, &pre_delete_map);
- sbitmap_vector_free (antloc);
- antloc = NULL;
- sbitmap_vector_free (ae_kill);
- ae_kill = NULL;
- sbitmap_free (trapping_expr);
-}
-\f
-/* PRE utilities */
-
-/* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
- block BB.
-
- VISITED is a pointer to a working buffer for tracking which BB's have
- been visited. It is NULL for the top-level call.
-
- We treat reaching expressions that go through blocks containing the same
- reaching expression as "not reaching". E.g. if EXPR is generated in blocks
- 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
- 2 as not reaching. The intent is to improve the probability of finding
- only one reaching expression and to reduce register lifetimes by picking
- the closest such expression. */
-
-static int
-pre_expr_reaches_here_p_work (basic_block occr_bb, struct expr *expr, basic_block bb, char *visited)
-{
- edge pred;
-
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
- {
- basic_block pred_bb = pred->src;
-
- if (pred->src == ENTRY_BLOCK_PTR
- /* Has predecessor has already been visited? */
- || visited[pred_bb->index])
- ;/* Nothing to do. */
-
- /* Does this predecessor generate this expression? */
- else if (TEST_BIT (comp[pred_bb->index], expr->bitmap_index))
- {
- /* Is this the occurrence we're looking for?
- Note that there's only one generating occurrence per block
- so we just need to check the block number. */
- if (occr_bb == pred_bb)
- return 1;
-
- visited[pred_bb->index] = 1;
- }
- /* Ignore this predecessor if it kills the expression. */
- else if (! TEST_BIT (transp[pred_bb->index], expr->bitmap_index))
- visited[pred_bb->index] = 1;
-
- /* Neither gen nor kill. */
- else
- {
- visited[pred_bb->index] = 1;
- if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
- return 1;
- }
- }
-
- /* All paths have been checked. */
- return 0;
-}
-
-/* The wrapper for pre_expr_reaches_here_work that ensures that any
- memory allocated for that function is returned. */
-
-static int
-pre_expr_reaches_here_p (basic_block occr_bb, struct expr *expr, basic_block bb)
-{
- int rval;
- char *visited = xcalloc (last_basic_block, 1);
-
- rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
-
- free (visited);
- return rval;
-}
-\f
-
-/* Given an expr, generate RTL which we can insert at the end of a BB,
- or on an edge. Set the block number of any insns generated to
- the value of BB. */
-
-static rtx
-process_insert_insn (struct expr *expr)
-{
- rtx reg = expr->reaching_reg;
- rtx exp = copy_rtx (expr->expr);
- rtx pat;
-
- start_sequence ();
-
- /* If the expression is something that's an operand, like a constant,
- just copy it to a register. */
- if (general_operand (exp, GET_MODE (reg)))
- emit_move_insn (reg, exp);
-
- /* Otherwise, make a new insn to compute this expression and make sure the
- insn will be recognized (this also adds any needed CLOBBERs). Copy the
- expression to make sure we don't have any sharing issues. */
- else if (insn_invalid_p (emit_insn (gen_rtx_SET (VOIDmode, reg, exp))))
- abort ();
-
- pat = get_insns ();
- end_sequence ();
-
- return pat;
-}
-
-/* Add EXPR to the end of basic block BB.
-
- This is used by both the PRE and code hoisting.
-
- For PRE, we want to verify that the expr is either transparent
- or locally anticipatable in the target block. This check makes
- no sense for code hoisting. */
-
-static void
-insert_insn_end_bb (struct expr *expr, basic_block bb, int pre)
-{
- rtx insn = bb->end;
- rtx new_insn;
- rtx reg = expr->reaching_reg;
- int regno = REGNO (reg);
- rtx pat, pat_end;
-
- pat = process_insert_insn (expr);
- if (pat == NULL_RTX || ! INSN_P (pat))
- abort ();
-
- pat_end = pat;
- while (NEXT_INSN (pat_end) != NULL_RTX)
- pat_end = NEXT_INSN (pat_end);
-
- /* If the last insn is a jump, insert EXPR in front [taking care to
- handle cc0, etc. properly]. Similarly we need to care trapping
- instructions in presence of non-call exceptions. */
-
- if (GET_CODE (insn) == JUMP_INSN
- || (GET_CODE (insn) == INSN
- && (bb->succ->succ_next || (bb->succ->flags & EDGE_ABNORMAL))))
- {
-#ifdef HAVE_cc0
- rtx note;
-#endif
- /* It should always be the case that we can put these instructions
- anywhere in the basic block with performing PRE optimizations.
- Check this. */
- if (GET_CODE (insn) == INSN && pre
- && !TEST_BIT (antloc[bb->index], expr->bitmap_index)
- && !TEST_BIT (transp[bb->index], expr->bitmap_index))
- abort ();
-
- /* If this is a jump table, then we can't insert stuff here. Since
- we know the previous real insn must be the tablejump, we insert
- the new instruction just before the tablejump. */
- if (GET_CODE (PATTERN (insn)) == ADDR_VEC
- || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
- insn = prev_real_insn (insn);
-
-#ifdef HAVE_cc0
- /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
- if cc0 isn't set. */
- note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
- if (note)
- insn = XEXP (note, 0);
- else
- {
- rtx maybe_cc0_setter = prev_nonnote_insn (insn);
- if (maybe_cc0_setter
- && INSN_P (maybe_cc0_setter)
- && sets_cc0_p (PATTERN (maybe_cc0_setter)))
- insn = maybe_cc0_setter;
- }
-#endif
- /* FIXME: What if something in cc0/jump uses value set in new insn? */
- new_insn = emit_insn_before (pat, insn);
- }
-
- /* Likewise if the last insn is a call, as will happen in the presence
- of exception handling. */
- else if (GET_CODE (insn) == CALL_INSN
- && (bb->succ->succ_next || (bb->succ->flags & EDGE_ABNORMAL)))
- {
- /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
- we search backward and place the instructions before the first
- parameter is loaded. Do this for everyone for consistency and a
- presumption that we'll get better code elsewhere as well.
-
- It should always be the case that we can put these instructions
- anywhere in the basic block with performing PRE optimizations.
- Check this. */
-
- if (pre
- && !TEST_BIT (antloc[bb->index], expr->bitmap_index)
- && !TEST_BIT (transp[bb->index], expr->bitmap_index))
- abort ();
-
- /* Since different machines initialize their parameter registers
- in different orders, assume nothing. Collect the set of all
- parameter registers. */
- insn = find_first_parameter_load (insn, bb->head);
-
- /* If we found all the parameter loads, then we want to insert
- before the first parameter load.
-
- If we did not find all the parameter loads, then we might have
- stopped on the head of the block, which could be a CODE_LABEL.
- If we inserted before the CODE_LABEL, then we would be putting
- the insn in the wrong basic block. In that case, put the insn
- after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
- while (GET_CODE (insn) == CODE_LABEL
- || NOTE_INSN_BASIC_BLOCK_P (insn))
- insn = NEXT_INSN (insn);
-
- new_insn = emit_insn_before (pat, insn);
- }
- else
- new_insn = emit_insn_after (pat, insn);
-
- while (1)
- {
- if (INSN_P (pat))
- {
- add_label_notes (PATTERN (pat), new_insn);
- note_stores (PATTERN (pat), record_set_info, pat);
- }
- if (pat == pat_end)
- break;
- pat = NEXT_INSN (pat);
- }
-
- gcse_create_count++;
-
- if (gcse_file)
- {
- fprintf (gcse_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",
- expr->bitmap_index, regno);
- }
-}
-
-/* Insert partially redundant expressions on edges in the CFG to make
- the expressions fully redundant. */
-
-static int
-pre_edge_insert (struct edge_list *edge_list, struct expr **index_map)
-{
- int e, i, j, num_edges, set_size, did_insert = 0;
- sbitmap *inserted;
-
- /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
- if it reaches any of the deleted expressions. */
-
- set_size = pre_insert_map[0]->size;
- num_edges = NUM_EDGES (edge_list);
- inserted = sbitmap_vector_alloc (num_edges, expr_hash_table.n_elems);
- sbitmap_vector_zero (inserted, num_edges);
-
- for (e = 0; e < num_edges; e++)
- {
- int indx;
- basic_block bb = INDEX_EDGE_PRED_BB (edge_list, e);
-
- for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
- {
- SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
-
- for (j = indx; insert && j < (int) expr_hash_table.n_elems; j++, insert >>= 1)
- if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
+ for (j = indx; insert && j < (int) expr_hash_table.n_elems; j++, insert >>= 1)
+ if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
{
struct expr *expr = index_map[j];
struct occr *occr;
- /* Now look at each deleted occurrence of this expression. */
- for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
- {
- if (! occr->deleted_p)
- continue;
-
- /* Insert this expression on this edge if if it would
- reach the deleted occurrence in BB. */
- if (!TEST_BIT (inserted[e], j))
- {
- rtx insn;
- edge eg = INDEX_EDGE (edge_list, e);
-
- /* We can't insert anything on an abnormal and
- critical edge, so we insert the insn at the end of
- the previous block. There are several alternatives
- detailed in Morgans book P277 (sec 10.5) for
- handling this situation. This one is easiest for
- now. */
-
- if ((eg->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
- insert_insn_end_bb (index_map[j], bb, 0);
- else
- {
- insn = process_insert_insn (index_map[j]);
- insert_insn_on_edge (insn, eg);
- }
-
- if (gcse_file)
- {
- fprintf (gcse_file, "PRE/HOIST: edge (%d,%d), ",
- bb->index,
- INDEX_EDGE_SUCC_BB (edge_list, e)->index);
- fprintf (gcse_file, "copy expression %d\n",
- expr->bitmap_index);
- }
-
- update_ld_motion_stores (expr);
- SET_BIT (inserted[e], j);
- did_insert = 1;
- gcse_create_count++;
- }
- }
- }
- }
- }
-
- sbitmap_vector_free (inserted);
- return did_insert;
-}
-
-/* Copy the result of INSN to REG. INDX is the expression number. */
-
-static void
-pre_insert_copy_insn (struct expr *expr, rtx insn)
-{
- rtx reg = expr->reaching_reg;
- int regno = REGNO (reg);
- int indx = expr->bitmap_index;
- rtx set = single_set (insn);
- rtx new_insn;
-
- if (!set)
- abort ();
-
- new_insn = emit_insn_after (gen_move_insn (reg, copy_rtx (SET_DEST (set))), insn);
-
- /* Keep register set table up to date. */
- record_one_set (regno, new_insn);
-
- gcse_create_count++;
-
- if (gcse_file)
- fprintf (gcse_file,
- "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
- BLOCK_NUM (insn), INSN_UID (new_insn), indx,
- INSN_UID (insn), regno);
- update_ld_motion_stores (expr);
-}
-
-/* Copy available expressions that reach the redundant expression
- to `reaching_reg'. */
-
-static void
-pre_insert_copies (void)
-{
- unsigned int i;
- struct expr *expr;
- struct occr *occr;
- struct occr *avail;
-
- /* For each available expression in the table, copy the result to
- `reaching_reg' if the expression reaches a deleted one.
-
- ??? The current algorithm is rather brute force.
- Need to do some profiling. */
-
- for (i = 0; i < expr_hash_table.size; i++)
- for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
- {
- /* If the basic block isn't reachable, PPOUT will be TRUE. However,
- we don't want to insert a copy here because the expression may not
- really be redundant. So only insert an insn if the expression was
- deleted. This test also avoids further processing if the
- expression wasn't deleted anywhere. */
- if (expr->reaching_reg == NULL)
- continue;
-
- for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
- {
- if (! occr->deleted_p)
- continue;
-
- for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
- {
- rtx insn = avail->insn;
-
- /* No need to handle this one if handled already. */
- if (avail->copied_p)
- continue;
-
- /* Don't handle this one if it's a redundant one. */
- if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
- continue;
-
- /* Or if the expression doesn't reach the deleted one. */
- if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail->insn),
- expr,
- BLOCK_FOR_INSN (occr->insn)))
- continue;
-
- /* Copy the result of avail to reaching_reg. */
- pre_insert_copy_insn (expr, insn);
- avail->copied_p = 1;
- }
- }
- }
-}
-
-/* Emit move from SRC to DEST noting the equivalence with expression computed
- in INSN. */
-static rtx
-gcse_emit_move_after (rtx src, rtx dest, rtx insn)
-{
- rtx new;
- rtx 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);
-
- /* Note the equivalence for local CSE pass. */
- set2 = single_set (new);
- if (!set2 || !rtx_equal_p (SET_DEST (set2), dest))
- return new;
- 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));
-
- return new;
-}
-
-/* Delete redundant computations.
- Deletion is done by changing the insn to copy the `reaching_reg' of
- the expression into the result of the SET. It is left to later passes
- (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
-
- Returns nonzero if a change is made. */
-
-static int
-pre_delete (void)
-{
- unsigned int i;
- int changed;
- struct expr *expr;
- struct occr *occr;
-
- changed = 0;
- for (i = 0; i < expr_hash_table.size; i++)
- for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
- {
- int indx = expr->bitmap_index;
-
- /* We only need to search antic_occr since we require
- ANTLOC != 0. */
-
- for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
- {
- rtx insn = occr->insn;
- rtx set;
- basic_block bb = BLOCK_FOR_INSN (insn);
-
- if (TEST_BIT (pre_delete_map[bb->index], indx))
- {
- set = single_set (insn);
- if (! set)
- abort ();
-
- /* Create a pseudo-reg to store the result of reaching
- expressions into. Get the mode for the new pseudo from
- the mode of the original destination pseudo. */
- if (expr->reaching_reg == NULL)
- expr->reaching_reg
- = gen_reg_rtx (GET_MODE (SET_DEST (set)));
-
- 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)
- {
- fprintf (gcse_file,
- "PRE: redundant insn %d (expression %d) in ",
- INSN_UID (insn), indx);
- fprintf (gcse_file, "bb %d, reaching reg is %d\n",
- bb->index, REGNO (expr->reaching_reg));
- }
- }
- }
- }
-
- return changed;
-}
-
-/* Perform GCSE optimizations using PRE.
- This is called by one_pre_gcse_pass after all the dataflow analysis
- has been done.
-
- This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
- lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
- Compiler Design and Implementation.
-
- ??? A new pseudo reg is created to hold the reaching expression. The nice
- thing about the classical approach is that it would try to use an existing
- reg. If the register can't be adequately optimized [i.e. we introduce
- reload problems], one could add a pass here to propagate the new register
- through the block.
-
- ??? We don't handle single sets in PARALLELs because we're [currently] not
- able to copy the rest of the parallel when we insert copies to create full
- redundancies from partial redundancies. However, there's no reason why we
- can't handle PARALLELs in the cases where there are no partial
- redundancies. */
-
-static int
-pre_gcse (void)
-{
- unsigned int i;
- int did_insert, changed;
- struct expr **index_map;
- struct expr *expr;
-
- /* Compute a mapping from expression number (`bitmap_index') to
- hash table entry. */
-
- index_map = xcalloc (expr_hash_table.n_elems, sizeof (struct expr *));
- for (i = 0; i < expr_hash_table.size; i++)
- for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
- index_map[expr->bitmap_index] = expr;
-
- /* 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
- specially allocated pseudo-reg that reaches the redundant expr. */
- pre_insert_copies ();
- if (did_insert)
- {
- commit_edge_insertions ();
- changed = 1;
- }
-
- free (index_map);
- sbitmap_free (pre_redundant_insns);
- return changed;
-}
-
-/* Top level routine to perform one PRE GCSE pass.
-
- Return nonzero if a change was made. */
-
-static int
-one_pre_gcse_pass (int pass)
-{
- int changed = 0;
-
- gcse_subst_count = 0;
- gcse_create_count = 0;
-
- alloc_hash_table (max_cuid, &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 (expr_hash_table.n_elems > 0)
- {
- alloc_pre_mem (last_basic_block, expr_hash_table.n_elems);
- compute_pre_data ();
- changed |= pre_gcse ();
- free_edge_list (edge_list);
- free_pre_mem ();
- }
-
- free_ldst_mems ();
- remove_fake_edges ();
- free_hash_table (&expr_hash_table);
-
- if (gcse_file)
- {
- fprintf (gcse_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
- current_function_name, pass, bytes_used);
- fprintf (gcse_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 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. */
-
-static void
-add_label_notes (rtx x, rtx insn)
-{
- enum rtx_code code = GET_CODE (x);
- int i, j;
- const char *fmt;
-
- if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
- {
- /* This code used to ignore labels that referred to dispatch tables to
- avoid flow generating (slightly) worse code.
-
- We no longer ignore such label references (see LABEL_REF handling in
- mark_jump_label for additional information). */
-
- REG_NOTES (insn) = gen_rtx_INSN_LIST (REG_LABEL, XEXP (x, 0),
- REG_NOTES (insn));
- if (LABEL_P (XEXP (x, 0)))
- LABEL_NUSES (XEXP (x, 0))++;
- return;
- }
-
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
- add_label_notes (XEXP (x, i), insn);
- else if (fmt[i] == 'E')
- for (j = XVECLEN (x, i) - 1; j >= 0; j--)
- add_label_notes (XVECEXP (x, i, j), insn);
- }
-}
-
-/* Compute transparent outgoing information for each block.
-
- An expression is transparent to an edge unless it is killed by
- the edge itself. This can only happen with abnormal control flow,
- when the edge is traversed through a call. This happens with
- non-local labels and exceptions.
-
- This would not be necessary if we split the edge. While this is
- normally impossible for abnormal critical edges, with some effort
- it should be possible with exception handling, since we still have
- control over which handler should be invoked. But due to increased
- EH table sizes, this may not be worthwhile. */
-
-static void
-compute_transpout (void)
-{
- basic_block bb;
- unsigned int i;
- struct expr *expr;
-
- sbitmap_vector_ones (transpout, last_basic_block);
-
- FOR_EACH_BB (bb)
- {
- /* Note that flow inserted a nop a the end of basic blocks that
- end in call instructions for reasons other than abnormal
- control flow. */
- if (GET_CODE (bb->end) != CALL_INSN)
- continue;
-
- for (i = 0; i < expr_hash_table.size; i++)
- for (expr = expr_hash_table.table[i]; expr ; expr = expr->next_same_hash)
- if (GET_CODE (expr->expr) == MEM)
- {
- if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
- && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
- continue;
-
- /* ??? Optimally, we would use interprocedural alias
- analysis to determine if this mem is actually killed
- by this call. */
- RESET_BIT (transpout[bb->index], expr->bitmap_index);
- }
- }
-}
-
-/* Removal of useless null pointer checks */
-
-/* Called via note_stores. X is set by SETTER. If X is a register we must
- invalidate nonnull_local and set nonnull_killed. DATA is really a
- `null_pointer_info *'.
-
- We ignore hard registers. */
-
-static void
-invalidate_nonnull_info (rtx x, rtx setter ATTRIBUTE_UNUSED, void *data)
-{
- unsigned int regno;
- struct null_pointer_info *npi = (struct null_pointer_info *) data;
-
- while (GET_CODE (x) == SUBREG)
- x = SUBREG_REG (x);
-
- /* Ignore anything that is not a register or is a hard register. */
- if (GET_CODE (x) != REG
- || REGNO (x) < npi->min_reg
- || REGNO (x) >= npi->max_reg)
- return;
-
- regno = REGNO (x) - npi->min_reg;
-
- RESET_BIT (npi->nonnull_local[npi->current_block->index], regno);
- SET_BIT (npi->nonnull_killed[npi->current_block->index], regno);
-}
-
-/* Do null-pointer check elimination for the registers indicated in
- NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
- they are not our responsibility to free. */
-
-static int
-delete_null_pointer_checks_1 (unsigned int *block_reg, sbitmap *nonnull_avin,
- sbitmap *nonnull_avout,
- struct null_pointer_info *npi)
-{
- basic_block bb, current_block;
- sbitmap *nonnull_local = npi->nonnull_local;
- sbitmap *nonnull_killed = npi->nonnull_killed;
- int something_changed = 0;
-
- /* Compute local properties, nonnull and killed. A register will have
- the nonnull property if at the end of the current block its value is
- known to be nonnull. The killed property indicates that somewhere in
- the block any information we had about the register is killed.
-
- Note that a register can have both properties in a single block. That
- indicates that it's killed, then later in the block a new value is
- computed. */
- sbitmap_vector_zero (nonnull_local, last_basic_block);
- sbitmap_vector_zero (nonnull_killed, last_basic_block);
-
- FOR_EACH_BB (current_block)
- {
- rtx insn, stop_insn;
-
- /* Set the current block for invalidate_nonnull_info. */
- npi->current_block = current_block;
-
- /* Scan each insn in the basic block looking for memory references and
- register sets. */
- stop_insn = NEXT_INSN (current_block->end);
- for (insn = current_block->head;
- insn != stop_insn;
- insn = NEXT_INSN (insn))
- {
- rtx set;
- rtx reg;
-
- /* Ignore anything that is not a normal insn. */
- if (! INSN_P (insn))
- continue;
-
- /* Basically ignore anything that is not a simple SET. We do have
- to make sure to invalidate nonnull_local and set nonnull_killed
- for such insns though. */
- set = single_set (insn);
- if (!set)
- {
- note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
- continue;
- }
-
- /* See if we've got a usable memory load. We handle it first
- in case it uses its address register as a dest (which kills
- the nonnull property). */
- if (GET_CODE (SET_SRC (set)) == MEM
- && GET_CODE ((reg = XEXP (SET_SRC (set), 0))) == REG
- && REGNO (reg) >= npi->min_reg
- && REGNO (reg) < npi->max_reg)
- SET_BIT (nonnull_local[current_block->index],
- REGNO (reg) - npi->min_reg);
-
- /* Now invalidate stuff clobbered by this insn. */
- note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
-
- /* And handle stores, we do these last since any sets in INSN can
- not kill the nonnull property if it is derived from a MEM
- appearing in a SET_DEST. */
- if (GET_CODE (SET_DEST (set)) == MEM
- && GET_CODE ((reg = XEXP (SET_DEST (set), 0))) == REG
- && REGNO (reg) >= npi->min_reg
- && REGNO (reg) < npi->max_reg)
- SET_BIT (nonnull_local[current_block->index],
- REGNO (reg) - npi->min_reg);
- }
- }
-
- /* Now compute global properties based on the local properties. This
- is a classic global availability algorithm. */
- compute_available (nonnull_local, nonnull_killed,
- nonnull_avout, nonnull_avin);
-
- /* Now look at each bb and see if it ends with a compare of a value
- against zero. */
- FOR_EACH_BB (bb)
- {
- rtx last_insn = bb->end;
- rtx condition, earliest;
- int compare_and_branch;
-
- /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
- since BLOCK_REG[BB] is zero if this block did not end with a
- comparison against zero, this condition works. */
- if (block_reg[bb->index] < npi->min_reg
- || block_reg[bb->index] >= npi->max_reg)
- continue;
-
- /* LAST_INSN is a conditional jump. Get its condition. */
- condition = get_condition (last_insn, &earliest);
-
- /* If we can't determine the condition then skip. */
- if (! condition)
- continue;
-
- /* Is the register known to have a nonzero value? */
- if (!TEST_BIT (nonnull_avout[bb->index], block_reg[bb->index] - npi->min_reg))
- continue;
-
- /* Try to compute whether the compare/branch at the loop end is one or
- two instructions. */
- if (earliest == last_insn)
- compare_and_branch = 1;
- else if (earliest == prev_nonnote_insn (last_insn))
- compare_and_branch = 2;
- else
- continue;
-
- /* We know the register in this comparison is nonnull at exit from
- this block. We can optimize this comparison. */
- if (GET_CODE (condition) == NE)
- {
- rtx new_jump;
-
- new_jump = emit_jump_insn_after (gen_jump (JUMP_LABEL (last_insn)),
- last_insn);
- JUMP_LABEL (new_jump) = JUMP_LABEL (last_insn);
- LABEL_NUSES (JUMP_LABEL (new_jump))++;
- emit_barrier_after (new_jump);
- }
-
- something_changed = 1;
- delete_insn (last_insn);
- if (compare_and_branch == 2)
- delete_insn (earliest);
- purge_dead_edges (bb);
-
- /* Don't check this block again. (Note that BLOCK_END is
- invalid here; we deleted the last instruction in the
- block.) */
- block_reg[bb->index] = 0;
- }
-
- return something_changed;
-}
-
-/* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
- at compile time.
-
- This is conceptually similar to global constant/copy propagation and
- classic global CSE (it even uses the same dataflow equations as cprop).
-
- If a register is used as memory address with the form (mem (reg)), then we
- know that REG can not be zero at that point in the program. Any instruction
- which sets REG "kills" this property.
-
- So, if every path leading to a conditional branch has an available memory
- reference of that form, then we know the register can not have the value
- zero at the conditional branch.
-
- So we merely need to compute the local properties and propagate that data
- around the cfg, then optimize where possible.
-
- We run this pass two times. Once before CSE, then again after CSE. This
- has proven to be the most profitable approach. It is rare for new
- optimization opportunities of this nature to appear after the first CSE
- pass.
-
- This could probably be integrated with global cprop with a little work. */
-
-int
-delete_null_pointer_checks (rtx f ATTRIBUTE_UNUSED)
-{
- sbitmap *nonnull_avin, *nonnull_avout;
- unsigned int *block_reg;
- basic_block bb;
- int reg;
- int regs_per_pass;
- int max_reg;
- struct null_pointer_info npi;
- int something_changed = 0;
-
- /* If we have only a single block, then there's nothing to do. */
- if (n_basic_blocks <= 1)
- return 0;
-
- /* Trying to perform global optimizations on flow graphs which have
- a high connectivity will take a long time and is unlikely to be
- particularly useful.
-
- In normal circumstances a cfg should have about twice as many edges
- as blocks. But we do not want to punish small functions which have
- a couple switch statements. So we require a relatively large number
- of basic blocks and the ratio of edges to blocks to be high. */
- if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
- return 0;
-
- /* We need four bitmaps, each with a bit for each register in each
- basic block. */
- max_reg = max_reg_num ();
- regs_per_pass = get_bitmap_width (4, last_basic_block, max_reg);
-
- /* Allocate bitmaps to hold local and global properties. */
- npi.nonnull_local = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
- npi.nonnull_killed = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
- nonnull_avin = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
- nonnull_avout = sbitmap_vector_alloc (last_basic_block, regs_per_pass);
-
- /* Go through the basic blocks, seeing whether or not each block
- ends with a conditional branch whose condition is a comparison
- against zero. Record the register compared in BLOCK_REG. */
- block_reg = xcalloc (last_basic_block, sizeof (int));
- FOR_EACH_BB (bb)
- {
- rtx last_insn = bb->end;
- rtx condition, earliest, reg;
-
- /* We only want conditional branches. */
- if (GET_CODE (last_insn) != JUMP_INSN
- || !any_condjump_p (last_insn)
- || !onlyjump_p (last_insn))
- continue;
-
- /* LAST_INSN is a conditional jump. Get its condition. */
- condition = get_condition (last_insn, &earliest);
-
- /* If we were unable to get the condition, or it is not an equality
- comparison against zero then there's nothing we can do. */
- if (!condition
- || (GET_CODE (condition) != NE && GET_CODE (condition) != EQ)
- || GET_CODE (XEXP (condition, 1)) != CONST_INT
- || (XEXP (condition, 1)
- != CONST0_RTX (GET_MODE (XEXP (condition, 0)))))
- continue;
-
- /* We must be checking a register against zero. */
- reg = XEXP (condition, 0);
- if (GET_CODE (reg) != REG)
- continue;
-
- block_reg[bb->index] = REGNO (reg);
- }
-
- /* Go through the algorithm for each block of registers. */
- for (reg = FIRST_PSEUDO_REGISTER; reg < max_reg; reg += regs_per_pass)
- {
- npi.min_reg = reg;
- npi.max_reg = MIN (reg + regs_per_pass, max_reg);
- something_changed |= delete_null_pointer_checks_1 (block_reg,
- nonnull_avin,
- nonnull_avout,
- &npi);
- }
-
- /* Free the table of registers compared at the end of every block. */
- free (block_reg);
-
- /* Free bitmaps. */
- sbitmap_vector_free (npi.nonnull_local);
- sbitmap_vector_free (npi.nonnull_killed);
- sbitmap_vector_free (nonnull_avin);
- sbitmap_vector_free (nonnull_avout);
-
- return something_changed;
-}
-
-/* Code Hoisting variables and subroutines. */
-
-/* Very busy expressions. */
-static sbitmap *hoist_vbein;
-static sbitmap *hoist_vbeout;
-
-/* Hoistable expressions. */
-static sbitmap *hoist_exprs;
-
-/* Dominator bitmaps. */
-dominance_info dominators;
-
-/* ??? We could compute post dominators and run this algorithm in
- reverse to perform tail merging, doing so would probably be
- more effective than the tail merging code in jump.c.
-
- It's unclear if tail merging could be run in parallel with
- code hoisting. It would be nice. */
-
-/* Allocate vars used for code hoisting analysis. */
-
-static void
-alloc_code_hoist_mem (int n_blocks, int n_exprs)
-{
- antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
- transp = sbitmap_vector_alloc (n_blocks, n_exprs);
- comp = sbitmap_vector_alloc (n_blocks, n_exprs);
-
- hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
- hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
- hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
- transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
-}
-
-/* Free vars used for code hoisting analysis. */
-
-static void
-free_code_hoist_mem (void)
-{
- sbitmap_vector_free (antloc);
- sbitmap_vector_free (transp);
- sbitmap_vector_free (comp);
-
- sbitmap_vector_free (hoist_vbein);
- sbitmap_vector_free (hoist_vbeout);
- sbitmap_vector_free (hoist_exprs);
- sbitmap_vector_free (transpout);
-
- free_dominance_info (dominators);
-}
-
-/* Compute the very busy expressions at entry/exit from each block.
-
- An expression is very busy if all paths from a given point
- compute the expression. */
-
-static void
-compute_code_hoist_vbeinout (void)
-{
- int changed, passes;
- basic_block bb;
-
- sbitmap_vector_zero (hoist_vbeout, last_basic_block);
- sbitmap_vector_zero (hoist_vbein, last_basic_block);
-
- passes = 0;
- changed = 1;
-
- while (changed)
- {
- changed = 0;
-
- /* We scan the blocks in the reverse order to speed up
- 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);
- }
-
- passes++;
- }
-
- if (gcse_file)
- fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
-}
-
-/* Top level routine to do the dataflow analysis needed by code hoisting. */
-
-static void
-compute_code_hoist_data (void)
-{
- compute_local_properties (transp, comp, antloc, &expr_hash_table);
- compute_transpout ();
- compute_code_hoist_vbeinout ();
- dominators = calculate_dominance_info (CDI_DOMINATORS);
- if (gcse_file)
- fprintf (gcse_file, "\n");
-}
-
-/* Determine if the expression identified by EXPR_INDEX would
- reach BB unimpared if it was placed at the end of EXPR_BB.
-
- It's unclear exactly what Muchnick meant by "unimpared". It seems
- to me that the expression must either be computed or transparent in
- *every* block in the path(s) from EXPR_BB to BB. Any other definition
- would allow the expression to be hoisted out of loops, even if
- the expression wasn't a loop invariant.
-
- Contrast this to reachability for PRE where an expression is
- considered reachable if *any* path reaches instead of *all*
- paths. */
-
-static int
-hoist_expr_reaches_here_p (basic_block expr_bb, int expr_index, basic_block bb, char *visited)
-{
- edge pred;
- int visited_allocated_locally = 0;
-
+ /* Now look at each deleted occurrence of this expression. */
+ for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
+ {
+ if (! occr->deleted_p)
+ continue;
- if (visited == NULL)
- {
- visited_allocated_locally = 1;
- visited = xcalloc (last_basic_block, 1);
- }
+ /* Insert this expression on this edge if it would
+ reach the deleted occurrence in BB. */
+ if (!TEST_BIT (inserted[e], j))
+ {
+ rtx insn;
+ edge eg = INDEX_EDGE (edge_list, e);
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
- {
- basic_block pred_bb = pred->src;
+ /* We can't insert anything on an abnormal and
+ critical edge, so we insert the insn at the end of
+ the previous block. There are several alternatives
+ detailed in Morgans book P277 (sec 10.5) for
+ handling this situation. This one is easiest for
+ now. */
- if (pred->src == ENTRY_BLOCK_PTR)
- break;
- else if (pred_bb == expr_bb)
- continue;
- else if (visited[pred_bb->index])
- continue;
+ if (eg->flags & EDGE_ABNORMAL)
+ insert_insn_end_basic_block (index_map[j], bb, 0);
+ else
+ {
+ insn = process_insert_insn (index_map[j]);
+ insert_insn_on_edge (insn, eg);
+ }
- /* Does this predecessor generate this expression? */
- else if (TEST_BIT (comp[pred_bb->index], expr_index))
- break;
- else if (! TEST_BIT (transp[pred_bb->index], expr_index))
- break;
+ if (dump_file)
+ {
+ fprintf (dump_file, "PRE: edge (%d,%d), ",
+ bb->index,
+ INDEX_EDGE_SUCC_BB (edge_list, e)->index);
+ fprintf (dump_file, "copy expression %d\n",
+ expr->bitmap_index);
+ }
- /* Not killed. */
- else
- {
- visited[pred_bb->index] = 1;
- if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
- pred_bb, visited))
- break;
+ update_ld_motion_stores (expr);
+ SET_BIT (inserted[e], j);
+ did_insert = 1;
+ gcse_create_count++;
+ }
+ }
+ }
}
}
- if (visited_allocated_locally)
- free (visited);
- return (pred == NULL);
+ sbitmap_vector_free (inserted);
+ return did_insert;
}
-\f
-/* Actually perform code hoisting. */
+
+/* Copy the result of EXPR->EXPR generated by INSN to EXPR->REACHING_REG.
+ Given "old_reg <- expr" (INSN), instead of adding after it
+ reaching_reg <- old_reg
+ it's better to do the following:
+ reaching_reg <- expr
+ old_reg <- reaching_reg
+ because this way copy propagation can discover additional PRE
+ opportunities. But if this fails, we try the old way.
+ When "expr" is a store, i.e.
+ given "MEM <- old_reg", instead of adding after it
+ reaching_reg <- old_reg
+ it's better to add it before as follows:
+ reaching_reg <- old_reg
+ MEM <- reaching_reg. */
static void
-hoist_code (void)
+pre_insert_copy_insn (struct expr *expr, rtx insn)
{
- basic_block bb, dominated;
- basic_block *domby;
- unsigned int domby_len;
- unsigned int i,j;
- struct expr **index_map;
- struct expr *expr;
-
- 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 *));
- 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;
+ rtx reg = expr->reaching_reg;
+ int regno = REGNO (reg);
+ int indx = expr->bitmap_index;
+ rtx pat = PATTERN (insn);
+ rtx set, first_set, new_insn;
+ rtx old_reg;
+ int i;
- /* Walk over each basic block looking for potentially hoistable
- expressions, nothing gets hoisted from the entry block. */
- FOR_EACH_BB (bb)
+ /* This block matches the logic in hash_scan_insn. */
+ switch (GET_CODE (pat))
{
- int found = 0;
- int insn_inserted_p;
+ case SET:
+ set = pat;
+ break;
- domby_len = get_dominated_by (dominators, bb, &domby);
- /* Examine each expression that is very busy at the exit of this
- block. These are the potentially hoistable expressions. */
- for (i = 0; i < hoist_vbeout[bb->index]->n_bits; i++)
+ 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++)
{
- int hoistable = 0;
-
- if (TEST_BIT (hoist_vbeout[bb->index], i)
- && TEST_BIT (transpout[bb->index], i))
+ rtx x = XVECEXP (pat, 0, i);
+ if (GET_CODE (x) == SET)
{
- /* 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++)
- {
- dominated = domby[j];
- /* Ignore self dominance. */
- if (bb == dominated)
- continue;
- /* We've found a dominated block, now see if it computes
- the busy expression and whether or not moving that
- expression to the "beginning" of that block is safe. */
- if (!TEST_BIT (antloc[dominated->index], i))
- continue;
-
- /* Note if the expression would reach the dominated block
- unimpared if it was placed at the end of BB.
-
- Keep track of how many times this expression is hoistable
- from a dominated block into BB. */
- if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
- hoistable++;
- }
-
- /* If we found more than one hoistable occurrence of this
- expression, then note it in the bitmap of expressions to
- hoist. It makes no sense to hoist things which are computed
- in only one BB, and doing so tends to pessimize register
- allocation. One could increase this value to try harder
- to avoid any possible code expansion due to register
- allocation issues; however experiments have shown that
- the vast majority of hoistable expressions are only movable
- from two successors, so raising this threshold is likely
- to nullify any benefit we get from code hoisting. */
- if (hoistable > 1)
- {
- SET_BIT (hoist_exprs[bb->index], i);
- found = 1;
- }
+ /* 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;
+ }
}
}
- /* If we found nothing to hoist, then quit now. */
- if (! found)
- {
- free (domby);
- continue;
- }
-
- /* Loop over all the hoistable expressions. */
- for (i = 0; i < hoist_exprs[bb->index]->n_bits; i++)
- {
- /* We want to insert the expression into BB only once, so
- note when we've inserted it. */
- insn_inserted_p = 0;
-
- /* These tests should be the same as the tests above. */
- if (TEST_BIT (hoist_vbeout[bb->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++)
- {
- dominated = domby[j];
- /* Ignore self dominance. */
- if (bb == dominated)
- continue;
-
- /* We've found a dominated block, now see if it computes
- the busy expression and whether or not moving that
- expression to the "beginning" of that block is safe. */
- if (!TEST_BIT (antloc[dominated->index], i))
- continue;
-
- /* The expression is computed in the dominated block and
- it would be safe to compute it at the start of the
- dominated block. Now we have to determine if the
- expression would reach the dominated block if it was
- placed at the end of BB. */
- if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
- {
- struct expr *expr = index_map[i];
- struct occr *occr = expr->antic_occr;
- rtx insn;
- rtx set;
-
- /* Find the right occurrence of this expression. */
- while (BLOCK_FOR_INSN (occr->insn) != dominated && occr)
- occr = occr->next;
-
- /* Should never happen. */
- if (!occr)
- abort ();
- insn = occr->insn;
+ gcc_assert (first_set);
+ if (set == NULL_RTX)
+ set = first_set;
+ break;
- set = single_set (insn);
- if (! set)
- abort ();
+ default:
+ gcc_unreachable ();
+ }
- /* 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)));
+ if (REG_P (SET_DEST (set)))
+ {
+ old_reg = SET_DEST (set);
+ /* Check if we can modify the set destination in the original insn. */
+ if (validate_change (insn, &SET_DEST (set), reg, 0))
+ {
+ new_insn = gen_move_insn (old_reg, reg);
+ new_insn = emit_insn_after (new_insn, insn);
+ }
+ else
+ {
+ new_insn = gen_move_insn (reg, old_reg);
+ new_insn = emit_insn_after (new_insn, insn);
+ }
+ }
+ else /* This is possible only in case of a store to memory. */
+ {
+ old_reg = SET_SRC (set);
+ new_insn = gen_move_insn (reg, old_reg);
- gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
- delete_insn (insn);
- occr->deleted_p = 1;
- if (!insn_inserted_p)
- {
- insert_insn_end_bb (index_map[i], bb, 0);
- insn_inserted_p = 1;
- }
- }
- }
- }
- }
- free (domby);
+ /* Check if we can modify the set source in the original insn. */
+ if (validate_change (insn, &SET_SRC (set), reg, 0))
+ new_insn = emit_insn_before (new_insn, insn);
+ else
+ new_insn = emit_insn_after (new_insn, insn);
}
- free (index_map);
-}
+ gcse_create_count++;
-/* Top level routine to perform one code hoisting (aka unification) pass
+ if (dump_file)
+ fprintf (dump_file,
+ "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
+ BLOCK_FOR_INSN (insn)->index, INSN_UID (new_insn), indx,
+ INSN_UID (insn), regno);
+}
- Return nonzero if a change was made. */
+/* Copy available expressions that reach the redundant expression
+ to `reaching_reg'. */
-static int
-one_code_hoisting_pass (void)
+static void
+pre_insert_copies (void)
{
- int changed = 0;
-
- alloc_hash_table (max_cuid, &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 (expr_hash_table.n_elems > 0)
- {
- alloc_code_hoist_mem (last_basic_block, expr_hash_table.n_elems);
- compute_code_hoist_data ();
- hoist_code ();
- free_code_hoist_mem ();
- }
+ unsigned int i, added_copy;
+ struct expr *expr;
+ struct occr *occr;
+ struct occr *avail;
- free_hash_table (&expr_hash_table);
+ /* For each available expression in the table, copy the result to
+ `reaching_reg' if the expression reaches a deleted one.
- return changed;
-}
-\f
-/* Here we provide the things required to do store motion towards
- the exit. In order for this to be effective, gcse also needed to
- be taught how to move a load when it is kill only by a store to itself.
+ ??? The current algorithm is rather brute force.
+ Need to do some profiling. */
- int i;
- float a[10];
+ for (i = 0; i < expr_hash_table.size; i++)
+ for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
+ {
+ /* If the basic block isn't reachable, PPOUT will be TRUE. However,
+ we don't want to insert a copy here because the expression may not
+ really be redundant. So only insert an insn if the expression was
+ deleted. This test also avoids further processing if the
+ expression wasn't deleted anywhere. */
+ if (expr->reaching_reg == NULL)
+ continue;
- void foo(float scale)
- {
- for (i=0; i<10; i++)
- a[i] *= scale;
- }
+ /* Set when we add a copy for that expression. */
+ added_copy = 0;
- 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
- the load out since its live around the loop, and stored at the bottom
- of the loop.
+ for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
+ {
+ if (! occr->deleted_p)
+ continue;
- The 'Load Motion' referred to and implemented in this file is
- an enhancement to gcse which when using edge based lcm, recognizes
- this situation and allows gcse to move the load out of the loop.
+ for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
+ {
+ rtx insn = avail->insn;
- Once gcse has hoisted the load, store motion can then push this
- load towards the exit, and we end up with no loads or stores of 'i'
- in the loop. */
+ /* No need to handle this one if handled already. */
+ if (avail->copied_p)
+ continue;
-/* This will search the ldst list for a matching expression. If it
- doesn't find one, we create one and initialize it. */
+ /* Don't handle this one if it's a redundant one. */
+ if (INSN_DELETED_P (insn))
+ continue;
-static struct ls_expr *
-ldst_entry (rtx x)
-{
- struct ls_expr * ptr;
+ /* Or if the expression doesn't reach the deleted one. */
+ if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail->insn),
+ expr,
+ BLOCK_FOR_INSN (occr->insn)))
+ continue;
- for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
- if (expr_equiv_p (ptr->pattern, x))
- break;
+ added_copy = 1;
- if (!ptr)
- {
- ptr = xmalloc (sizeof (struct ls_expr));
-
- ptr->next = pre_ldst_mems;
- ptr->expr = NULL;
- ptr->pattern = x;
- ptr->pattern_regs = NULL_RTX;
- ptr->loads = NULL_RTX;
- ptr->stores = NULL_RTX;
- ptr->reaching_reg = NULL_RTX;
- ptr->invalid = 0;
- ptr->index = 0;
- ptr->hash_index = 0;
- pre_ldst_mems = ptr;
- }
+ /* Copy the result of avail to reaching_reg. */
+ pre_insert_copy_insn (expr, insn);
+ avail->copied_p = 1;
+ }
+ }
- return ptr;
+ if (added_copy)
+ update_ld_motion_stores (expr);
+ }
}
-/* Free up an individual ldst entry. */
-
-static void
-free_ldst_entry (struct ls_expr * ptr)
+/* Emit move from SRC to DEST noting the equivalence with expression computed
+ in INSN. */
+static rtx
+gcse_emit_move_after (rtx src, rtx dest, rtx insn)
{
- free_INSN_LIST_list (& ptr->loads);
- free_INSN_LIST_list (& ptr->stores);
-
- free (ptr);
-}
+ rtx new_rtx;
+ rtx set = single_set (insn), set2;
+ rtx note;
+ rtx eqv;
-/* Free up all memory associated with the ldst list. */
+ /* This should never fail since we're creating a reg->reg copy
+ we've verified to be valid. */
-static void
-free_ldst_mems (void)
-{
- while (pre_ldst_mems)
- {
- struct ls_expr * tmp = pre_ldst_mems;
+ new_rtx = emit_insn_after (gen_move_insn (dest, src), insn);
- pre_ldst_mems = pre_ldst_mems->next;
+ /* Note the equivalence for local CSE pass. */
+ set2 = single_set (new_rtx);
+ if (!set2 || !rtx_equal_p (SET_DEST (set2), dest))
+ return new_rtx;
+ if ((note = find_reg_equal_equiv_note (insn)))
+ eqv = XEXP (note, 0);
+ else
+ eqv = SET_SRC (set);
- free_ldst_entry (tmp);
- }
+ set_unique_reg_note (new_rtx, REG_EQUAL, copy_insn_1 (eqv));
- pre_ldst_mems = NULL;
+ return new_rtx;
}
-/* Dump debugging info about the ldst list. */
-
-static void
-print_ldst_list (FILE * file)
-{
- struct ls_expr * ptr;
+/* Delete redundant computations.
+ Deletion is done by changing the insn to copy the `reaching_reg' of
+ the expression into the result of the SET. It is left to later passes
+ (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
- fprintf (file, "LDST list: \n");
+ Returns nonzero if a change is made. */
- for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
- {
- fprintf (file, " Pattern (%3d): ", ptr->index);
+static int
+pre_delete (void)
+{
+ unsigned int i;
+ int changed;
+ struct expr *expr;
+ struct occr *occr;
- print_rtl (file, ptr->pattern);
+ changed = 0;
+ for (i = 0; i < expr_hash_table.size; i++)
+ for (expr = expr_hash_table.table[i];
+ expr != NULL;
+ expr = expr->next_same_hash)
+ {
+ int indx = expr->bitmap_index;
- fprintf (file, "\n Loads : ");
+ /* We only need to search antic_occr since we require
+ ANTLOC != 0. */
- if (ptr->loads)
- print_rtl (file, ptr->loads);
- else
- fprintf (file, "(nil)");
+ for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
+ {
+ rtx insn = occr->insn;
+ rtx set;
+ basic_block bb = BLOCK_FOR_INSN (insn);
- fprintf (file, "\n Stores : ");
+ /* We only delete insns that have a single_set. */
+ if (TEST_BIT (pre_delete_map[bb->index], indx)
+ && (set = single_set (insn)) != 0
+ && dbg_cnt (pre_insn))
+ {
+ /* 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_and_attrs (SET_DEST (set));
- if (ptr->stores)
- print_rtl (file, ptr->stores);
- else
- fprintf (file, "(nil)");
+ gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
+ delete_insn (insn);
+ occr->deleted_p = 1;
+ changed = 1;
+ gcse_subst_count++;
- fprintf (file, "\n\n");
- }
+ if (dump_file)
+ {
+ fprintf (dump_file,
+ "PRE: redundant insn %d (expression %d) in ",
+ INSN_UID (insn), indx);
+ fprintf (dump_file, "bb %d, reaching reg is %d\n",
+ bb->index, REGNO (expr->reaching_reg));
+ }
+ }
+ }
+ }
- fprintf (file, "\n");
+ return changed;
}
-/* Returns 1 if X is in the list of ldst only expressions. */
-
-static struct ls_expr *
-find_rtx_in_ldst (rtx x)
-{
- struct ls_expr * ptr;
+/* Perform GCSE optimizations using PRE.
+ This is called by one_pre_gcse_pass after all the dataflow analysis
+ has been done.
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- if (expr_equiv_p (ptr->pattern, x) && ! ptr->invalid)
- return ptr;
+ This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
+ lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
+ Compiler Design and Implementation.
- return NULL;
-}
+ ??? A new pseudo reg is created to hold the reaching expression. The nice
+ thing about the classical approach is that it would try to use an existing
+ reg. If the register can't be adequately optimized [i.e. we introduce
+ reload problems], one could add a pass here to propagate the new register
+ through the block.
-/* Assign each element of the list of mems a monotonically increasing value. */
+ ??? We don't handle single sets in PARALLELs because we're [currently] not
+ able to copy the rest of the parallel when we insert copies to create full
+ redundancies from partial redundancies. However, there's no reason why we
+ can't handle PARALLELs in the cases where there are no partial
+ redundancies. */
static int
-enumerate_ldsts (void)
+pre_gcse (void)
{
- struct ls_expr * ptr;
- int n = 0;
+ unsigned int i;
+ int did_insert, changed;
+ struct expr **index_map;
+ struct expr *expr;
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- ptr->index = n++;
+ /* Compute a mapping from expression number (`bitmap_index') to
+ hash table entry. */
- return n;
-}
+ 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;
-/* Return first item in the list. */
+ /* 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 */
-static inline struct ls_expr *
-first_ls_expr (void)
-{
- return pre_ldst_mems;
-}
+ changed = pre_delete ();
+ did_insert = pre_edge_insert (edge_list, index_map);
-/* Return the next item in the list after the specified one. */
+ /* In other places with reaching expressions, copy the expression to the
+ specially allocated pseudo-reg that reaches the redundant expr. */
+ pre_insert_copies ();
+ if (did_insert)
+ {
+ commit_edge_insertions ();
+ changed = 1;
+ }
-static inline struct ls_expr *
-next_ls_expr (struct ls_expr * ptr)
-{
- return ptr->next;
+ free (index_map);
+ return changed;
}
-\f
-/* Load Motion for loads which only kill themselves. */
-/* Return true if x is a simple MEM operation, with no registers or
- side effects. These are the types of loads we consider for the
- ld_motion list, otherwise we let the usual aliasing take care of it. */
+/* Top level routine to perform one PRE GCSE pass.
+
+ Return nonzero if a change was made. */
static int
-simple_mem (rtx x)
+one_pre_gcse_pass (void)
{
- if (GET_CODE (x) != MEM)
- return 0;
+ int changed = 0;
- if (MEM_VOLATILE_P (x))
- return 0;
+ gcse_subst_count = 0;
+ gcse_create_count = 0;
- if (GET_MODE (x) == BLKmode)
+ /* 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;
- /* If we are handling exceptions, we must be careful with memory references
- that may trap. If we are not, the behavior is undefined, so we may just
- continue. */
- if (flag_non_call_exceptions && may_trap_p (x))
- return 0;
+ /* We need alias. */
+ init_alias_analysis ();
- if (side_effects_p (x))
- return 0;
+ bytes_used = 0;
+ gcc_obstack_init (&gcse_obstack);
+ alloc_gcse_mem ();
+
+ alloc_hash_table (&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 (dump_file)
+ dump_hash_table (dump_file, "Expression", &expr_hash_table);
+
+ if (expr_hash_table.n_elems > 0)
+ {
+ alloc_pre_mem (last_basic_block, expr_hash_table.n_elems);
+ compute_pre_data ();
+ changed |= pre_gcse ();
+ free_edge_list (edge_list);
+ free_pre_mem ();
+ }
+
+ free_ldst_mems ();
+ remove_fake_exit_edges ();
+ free_hash_table (&expr_hash_table);
- /* Do not consider function arguments passed on stack. */
- if (reg_mentioned_p (stack_pointer_rtx, x))
- return 0;
+ free_gcse_mem ();
+ obstack_free (&gcse_obstack, NULL);
- if (flag_float_store && FLOAT_MODE_P (GET_MODE (x)))
- return 0;
+ /* We are finished with alias. */
+ end_alias_analysis ();
- return 1;
+ if (dump_file)
+ {
+ 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_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. */
-/* Make sure there isn't a buried reference in this pattern anywhere.
- If there is, invalidate the entry for it since we're not capable
- of fixing it up just yet.. We have to be sure we know about ALL
- loads since the aliasing code will allow all entries in the
- ld_motion list to not-alias itself. If we miss a load, we will get
- the wrong value since gcse might common it and we won't know to
- fix it up. */
+/* ??? 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_OPERAND and REG_LABEL_TARGET notes. */
static void
-invalidate_any_buried_refs (rtx x)
+add_label_notes (rtx x, rtx insn)
{
- const char * fmt;
+ enum rtx_code code = GET_CODE (x);
int i, j;
- struct ls_expr * ptr;
+ const char *fmt;
- /* Invalidate it in the list. */
- if (GET_CODE (x) == MEM && simple_mem (x))
+ if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
{
- ptr = ldst_entry (x);
- ptr->invalid = 1;
- }
+ /* This code used to ignore labels that referred to dispatch tables to
+ avoid flow generating (slightly) worse code.
- /* Recursively process the insn. */
- fmt = GET_RTX_FORMAT (GET_CODE (x));
+ We no longer ignore such label references (see LABEL_REF handling in
+ mark_jump_label for additional information). */
- for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
+ /* 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 (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
{
if (fmt[i] == 'e')
- invalidate_any_buried_refs (XEXP (x, i));
+ add_label_notes (XEXP (x, i), insn);
else if (fmt[i] == 'E')
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
- invalidate_any_buried_refs (XVECEXP (x, i, j));
+ add_label_notes (XVECEXP (x, i, j), insn);
}
}
-/* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
- being defined as MEM loads and stores to symbols, with no side effects
- and no registers in the expression. For a MEM destination, we also
- check that the insn is still valid if we replace the destination with a
- REG, as is done in update_ld_motion_stores. If there are any uses/defs
- which don't match this criteria, they are invalidated and trimmed out
- later. */
+/* Compute transparent outgoing information for each block.
+
+ An expression is transparent to an edge unless it is killed by
+ the edge itself. This can only happen with abnormal control flow,
+ when the edge is traversed through a call. This happens with
+ non-local labels and exceptions.
+
+ This would not be necessary if we split the edge. While this is
+ normally impossible for abnormal critical edges, with some effort
+ it should be possible with exception handling, since we still have
+ control over which handler should be invoked. But due to increased
+ EH table sizes, this may not be worthwhile. */
static void
-compute_ld_motion_mems (void)
+compute_transpout (void)
{
- struct ls_expr * ptr;
basic_block bb;
- rtx insn;
+ unsigned int i;
+ struct expr *expr;
- pre_ldst_mems = NULL;
+ sbitmap_vector_ones (transpout, last_basic_block);
FOR_EACH_BB (bb)
{
- for (insn = bb->head;
- insn && insn != NEXT_INSN (bb->end);
- insn = NEXT_INSN (insn))
- {
- if (INSN_P (insn))
- {
- if (GET_CODE (PATTERN (insn)) == SET)
- {
- rtx src = SET_SRC (PATTERN (insn));
- rtx dest = SET_DEST (PATTERN (insn));
-
- /* Check for a simple LOAD... */
- if (GET_CODE (src) == MEM && simple_mem (src))
- {
- ptr = ldst_entry (src);
- if (GET_CODE (dest) == REG)
- ptr->loads = alloc_INSN_LIST (insn, ptr->loads);
- else
- ptr->invalid = 1;
- }
- else
- {
- /* Make sure there isn't a buried load somewhere. */
- invalidate_any_buried_refs (src);
- }
+ /* 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)))
+ continue;
- /* Check for stores. Don't worry about aliased ones, they
- will block any movement we might do later. We only care
- about this exact pattern since those are the only
- circumstance that we will ignore the aliasing info. */
- if (GET_CODE (dest) == MEM && simple_mem (dest))
- {
- ptr = ldst_entry (dest);
+ for (i = 0; i < expr_hash_table.size; i++)
+ for (expr = expr_hash_table.table[i]; expr ; expr = expr->next_same_hash)
+ if (MEM_P (expr->expr))
+ {
+ if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
+ && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
+ continue;
- if (GET_CODE (src) != MEM
- && GET_CODE (src) != ASM_OPERANDS
- /* Check for REG manually since want_to_gcse_p
- returns 0 for all REGs. */
- && (REG_P (src) || want_to_gcse_p (src)))
- ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
- else
- ptr->invalid = 1;
- }
- }
- else
- invalidate_any_buried_refs (PATTERN (insn));
+ /* ??? Optimally, we would use interprocedural alias
+ analysis to determine if this mem is actually killed
+ by this call. */
+ RESET_BIT (transpout[bb->index], expr->bitmap_index);
}
- }
}
}
-/* Remove any references that have been either invalidated or are not in the
- expression list for pre gcse. */
-
-static void
-trim_ld_motion_mems (void)
-{
- struct ls_expr * last = NULL;
- struct ls_expr * ptr = first_ls_expr ();
-
- while (ptr != NULL)
- {
- int del = ptr->invalid;
- struct expr * expr = NULL;
+/* Code Hoisting variables and subroutines. */
- /* Delete if entry has been made invalid. */
- if (!del)
- {
- unsigned int i;
+/* Very busy expressions. */
+static sbitmap *hoist_vbein;
+static sbitmap *hoist_vbeout;
- del = 1;
- /* Delete if we cannot find this mem in the expression list. */
- for (i = 0; i < expr_hash_table.size && del; i++)
- {
- for (expr = expr_hash_table.table[i];
- expr != NULL;
- expr = expr->next_same_hash)
- if (expr_equiv_p (expr->expr, ptr->pattern))
- {
- del = 0;
- break;
- }
- }
- }
+/* Hoistable expressions. */
+static sbitmap *hoist_exprs;
- if (del)
- {
- if (last != NULL)
- {
- last->next = ptr->next;
- free_ldst_entry (ptr);
- ptr = last->next;
- }
- else
- {
- pre_ldst_mems = pre_ldst_mems->next;
- free_ldst_entry (ptr);
- ptr = pre_ldst_mems;
- }
- }
- else
- {
- /* Set the expression field if we are keeping it. */
- last = ptr;
- ptr->expr = expr;
- ptr = ptr->next;
- }
- }
+/* ??? We could compute post dominators and run this algorithm in
+ reverse to perform tail merging, doing so would probably be
+ more effective than the tail merging code in jump.c.
- /* Show the world what we've found. */
- if (gcse_file && pre_ldst_mems != NULL)
- print_ldst_list (gcse_file);
-}
+ It's unclear if tail merging could be run in parallel with
+ code hoisting. It would be nice. */
-/* This routine will take an expression which we are replacing with
- a reaching register, and update any stores that are needed if
- that expression is in the ld_motion list. Stores are updated by
- copying their SRC to the reaching register, and then storeing
- the reaching register into the store location. These keeps the
- correct value in the reaching register for the loads. */
+/* Allocate vars used for code hoisting analysis. */
static void
-update_ld_motion_stores (struct expr * expr)
+alloc_code_hoist_mem (int n_blocks, int n_exprs)
{
- struct ls_expr * mem_ptr;
+ antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
+ transp = sbitmap_vector_alloc (n_blocks, n_exprs);
+ comp = sbitmap_vector_alloc (n_blocks, n_exprs);
- if ((mem_ptr = find_rtx_in_ldst (expr->expr)))
- {
- /* We can try to find just the REACHED stores, but is shouldn't
- matter to set the reaching reg everywhere... some might be
- dead and should be eliminated later. */
+ hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
+ hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
+ hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
+ transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
+}
- /* We replace (set mem expr) with (set reg expr) (set mem reg)
- where reg is the reaching reg used in the load. We checked in
- compute_ld_motion_mems that we can replace (set mem expr) with
- (set reg expr) in that insn. */
- rtx list = mem_ptr->stores;
+/* Free vars used for code hoisting analysis. */
- for ( ; list != NULL_RTX; list = XEXP (list, 1))
- {
- rtx insn = XEXP (list, 0);
- rtx pat = PATTERN (insn);
- rtx src = SET_SRC (pat);
- rtx reg = expr->reaching_reg;
- rtx copy, new;
+static void
+free_code_hoist_mem (void)
+{
+ sbitmap_vector_free (antloc);
+ sbitmap_vector_free (transp);
+ sbitmap_vector_free (comp);
- /* If we've already copied it, continue. */
- if (expr->reaching_reg == src)
- continue;
+ sbitmap_vector_free (hoist_vbein);
+ sbitmap_vector_free (hoist_vbeout);
+ sbitmap_vector_free (hoist_exprs);
+ sbitmap_vector_free (transpout);
- if (gcse_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");
- }
+ free_dominance_info (CDI_DOMINATORS);
+}
- copy = gen_move_insn ( reg, copy_rtx (SET_SRC (pat)));
- new = emit_insn_before (copy, insn);
- record_one_set (REGNO (reg), new);
- SET_SRC (pat) = reg;
+/* Compute the very busy expressions at entry/exit from each block.
- /* un-recognize this pattern since it's probably different now. */
- INSN_CODE (insn) = -1;
- gcse_create_count++;
- }
- }
-}
-\f
-/* Store motion code. */
+ An expression is very busy if all paths from a given point
+ compute the expression. */
-#define ANTIC_STORE_LIST(x) ((x)->loads)
-#define AVAIL_STORE_LIST(x) ((x)->stores)
-#define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
+static void
+compute_code_hoist_vbeinout (void)
+{
+ int changed, passes;
+ basic_block bb;
-/* 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;
+ sbitmap_vector_zero (hoist_vbeout, last_basic_block);
+ sbitmap_vector_zero (hoist_vbein, last_basic_block);
-/* And current insn, for the same routine. */
-static rtx compute_store_table_current_insn;
+ passes = 0;
+ changed = 1;
-/* Used in computing the reverse edge graph bit vectors. */
-static sbitmap * st_antloc;
+ while (changed)
+ {
+ changed = 0;
-/* Global holding the number of store expressions we are dealing with. */
-static int num_stores;
+ /* We scan the blocks in the reverse order to speed up
+ the convergence. */
+ FOR_EACH_BB_REVERSE (bb)
+ {
+ if (bb->next_bb != EXIT_BLOCK_PTR)
+ sbitmap_intersection_of_succs (hoist_vbeout[bb->index],
+ hoist_vbein, bb->index);
-/* Checks to set if we need to mark a register set. Called from note_stores. */
+ changed |= sbitmap_a_or_b_and_c_cg (hoist_vbein[bb->index],
+ antloc[bb->index],
+ hoist_vbeout[bb->index],
+ transp[bb->index]);
+ }
-static void
-reg_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED,
- void *data ATTRIBUTE_UNUSED)
-{
- if (GET_CODE (dest) == SUBREG)
- dest = SUBREG_REG (dest);
+ passes++;
+ }
- if (GET_CODE (dest) == REG)
- regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
+ if (dump_file)
+ fprintf (dump_file, "hoisting vbeinout computation: %d passes\n", passes);
}
-/* Return zero if some of the registers in list X are killed
- due to set of registers in bitmap REGS_SET. */
+/* Top level routine to do the dataflow analysis needed by code hoisting. */
-static bool
-store_ops_ok (rtx x, int *regs_set)
+static void
+compute_code_hoist_data (void)
{
- rtx reg;
+ compute_local_properties (transp, comp, antloc, &expr_hash_table);
+ compute_transpout ();
+ compute_code_hoist_vbeinout ();
+ calculate_dominance_info (CDI_DOMINATORS);
+ if (dump_file)
+ fprintf (dump_file, "\n");
+}
- for (; x; x = XEXP (x, 1))
- {
- reg = XEXP (x, 0);
- if (regs_set[REGNO(reg)])
- return false;
- }
+/* Determine if the expression identified by EXPR_INDEX would
+ reach BB unimpared if it was placed at the end of EXPR_BB.
- return true;
-}
+ It's unclear exactly what Muchnick meant by "unimpared". It seems
+ to me that the expression must either be computed or transparent in
+ *every* block in the path(s) from EXPR_BB to BB. Any other definition
+ would allow the expression to be hoisted out of loops, even if
+ the expression wasn't a loop invariant.
-/* Returns a list of registers mentioned in X. */
-static rtx
-extract_mentioned_regs (rtx x)
-{
- return extract_mentioned_regs_helper (x, NULL_RTX);
-}
+ Contrast this to reachability for PRE where an expression is
+ considered reachable if *any* path reaches instead of *all*
+ paths. */
-/* Helper for extract_mentioned_regs; ACCUM is used to accumulate used
- registers. */
-static rtx
-extract_mentioned_regs_helper (rtx x, rtx accum)
+static int
+hoist_expr_reaches_here_p (basic_block expr_bb, int expr_index, basic_block bb, char *visited)
{
- int i;
- enum rtx_code code;
- const char * fmt;
-
- /* Repeat is used to turn tail-recursion into iteration. */
- repeat:
+ edge pred;
+ edge_iterator ei;
+ int visited_allocated_locally = 0;
- if (x == 0)
- return accum;
- code = GET_CODE (x);
- switch (code)
+ if (visited == NULL)
{
- case REG:
- return alloc_EXPR_LIST (0, x, accum);
+ visited_allocated_locally = 1;
+ visited = XCNEWVEC (char, last_basic_block);
+ }
- case MEM:
- x = XEXP (x, 0);
- goto repeat;
+ FOR_EACH_EDGE (pred, ei, bb->preds)
+ {
+ basic_block pred_bb = pred->src;
- case PRE_DEC:
- case PRE_INC:
- case POST_DEC:
- case POST_INC:
- /* We do not run this function with arguments having side effects. */
- abort ();
+ if (pred->src == ENTRY_BLOCK_PTR)
+ break;
+ else if (pred_bb == expr_bb)
+ continue;
+ else if (visited[pred_bb->index])
+ continue;
- 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;
+ /* Does this predecessor generate this expression? */
+ else if (TEST_BIT (comp[pred_bb->index], expr_index))
+ break;
+ else if (! TEST_BIT (transp[pred_bb->index], expr_index))
+ break;
- default:
- break;
+ /* Not killed. */
+ else
+ {
+ visited[pred_bb->index] = 1;
+ if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
+ pred_bb, visited))
+ break;
+ }
}
+ if (visited_allocated_locally)
+ free (visited);
- i = GET_RTX_LENGTH (code) - 1;
- fmt = GET_RTX_FORMAT (code);
+ return (pred == NULL);
+}
+\f
+/* Actually perform code hoisting. */
- for (; i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- rtx tem = XEXP (x, i);
+static int
+hoist_code (void)
+{
+ basic_block bb, dominated;
+ VEC (basic_block, heap) *domby;
+ unsigned int i,j;
+ struct expr **index_map;
+ struct expr *expr;
+ int changed = 0;
- /* 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;
- }
+ sbitmap_vector_zero (hoist_exprs, last_basic_block);
- accum = extract_mentioned_regs_helper (tem, accum);
- }
- else if (fmt[i] == 'E')
- {
- int j;
+ /* Compute a mapping from expression number (`bitmap_index') to
+ hash table entry. */
- for (j = 0; j < XVECLEN (x, i); j++)
- accum = extract_mentioned_regs_helper (XVECEXP (x, i, j), accum);
- }
- }
+ 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;
- return accum;
-}
+ /* Walk over each basic block looking for potentially hoistable
+ expressions, nothing gets hoisted from the entry block. */
+ FOR_EACH_BB (bb)
+ {
+ int found = 0;
+ int insn_inserted_p;
-/* 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.
+ 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++)
+ {
+ int hoistable = 0;
- The results are stored this way:
+ if (TEST_BIT (hoist_vbeout[bb->index], i)
+ && TEST_BIT (transpout[bb->index], i))
+ {
+ /* We've found a potentially hoistable expression, now
+ we look at every block BB dominates to see if it
+ computes the expression. */
+ for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
+ {
+ /* Ignore self dominance. */
+ if (bb == dominated)
+ continue;
+ /* We've found a dominated block, now see if it computes
+ the busy expression and whether or not moving that
+ expression to the "beginning" of that block is safe. */
+ if (!TEST_BIT (antloc[dominated->index], i))
+ continue;
- -- 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.
+ /* Note if the expression would reach the dominated block
+ unimpared if it was placed at the end of BB.
- 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.
- */
+ Keep track of how many times this expression is hoistable
+ from a dominated block into BB. */
+ if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
+ hoistable++;
+ }
-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);
+ /* If we found more than one hoistable occurrence of this
+ expression, then note it in the bitmap of expressions to
+ hoist. It makes no sense to hoist things which are computed
+ in only one BB, and doing so tends to pessimize register
+ allocation. One could increase this value to try harder
+ to avoid any possible code expansion due to register
+ allocation issues; however experiments have shown that
+ the vast majority of hoistable expressions are only movable
+ from two successors, so raising this threshold is likely
+ to nullify any benefit we get from code hoisting. */
+ if (hoistable > 1)
+ {
+ SET_BIT (hoist_exprs[bb->index], i);
+ found = 1;
+ }
+ }
+ }
+ /* If we found nothing to hoist, then quit now. */
+ if (! found)
+ {
+ VEC_free (basic_block, heap, domby);
+ continue;
+ }
- set = single_set (insn);
- if (!set)
- return;
+ /* Loop over all the hoistable expressions. */
+ for (i = 0; i < hoist_exprs[bb->index]->n_bits; i++)
+ {
+ /* We want to insert the expression into BB only once, so
+ note when we've inserted it. */
+ insn_inserted_p = 0;
+
+ /* These tests should be the same as the tests above. */
+ if (TEST_BIT (hoist_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; VEC_iterate (basic_block, domby, j, dominated); j++)
+ {
+ /* Ignore self dominance. */
+ if (bb == dominated)
+ continue;
- dest = SET_DEST (set);
+ /* We've found a dominated block, now see if it computes
+ the busy expression and whether or not moving that
+ expression to the "beginning" of that block is safe. */
+ if (!TEST_BIT (antloc[dominated->index], i))
+ continue;
- if (GET_CODE (dest) != MEM || MEM_VOLATILE_P (dest)
- || GET_MODE (dest) == BLKmode)
- return;
+ /* The expression is computed in the dominated block and
+ it would be safe to compute it at the start of the
+ dominated block. Now we have to determine if the
+ expression would reach the dominated block if it was
+ placed at the end of BB. */
+ if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
+ {
+ struct expr *expr = index_map[i];
+ struct occr *occr = expr->antic_occr;
+ rtx insn;
+ rtx set;
- if (side_effects_p (dest))
- return;
+ /* Find the right occurrence of this expression. */
+ while (BLOCK_FOR_INSN (occr->insn) != dominated && occr)
+ occr = occr->next;
- /* 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;
+ gcc_assert (occr);
+ insn = occr->insn;
+ set = single_set (insn);
+ gcc_assert (set);
- ptr = ldst_entry (dest);
- if (!ptr->pattern_regs)
- ptr->pattern_regs = extract_mentioned_regs (dest);
+ /* 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_and_attrs (SET_DEST (set));
- /* Do not check for anticipatability if we either found one anticipatable
- store already, or tested for one and found out that it was killed. */
- check_anticipatable = 0;
- if (!ANTIC_STORE_LIST (ptr))
- check_anticipatable = 1;
- else
- {
- tmp = XEXP (ANTIC_STORE_LIST (ptr), 0);
- if (tmp != NULL_RTX
- && BLOCK_FOR_INSN (tmp) != bb)
- check_anticipatable = 1;
- }
- if (check_anticipatable)
- {
- if (store_killed_before (dest, ptr->pattern_regs, insn, bb, regs_set_before))
- tmp = NULL_RTX;
- else
- tmp = insn;
- ANTIC_STORE_LIST (ptr) = alloc_INSN_LIST (tmp,
- ANTIC_STORE_LIST (ptr));
- }
+ gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
+ delete_insn (insn);
+ occr->deleted_p = 1;
+ changed = 1;
+ gcse_subst_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;
- tmp != insn && tmp != LAST_AVAIL_CHECK_FAILURE (ptr);
- tmp = PREV_INSN (tmp))
- continue;
- if (tmp == insn)
- check_available = 0;
+ if (!insn_inserted_p)
+ {
+ insert_insn_end_basic_block (index_map[i], bb, 0);
+ insn_inserted_p = 1;
+ }
+ }
+ }
+ }
}
- else
- check_available = store_killed_after (dest, ptr->pattern_regs, insn,
- bb, regs_set_after,
- &LAST_AVAIL_CHECK_FAILURE (ptr));
+ VEC_free (basic_block, heap, domby);
}
- if (!check_available)
- AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn, AVAIL_STORE_LIST (ptr));
+
+ free (index_map);
+
+ return changed;
}
-/* Find available and anticipatable stores. */
+/* Top level routine to perform one code hoisting (aka unification) pass
+
+ Return nonzero if a change was made. */
static int
-compute_store_table (void)
+one_code_hoisting_pass (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;
+ int changed = 0;
+
+ gcse_subst_count = 0;
+ gcse_create_count = 0;
- max_gcse_regno = max_reg_num ();
+ /* 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;
- reg_set_in_block = sbitmap_vector_alloc (last_basic_block,
- max_gcse_regno);
- sbitmap_vector_zero (reg_set_in_block, last_basic_block);
- pre_ldst_mems = 0;
- last_set_in = xmalloc (sizeof (int) * max_gcse_regno);
- already_set = xmalloc (sizeof (int) * max_gcse_regno);
+ /* We need alias. */
+ init_alias_analysis ();
- /* Find all the stores we care about. */
- FOR_EACH_BB (bb)
+ bytes_used = 0;
+ gcc_obstack_init (&gcse_obstack);
+ alloc_gcse_mem ();
+
+ alloc_hash_table (&expr_hash_table, 0);
+ compute_hash_table (&expr_hash_table);
+ if (dump_file)
+ dump_hash_table (dump_file, "Code Hosting Expressions", &expr_hash_table);
+
+ if (expr_hash_table.n_elems > 0)
{
- /* First compute the registers set in this block. */
- memset (last_set_in, 0, sizeof (int) * max_gcse_regno);
- regvec = last_set_in;
+ alloc_code_hoist_mem (last_basic_block, expr_hash_table.n_elems);
+ compute_code_hoist_data ();
+ changed = hoist_code ();
+ free_code_hoist_mem ();
+ }
- for (insn = bb->head;
- insn != NEXT_INSN (bb->end);
- insn = NEXT_INSN (insn))
- {
- if (! INSN_P (insn))
- continue;
+ free_hash_table (&expr_hash_table);
+ free_gcse_mem ();
+ obstack_free (&gcse_obstack, NULL);
- if (GET_CODE (insn) == CALL_INSN)
- {
- bool clobbers_all = false;
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
+ /* We are finished with alias. */
+ end_alias_analysis ();
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- last_set_in[regno] = INSN_UID (insn);
- }
+ if (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);
+ }
- pat = PATTERN (insn);
- compute_store_table_current_insn = insn;
- note_stores (pat, reg_set_info, NULL);
- }
+ return changed;
+}
+\f
+/* Here we provide the things required to do store motion towards
+ the exit. In order for this to be effective, gcse also needed to
+ be taught how to move a load when it is kill only by a store to itself.
- /* Record the set registers. */
- for (regno = 0; regno < max_gcse_regno; regno++)
- if (last_set_in[regno])
- SET_BIT (reg_set_in_block[bb->index], regno);
-
- /* Now find the stores. */
- memset (already_set, 0, sizeof (int) * max_gcse_regno);
- regvec = already_set;
- for (insn = bb->head;
- insn != NEXT_INSN (bb->end);
- insn = NEXT_INSN (insn))
- {
- if (! INSN_P (insn))
- continue;
+ int i;
+ float a[10];
- if (GET_CODE (insn) == CALL_INSN)
+ void foo(float scale)
{
- bool clobbers_all = false;
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
-
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- already_set[regno] = 1;
+ for (i=0; i<10; i++)
+ a[i] *= scale;
}
- pat = PATTERN (insn);
- note_stores (pat, reg_set_info, NULL);
+ 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
+ the load out since its live around the loop, and stored at the bottom
+ of the loop.
- /* Now that we've marked regs, look for stores. */
- find_moveable_store (insn, already_set, last_set_in);
+ The 'Load Motion' referred to and implemented in this file is
+ an enhancement to gcse which when using edge based lcm, recognizes
+ this situation and allows gcse to move the load out of the loop.
- /* Unmark regs that are no longer set. */
- for (regno = 0; regno < max_gcse_regno; regno++)
- if (last_set_in[regno] == INSN_UID (insn))
- last_set_in[regno] = 0;
- }
+ Once gcse has hoisted the load, store motion can then push this
+ load towards the exit, and we end up with no loads or stores of 'i'
+ in the loop. */
- /* 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);
- }
- }
+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);
+}
- /* 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;
- }
+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);
+}
- ret = enumerate_ldsts ();
+/* This will search the ldst list for a matching expression. If it
+ doesn't find one, we create one and initialize it. */
- if (gcse_file)
- {
- fprintf (gcse_file, "ST_avail and ST_antic (shown under loads..)\n");
- print_ldst_list (gcse_file);
- }
+static struct ls_expr *
+ldst_entry (rtx x)
+{
+ int do_not_record_p = 0;
+ struct ls_expr * ptr;
+ unsigned int hash;
+ void **slot;
+ struct ls_expr e;
+
+ hash = hash_rtx (x, GET_MODE (x), &do_not_record_p,
+ NULL, /*have_reg_qty=*/false);
+
+ e.pattern = x;
+ slot = htab_find_slot_with_hash (pre_ldst_table, &e, hash, INSERT);
+ if (*slot)
+ return (struct ls_expr *)*slot;
+
+ ptr = XNEW (struct ls_expr);
+
+ ptr->next = pre_ldst_mems;
+ ptr->expr = NULL;
+ ptr->pattern = x;
+ ptr->pattern_regs = NULL_RTX;
+ ptr->loads = NULL_RTX;
+ ptr->stores = NULL_RTX;
+ ptr->reaching_reg = NULL_RTX;
+ ptr->invalid = 0;
+ ptr->index = 0;
+ ptr->hash_index = hash;
+ pre_ldst_mems = ptr;
+ *slot = ptr;
- free (last_set_in);
- free (already_set);
- return ret;
+ return ptr;
}
-/* 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. */
+/* Free up an individual ldst entry. */
-static bool
-load_kills_store (rtx x, rtx store_pattern, int after)
+static void
+free_ldst_entry (struct ls_expr * ptr)
{
- if (after)
- return anti_dependence (x, store_pattern);
- else
- return true_dependence (store_pattern, GET_MODE (store_pattern), x,
- rtx_addr_varies_p);
+ free_INSN_LIST_list (& ptr->loads);
+ free_INSN_LIST_list (& ptr->stores);
+
+ free (ptr);
}
-/* 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. */
+/* Free up all memory associated with the ldst list. */
-static bool
-find_loads (rtx x, rtx store_pattern, int after)
+static void
+free_ldst_mems (void)
{
- const char * fmt;
- int i, j;
- int ret = false;
-
- if (!x)
- return false;
+ if (pre_ldst_table)
+ htab_delete (pre_ldst_table);
+ pre_ldst_table = NULL;
- if (GET_CODE (x) == SET)
- x = SET_SRC (x);
-
- if (GET_CODE (x) == MEM)
+ while (pre_ldst_mems)
{
- if (load_kills_store (x, store_pattern, after))
- return true;
- }
+ struct ls_expr * tmp = pre_ldst_mems;
- /* Recursively process the insn. */
- fmt = GET_RTX_FORMAT (GET_CODE (x));
+ pre_ldst_mems = pre_ldst_mems->next;
- 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);
+ free_ldst_entry (tmp);
}
- return ret;
+
+ pre_ldst_mems = NULL;
}
-/* 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. */
+/* Dump debugging info about the ldst list. */
-static bool
-store_killed_in_insn (rtx x, rtx x_regs, rtx insn, int after)
+static void
+print_ldst_list (FILE * file)
{
- rtx reg, base;
+ struct ls_expr * ptr;
- if (!INSN_P (insn))
- return false;
+ fprintf (file, "LDST list: \n");
- if (GET_CODE (insn) == CALL_INSN)
+ for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
{
- /* 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;
- }
+ fprintf (file, " Pattern (%3d): ", ptr->index);
- if (GET_CODE (PATTERN (insn)) == SET)
- {
- rtx pat = PATTERN (insn);
- rtx dest = SET_DEST (pat);
+ print_rtl (file, ptr->pattern);
- if (GET_CODE (dest) == SIGN_EXTRACT
- || GET_CODE (dest) == ZERO_EXTRACT)
- dest = XEXP (dest, 0);
+ fprintf (file, "\n Loads : ");
- /* Check for memory stores to aliased objects. */
- if (GET_CODE (dest) == MEM
- && !expr_equiv_p (dest, x))
- {
- if (after)
- {
- if (output_dependence (dest, x))
- return true;
- }
- else
- {
- if (output_dependence (x, dest))
- return true;
- }
- }
- return find_loads (SET_SRC (pat), x, after);
- }
- else
- return find_loads (PATTERN (insn), x, after);
-}
+ if (ptr->loads)
+ print_rtl (file, ptr->loads);
+ else
+ fprintf (file, "(nil)");
-/* 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. */
+ fprintf (file, "\n Stores : ");
-static bool
-store_killed_after (rtx x, rtx x_regs, rtx insn, basic_block bb,
- int *regs_set_after, rtx *fail_insn)
-{
- rtx last = bb->end, act;
+ if (ptr->stores)
+ print_rtl (file, ptr->stores);
+ else
+ fprintf (file, "(nil)");
- 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;
+ fprintf (file, "\n\n");
}
- /* 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;
+ fprintf (file, "\n");
}
-/* 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)
-{
- rtx first = bb->head;
+/* Returns 1 if X is in the list of ldst only expressions. */
- if (!store_ops_ok (x_regs, regs_set_before))
- return true;
+static struct ls_expr *
+find_rtx_in_ldst (rtx x)
+{
+ 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;
+}
- for ( ; insn != PREV_INSN (first); insn = PREV_INSN (insn))
- if (store_killed_in_insn (x, x_regs, insn, true))
- return true;
+/* Return first item in the list. */
- return false;
+static inline struct ls_expr *
+first_ls_expr (void)
+{
+ return pre_ldst_mems;
}
-/* 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)
-{
- basic_block bb;
- int *regs_set_in_block;
- rtx insn, st;
- struct ls_expr * ptr;
- unsigned regno;
+/* Return the next item in the list after the specified one. */
- /* 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);
+static inline struct ls_expr *
+next_ls_expr (struct ls_expr * ptr)
+{
+ return ptr->next;
+}
+\f
+/* Load Motion for loads which only kill themselves. */
- st_antloc = sbitmap_vector_alloc (last_basic_block, num_stores);
- sbitmap_vector_zero (st_antloc, last_basic_block);
+/* Return true if x is a simple MEM operation, with no registers or
+ side effects. These are the types of loads we consider for the
+ ld_motion list, otherwise we let the usual aliasing take care of it. */
- 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);
- continue;
- }
- SET_BIT (ae_gen[bb->index], ptr->index);
- }
+static int
+simple_mem (const_rtx x)
+{
+ if (! MEM_P (x))
+ return 0;
- 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);
- }
- }
+ if (MEM_VOLATILE_P (x))
+ return 0;
- ae_kill = sbitmap_vector_alloc (last_basic_block, num_stores);
- sbitmap_vector_zero (ae_kill, last_basic_block);
+ if (GET_MODE (x) == BLKmode)
+ return 0;
- 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);
+ /* If we are handling exceptions, we must be careful with memory references
+ that may trap. If we are not, the behavior is undefined, so we may just
+ continue. */
+ if (flag_non_call_exceptions && may_trap_p (x))
+ return 0;
- 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);
+ if (side_effects_p (x))
+ return 0;
- for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
- {
- if (store_killed_after (ptr->pattern, ptr->pattern_regs, bb->head,
- 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);
- }
- }
+ /* Do not consider function arguments passed on stack. */
+ if (reg_mentioned_p (stack_pointer_rtx, x))
+ return 0;
- free (regs_set_in_block);
+ if (flag_float_store && FLOAT_MODE_P (GET_MODE (x)))
+ return 0;
- 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);
- }
+ return 1;
}
-/* Insert an instruction at the beginning of a basic block, and update
- the BLOCK_HEAD if needed. */
+/* Make sure there isn't a buried reference in this pattern anywhere.
+ If there is, invalidate the entry for it since we're not capable
+ of fixing it up just yet.. We have to be sure we know about ALL
+ loads since the aliasing code will allow all entries in the
+ ld_motion list to not-alias itself. If we miss a load, we will get
+ the wrong value since gcse might common it and we won't know to
+ fix it up. */
static void
-insert_insn_start_bb (rtx insn, basic_block bb)
+invalidate_any_buried_refs (rtx x)
{
- /* Insert at start of successor block. */
- rtx prev = PREV_INSN (bb->head);
- rtx before = bb->head;
- while (before != 0)
+ const char * fmt;
+ int i, j;
+ struct ls_expr * ptr;
+
+ /* Invalidate it in the list. */
+ if (MEM_P (x) && simple_mem (x))
{
- if (GET_CODE (before) != CODE_LABEL
- && (GET_CODE (before) != NOTE
- || NOTE_LINE_NUMBER (before) != NOTE_INSN_BASIC_BLOCK))
- break;
- prev = before;
- if (prev == bb->end)
- break;
- before = NEXT_INSN (before);
+ ptr = ldst_entry (x);
+ ptr->invalid = 1;
}
- insn = emit_insn_after (insn, prev);
+ /* Recursively process the insn. */
+ fmt = GET_RTX_FORMAT (GET_CODE (x));
- if (gcse_file)
+ for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
{
- 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");
+ if (fmt[i] == 'e')
+ invalidate_any_buried_refs (XEXP (x, i));
+ else if (fmt[i] == 'E')
+ for (j = XVECLEN (x, i) - 1; j >= 0; j--)
+ invalidate_any_buried_refs (XVECEXP (x, i, j));
}
}
-/* 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. */
+/* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
+ being defined as MEM loads and stores to symbols, with no side effects
+ and no registers in the expression. For a MEM destination, we also
+ check that the insn is still valid if we replace the destination with a
+ REG, as is done in update_ld_motion_stores. If there are any uses/defs
+ which don't match this criteria, they are invalidated and trimmed out
+ later. */
-static int
-insert_store (struct ls_expr * expr, edge e)
+static void
+compute_ld_motion_mems (void)
{
- rtx reg, insn;
+ struct ls_expr * ptr;
basic_block bb;
- edge tmp;
-
- /* 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;
-
- reg = expr->reaching_reg;
- insn = gen_move_insn (copy_rtx (expr->pattern), reg);
+ rtx insn;
- /* If we are inserting this expression on ALL predecessor edges of a BB,
- insert it at the start of the BB, and reset the insert bits on the other
- edges so we don't try to insert it on the other edges. */
- bb = e->dest;
- for (tmp = e->dest->pred; tmp ; tmp = tmp->pred_next)
- {
- int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
- if (index == EDGE_INDEX_NO_EDGE)
- abort ();
- if (! TEST_BIT (pre_insert_map[index], expr->index))
- break;
- }
+ pre_ldst_mems = NULL;
+ pre_ldst_table = htab_create (13, pre_ldst_expr_hash,
+ pre_ldst_expr_eq, NULL);
- /* 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_BB (bb)
{
- for (tmp = e->dest->pred; tmp ; tmp = tmp->pred_next)
+ FOR_BB_INSNS (bb, insn)
{
- 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;
- }
+ if (NONDEBUG_INSN_P (insn))
+ {
+ if (GET_CODE (PATTERN (insn)) == SET)
+ {
+ rtx src = SET_SRC (PATTERN (insn));
+ rtx dest = SET_DEST (PATTERN (insn));
- /* We can't insert on this edge, so we'll insert at the head of the
- successors block. See Morgan, sec 10.5. */
- if ((e->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
- {
- insert_insn_start_bb (insn, bb);
- return 0;
- }
+ /* Check for a simple LOAD... */
+ if (MEM_P (src) && simple_mem (src))
+ {
+ ptr = ldst_entry (src);
+ if (REG_P (dest))
+ ptr->loads = alloc_INSN_LIST (insn, ptr->loads);
+ else
+ ptr->invalid = 1;
+ }
+ else
+ {
+ /* Make sure there isn't a buried load somewhere. */
+ invalidate_any_buried_refs (src);
+ }
- insert_insn_on_edge (insn, e);
+ /* Check for stores. Don't worry about aliased ones, they
+ will block any movement we might do later. We only care
+ about this exact pattern since those are the only
+ circumstance that we will ignore the aliasing info. */
+ if (MEM_P (dest) && simple_mem (dest))
+ {
+ ptr = ldst_entry (dest);
- 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");
+ if (! MEM_P (src)
+ && GET_CODE (src) != ASM_OPERANDS
+ /* Check for REG manually since want_to_gcse_p
+ returns 0 for all REGs. */
+ && can_assign_to_reg_without_clobbers_p (src))
+ ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
+ else
+ ptr->invalid = 1;
+ }
+ }
+ else
+ invalidate_any_buried_refs (PATTERN (insn));
+ }
+ }
}
-
- return 1;
}
-/* This routine will replace a store with a SET to a specified register. */
+/* Remove any references that have been either invalidated or are not in the
+ expression list for pre gcse. */
static void
-replace_store_insn (rtx reg, rtx del, basic_block bb)
+trim_ld_motion_mems (void)
{
- rtx insn;
-
- insn = gen_move_insn (reg, SET_SRC (single_set (del)));
- insn = emit_insn_after (insn, del);
+ struct ls_expr * * last = & pre_ldst_mems;
+ struct ls_expr * ptr = pre_ldst_mems;
- if (gcse_file)
+ while (ptr != NULL)
{
- 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");
+ struct expr * expr;
+
+ /* Delete if entry has been made invalid. */
+ if (! ptr->invalid)
+ {
+ /* Delete if we cannot find this mem in the expression list. */
+ unsigned int hash = ptr->hash_index % expr_hash_table.size;
+
+ for (expr = expr_hash_table.table[hash];
+ expr != NULL;
+ expr = expr->next_same_hash)
+ if (expr_equiv_p (expr->expr, ptr->pattern))
+ break;
+ }
+ else
+ expr = (struct expr *) 0;
+
+ if (expr)
+ {
+ /* Set the expression field if we are keeping it. */
+ ptr->expr = expr;
+ last = & ptr->next;
+ ptr = ptr->next;
+ }
+ else
+ {
+ *last = ptr->next;
+ htab_remove_elt_with_hash (pre_ldst_table, ptr, ptr->hash_index);
+ free_ldst_entry (ptr);
+ ptr = * last;
+ }
}
- delete_insn (del);
+ /* Show the world what we've found. */
+ if (dump_file && pre_ldst_mems != NULL)
+ print_ldst_list (dump_file);
}
-
-/* Delete a store, but copy the value that would have been stored into
- the reaching_reg for later storing. */
+/* This routine will take an expression which we are replacing with
+ a reaching register, and update any stores that are needed if
+ that expression is in the ld_motion list. Stores are updated by
+ copying their SRC to the reaching register, and then storing
+ the reaching register into the store location. These keeps the
+ correct value in the reaching register for the loads. */
static void
-delete_store (struct ls_expr * expr, basic_block bb)
+update_ld_motion_stores (struct expr * expr)
{
- rtx reg, i, del;
+ struct ls_expr * mem_ptr;
- if (expr->reaching_reg == NULL_RTX)
- expr->reaching_reg = gen_reg_rtx (GET_MODE (expr->pattern));
+ if ((mem_ptr = find_rtx_in_ldst (expr->expr)))
+ {
+ /* We can try to find just the REACHED stores, but is shouldn't
+ matter to set the reaching reg everywhere... some might be
+ dead and should be eliminated later. */
- reg = expr->reaching_reg;
+ /* We replace (set mem expr) with (set reg expr) (set mem reg)
+ where reg is the reaching reg used in the load. We checked in
+ compute_ld_motion_mems that we can replace (set mem expr) with
+ (set reg expr) in that insn. */
+ rtx list = mem_ptr->stores;
- for (i = AVAIL_STORE_LIST (expr); i; i = XEXP (i, 1))
- {
- del = XEXP (i, 0);
- if (BLOCK_FOR_INSN (del) == bb)
+ for ( ; list != NULL_RTX; list = XEXP (list, 1))
{
- /* We know there is only one since we deleted redundant
- ones during the available computation. */
- replace_store_insn (reg, del, bb);
- break;
- }
- }
-}
+ rtx insn = XEXP (list, 0);
+ rtx pat = PATTERN (insn);
+ rtx src = SET_SRC (pat);
+ rtx reg = expr->reaching_reg;
+ rtx copy;
-/* Free memory used by store motion. */
+ /* If we've already copied it, continue. */
+ if (expr->reaching_reg == src)
+ continue;
-static void
-free_store_memory (void)
-{
- free_ldst_mems ();
+ if (dump_file)
+ {
+ 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");
+ }
- 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);
+ copy = gen_move_insn (reg, copy_rtx (SET_SRC (pat)));
+ emit_insn_before (copy, insn);
+ SET_SRC (pat) = reg;
+ df_insn_rescan (insn);
- ae_gen = ae_kill = transp = st_antloc = NULL;
- pre_insert_map = pre_delete_map = reg_set_in_block = NULL;
+ /* un-recognize this pattern since it's probably different now. */
+ INSN_CODE (insn) = -1;
+ gcse_create_count++;
+ }
+ }
}
+\f
+/* Return true if the graph is too expensive to optimize. PASS is the
+ optimization about to be performed. */
-/* 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)
+static bool
+is_too_expensive (const char *pass)
{
- basic_block bb;
- int x;
- struct ls_expr * ptr;
- int update_flow = 0;
+ /* Trying to perform global optimizations on flow graphs which have
+ a high connectivity will take a long time and is unlikely to be
+ particularly useful.
- if (gcse_file)
+ 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)
{
- fprintf (gcse_file, "before store motion\n");
- print_rtl (gcse_file, get_insns ());
- }
+ warning (OPT_Wdisabled_optimization,
+ "%s: %d basic blocks and %d edges/basic block",
+ pass, n_basic_blocks, n_edges / n_basic_blocks);
- 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;
+ return true;
}
- /* Now compute kill & transp vectors. */
- build_store_vectors ();
- add_noreturn_fake_exit_edges ();
-
- 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 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)
{
- FOR_EACH_BB (bb)
- if (TEST_BIT (pre_delete_map[bb->index], ptr->index))
- delete_store (ptr, bb);
+ warning (OPT_Wdisabled_optimization,
+ "%s: %d basic blocks and %d registers",
+ pass, n_basic_blocks, max_reg_num ());
- 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));
+ return true;
}
- if (update_flow)
- commit_edge_insertions ();
-
- free_store_memory ();
- free_edge_list (edge_list);
- remove_fake_edges ();
- end_alias_analysis ();
+ return false;
}
\f
-/* Entry point for jump bypassing optimization pass. */
+/* Main function for the CPROP pass. */
-int
-bypass_jumps (FILE *file)
+static int
+one_cprop_pass (void)
{
- int changed;
+ int changed = 0;
- /* We do not construct an accurate cfg in functions which call
- setjmp, so just punt to be safe. */
- if (current_function_calls_setjmp)
+ /* 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;
- /* 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 ();
+ global_const_prop_count = local_const_prop_count = 0;
+ global_copy_prop_count = local_copy_prop_count = 0;
- if (file)
- dump_flow_info (file);
+ bytes_used = 0;
+ gcc_obstack_init (&gcse_obstack);
+ alloc_gcse_mem ();
- /* Return if there's nothing to do. */
- if (n_basic_blocks <= 1)
- return 0;
+ /* 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.
- /* 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.
+ 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.
- In normal circumstances a cfg should have about twice as many edges
- as blocks. But we do not want to punish small functions which have
- a couple switch statements. So we require a relatively large number
- of basic blocks and the ratio of edges to blocks to be high. */
- if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
+ 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 ())
{
- if (warn_disabled_optimization)
- warning ("BYPASS disabled: %d > 1000 basic blocks and %d >= 20 edges/basic block",
- n_basic_blocks, n_edges / n_basic_blocks);
- return 0;
+ delete_unreachable_blocks ();
+ df_analyze ();
}
- /* If allocating memory for the cprop bitmap would take up too much
- storage it's better just to disable the optimization. */
- if ((n_basic_blocks
- * SBITMAP_SET_SIZE (max_gcse_regno)
- * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
- {
- if (warn_disabled_optimization)
- warning ("GCSE disabled: %d basic blocks and %d registers",
- n_basic_blocks, max_gcse_regno);
+ /* Determine implicit sets. */
+ implicit_sets = XCNEWVEC (rtx, last_basic_block);
+ find_implicit_sets ();
- return 0;
- }
+ alloc_hash_table (&set_hash_table, 1);
+ compute_hash_table (&set_hash_table);
- gcc_obstack_init (&gcse_obstack);
- bytes_used = 0;
+ /* Free implicit_sets before peak usage. */
+ free (implicit_sets);
+ implicit_sets = NULL;
- /* We need alias. */
- init_alias_analysis ();
+ if (dump_file)
+ dump_hash_table (dump_file, "SET", &set_hash_table);
+ if (set_hash_table.n_elems > 0)
+ {
+ basic_block bb;
+ rtx insn;
+
+ alloc_cprop_mem (last_basic_block, set_hash_table.n_elems);
+ compute_cprop_data ();
- /* 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.
+ 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 ();
- 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. */
+ 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);
+ }
+ }
- alloc_reg_set_mem (max_gcse_regno);
- compute_sets (get_insns ());
+ changed |= bypass_conditional_jumps ();
+ free_cprop_mem ();
+ }
- max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (get_insns ());
- changed = one_cprop_pass (1, 1, 1);
+ free_hash_table (&set_hash_table);
free_gcse_mem ();
+ obstack_free (&gcse_obstack, NULL);
- if (file)
+ if (dump_file)
{
- fprintf (file, "BYPASS of %s: %d basic blocks, ",
- current_function_name, n_basic_blocks);
- fprintf (file, "%d bytes\n\n", bytes_used);
+ 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);
}
- obstack_free (&gcse_obstack, NULL);
- free_reg_set_mem ();
+ return changed;
+}
- /* We are finished with alias. */
- end_alias_analysis ();
- allocate_reg_info (max_reg_num (), FALSE, FALSE);
+\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.
- return changed;
+ 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
+gate_rtl_cprop (void)
+{
+ return optimize > 0 && flag_gcse
+ && !cfun->calls_setjmp
+ && dbg_cnt (cprop);
+}
+
+static unsigned int
+execute_rtl_cprop (void)
+{
+ delete_unreachable_blocks ();
+ df_set_flags (DF_LR_RUN_DCE);
+ df_analyze ();
+ flag_rerun_cse_after_global_opts |= one_cprop_pass ();
+ return 0;
+}
+
+static bool
+gate_rtl_pre (void)
+{
+ return optimize > 0 && flag_gcse
+ && !cfun->calls_setjmp
+ && optimize_function_for_speed_p (cfun)
+ && dbg_cnt (pre);
+}
+
+static unsigned int
+execute_rtl_pre (void)
+{
+ delete_unreachable_blocks ();
+ df_analyze ();
+ flag_rerun_cse_after_global_opts |= one_pre_gcse_pass ();
+ return 0;
+}
+
+static bool
+gate_rtl_hoist (void)
+{
+ 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);
+}
+
+static unsigned int
+execute_rtl_hoist (void)
+{
+ delete_unreachable_blocks ();
+ df_analyze ();
+ flag_rerun_cse_after_global_opts |= one_code_hoisting_pass ();
+ return 0;
}
+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 */
+ }
+};
+
+struct rtl_opt_pass pass_rtl_pre =
+{
+ {
+ RTL_PASS,
+ "rtl 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 */
+ }
+};
+
+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
