Subversion Repositories openrisc_2011-10-31

/* Global common subexpression elimination/Partial redundancy elimination

   and global constant/copy propagation for GNU compiler.

   Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2007

   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 3, or (at your option) any later

version.

GCC is distributed in the hope that it will be useful, but WITHOUT ANY

WARRANTY; without even the implied warranty of MERCHANTABILITY or

FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License

for more details.

You should have received a copy of the GNU General Public License

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

   - do rough calc of how many regs are needed in each block, and a rough

     calc of how many regs are available in each class and use that to

     throttle back the code in cases where RTX_COST is minimal.

   - 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

*/

/* References searched while implementing this.

   Compilers Principles, Techniques and Tools

   Aho, Sethi, Ullman

   Addison-Wesley, 1988

   Global Optimization by Suppression of Partial Redundancies

   E. Morel, C. Renvoise

   communications of the acm, Vol. 22, Num. 2, Feb. 1979

   A Portable Machine-Independent Global Optimizer - Design and Measurements

   Frederick Chow

   Stanford Ph.D. thesis, Dec. 1983

   A Fast Algorithm for Code Movement Optimization

   D.M. Dhamdhere

   SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988

   A Solution to a Problem with Morel and Renvoise's

   Global Optimization by Suppression of Partial Redundancies

   K-H Drechsler, M.P. Stadel

   ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988

   Practical Adaptation of the Global Optimization

   Algorithm of Morel and Renvoise

   D.M. Dhamdhere

   ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991

   Efficiently Computing Static Single Assignment Form and the Control

   Dependence Graph

   R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck

   ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991

   Lazy Code Motion

   J. Knoop, O. Ruthing, B. Steffen

   ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI

   What's In a Region?  Or Computing Control Dependence Regions in Near-Linear

   Time for Reducible Flow Control

   Thomas Ball

   ACM Letters on Programming Languages and Systems,

   Vol. 2, Num. 1-4, Mar-Dec 1993

   An Efficient Representation for Sparse Sets

   Preston Briggs, Linda Torczon

   ACM Letters on Programming Languages and Systems,

   Vol. 2, Num. 1-4, Mar-Dec 1993

   A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion

   K-H Drechsler, M.P. Stadel

   ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993

   Partial Dead Code Elimination

   J. Knoop, O. Ruthing, B. Steffen

   ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994

   Effective Partial Redundancy Elimination

   P. Briggs, K.D. Cooper

   ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994

   The Program Structure Tree: Computing Control Regions in Linear Time

   R. Johnson, D. Pearson, K. Pingali

   ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994

   Optimal Code Motion: Theory and Practice

   J. Knoop, O. Ruthing, B. Steffen

   ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994

   The power of assignment motion

   J. Knoop, O. Ruthing, B. Steffen

   ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI

   Global code motion / global value numbering

   C. Click

   ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI

   Value Driven Redundancy Elimination

   L.T. Simpson

   Rice University Ph.D. thesis, Apr. 1996

   Value Numbering

   L.T. Simpson

   Massively Scalar Compiler Project, Rice University, Sep. 1996

   High Performance Compilers for Parallel Computing

   Michael Wolfe

   Addison-Wesley, 1996

   Advanced Compiler Design and Implementation

   Steven Muchnick

   Morgan Kaufmann, 1997

   Building an Optimizing Compiler

   Robert Morgan

   Digital Press, 1998

   People wishing to speed up the code here should read:

     Elimination Algorithms for Data Flow Analysis

     B.G. Ryder, M.C. Paull

     ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986

     How to Analyze Large Programs Efficiently and Informatively

     D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck

     ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI

   People wishing to do something different can find various possibilities

   in the above papers and elsewhere.

*/

#include "config.h"

#include "system.h"

#include "coretypes.h"

#include "tm.h"

#include "toplev.h"

#include "rtl.h"

#include "tree.h"

#include "tm_p.h"

#include "regs.h"

#include "hard-reg-set.h"

#include "flags.h"

#include "real.h"

#include "insn-config.h"

#include "recog.h"

#include "basic-block.h"

#include "output.h"

#include "function.h"

#include "expr.h"

#include "except.h"

#include "ggc.h"

#include "params.h"

#include "cselib.h"

#include "intl.h"

#include "obstack.h"

#include "timevar.h"

#include "tree-pass.h"

#include "hashtab.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.  */

/* We support GCSE via Partial Redundancy Elimination.  PRE optimizations

   are a superset of those done by GCSE.

   We perform the following steps:

   1) Compute basic block information.

   2) Compute table of places where registers are set.

   3) Perform copy/constant propagation.

   4) Perform global cse using lazy code motion if not optimizing

      for size, or code hoisting if we are.

   5) Perform another pass of copy/constant propagation.

   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

   GCSE creates.  This is needed more for PRE than for Classic because Classic

   GCSE will try to use an existing register containing the common

   subexpression rather than create a new one.  This is harder to do for PRE

   because of the code motion (which Classic GCSE doesn't do).

   Expressions we are interested in GCSE-ing are of the form

   (set (pseudo-reg) (expression)).

   Function want_to_gcse_p says what these are.

   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

   doing so.  The first pass usually makes 90% of the changes that are doable.

   A second pass can make a few more changes made possible by the first pass.

   Experiments show any further passes don't make enough changes to justify

   the expense.

   A study of spec92 using an unlimited number of passes:

   [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,

   [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,

   [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1

   It was found doing copy propagation between each pass enables further

   substitutions.

   PRE is quite expensive in complicated functions because the DFA can take

   a while to converge.  Hence we only perform one pass.  The parameter

   max-gcse-passes can be modified if one wants to experiment.

   **********************

   The steps for PRE are:

   1) Build the hash table of expressions we wish to GCSE (expr_hash_table).

   2) Perform the data flow analysis for PRE.

   3) Delete the redundant instructions

   4) Insert the required copies [if any] that make the partially

      redundant instructions fully redundant.

   5) For other reaching expressions, insert an instruction to copy the value

      to a newly created pseudo that will reach the redundant instruction.

   The deletion is done first so that when we do insertions we

   know which pseudo reg to use.

   Various papers have argued that PRE DFA is expensive (O(n^2)) and others

   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

   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.  */

/* GCSE global vars.  */

/* 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;

/* An obstack for our working variables.  */

static struct obstack gcse_obstack;

struct reg_use {rtx reg_rtx; };

/* Hash table of expressions.  */

struct expr

{

  /* The expression (SET_SRC for expressions, PATTERN for assignments).  */

  rtx expr;

  /* Index in the available expression bitmaps.  */

  int bitmap_index;

  /* Next entry with the same hash.  */

  struct expr *next_same_hash;

  /* List of anticipatable occurrences in basic blocks in the function.

     An "anticipatable occurrence" is one that is the first occurrence in the

     basic block, the operands are not modified in the basic block prior

     to the occurrence and the output is not used between the start of

     the block and the occurrence.  */

  struct occr *antic_occr;

  /* List of available occurrence in basic blocks in the function.

     An "available occurrence" is one that is the last occurrence in the

     basic block and the operands are not modified by following statements in

     the basic block [including this insn].  */

  struct occr *avail_occr;

  /* Non-null if the computation is PRE redundant.

     The value is the newly created pseudo-reg to record a copy of the

     expression in all the places that reach the redundant copy.  */

  rtx reaching_reg;

};

/* Occurrence of an expression.

   There is one per basic block.  If a pattern appears more than once the

   last appearance is used [or first for anticipatable expressions].  */

struct occr

{

  /* Next occurrence of this expression.  */

  struct occr *next;

  /* The insn that computes the expression.  */

  rtx insn;

  /* Nonzero if this [anticipatable] occurrence has been deleted.  */

  char deleted_p;

  /* Nonzero if this [available] occurrence has been copied to

     reaching_reg.  */

  /* ??? This is mutually exclusive with deleted_p, so they could share

     the same byte.  */

  char copied_p;

};

/* Expression and copy propagation hash tables.

   Each hash table is an array of buckets.

   ??? It is known that if it were an array of entries, structure elements

   `next_same_hash' and `bitmap_index' wouldn't be necessary.  However, it is

   not clear whether in the final analysis a sufficient amount of memory would

   be saved as the size of the available expression bitmaps would be larger

   [one could build a mapping table without holes afterwards though].

   Someday I'll perform the computation and figure it out.  */

struct hash_table

{

  /* The table itself.

     This is an array of `expr_hash_table_size' elements.  */

  struct expr **table;

  /* Size of the hash table, in elements.  */

  unsigned int size;

  /* Number of hash table elements.  */

  unsigned int n_elems;

  /* Whether the table is expression of copy propagation one.  */

  int set_p;

};

/* Expression hash table.  */

static struct hash_table expr_hash_table;

/* Copy propagation hash table.  */

static struct hash_table set_hash_table;

/* Mapping of uids to cuids.

   Only real insns get cuids.  */

static int *uid_cuid;

/* Highest UID in UID_CUID.  */

static int max_uid;

/* Get the cuid of an insn.  */

#ifdef ENABLE_CHECKING

#define INSN_CUID(INSN) \

  (gcc_assert (INSN_UID (INSN) <= max_uid), uid_cuid[INSN_UID (INSN)])

#else

#define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])

#endif

/* Number of cuids.  */

static int max_cuid;

/* Mapping of cuids to insns.  */

static rtx *cuid_insn;

/* Get insn from cuid.  */

#define CUID_INSN(CUID) (cuid_insn[CUID])

/* Maximum register number in function prior to doing gcse + 1.

   Registers created during this pass have regno >= max_gcse_regno.

   This is named with "gcse" to not collide with global of same name.  */

static unsigned int max_gcse_regno;

/* Table of registers that are modified.

   For each register, each element is a list of places where the pseudo-reg

   is set.

   For simplicity, GCSE is done on sets of pseudo-regs only.  PRE GCSE only

   requires knowledge of which blocks kill which regs [and thus could use

   a bitmap instead of the lists `reg_set_table' uses].

   `reg_set_table' and could be turned into an array of bitmaps (num-bbs x

   num-regs) [however perhaps it may be useful to keep the data as is].  One

   advantage of recording things this way is that `reg_set_table' is fairly

   sparse with respect to pseudo regs but for hard regs could be fairly dense

   [relatively speaking].  And recording sets of pseudo-regs in lists speeds

   up functions like compute_transp since in the case of pseudo-regs we only

   need to iterate over the number of times a pseudo-reg is set, not over the

   number of basic blocks [clearly there is a bit of a slow down in the cases

   where a pseudo is set more than once in a block, however it is believed

   that the net effect is to speed things up].  This isn't done for hard-regs

   because recording call-clobbered hard-regs in `reg_set_table' at each

   function call can consume a fair bit of memory, and iterating over

   hard-regs stored this way in compute_transp will be more expensive.  */

typedef struct reg_set

{

  /* The next setting of this register.  */

  struct reg_set *next;

  /* The index of the block where it was set.  */

  int bb_index;

} 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

/* 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. (i.e., loads and stores to a single location).

   We can then allow movement of these MEM refs with a little special

   allowance. (all stores copy the same value to the reaching reg used

   for the loads).  This means all values used to store into memory must have

   no side effects so we can re-issue the setter value.

   Store Motion uses this structure as an expression table to track stores

   which look interesting, and might be moveable towards the exit block.  */

struct ls_expr

{

  struct expr * expr;           /* Gcse expression reference for LM.  */

  rtx pattern;                  /* Pattern of this mem.  */

  rtx pattern_regs;             /* List of registers mentioned by the mem.  */

  rtx loads;                    /* INSN list of loads seen.  */

  rtx stores;                   /* INSN list of stores seen.  */

  struct ls_expr * next;        /* Next in the list.  */

  int invalid;                  /* Invalid for some reason.  */

  int index;                    /* If it maps to a bitmap index.  */

  unsigned int hash_index;      /* Index when in a hash table.  */

  rtx reaching_reg;             /* Register to use when re-writing.  */

};

/* Array of implicit set patterns indexed by basic block index.  */

static rtx *implicit_sets;

/* Head of the list of load/store memory refs.  */

static struct ls_expr * pre_ldst_mems = NULL;

/* Hashtable for the load/store memory refs.  */

static htab_t pre_ldst_table = NULL;

/* Bitmap containing one bit for each register in the program.

   Used when performing GCSE to track which registers have been set since

   the start of the basic block.  */

static regset reg_set_bitmap;

/* For each block, a bitmap of registers set in the block.

   This is used by 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;

static bitmap modify_mem_list_set;

/* This array parallels modify_mem_list, but is kept canonicalized.  */

static rtx * canon_modify_mem_list;

/* 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.

   This isn't intended to be absolutely precise.  Its intent is only

   to keep an eye on memory usage.  */

static int bytes_used;

/* GCSE substitutions made.  */

static int gcse_subst_count;

/* Number of copy instructions created.  */

static int gcse_create_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;

/* For available exprs */

static sbitmap *ae_kill, *ae_gen;

static void compute_can_copy (void);

static void *gmalloc (size_t) ATTRIBUTE_MALLOC;

static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC;

static void *grealloc (void *, size_t);

static void *gcse_alloc (unsigned long);

static void alloc_gcse_mem (void);

static void free_gcse_mem (void);

static void alloc_reg_set_mem (int);

static void free_reg_set_mem (void);

static void record_one_set (int, rtx);

static void record_set_info (rtx, rtx, void *);

static void compute_sets (void);

static void hash_scan_insn (rtx, struct hash_table *, int);

static void hash_scan_set (rtx, rtx, struct hash_table *);

static void hash_scan_clobber (rtx, rtx, struct hash_table *);

static void hash_scan_call (rtx, rtx, struct hash_table *);

static int want_to_gcse_p (rtx);

static bool can_assign_to_reg_p (rtx);

static bool gcse_constant_p (rtx);

static int oprs_unchanged_p (rtx, rtx, int);

static int oprs_anticipatable_p (rtx, rtx);

static int oprs_available_p (rtx, rtx);

static void insert_expr_in_table (rtx, enum machine_mode, rtx, int, int,

                                  struct hash_table *);

static void insert_set_in_table (rtx, rtx, struct hash_table *);

static unsigned int hash_expr (rtx, enum machine_mode, int *, int);

static unsigned int hash_set (int, int);

static int expr_equiv_p (rtx, rtx);

static void record_last_reg_set_info (rtx, int);

static void record_last_mem_set_info (rtx);

static void record_last_set_info (rtx, rtx, void *);

static void compute_hash_table (struct hash_table *);

static void alloc_hash_table (int, struct hash_table *, int);

static void free_hash_table (struct hash_table *);

static void compute_hash_table_work (struct hash_table *);

static void dump_hash_table (FILE *, const char *, struct hash_table *);

static struct expr *lookup_set (unsigned int, struct hash_table *);

static struct expr *next_set (unsigned int, struct expr *);

static void reset_opr_set_tables (void);

static int oprs_not_set_p (rtx, rtx);

static void mark_call (rtx);

static void mark_set (rtx, rtx);

static void mark_clobber (rtx, rtx);

static void mark_oprs_set (rtx);

static void alloc_cprop_mem (int, int);

static void free_cprop_mem (void);

static void compute_transp (rtx, int, sbitmap *, int);

static void compute_transpout (void);

static void compute_local_properties (sbitmap *, sbitmap *, sbitmap *,

                                      struct hash_table *);

static void compute_cprop_data (void);

static void find_used_regs (rtx *, void *);

static int try_replace_reg (rtx, rtx, rtx);

static struct expr *find_avail_set (int, rtx);

static int cprop_jump (basic_block, rtx, rtx, rtx, rtx);

static void mems_conflict_for_gcse_p (rtx, rtx, void *);

static int load_killed_in_block_p (basic_block, int, rtx, int);

static void canon_list_insert (rtx, rtx, void *);

static int cprop_insn (rtx, int);

static int cprop (int);

static void find_implicit_sets (void);

static int one_cprop_pass (int, bool, bool);

static bool constprop_register (rtx, rtx, rtx, bool);

static struct expr *find_bypass_set (int, int);

static bool reg_killed_on_edge (rtx, edge);

static int bypass_block (basic_block, rtx, rtx);

static int bypass_conditional_jumps (void);

static void alloc_pre_mem (int, int);

static void free_pre_mem (void);

static void compute_pre_data (void);

static int pre_expr_reaches_here_p (basic_block, struct expr *,

                                    basic_block);

static void insert_insn_end_bb (struct expr *, basic_block, int);

static void pre_insert_copy_insn (struct expr *, rtx);

static void pre_insert_copies (void);

static int pre_delete (void);

static int pre_gcse (void);

static int one_pre_gcse_pass (int);

static void add_label_notes (rtx, rtx);

static void alloc_code_hoist_mem (int, int);

static void free_code_hoist_mem (void);

static void compute_code_hoist_vbeinout (void);

static void compute_code_hoist_data (void);

static int hoist_expr_reaches_here_p (basic_block, int, basic_block, char *);

static void hoist_code (void);

static int one_code_hoisting_pass (void);

static rtx process_insert_insn (struct expr *);

static int pre_edge_insert (struct edge_list *, struct expr **);

static int pre_expr_reaches_here_p_work (basic_block, struct expr *,

                                         basic_block, char *);

static struct ls_expr * ldst_entry (rtx);

static void free_ldst_entry (struct ls_expr *);

static void free_ldst_mems (void);

static void print_ldst_list (FILE *);

static struct ls_expr * find_rtx_in_ldst (rtx);

static int enumerate_ldsts (void);

static inline struct ls_expr * first_ls_expr (void);

static inline struct ls_expr * next_ls_expr (struct ls_expr *);

static int simple_mem (rtx);

static void invalidate_any_buried_refs (rtx);

static void compute_ld_motion_mems (void);

static void trim_ld_motion_mems (void);

static void update_ld_motion_stores (struct expr *);

static void reg_set_info (rtx, rtx, void *);

static void reg_clear_last_set (rtx, rtx, void *);

static bool store_ops_ok (rtx, int *);

static rtx extract_mentioned_regs (rtx);

static rtx extract_mentioned_regs_helper (rtx, rtx);

static void find_moveable_store (rtx, int *, int *);

static int compute_store_table (void);

static bool load_kills_store (rtx, rtx, int);

static bool find_loads (rtx, rtx, int);

static bool store_killed_in_insn (rtx, rtx, rtx, int);

static bool store_killed_after (rtx, rtx, rtx, basic_block, int *, rtx *);

static bool store_killed_before (rtx, rtx, rtx, basic_block, int *);

static void build_store_vectors (void);

static void insert_insn_start_bb (rtx, basic_block);

static int insert_store (struct ls_expr *, edge);

static void remove_reachable_equiv_notes (basic_block, struct ls_expr *);

static void replace_store_insn (rtx, rtx, basic_block, struct ls_expr *);

static void delete_store (struct ls_expr *, basic_block);

static void free_store_memory (void);

static void store_motion (void);

static void free_insn_expr_list_list (rtx *);

static void clear_modify_mem_tables (void);

static void free_modify_mem_tables (void);

static rtx gcse_emit_move_after (rtx, rtx, rtx);

static void local_cprop_find_used_regs (rtx *, void *);

static bool do_local_cprop (rtx, rtx, bool, rtx*);

static bool adjust_libcall_notes (rtx, rtx, rtx, rtx*);

static void local_cprop_pass (bool);

static bool is_too_expensive (const char *);

/* Entry point for global common subexpression elimination.

   F is the first instruction in the function.  Return nonzero if a

   change is mode.  */

static int

gcse_main (rtx f ATTRIBUTE_UNUSED)

{

  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;

  /* Identify the basic block information for this function, including

     successors and predecessors.  */

  max_gcse_regno = max_reg_num ();

  if (dump_file)

    dump_flow_info (dump_file, dump_flags);

  /* Return if there's nothing to do, or it is too expensive.  */

  if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1

      || is_too_expensive (_("GCSE disabled")))

    return 0;

  gcc_obstack_init (&gcse_obstack);

  bytes_used = 0;

  /* We need alias.  */

  init_alias_analysis ();

  /* Record where pseudo-registers are set.  This data is kept accurate

     during each pass.  ??? We could also record hard-reg information here

     [since it's unchanging], however it is currently done during hash table

     computation.

     It may be tempting to compute MEM set information here too, but MEM sets

     will be subject to code motion one day and thus we need to compute

     information about memory sets when we build the hash tables.  */

  alloc_reg_set_mem (max_gcse_regno);

  compute_sets ();

  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 (dump_file)

        fprintf (dump_file, "GCSE pass %d\n\n", pass + 1);

      /* Initialize bytes_used to the space for the pred/succ lists,

         and the reg_set_table data.  */

      bytes_used = initial_bytes_used;

      /* Each pass may create new registers, so recalculate each time.  */

      max_gcse_regno = max_reg_num ();

      alloc_gcse_mem ();

      /* Don't allow constant propagation to modify jumps

         during this pass.  */

      timevar_push (TV_CPROP1);

      changed = one_cprop_pass (pass + 1, false, false);

      timevar_pop (TV_CPROP1);

      if (optimize_size)

        /* Do nothing.  */ ;

      else

        {

          timevar_push (TV_PRE);

          changed |= one_pre_gcse_pass (pass + 1);

          /* We may have just created new basic blocks.  Release and

             recompute various things which are sized on the number of

             basic blocks.  */

          if (changed)

            {

              free_modify_mem_tables ();

              modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));

              canon_modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));

            }