/* 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
+ Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005
Free Software Foundation, Inc.
This file is part of GCC.
#include "cselib.h"
#include "intl.h"
#include "obstack.h"
+#include "timevar.h"
/* Propagate flow information through back edges and thus enable PRE's
moving loop invariant calculations out of loops.
3) Perform copy/constant propagation.
- 4) Perform global cse.
+ 4) Perform global cse using lazy code motion if not optimizing
+ for size, or code hoisting if we are.
5) Perform another pass of copy/constant propagation.
substitutions.
PRE is quite expensive in complicated functions because the DFA can take
- awhile to converge. Hence we only perform one pass. The parameter max-gcse-passes can
- be modified if one wants to experiment.
+ a while to converge. Hence we only perform one pass. The parameter
+ max-gcse-passes can be modified if one wants to experiment.
**********************
the result of the expression is copied to a new register, and the redundant
expression is deleted by replacing it with this new register. Classic GCSE
doesn't have this problem as much as it computes the reaching defs of
- each register in each block and thus can try to use an existing register.
-
- **********************
-
- A fair bit of simplicity is created by creating small functions for simple
- tasks, even when the function is only called in one place. This may
- measurably slow things down [or may not] by creating more function call
- overhead than is necessary. The source is laid out so that it's trivial
- to make the affected functions inline so that one can measure what speed
- up, if any, can be achieved, and maybe later when things settle things can
- be rearranged.
-
- Help stamp out big monolithic functions! */
+ each register in each block and thus can try to use an existing
+ register. */
\f
/* GCSE global vars. */
* If we changed any jumps via cprop.
* If we added any labels via edge splitting. */
-
static int run_jump_opt_after_gcse;
/* Bitmaps are normally not included in debugging dumps.
/* Get the cuid of an insn. */
#ifdef ENABLE_CHECKING
-#define INSN_CUID(INSN) (INSN_UID (INSN) > max_uid ? (abort (), 0) : uid_cuid[INSN_UID (INSN)])
+#define INSN_CUID(INSN) \
+ (gcc_assert (INSN_UID (INSN) <= max_uid), uid_cuid[INSN_UID (INSN)])
#else
#define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
#endif
{
/* The next setting of this register. */
struct reg_set *next;
- /* The insn where it was set. */
- rtx insn;
+ /* The index of the block where it was set. */
+ int bb_index;
} reg_set;
static reg_set **reg_set_table;
/* This is a list of expressions which are MEMs and will be used by load
or store motion.
Load motion tracks MEMs which aren't killed by
- anything except itself. (ie, loads and stores to a single location).
+ anything except itself. (i.e., loads and stores to a single location).
We can then allow movement of these MEM refs with a little special
allowance. (all stores copy the same value to the reaching reg used
for the loads). This means all values used to store into memory must have
static regset reg_set_bitmap;
/* For each block, a bitmap of registers set in the block.
- This is used by expr_killed_p and compute_transp.
+ This is used by compute_transp.
It is computed during hash table computation and not by compute_sets
as it includes registers added since the last pass (or between cprop and
gcse) and it's currently not easy to realloc sbitmap vectors. */
/* Array, indexed by basic block number for a list of insns which modify
memory within that block. */
static rtx * modify_mem_list;
-bitmap modify_mem_list_set;
+static bitmap modify_mem_list_set;
/* This array parallels modify_mem_list, but is kept canonicalized. */
static rtx * canon_modify_mem_list;
-bitmap canon_modify_mem_list_set;
+
+/* Bitmap indexed by block numbers to record which blocks contain
+ function calls. */
+static bitmap blocks_with_calls;
+
/* Various variables for statistics gathering. */
/* Memory used in a pass.
static int gcse_subst_count;
/* Number of copy instructions created. */
static int gcse_create_count;
-/* Number of constants propagated. */
-static int const_prop_count;
-/* Number of copys propagated. */
-static int copy_prop_count;
+/* Number of local constants propagated. */
+static int local_const_prop_count;
+/* Number of local copys propagated. */
+static int local_copy_prop_count;
+/* Number of global constants propagated. */
+static int global_const_prop_count;
+/* Number of global copys propagated. */
+static int global_copy_prop_count;
\f
-/* These variables are used by classic GCSE.
- Normally they'd be defined a bit later, but `rd_gen' needs to
- be declared sooner. */
-
-/* Each block has a bitmap of each type.
- The length of each blocks bitmap is:
-
- max_cuid - for reaching definitions
- n_exprs - for available expressions
-
- Thus we view the bitmaps as 2 dimensional arrays. i.e.
- rd_kill[block_num][cuid_num]
- ae_kill[block_num][expr_num] */
-
-/* For reaching defs */
-static sbitmap *rd_kill, *rd_gen, *reaching_defs, *rd_out;
-
-/* for available exprs */
-static sbitmap *ae_kill, *ae_gen, *ae_in, *ae_out;
-
-/* Objects of this type are passed around by the null-pointer check
- removal routines. */
-struct null_pointer_info
-{
- /* The basic block being processed. */
- basic_block current_block;
- /* The first register to be handled in this pass. */
- unsigned int min_reg;
- /* One greater than the last register to be handled in this pass. */
- unsigned int max_reg;
- sbitmap *nonnull_local;
- sbitmap *nonnull_killed;
-};
+/* For available exprs */
+static sbitmap *ae_kill, *ae_gen;
\f
static void compute_can_copy (void);
static void *gmalloc (size_t) ATTRIBUTE_MALLOC;
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 replace_one_set (int, rtx, rtx);
static void record_set_info (rtx, rtx, void *);
static void compute_sets (rtx);
static void hash_scan_insn (rtx, struct hash_table *, int);
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 *);
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 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 struct expr *next_set (unsigned int, struct expr *);
static void reset_opr_set_tables (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 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);
\f
/* Entry point for global common subexpression elimination.
- F is the first instruction in the function. */
+ F is the first instruction in the function. Return nonzero if a
+ change is mode. */
int
gcse_main (rtx f, FILE *file)
/* Return if there's nothing to do, or it is too expensive. */
if (n_basic_blocks <= 1 || is_too_expensive (_("GCSE disabled")))
return 0;
-
+
gcc_obstack_init (&gcse_obstack);
bytes_used = 0;
/* Don't allow constant propagation to modify jumps
during this pass. */
+ timevar_push (TV_CPROP1);
changed = one_cprop_pass (pass + 1, 0, 0);
+ timevar_pop (TV_CPROP1);
if (optimize_size)
- changed |= one_classic_gcse_pass (pass + 1);
+ /* Do nothing. */ ;
else
{
+ timevar_push (TV_PRE);
changed |= one_pre_gcse_pass (pass + 1);
/* We may have just created new basic blocks. Release and
recompute various things which are sized on the number of
alloc_reg_set_mem (max_reg_num ());
compute_sets (f);
run_jump_opt_after_gcse = 1;
+ timevar_pop (TV_PRE);
}
if (max_pass_bytes < bytes_used)
/* It does not make sense to run code hoisting unless we are optimizing
for code size -- it rarely makes programs faster, and can make
them bigger if we did partial redundancy elimination (when optimizing
- for space, we use a classic gcse algorithm instead of partial
- redundancy algorithms). */
+ for space, we don't run the partial redundancy algorithms). */
if (optimize_size)
{
+ timevar_push (TV_HOIST);
max_gcse_regno = max_reg_num ();
alloc_gcse_mem (f);
changed |= one_code_hoisting_pass ();
if (max_pass_bytes < bytes_used)
max_pass_bytes = bytes_used;
+ timevar_pop (TV_HOIST);
}
if (file)
max_gcse_regno = max_reg_num ();
alloc_gcse_mem (f);
/* This time, go ahead and allow cprop to alter jumps. */
+ timevar_push (TV_CPROP2);
one_cprop_pass (pass + 1, 1, 0);
+ timevar_pop (TV_CPROP2);
free_gcse_mem ();
if (file)
obstack_free (&gcse_obstack, NULL);
free_reg_set_mem ();
+
/* We are finished with alias. */
end_alias_analysis ();
allocate_reg_info (max_reg_num (), FALSE, FALSE);
if (!optimize_size && flag_gcse_sm)
- store_motion ();
+ {
+ timevar_push (TV_LSM);
+ store_motion ();
+ timevar_pop (TV_LSM);
+ }
/* Record where pseudo-registers are set. */
return run_jump_opt_after_gcse;
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);
basic block. */
modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
canon_modify_mem_list = gcalloc (last_basic_block, sizeof (rtx));
- modify_mem_list_set = BITMAP_XMALLOC ();
- canon_modify_mem_list_set = BITMAP_XMALLOC ();
+ modify_mem_list_set = BITMAP_ALLOC (NULL);
+ blocks_with_calls = BITMAP_ALLOC (NULL);
}
/* Free memory allocated by alloc_gcse_mem. */
free (uid_cuid);
free (cuid_insn);
- BITMAP_XFREE (reg_set_bitmap);
+ BITMAP_FREE (reg_set_bitmap);
sbitmap_vector_free (reg_set_in_block);
free_modify_mem_tables ();
- BITMAP_XFREE (modify_mem_list_set);
- BITMAP_XFREE (canon_modify_mem_list_set);
-}
-
-/* Many of the global optimization algorithms work by solving dataflow
- equations for various expressions. Initially, some local value is
- computed for each expression in each block. Then, the values across the
- various blocks are combined (by following flow graph edges) to arrive at
- global values. Conceptually, each set of equations is independent. We
- may therefore solve all the equations in parallel, solve them one at a
- time, or pick any intermediate approach.
-
- When you're going to need N two-dimensional bitmaps, each X (say, the
- number of blocks) by Y (say, the number of expressions), call this
- function. It's not important what X and Y represent; only that Y
- correspond to the things that can be done in parallel. This function will
- return an appropriate chunking factor C; you should solve C sets of
- equations in parallel. By going through this function, we can easily
- trade space against time; by solving fewer equations in parallel we use
- less space. */
-
-static int
-get_bitmap_width (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 *table)
{
unsigned int i;
obstack_free (®_set_obstack, NULL);
}
-/* An OLD_INSN that used to set REGNO was replaced by NEW_INSN.
- Update the corresponding `reg_set_table' entry accordingly.
- We assume that NEW_INSN is not already recorded in reg_set_table[regno]. */
-
-static void
-replace_one_set (int regno, rtx old_insn, rtx new_insn)
-{
- struct reg_set *reg_info;
- if (regno >= reg_set_table_size)
- return;
- for (reg_info = reg_set_table[regno]; reg_info; reg_info = reg_info->next)
- if (reg_info->insn == old_insn)
- {
- reg_info->insn = new_insn;
- break;
- }
-}
-
/* Record REGNO in the reg_set table. */
static void
new_reg_info = obstack_alloc (®_set_obstack, sizeof (struct reg_set));
bytes_used += sizeof (struct reg_set);
- new_reg_info->insn = insn;
+ new_reg_info->bb_index = BLOCK_NUM (insn);
new_reg_info->next = reg_set_table[regno];
reg_set_table[regno] = new_reg_info;
}
{
rtx record_set_insn = (rtx) data;
- if (GET_CODE (dest) == REG && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
+ if (REG_P (dest) && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
record_one_set (REGNO (dest), record_set_insn);
}
case CONST_DOUBLE:
case CONST_VECTOR:
case CALL:
- case CONSTANT_P_RTX:
return 0;
default:
{
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,
/* 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
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. HASH_TABLE_SIZE is
- the current size of the hash table to be probed.
-
- ??? One might want to merge this with canon_hash. Later. */
+ 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,
*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++;
-
- return hash;
-}
-
-/* Subroutine of hash_expr to do the actual work. */
-
-static unsigned int
-hash_expr_1 (rtx x, enum machine_mode mode, int *do_not_record_p)
-{
- int i, j;
- unsigned hash = 0;
- enum rtx_code code;
- const char *fmt;
-
- /* Used to turn recursion into iteration. We can't rely on GCC's
- tail-recursion elimination since we need to keep accumulating values
- in HASH. */
-
- if (x == 0)
- return hash;
-
- repeat:
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- hash += ((unsigned int) REG << 7) + REGNO (x);
- return hash;
-
- case CONST_INT:
- hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
- + (unsigned int) INTVAL (x));
- return hash;
-
- case CONST_DOUBLE:
- /* This is like the general case, except that it only counts
- the integers representing the constant. */
- hash += (unsigned int) code + (unsigned int) GET_MODE (x);
- if (GET_MODE (x) != VOIDmode)
- for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
- hash += (unsigned int) XWINT (x, i);
- else
- hash += ((unsigned int) CONST_DOUBLE_LOW (x)
- + (unsigned int) CONST_DOUBLE_HIGH (x));
- return hash;
-
- case CONST_VECTOR:
- {
- int units;
- rtx elt;
-
- units = CONST_VECTOR_NUNITS (x);
-
- for (i = 0; i < units; ++i)
- {
- elt = CONST_VECTOR_ELT (x, i);
- hash += hash_expr_1 (elt, GET_MODE (elt), do_not_record_p);
- }
-
- return hash;
- }
-
- /* Assume there is only one rtx object for any given label. */
- case LABEL_REF:
- /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
- differences and differences between each stage's debugging dumps. */
- hash += (((unsigned int) LABEL_REF << 7)
- + CODE_LABEL_NUMBER (XEXP (x, 0)));
- return hash;
-
- case SYMBOL_REF:
- {
- /* Don't hash on the symbol's address to avoid bootstrap differences.
- Different hash values may cause expressions to be recorded in
- different orders and thus different registers to be used in the
- final assembler. This also avoids differences in the dump files
- between various stages. */
- unsigned int h = 0;
- const unsigned char *p = (const unsigned char *) XSTR (x, 0);
-
- while (*p)
- h += (h << 7) + *p++; /* ??? revisit */
-
- hash += ((unsigned int) SYMBOL_REF << 7) + h;
- return hash;
- }
-
- case MEM:
- if (MEM_VOLATILE_P (x))
- {
- *do_not_record_p = 1;
- return 0;
- }
-
- hash += (unsigned int) MEM;
- /* We used alias set for hashing, but this is not good, since the alias
- set may differ in -fprofile-arcs and -fbranch-probabilities compilation
- causing the profiles to fail to match. */
- x = XEXP (x, 0);
- goto repeat;
-
- case PRE_DEC:
- case PRE_INC:
- case POST_DEC:
- case POST_INC:
- case PC:
- case CC0:
- case CALL:
- case UNSPEC_VOLATILE:
- *do_not_record_p = 1;
- return 0;
-
- case ASM_OPERANDS:
- if (MEM_VOLATILE_P (x))
- {
- *do_not_record_p = 1;
- return 0;
- }
- else
- {
- /* We don't want to take the filename and line into account. */
- hash += (unsigned) code + (unsigned) GET_MODE (x)
- + hash_string_1 (ASM_OPERANDS_TEMPLATE (x))
- + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
- + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
-
- if (ASM_OPERANDS_INPUT_LENGTH (x))
- {
- for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
- {
- hash += (hash_expr_1 (ASM_OPERANDS_INPUT (x, i),
- GET_MODE (ASM_OPERANDS_INPUT (x, i)),
- do_not_record_p)
- + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
- (x, i)));
- }
-
- hash += hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
- x = ASM_OPERANDS_INPUT (x, 0);
- mode = GET_MODE (x);
- goto repeat;
- }
- return hash;
- }
-
- default:
- break;
- }
-
- hash += (unsigned) code + (unsigned) GET_MODE (x);
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- /* If we are about to do the last recursive call
- needed at this level, change it into iteration.
- This function is called enough to be worth it. */
- if (i == 0)
- {
- x = XEXP (x, i);
- goto repeat;
- }
-
- hash += hash_expr_1 (XEXP (x, i), 0, do_not_record_p);
- if (*do_not_record_p)
- return 0;
- }
-
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- {
- hash += hash_expr_1 (XVECEXP (x, i, j), 0, do_not_record_p);
- if (*do_not_record_p)
- return 0;
- }
-
- else if (fmt[i] == 's')
- hash += hash_string_1 (XSTR (x, i));
- else if (fmt[i] == 'i')
- hash += (unsigned int) XINT (x, i);
- else
- abort ();
- }
-
- return hash;
-}
-
/* Hash a set of register REGNO.
Sets are hashed on the register that is set. This simplifies the PRE copy
return hash % hash_table_size;
}
-/* Return nonzero if exp1 is equivalent to exp2.
- ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
+/* Return nonzero if exp1 is equivalent to exp2. */
static int
expr_equiv_p (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;
-
- /* A volatile mem should not be considered equivalent to any other. */
- if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
- return 0;
- break;
-
- /* For commutative operations, check both orders. */
- case PLUS:
- case MULT:
- case AND:
- case IOR:
- case XOR:
- case NE:
- case EQ:
- return ((expr_equiv_p (XEXP (x, 0), XEXP (y, 0))
- && expr_equiv_p (XEXP (x, 1), XEXP (y, 1)))
- || (expr_equiv_p (XEXP (x, 0), XEXP (y, 1))
- && expr_equiv_p (XEXP (x, 1), XEXP (y, 0))));
-
- case ASM_OPERANDS:
- /* We don't use the generic code below because we want to
- disregard filename and line numbers. */
-
- /* A volatile asm isn't equivalent to any other. */
- if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
- return 0;
-
- if (GET_MODE (x) != GET_MODE (y)
- || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
- || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
- ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
- || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
- || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
- return 0;
-
- if (ASM_OPERANDS_INPUT_LENGTH (x))
- {
- for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
- if (! expr_equiv_p (ASM_OPERANDS_INPUT (x, i),
- ASM_OPERANDS_INPUT (y, i))
- || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
- ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
- return 0;
- }
-
- return 1;
-
- default:
- break;
- }
-
- /* Compare the elements. If any pair of corresponding elements
- fail to match, return 0 for the whole thing. */
-
- fmt = GET_RTX_FORMAT (code);
- for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
- {
- switch (fmt[i])
- {
- case 'e':
- if (! expr_equiv_p (XEXP (x, i), XEXP (y, i)))
- return 0;
- break;
-
- case 'E':
- if (XVECLEN (x, i) != XVECLEN (y, i))
- return 0;
- for (j = 0; j < XVECLEN (x, i); j++)
- if (! expr_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j)))
- return 0;
- break;
-
- case 's':
- if (strcmp (XSTR (x, i), XSTR (y, i)))
- return 0;
- break;
-
- case 'i':
- if (XINT (x, i) != XINT (y, i))
- return 0;
- break;
-
- case 'w':
- if (XWINT (x, i) != XWINT (y, i))
- return 0;
- break;
-
- case '0':
- break;
-
- default:
- abort ();
- }
- }
-
- return 1;
+ return exp_equiv_p (x, y, 0, true);
}
/* Insert expression X in INSN in the hash TABLE.
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);
{
antic_occr = cur_expr->antic_occr;
- /* Search for another occurrence in the same basic block. */
- while (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
- {
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = antic_occr;
- antic_occr = antic_occr->next;
- }
+ if (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
+ antic_occr = NULL;
if (antic_occr)
/* Found another instance of the expression in the same basic block.
/* First occurrence of this expression in this basic block. */
antic_occr = gcse_alloc (sizeof (struct occr));
bytes_used += sizeof (struct occr);
- /* First occurrence of this expression in any block? */
- if (cur_expr->antic_occr == NULL)
- cur_expr->antic_occr = antic_occr;
- else
- last_occr->next = antic_occr;
-
antic_occr->insn = insn;
- antic_occr->next = NULL;
+ antic_occr->next = cur_expr->antic_occr;
antic_occr->deleted_p = 0;
+ cur_expr->antic_occr = antic_occr;
}
}
{
avail_occr = cur_expr->avail_occr;
- /* Search for another occurrence in the same basic block. */
- while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))
+ if (avail_occr && BLOCK_NUM (avail_occr->insn) == BLOCK_NUM (insn))
{
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = avail_occr;
- avail_occr = avail_occr->next;
+ /* Found another instance of the expression in the same basic block.
+ Prefer this occurrence to the currently recorded one. We want
+ the last one in the block and the block is scanned from start
+ to end. */
+ avail_occr->insn = insn;
}
-
- if (avail_occr)
- /* Found another instance of the expression in the same basic block.
- Prefer this occurrence to the currently recorded one. We want
- the last one in the block and the block is scanned from start
- to end. */
- avail_occr->insn = insn;
else
{
/* First occurrence of this expression in this basic block. */
avail_occr = gcse_alloc (sizeof (struct occr));
bytes_used += sizeof (struct occr);
-
- /* First occurrence of this expression in any block? */
- if (cur_expr->avail_occr == NULL)
- cur_expr->avail_occr = avail_occr;
- else
- last_occr->next = avail_occr;
-
avail_occr->insn = insn;
- avail_occr->next = NULL;
+ avail_occr->next = cur_expr->avail_occr;
avail_occr->deleted_p = 0;
+ cur_expr->avail_occr = avail_occr;
}
}
}
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);
/* Now record the occurrence. */
cur_occr = cur_expr->avail_occr;
- /* Search for another occurrence in the same basic block. */
- while (cur_occr && BLOCK_NUM (cur_occr->insn) != BLOCK_NUM (insn))
+ if (cur_occr && BLOCK_NUM (cur_occr->insn) == BLOCK_NUM (insn))
{
- /* If an occurrence isn't found, save a pointer to the end of
- the list. */
- last_occr = cur_occr;
- cur_occr = cur_occr->next;
+ /* Found another instance of the expression in the same basic block.
+ Prefer this occurrence to the currently recorded one. We want
+ the last one in the block and the block is scanned from start
+ to end. */
+ cur_occr->insn = insn;
}
-
- if (cur_occr)
- /* Found another instance of the expression in the same basic block.
- Prefer this occurrence to the currently recorded one. We want the
- last one in the block and the block is scanned from start to end. */
- cur_occr->insn = insn;
else
{
/* First occurrence of this expression in this basic block. */
cur_occr = gcse_alloc (sizeof (struct occr));
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->deleted_p = 0;
+ cur_occr->insn = insn;
+ cur_occr->next = cur_expr->avail_occr;
+ cur_occr->deleted_p = 0;
+ cur_expr->avail_occr = cur_occr;
}
}
&& GET_CODE (XEXP (x, 1)) == CONST_INT)
return true;
-
/* Consider a COMPARE of the same registers is a constant
- if they are not floating point registers. */
+ if 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);
}
if (GET_CODE (src) == CALL)
hash_scan_call (src, insn, table);
- else if (GET_CODE (dest) == REG)
+ else if (REG_P (dest))
{
unsigned int regno = REGNO (dest);
rtx tmp;
explicitly, it means address of parameter has been taken,
so we should not extend the lifetime of the pseudo. */
&& ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
- || GET_CODE (XEXP (note, 0)) != MEM))
+ || ! MEM_P (XEXP (note, 0))))
{
/* An expression is not anticipatable if its operands are
modified before this insn or if this is not the only SET in
/* Record sets for constant/copy propagation. */
else if (table->set_p
&& regno >= FIRST_PSEUDO_REGISTER
- && ((GET_CODE (src) == REG
+ && ((REG_P (src)
&& REGNO (src) >= FIRST_PSEUDO_REGISTER
&& can_copy_p (GET_MODE (dest))
&& REGNO (src) != regno)
/* 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 && GET_CODE (src) == REG && GET_CODE (dest) == MEM)
+ else if (flag_gcse_las && REG_P (src) && MEM_P (dest))
{
unsigned int regno = REGNO (src);
explicitly, it means address of parameter has been taken,
so we should not extend the lifetime of the pseudo. */
&& ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
- || GET_CODE (XEXP (note, 0)) != MEM))
+ || ! MEM_P (XEXP (note, 0))))
{
/* Stores are never anticipatable. */
int antic_p = 0;
while (GET_CODE (dest) == SUBREG
|| GET_CODE (dest) == ZERO_EXTRACT
- || GET_CODE (dest) == SIGN_EXTRACT
|| GET_CODE (dest) == STRICT_LOW_PART)
dest = XEXP (dest, 0);
that function calls are assumed to clobber memory, but are handled
elsewhere. */
- if (GET_CODE (dest) != MEM)
+ if (! MEM_P (dest))
return;
dest_addr = get_addr (XEXP (dest, 0));
alloc_EXPR_LIST (VOIDmode, dest_addr, canon_modify_mem_list[bb]);
canon_modify_mem_list[bb] =
alloc_EXPR_LIST (VOIDmode, dest, canon_modify_mem_list[bb]);
- bitmap_set_bit (canon_modify_mem_list_set, bb);
}
/* Record memory modification information for INSN. We do not actually care
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);
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);
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);
\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 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)
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. */
-
-/* Allocate reaching def variables. */
+/* Compute copy/constant propagation working variables. */
-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);
+/* Local properties of assignments. */
+static sbitmap *cprop_pavloc;
+static sbitmap *cprop_absaltered;
- rd_gen = sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_gen, n_blocks);
+/* Global properties of assignments (computed from the local properties). */
+static sbitmap *cprop_avin;
+static sbitmap *cprop_avout;
- reaching_defs = sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (reaching_defs, n_blocks);
-
- rd_out = sbitmap_vector_alloc (n_blocks, n_insns);
- sbitmap_vector_zero (rd_out, n_blocks);
-}
-
-/* Free reaching def variables. */
+/* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
+ basic blocks. N_SETS is the number of sets. */
static void
-free_rd_mem (void)
+alloc_cprop_mem (int n_blocks, int n_sets)
{
- sbitmap_vector_free (rd_kill);
- sbitmap_vector_free (rd_gen);
- sbitmap_vector_free (reaching_defs);
- sbitmap_vector_free (rd_out);
+ cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
+ cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
+
+ cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
+ cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
}
-/* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
+/* Free vars used by copy/const propagation. */
static void
-handle_rd_kill_set (rtx insn, int regno, basic_block bb)
+free_cprop_mem (void)
{
- 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));
+ sbitmap_vector_free (cprop_pavloc);
+ sbitmap_vector_free (cprop_absaltered);
+ sbitmap_vector_free (cprop_avin);
+ sbitmap_vector_free (cprop_avout);
}
-/* Compute the set of kill's for reaching definitions. */
+/* For each block, compute whether X is transparent. X is either an
+ expression or an assignment [though we don't care which, for this context
+ an assignment is treated as an expression]. For each block where an
+ element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
+ bit in BMAP. */
static void
-compute_kill_rd (void)
+compute_transp (rtx x, int indx, sbitmap *bmap, int set_p)
{
- int cuid;
- unsigned int regno;
- int i;
+ int i, j;
basic_block bb;
+ enum rtx_code code;
+ reg_set *r;
+ const char *fmt;
- /* For each block
- For each set bit in `gen' of the block (i.e each insn which
- generates a definition in the block)
- Call the reg set by the insn corresponding to that bit regx
- Look at the linked list starting at reg_set_table[regx]
- For each setting of regx in the linked list, which is not in
- this block
- Set the bit in `kill' corresponding to that insn. */
- FOR_EACH_BB (bb)
- for (cuid = 0; cuid < max_cuid; cuid++)
- if (TEST_BIT (rd_gen[bb->index], cuid))
- {
- rtx insn = CUID_INSN (cuid);
- rtx pat = PATTERN (insn);
+ /* 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;
- if (GET_CODE (insn) == CALL_INSN)
+ code = GET_CODE (x);
+ switch (code)
+ {
+ case REG:
+ if (set_p)
+ {
+ if (REGNO (x) < FIRST_PSEUDO_REGISTER)
{
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- handle_rd_kill_set (insn, regno, bb);
+ FOR_EACH_BB (bb)
+ if (TEST_BIT (reg_set_in_block[bb->index], REGNO (x)))
+ SET_BIT (bmap[bb->index], indx);
}
-
- if (GET_CODE (pat) == PARALLEL)
+ else
{
- for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
- {
- enum rtx_code code = GET_CODE (XVECEXP (pat, 0, i));
-
- if ((code == SET || code == CLOBBER)
- && GET_CODE (XEXP (XVECEXP (pat, 0, i), 0)) == REG)
- handle_rd_kill_set (insn,
- REGNO (XEXP (XVECEXP (pat, 0, i), 0)),
- bb);
- }
+ for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
+ SET_BIT (bmap[r->bb_index], indx);
}
- else if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == REG)
- /* Each setting of this register outside of this block
- must be marked in the set of kills in this block. */
- handle_rd_kill_set (insn, REGNO (SET_DEST (pat)), bb);
}
-}
-
-/* Compute the reaching definitions as in
- Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
- Chapter 10. It is the same algorithm as used for computing available
- expressions but applied to the gens and kills of reaching definitions. */
-
-static void
-compute_rd (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]);
+ if (REGNO (x) < FIRST_PSEUDO_REGISTER)
+ {
+ FOR_EACH_BB (bb)
+ if (TEST_BIT (reg_set_in_block[bb->index], REGNO (x)))
+ RESET_BIT (bmap[bb->index], indx);
+ }
+ else
+ {
+ for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
+ RESET_BIT (bmap[r->bb_index], indx);
+ }
}
- passes++;
- }
-
- if (gcse_file)
- fprintf (gcse_file, "reaching def computation: %d passes\n", passes);
-}
-\f
-/* Classic GCSE available expression support. */
-
-/* Allocate memory for available expression computation. */
-
-static void
-alloc_avail_expr_mem (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:
+ {
+ 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, 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 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.
-
- Called only by handle_avail_expr. */
+ rtx note = find_reg_equal_equiv_note (insn);
+ rtx src = 0;
+ int success = 0;
+ rtx set = single_set (insn);
-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
- {
- /* Pattern is computed more than once.
- Search backwards from this insn to see how many of these
- computations actually reach this insn. */
- struct occr *occr;
- rtx insn_computes_expr = NULL;
- int can_reach = 0;
-
- for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
- {
- if (BLOCK_FOR_INSN (occr->insn) == bb)
- {
- /* The expression is generated in this block.
- The only time we care about this is when the expression
- is generated later in the block [and thus there's a loop].
- We let the normal cse pass handle the other cases. */
- if (INSN_CUID (insn) < INSN_CUID (occr->insn)
- && expr_reaches_here_p (occr, expr, bb, 1))
- {
- can_reach++;
- if (can_reach > 1)
- return NULL;
- insn_computes_expr = occr->insn;
- }
- }
- else if (expr_reaches_here_p (occr, expr, bb, 0))
- {
- can_reach++;
- if (can_reach > 1)
- return NULL;
+ /* If there is already a NOTE, update the expression in it with our
+ replacement. */
+ if (note != 0)
+ XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
- insn_computes_expr = occr->insn;
- }
- }
+ 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 (insn_computes_expr == NULL)
- abort ();
+ if (!rtx_equal_p (src, SET_SRC (set))
+ && validate_change (insn, &SET_SRC (set), src, 0))
+ success = 1;
- return insn_computes_expr;
+ /* If we've failed to do replacement, have a single SET, don't already
+ have a note, and have no special SET, add a REG_EQUAL note to not
+ lose information. */
+ if (!success && note == 0 && set != 0
+ && GET_CODE (SET_DEST (set)) != ZERO_EXTRACT)
+ note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
}
+
+ /* 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);
+
+ 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_NUM (insn)], 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);
- /* The set might be in a parallel with multiple sets; we could
- probably handle that, but there's currently no easy way to find
- the relevant sub-expression. */
- if (!expr_set)
- return 0;
-
- found_setting = 0;
- use_src = 0;
+ rtx new, 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));
+ /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
+ set_src = note_src ? note_src : SET_SRC (set);
- /* This shouldn't happen. */
- if (regnum_for_replacing >= max_gcse_regno)
- abort ();
+ /* 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;
- this_reg = reg_set_table[regnum_for_replacing];
+ new = simplify_replace_rtx (set_src, from, src);
- /* If the register the expression is computed into is set only once,
- or only one set reaches this insn, use it. */
- if (this_reg->next == NULL
- || can_disregard_other_sets (&this_reg, insn, 0))
- found_setting = 1;
- }
+ /* If no simplification can be made, then try the next register. */
+ if (rtx_equal_p (new, SET_SRC (set)))
+ return 0;
- if (found_setting)
+ /* If this is now a no-op delete it, otherwise this must be a valid insn. */
+ if (new == pc_rtx)
+ delete_insn (jump);
+ else
{
- pat = PATTERN (insn);
- if (use_src)
- to = SET_SRC (expr_set);
- else
- to = SET_DEST (expr_set);
- changed = validate_change (insn, &SET_SRC (pat), to, 0);
-
- /* We should be able to ignore the return code from validate_change but
- to play it safe we check. */
- if (changed)
+ /* 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))
{
- 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));
- }
- }
- }
+ /* 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. */
- /* The register that the expr is computed into is set more than once. */
- else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
- {
- /* Insert an insn after insnx that copies the reg set in insnx
- into a new pseudo register call this new register REGN.
- From insnb until end of basic block or until REGB is set
- replace all uses of REGB with REGN. */
- rtx new_insn;
-
- to = gen_reg_rtx (GET_MODE (SET_DEST (expr_set)));
-
- /* Generate the new insn. */
- /* ??? If the change fails, we return 0, even though we created
- an insn. I think this is ok. */
- new_insn
- = emit_insn_after (gen_rtx_SET (VOIDmode, to,
- SET_DEST (expr_set)),
- insn_computes_expr);
+ if (!rtx_equal_p (new, note_src))
+ set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new));
+ return 0;
+ }
- /* Keep register set table up to date. */
- record_one_set (REGNO (to), new_insn);
+ /* Remove REG_EQUAL note after simplification. */
+ if (note_src)
+ remove_note (jump, note);
- gcse_create_count++;
- if (gcse_file != NULL)
- {
- fprintf (gcse_file, "GCSE: Creating insn %d to copy value of reg %d",
- INSN_UID (NEXT_INSN (insn_computes_expr)),
- REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr)))));
- fprintf (gcse_file, ", computed in insn %d,\n",
- INSN_UID (insn_computes_expr));
- fprintf (gcse_file, " into newly allocated reg %d\n",
- REGNO (to));
- }
+ /* 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);
+ }
- pat = PATTERN (insn);
+#ifdef HAVE_cc0
+ /* Delete the cc0 setter. */
+ if (setcc != NULL && CC0_P (SET_DEST (single_set (setcc))))
+ delete_insn (setcc);
+#endif
- /* Do register replacement for INSN. */
- changed = validate_change (insn, &SET_SRC (pat),
- SET_DEST (PATTERN
- (NEXT_INSN (insn_computes_expr))),
- 0);
+ run_jump_opt_after_gcse = 1;
- /* 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 (gcse_file != NULL)
+ {
+ fprintf (gcse_file,
+ "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
+ REGNO (from), INSN_UID (jump));
+ print_rtl (gcse_file, src);
+ fprintf (gcse_file, "\n");
}
+ purge_dead_edges (bb);
- return changed;
+ return 1;
}
-/* Perform classic GCSE. This is called by one_classic_gcse_pass after all
- the dataflow analysis has been done.
-
- The result is nonzero if a change was made. */
-
-static int
-classic_gcse (void)
+static bool
+constprop_register (rtx insn, rtx from, rtx to, int alter_jumps)
{
- 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 (alter_jumps
+ && (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 (bb);
- insn != NULL && insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
- {
- /* Is insn of form (set (pseudo-reg) ...)? */
- if (GET_CODE (insn) == INSN
- && GET_CODE (PATTERN (insn)) == SET
- && GET_CODE (SET_DEST (PATTERN (insn))) == REG
- && REGNO (SET_DEST (PATTERN (insn))) >= FIRST_PSEUDO_REGISTER)
- {
- rtx pat = PATTERN (insn);
- rtx src = SET_SRC (pat);
- struct expr *expr;
-
- if (want_to_gcse_p (src)
- /* Is the expression recorded? */
- && ((expr = lookup_expr (src, &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 (alter_jumps && 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, int alter_jumps)
{
+ 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);
- }
+ unsigned int regno = REGNO (reg_used->reg_rtx);
+ rtx pat, src;
+ struct expr *set;
- return changed;
-}
-\f
-/* Compute copy/constant propagation working variables. */
+ /* Ignore registers created by GCSE.
+ We do this because ... */
+ if (regno >= max_gcse_regno)
+ continue;
-/* 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, alter_jumps))
+ {
+ changed = 1;
+ global_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 (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 (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));
+ }
-/* 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);
+ 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)
- {
- 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);
- }
- }
- else
- {
- if (REGNO (x) < FIRST_PSEUDO_REGISTER)
- {
- FOR_EACH_BB (bb)
- if (TEST_BIT (reg_set_in_block[bb->index], REGNO (x)))
- RESET_BIT (bmap[bb->index], indx);
- }
- else
- {
- for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
- RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
- }
- }
+ 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 MEM:
- FOR_EACH_BB (bb)
- {
- rtx list_entry = canon_modify_mem_list[bb->index];
+ 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;
- while (list_entry)
- {
- rtx dest, dest_addr;
+ default:
+ break;
+ }
- 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. */
+ find_used_regs (xptr, data);
+}
- dest = XEXP (list_entry, 0);
- list_entry = XEXP (list_entry, 1);
- dest_addr = XEXP (list_entry, 0);
+/* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
+ their REG_EQUAL notes need updating. */
- if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
- x, rtx_addr_varies_p))
- {
- if (set_p)
- SET_BIT (bmap[bb->index], indx);
- else
- RESET_BIT (bmap[bb->index], indx);
- break;
- }
- list_entry = XEXP (list_entry, 1);
- }
- }
+static bool
+do_local_cprop (rtx x, rtx insn, int alter_jumps, rtx *libcall_sp)
+{
+ rtx newreg = NULL, newcnst = NULL;
- x = XEXP (x, 0);
- goto repeat;
+ /* 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;
- 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;
+ if (!val)
+ return false;
+ for (l = val->locs; l; l = l->next)
+ {
+ rtx this_rtx = l->loc;
+ rtx note;
- default:
- break;
- }
+ /* Don't CSE non-constant values out of libcall blocks. */
+ if (l->in_libcall && ! CONSTANT_P (this_rtx))
+ continue;
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
- {
- if (fmt[i] == 'e')
+ if (gcse_constant_p (this_rtx))
+ newcnst = this_rtx;
+ if (REG_P (this_rtx) && REGNO (this_rtx) >= FIRST_PSEUDO_REGISTER
+ /* Don't copy propagate if it has attached REG_EQUIV note.
+ At this point this only function parameters should have
+ REG_EQUIV notes and if the argument slot is used somewhere
+ explicitly, it means address of parameter has been taken,
+ so we should not extend the lifetime of the pseudo. */
+ && (!(note = find_reg_note (l->setting_insn, REG_EQUIV, NULL_RTX))
+ || ! MEM_P (XEXP (note, 0))))
+ newreg = this_rtx;
+ }
+ if (newcnst && constprop_register (insn, x, newcnst, alter_jumps))
{
- /* If 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)
+ /* If we find a case where we can't fix the retval REG_EQUAL notes
+ match the new register, we either have to abandon this replacement
+ or fix delete_trivially_dead_insns to preserve the setting insn,
+ or make it delete the REG_EUAQL note, and fix up all passes that
+ require the REG_EQUAL note there. */
+ bool adjusted;
+
+ adjusted = adjust_libcall_notes (x, newcnst, insn, libcall_sp);
+ gcc_assert (adjusted);
+
+ if (gcse_file != NULL)
{
- x = XEXP (x, i);
- goto repeat;
+ 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");
}
-
- compute_transp (XEXP (x, i), indx, bmap, set_p);
+ local_const_prop_count++;
+ return true;
+ }
+ else if (newreg && newreg != x && try_replace_reg (x, newreg, insn))
+ {
+ adjust_libcall_notes (x, newreg, insn, libcall_sp);
+ if (gcse_file != NULL)
+ {
+ 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));
+ }
+ local_copy_prop_count++;
+ return true;
}
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
}
+ return false;
}
-/* Top level routine to do the dataflow analysis needed by copy/const
- propagation. */
-
-static void
-compute_cprop_data (void)
+/* 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)
{
- compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, &set_hash_table);
- compute_available (cprop_pavloc, cprop_absaltered,
- cprop_avout, cprop_avin);
-}
-\f
-/* Copy/constant propagation. */
-
-/* Maximum number of register uses in an insn that we handle. */
-#define MAX_USES 8
+ rtx end;
-/* 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];
+ while ((end = *libcall_sp++))
+ {
+ rtx note = find_reg_equal_equiv_note (end);
-/* Index into `reg_use_table' while building it. */
-static int reg_use_count;
+ if (! note)
+ continue;
-/* Set up a list of register numbers used in INSN. The found uses are stored
- in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
- and contains the number of uses in the table upon exit.
+ if (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) = simplify_replace_rtx (XEXP (note, 0), oldreg, newval);
+ insn = end;
+ }
+ return true;
+}
- ??? If a register appears multiple times we will record it multiple times.
- This doesn't hurt anything but it will slow things down. */
+#define MAX_NESTED_LIBCALLS 9
static void
-find_used_regs (rtx *xptr, void *data ATTRIBUTE_UNUSED)
+local_cprop_pass (int alter_jumps)
{
- int i, j;
- enum rtx_code code;
- const char *fmt;
- rtx x = *xptr;
-
- /* repeat is used to turn tail-recursion into iteration since GCC
- can't do it when there's no return value. */
- repeat:
- if (x == 0)
- return;
-
- code = GET_CODE (x);
- if (REG_P (x))
- {
- if (reg_use_count == MAX_USES)
- return;
-
- reg_use_table[reg_use_count].reg_rtx = x;
- reg_use_count++;
- }
-
- /* Recursively scan the operands of this expression. */
+ rtx insn;
+ struct reg_use *reg_used;
+ rtx libcall_stack[MAX_NESTED_LIBCALLS + 1], *libcall_sp;
+ bool changed = false;
- for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
+ cselib_init (false);
+ libcall_sp = &libcall_stack[MAX_NESTED_LIBCALLS];
+ *libcall_sp = 0;
+ for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
{
- if (fmt[i] == 'e')
+ if (INSN_P (insn))
{
- /* 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)
+ rtx note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
+
+ if (note)
{
- x = XEXP (x, 0);
- goto repeat;
+ gcc_assert (libcall_sp != libcall_stack);
+ *--libcall_sp = XEXP (note, 0);
}
+ note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
+ if (note)
+ libcall_sp++;
+ note = find_reg_equal_equiv_note (insn);
+ do
+ {
+ reg_use_count = 0;
+ note_uses (&PATTERN (insn), local_cprop_find_used_regs, NULL);
+ if (note)
+ local_cprop_find_used_regs (&XEXP (note, 0), NULL);
- find_used_regs (&XEXP (x, i), data);
+ 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);
}
- else if (fmt[i] == 'E')
- for (j = 0; j < XVECLEN (x, i); j++)
- find_used_regs (&XVECEXP (x, i, j), data);
+ cselib_process_insn (insn);
+ }
+ cselib_finish ();
+ /* Global analysis may get into infinite loops for unreachable blocks. */
+ if (changed && alter_jumps)
+ {
+ delete_unreachable_blocks ();
+ free_reg_set_mem ();
+ alloc_reg_set_mem (max_reg_num ());
+ compute_sets (get_insns ());
}
}
-/* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
- Returns nonzero is successful. */
+/* Forward propagate copies. This includes copies and constants. Return
+ nonzero if a change was made. */
static int
-try_replace_reg (rtx from, rtx to, rtx insn)
+cprop (int alter_jumps)
{
- rtx note = find_reg_equal_equiv_note (insn);
- rtx src = 0;
- int success = 0;
- rtx set = single_set (insn);
+ int changed;
+ basic_block bb;
+ rtx insn;
- validate_replace_src_group (from, to, insn);
- if (num_changes_pending () && apply_change_group ())
- success = 1;
+ /* 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;
+ }
- /* Try to simplify SET_SRC if we have substituted a constant. */
- if (success && set && CONSTANT_P (to))
+ changed = 0;
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
{
- src = simplify_rtx (SET_SRC (set));
+ /* Reset tables used to keep track of what's still valid [since the
+ start of the block]. */
+ reset_opr_set_tables ();
- if (src)
- validate_change (insn, &SET_SRC (set), src, 0);
+ for (insn = BB_HEAD (bb);
+ insn != NULL && insn != NEXT_INSN (BB_END (bb));
+ insn = NEXT_INSN (insn))
+ if (INSN_P (insn))
+ {
+ changed |= cprop_insn (insn, alter_jumps);
+
+ /* Keep track of everything modified by this insn. */
+ /* ??? Need to be careful w.r.t. mods done to INSN. Don't
+ call mark_oprs_set if we turned the insn into a NOTE. */
+ if (! NOTE_P (insn))
+ mark_oprs_set (insn);
+ }
}
- /* If there is already a NOTE, update the expression in it with our
- replacement. */
- if (note != 0)
- XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
+ if (gcse_file != NULL)
+ fprintf (gcse_file, "\n");
- 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);
+ return changed;
+}
- if (!rtx_equal_p (src, SET_SRC (set))
- && validate_change (insn, &SET_SRC (set), src, 0))
- success = 1;
+/* Similar to get_condition, only the resulting condition must be
+ valid at JUMP, instead of at EARLIEST.
- /* 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));
+ 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. */
+
+rtx
+fis_get_condition (rtx jump)
+{
+ return get_condition (jump, NULL, false, true);
+}
+
+/* Check the comparison COND to see if we can safely form an implicit set from
+ it. COND is either an EQ or NE comparison. */
+
+static bool
+implicit_set_cond_p (rtx cond)
+{
+ enum machine_mode mode = GET_MODE (XEXP (cond, 0));
+ rtx cst = XEXP (cond, 1);
+
+ /* We can't perform this optimization if either operand might be or might
+ contain a signed zero. */
+ if (HONOR_SIGNED_ZEROS (mode))
+ {
+ /* It is sufficient to check if CST is or contains a zero. We must
+ handle float, complex, and vector. If any subpart is a zero, then
+ the optimization can't be performed. */
+ /* ??? The complex and vector checks are not implemented yet. We just
+ always return zero for them. */
+ if (GET_CODE (cst) == CONST_DOUBLE)
+ {
+ REAL_VALUE_TYPE d;
+ REAL_VALUE_FROM_CONST_DOUBLE (d, cst);
+ if (REAL_VALUES_EQUAL (d, dconst0))
+ return 0;
+ }
+ else
+ return 0;
}
- /* 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);
+ return gcse_constant_p (cst);
+}
- return success;
+/* Find the implicit sets of a function. An "implicit set" is a constraint
+ on the value of a variable, implied by a conditional jump. For example,
+ following "if (x == 2)", the then branch may be optimized as though the
+ conditional performed an "explicit set", in this example, "x = 2". This
+ function records the set patterns that are implicit at the start of each
+ basic block. */
+
+static void
+find_implicit_sets (void)
+{
+ basic_block bb, dest;
+ unsigned int count;
+ rtx cond, new;
+
+ count = 0;
+ FOR_EACH_BB (bb)
+ /* Check for more than one successor. */
+ if (EDGE_COUNT (bb->succs) > 1)
+ {
+ cond = fis_get_condition (BB_END (bb));
+
+ if (cond
+ && (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
+ && REG_P (XEXP (cond, 0))
+ && REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
+ && implicit_set_cond_p (cond))
+ {
+ dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
+ : FALLTHRU_EDGE (bb)->dest;
+
+ if (dest && single_pred_p (dest)
+ && dest != EXIT_BLOCK_PTR)
+ {
+ new = 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 (gcse_file)
+ fprintf (gcse_file, "Found %d implicit sets\n", count);
}
-/* Find a set of REGNOs that are available on entry to INSN's block. Returns
- NULL no such set is found. */
+/* Perform one copy/constant propagation pass.
+ PASS is the pass count. If CPROP_JUMPS is true, perform constant
+ propagation into conditional jumps. If BYPASS_JUMPS is true,
+ perform conditional jump bypassing optimizations. */
-static struct expr *
-find_avail_set (int regno, rtx insn)
+static int
+one_cprop_pass (int pass, int cprop_jumps, int bypass_jumps)
{
- /* SET1 contains the last set found that can be returned to the caller for
- use in a substitution. */
- struct expr *set1 = 0;
+ int changed = 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:
+ global_const_prop_count = local_const_prop_count = 0;
+ global_copy_prop_count = local_copy_prop_count = 0;
- (set (reg X) (reg Y))
- (set (reg Y) (reg X))
+ local_cprop_pass (cprop_jumps);
- 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)
+ /* Determine implicit sets. */
+ implicit_sets = xcalloc (last_basic_block, sizeof (rtx));
+ find_implicit_sets ();
+
+ alloc_hash_table (max_cuid, &set_hash_table, 1);
+ compute_hash_table (&set_hash_table);
+
+ /* Free implicit_sets before peak usage. */
+ free (implicit_sets);
+ implicit_sets = NULL;
+
+ if (gcse_file)
+ dump_hash_table (gcse_file, "SET", &set_hash_table);
+ if (set_hash_table.n_elems > 0)
+ {
+ alloc_cprop_mem (last_basic_block, set_hash_table.n_elems);
+ compute_cprop_data ();
+ changed = cprop (cprop_jumps);
+ if (bypass_jumps)
+ changed |= bypass_conditional_jumps ();
+ free_cprop_mem ();
+ }
+
+ free_hash_table (&set_hash_table);
+
+ if (gcse_file)
+ {
+ fprintf (gcse_file, "CPROP of %s, pass %d: %d bytes needed, ",
+ current_function_name (), pass, bytes_used);
+ fprintf (gcse_file, "%d local const props, %d local copy props\n\n",
+ local_const_prop_count, local_copy_prop_count);
+ fprintf (gcse_file, "%d global const props, %d global copy props\n\n",
+ global_const_prop_count, global_copy_prop_count);
+ }
+ /* Global analysis may get into infinite loops for unreachable blocks. */
+ if (changed && cprop_jumps)
+ delete_unreachable_blocks ();
+
+ return changed;
+}
+\f
+/* Bypass conditional jumps. */
+
+/* The value of last_basic_block at the beginning of the jump_bypass
+ pass. The use of redirect_edge_and_branch_force may introduce new
+ basic blocks, but the data flow analysis is only valid for basic
+ block indices less than bypass_last_basic_block. */
+
+static int bypass_last_basic_block;
+
+/* Find a set of REGNO to a constant that is available at the end of basic
+ block BB. Returns NULL if no such set is found. Based heavily upon
+ find_avail_set. */
+
+static struct expr *
+find_bypass_set (int regno, int bb)
+{
+ struct expr *result = 0;
+
+ for (;;)
{
rtx src;
struct expr *set = lookup_set (regno, &set_hash_table);
- /* 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);
}
-
- /* SET1 holds the last set that was available and anticipatable at
- INSN. */
- return set1;
+ return result;
}
-/* Subroutine of cprop_insn that tries to propagate constants into
- JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
- it is the instruction that immediately precedes JUMP, and must be a
- single SET of a register. FROM is what we will try to replace,
- SRC is the constant we will try to substitute for it. Returns nonzero
- if a change was made. */
-static int
-cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
+/* 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 (rtx reg, edge e)
{
- rtx new, set_src, note_src;
- rtx set = pc_set (jump);
- rtx note = find_reg_equal_equiv_note (jump);
+ rtx insn;
- if (note)
- {
- note_src = XEXP (note, 0);
- if (GET_CODE (note_src) == EXPR_LIST)
- note_src = NULL_RTX;
- }
- else note_src = NULL_RTX;
+ for (insn = e->insns.r; insn; insn = NEXT_INSN (insn))
+ if (INSN_P (insn) && reg_set_p (reg, insn))
+ return true;
- /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
- set_src = note_src ? note_src : SET_SRC (set);
+ return false;
+}
- /* First substitute the SETCC condition into the JUMP instruction,
- then substitute that given values into this expanded JUMP. */
- if (setcc != NULL_RTX
- && !modified_between_p (from, setcc, jump)
- && !modified_between_p (src, setcc, jump))
- {
- rtx setcc_src;
- rtx setcc_set = single_set (setcc);
- rtx setcc_note = find_reg_equal_equiv_note (setcc);
- setcc_src = (setcc_note && GET_CODE (XEXP (setcc_note, 0)) != EXPR_LIST)
- ? XEXP (setcc_note, 0) : SET_SRC (setcc_set);
- set_src = simplify_replace_rtx (set_src, SET_DEST (setcc_set),
- setcc_src);
- }
- else
- setcc = NULL_RTX;
-
- new = 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 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))
- {
- /* 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. */
+/* 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.
- if (!rtx_equal_p (new, note_src))
- set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new));
- return 0;
- }
+ 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. */
- /* Remove REG_EQUAL note after simplification. */
- if (note_src)
- remove_note (jump, note);
+static int
+bypass_block (basic_block bb, rtx setcc, rtx jump)
+{
+ rtx insn, note;
+ edge e, edest;
+ int i, change;
+ int may_be_loop_header;
+ unsigned removed_p;
+ edge_iterator ei;
- /* 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);
- }
+ insn = (setcc != NULL) ? setcc : jump;
-#ifdef HAVE_cc0
- /* Delete the cc0 setter. */
- if (setcc != NULL && CC0_P (SET_DEST (single_set (setcc))))
- delete_insn (setcc);
-#endif
+ /* 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);
- run_jump_opt_after_gcse = 1;
+ may_be_loop_header = false;
+ FOR_EACH_EDGE (e, ei, bb->preds)
+ if (e->flags & EDGE_DFS_BACK)
+ {
+ may_be_loop_header = true;
+ break;
+ }
- const_prop_count++;
- if (gcse_file != NULL)
+ change = 0;
+ for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
{
- 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);
-
- return 1;
-}
-
-static bool
-constprop_register (rtx insn, rtx from, rtx to, int alter_jumps)
-{
- rtx sset;
+ removed_p = 0;
+
+ if (e->flags & EDGE_COMPLEX)
+ {
+ ei_next (&ei);
+ continue;
+ }
- /* 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;
- }
+ /* We can't redirect edges from new basic blocks. */
+ if (e->src->index >= bypass_last_basic_block)
+ {
+ ei_next (&ei);
+ continue;
+ }
- /* Handle normal insns next. */
- if (GET_CODE (insn) == INSN
- && try_replace_reg (from, to, insn))
- return 1;
+ /* 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;
+ }
- /* 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.
+ 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;
- 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 (regno >= max_gcse_regno)
+ continue;
-/* Perform constant and copy propagation on INSN.
- The result is nonzero if a change was made. */
+ set = find_bypass_set (regno, e->src->index);
-static int
-cprop_insn (rtx insn, int alter_jumps)
-{
- struct reg_use *reg_used;
- int changed = 0;
- rtx note;
+ if (! set)
+ continue;
- if (!INSN_P (insn))
- return 0;
+ /* Check the data flow is valid after edge insertions. */
+ if (e->insns.r && reg_killed_on_edge (reg_used->reg_rtx, e))
+ continue;
- reg_use_count = 0;
- note_uses (&PATTERN (insn), find_used_regs, NULL);
+ src = SET_SRC (pc_set (jump));
- note = find_reg_equal_equiv_note (insn);
+ if (setcc != NULL)
+ src = simplify_replace_rtx (src,
+ SET_DEST (PATTERN (setcc)),
+ SET_SRC (PATTERN (setcc)));
- /* We may win even when propagating constants into notes. */
- if (note)
- find_used_regs (&XEXP (note, 0), NULL);
+ new = simplify_replace_rtx (src, reg_used->reg_rtx,
+ SET_SRC (set->expr));
- for (reg_used = ®_use_table[0]; reg_use_count > 0;
- reg_used++, reg_use_count--)
- {
- unsigned int regno = REGNO (reg_used->reg_rtx);
- rtx pat, src;
- struct expr *set;
+ /* 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. */
- /* Ignore registers created by GCSE.
- We do this because ... */
- if (regno >= max_gcse_regno)
- continue;
+ if (new == pc_rtx)
+ {
+ edest = FALLTHRU_EDGE (bb);
+ dest = edest->insns.r ? NULL : edest->dest;
+ }
+ else if (GET_CODE (new) == LABEL_REF)
+ {
+ dest = BLOCK_FOR_INSN (XEXP (new, 0));
+ /* Don't bypass edges containing instructions. */
+ edest = find_edge (bb, dest);
+ if (edest && edest->insns.r)
+ dest = NULL;
+ }
+ else
+ dest = NULL;
- /* 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;
+ /* Avoid unification of the edge with other edges from original
+ branch. We would end up emitting the instruction on "both"
+ edges. */
- /* 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;
+ if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc)))
+ && find_edge (e->src, dest))
+ dest = NULL;
- pat = set->expr;
- /* ??? We might be able to handle PARALLELs. Later. */
- if (GET_CODE (pat) != SET)
- abort ();
+ old_dest = e->dest;
+ if (dest != NULL
+ && dest != old_dest
+ && dest != EXIT_BLOCK_PTR)
+ {
+ redirect_edge_and_branch_force (e, dest);
- src = SET_SRC (pat);
+ /* 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);
+ }
- /* 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");
+ 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);
}
- if (INSN_DELETED_P (insn))
- return 1;
- }
- }
- else if (GET_CODE (src) == REG
- && REGNO (src) >= FIRST_PSEUDO_REGISTER
- && REGNO (src) != regno)
- {
- if (try_replace_reg (reg_used->reg_rtx, src, insn))
- {
- changed = 1;
- copy_prop_count++;
- if (gcse_file != NULL)
- {
- 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));
- }
-
- /* 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. */
+ change = 1;
+ removed_p = 1;
+ break;
}
}
+ if (!removed_p)
+ ei_next (&ei);
}
-
- return changed;
+ return change;
}
-/* 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. */
+/* 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.
-static void
-local_cprop_find_used_regs (rtx *xptr, void *data)
-{
- rtx x = *xptr;
+ This function is now mis-named, because we also handle indirect jumps. */
- if (x == 0)
- return;
+static int
+bypass_conditional_jumps (void)
+{
+ basic_block bb;
+ int changed;
+ rtx setcc;
+ rtx insn;
+ rtx dest;
- switch (GET_CODE (x))
- {
- case ZERO_EXTRACT:
- case SIGN_EXTRACT:
- case STRICT_LOW_PART:
- return;
+ /* Note we start at block 1. */
+ if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
+ return 0;
- 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;
+ bypass_last_basic_block = last_basic_block;
+ mark_dfs_back_edges ();
- 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;
+ 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 (!single_pred_p (bb))
+ {
+ setcc = NULL_RTX;
+ for (insn = BB_HEAD (bb);
+ insn != NULL && insn != NEXT_INSN (BB_END (bb));
+ insn = NEXT_INSN (insn))
+ if (NONJUMP_INSN_P (insn))
+ {
+ if (setcc)
+ break;
+ if (GET_CODE (PATTERN (insn)) != SET)
+ break;
- default:
- break;
+ dest = SET_DEST (PATTERN (insn));
+ if (REG_P (dest) || CC0_P (dest))
+ setcc = insn;
+ else
+ break;
+ }
+ else if (JUMP_P (insn))
+ {
+ if ((any_condjump_p (insn) || computed_jump_p (insn))
+ && onlyjump_p (insn))
+ changed |= bypass_block (bb, setcc, insn);
+ break;
+ }
+ else if (INSN_P (insn))
+ break;
+ }
}
- find_used_regs (xptr, data);
+ /* 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. */
-/* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
- their REG_EQUAL notes need updating. */
+/* Local properties of expressions. */
+/* Nonzero for expressions that are transparent in the block. */
+static sbitmap *transp;
-static bool
-do_local_cprop (rtx x, rtx insn, int alter_jumps, rtx *libcall_sp)
-{
- rtx newreg = NULL, newcnst = NULL;
+/* 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;
- /* 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;
+/* Nonzero for expressions that are computed (available) in the block. */
+static sbitmap *comp;
- if (!val)
- return false;
- for (l = val->locs; l; l = l->next)
- {
- rtx this_rtx = l->loc;
- rtx note;
+/* Nonzero for expressions that are locally anticipatable in the block. */
+static sbitmap *antloc;
- if (l->in_libcall)
- continue;
+/* Nonzero for expressions where this block is an optimal computation
+ point. */
+static sbitmap *pre_optimal;
- 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;
-}
+/* Nonzero for expressions which are redundant in a particular block. */
+static sbitmap *pre_redundant;
-/* 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)
-{
- rtx end;
+/* Nonzero for expressions which should be inserted on a specific edge. */
+static sbitmap *pre_insert_map;
- while ((end = *libcall_sp++))
- {
- rtx note = find_reg_equal_equiv_note (end);
+/* Nonzero for expressions which should be deleted in a specific block. */
+static sbitmap *pre_delete_map;
- if (! note)
- continue;
+/* Contains the edge_list returned by pre_edge_lcm. */
+static struct edge_list *edge_list;
- 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;
-}
+/* Redundant insns. */
+static sbitmap pre_redundant_insns;
-#define MAX_NESTED_LIBCALLS 9
+/* Allocate vars used for PRE analysis. */
static void
-local_cprop_pass (int alter_jumps)
+alloc_pre_mem (int n_blocks, int n_exprs)
{
- rtx insn;
- struct reg_use *reg_used;
- rtx libcall_stack[MAX_NESTED_LIBCALLS + 1], *libcall_sp;
- bool changed = false;
+ transp = sbitmap_vector_alloc (n_blocks, n_exprs);
+ comp = sbitmap_vector_alloc (n_blocks, n_exprs);
+ antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
- cselib_init (false);
- libcall_sp = &libcall_stack[MAX_NESTED_LIBCALLS];
- *libcall_sp = 0;
- for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
- {
- if (INSN_P (insn))
- {
- rtx note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
+ pre_optimal = NULL;
+ pre_redundant = NULL;
+ pre_insert_map = NULL;
+ pre_delete_map = NULL;
+ ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
- 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);
+ /* pre_insert and pre_delete are allocated later. */
+}
- 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)
- {
- delete_unreachable_blocks ();
- free_reg_set_mem ();
- alloc_reg_set_mem (max_reg_num ());
- compute_sets (get_insns ());
- }
+/* 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);
+
+ transp = comp = NULL;
+ pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
}
-/* Forward propagate copies. This includes copies and constants. Return
- nonzero if a change was made. */
+/* Top level routine to do the dataflow analysis needed by PRE. */
-static int
-cprop (int alter_jumps)
+static void
+compute_pre_data (void)
{
- int changed;
+ sbitmap trapping_expr;
basic_block bb;
- rtx insn;
+ unsigned int ui;
- /* Note we start at block 1. */
- if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
+ 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++)
{
- if (gcse_file != NULL)
- fprintf (gcse_file, "\n");
- return 0;
+ 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);
}
- changed = 0;
- FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
+ /* Compute ae_kill for each basic block using:
+
+ ~(TRANSP | COMP)
+ */
+
+ 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 (bb);
- insn != NULL && insn != NEXT_INSN (BB_END (bb));
- 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 (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 */
-/* Similar to get_condition, only the resulting condition must be
- valid at JUMP, instead of at EARLIEST.
+/* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
+ block BB.
- 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. */
+ VISITED is a pointer to a working buffer for tracking which BB's have
+ been visited. It is NULL for the top-level call.
-rtx
-fis_get_condition (rtx jump)
+ 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)
{
- rtx cond, set, tmp, insn, earliest;
- bool reverse;
+ edge pred;
+ edge_iterator ei;
+
+ FOR_EACH_EDGE (pred, ei, bb->preds)
+ {
+ 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. */
- if (! any_condjump_p (jump))
- return NULL_RTX;
+ /* 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;
- set = pc_set (jump);
- cond = XEXP (SET_SRC (set), 0);
+ 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;
- /* 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));
+ /* 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;
+ }
+ }
- /* 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,
- false);
- if (!tmp)
- return NULL_RTX;
+ /* All paths have been checked. */
+ return 0;
+}
- /* 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 wrapper for pre_expr_reaches_here_work that ensures that any
+ memory allocated for that function is returned. */
- /* 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,
- false);
- if (!tmp)
- return NULL_RTX;
+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);
- /* 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;
+ rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
- return tmp;
+ free (visited);
+ return rval;
}
+\f
-/* Check the comparison COND to see if we can safely form an implicit set from
- it. COND is either an EQ or NE comparison. */
+/* 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 bool
-implicit_set_cond_p (rtx cond)
+static rtx
+process_insert_insn (struct expr *expr)
{
- enum machine_mode mode = GET_MODE (XEXP (cond, 0));
- rtx cst = XEXP (cond, 1);
+ rtx reg = expr->reaching_reg;
+ rtx exp = copy_rtx (expr->expr);
+ rtx pat;
- /* We can't perform this optimization if either operand might be or might
- contain a signed zero. */
- if (HONOR_SIGNED_ZEROS (mode))
+ 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
{
- /* 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;
+ rtx insn = emit_insn (gen_rtx_SET (VOIDmode, reg, exp));
+
+ if (insn_invalid_p (insn))
+ gcc_unreachable ();
}
+
- return gcse_constant_p (cst);
+ pat = get_insns ();
+ end_sequence ();
+
+ return pat;
}
-/* 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. */
+/* Add EXPR to the end of basic block BB.
-static void
-find_implicit_sets (void)
-{
- basic_block bb, dest;
- unsigned int count;
- rtx cond, new;
-
- count = 0;
- FOR_EACH_BB (bb)
- /* Check for more than one successor. */
- if (bb->succ && bb->succ->succ_next)
- {
- cond = fis_get_condition (BB_END (bb));
-
- if (cond
- && (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
- && GET_CODE (XEXP (cond, 0)) == REG
- && REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
- && implicit_set_cond_p (cond))
- {
- dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
- : FALLTHRU_EDGE (bb)->dest;
-
- if (dest && ! 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 (gcse_file)
- fprintf (gcse_file, "Found %d implicit sets\n", count);
-}
+ This is used by both the PRE and code hoisting.
-/* 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. */
+ 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 int
-one_cprop_pass (int pass, int cprop_jumps, int bypass_jumps)
+static void
+insert_insn_end_bb (struct expr *expr, basic_block bb, int pre)
{
- int changed = 0;
+ rtx insn = BB_END (bb);
+ rtx new_insn;
+ rtx reg = expr->reaching_reg;
+ int regno = REGNO (reg);
+ rtx pat, pat_end;
- const_prop_count = 0;
- copy_prop_count = 0;
+ pat = process_insert_insn (expr);
+ gcc_assert (pat && INSN_P (pat));
- local_cprop_pass (cprop_jumps);
+ pat_end = pat;
+ while (NEXT_INSN (pat_end) != NULL_RTX)
+ pat_end = NEXT_INSN (pat_end);
- /* Determine implicit sets. */
- implicit_sets = xcalloc (last_basic_block, sizeof (rtx));
- find_implicit_sets ();
+ /* 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. */
- alloc_hash_table (max_cuid, &set_hash_table, 1);
- compute_hash_table (&set_hash_table);
+ if (JUMP_P (insn)
+ || (NONJUMP_INSN_P (insn)
+ && (!single_succ_p (bb)
+ || single_succ_edge (bb)->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. */
+ gcc_assert (!NONJUMP_INSN_P (insn) || !pre
+ || TEST_BIT (antloc[bb->index], expr->bitmap_index)
+ || TEST_BIT (transp[bb->index], expr->bitmap_index));
- /* Free implicit_sets before peak usage. */
- free (implicit_sets);
- implicit_sets = NULL;
+ /* 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);
- if (gcse_file)
- dump_hash_table (gcse_file, "SET", &set_hash_table);
- if (set_hash_table.n_elems > 0)
- {
- alloc_cprop_mem (last_basic_block, set_hash_table.n_elems);
- compute_cprop_data ();
- changed = cprop (cprop_jumps);
- if (bypass_jumps)
- changed |= bypass_conditional_jumps ();
- free_cprop_mem ();
+#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_noloc (pat, insn);
}
- free_hash_table (&set_hash_table);
-
- if (gcse_file)
+ /* 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))
{
- 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 ();
+ /* 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.
- return changed;
-}
-\f
-/* Bypass conditional jumps. */
+ It should always be the case that we can put these instructions
+ anywhere in the basic block with performing PRE optimizations.
+ Check this. */
-/* 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. */
+ gcc_assert (!pre
+ || TEST_BIT (antloc[bb->index], expr->bitmap_index)
+ || TEST_BIT (transp[bb->index], expr->bitmap_index));
-static int bypass_last_basic_block;
+ /* 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));
-/* 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. */
+ /* If we found all the parameter loads, then we want to insert
+ before the first parameter load.
-static struct expr *
-find_bypass_set (int regno, int bb)
-{
- struct expr *result = 0;
+ 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);
- for (;;)
- {
- rtx src;
- struct expr *set = lookup_set (regno, &set_hash_table);
+ new_insn = emit_insn_before_noloc (pat, insn);
+ }
+ else
+ new_insn = emit_insn_after_noloc (pat, insn);
- while (set)
+ while (1)
+ {
+ if (INSN_P (pat))
{
- if (TEST_BIT (cprop_avout[bb], set->bitmap_index))
- break;
- set = next_set (regno, set);
+ add_label_notes (PATTERN (pat), new_insn);
+ note_stores (PATTERN (pat), record_set_info, pat);
}
-
- if (set == 0)
+ if (pat == pat_end)
break;
+ pat = NEXT_INSN (pat);
+ }
- if (GET_CODE (set->expr) != SET)
- abort ();
-
- src = SET_SRC (set->expr);
- if (gcse_constant_p (src))
- result = set;
-
- if (GET_CODE (src) != REG)
- break;
+ gcse_create_count++;
- regno = REGNO (src);
+ 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);
}
- return result;
}
+/* Insert partially redundant expressions on edges in the CFG to make
+ the expressions fully redundant. */
-/* 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 (rtx reg, edge e)
+static int
+pre_edge_insert (struct edge_list *edge_list, struct expr **index_map)
{
- rtx insn;
-
- for (insn = e->insns; insn; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && reg_set_p (reg, insn))
- return true;
+ int e, i, j, num_edges, set_size, did_insert = 0;
+ sbitmap *inserted;
- return false;
-}
+ /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
+ if it reaches any of the deleted expressions. */
-/* 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.
+ 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);
- 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. */
+ for (e = 0; e < num_edges; e++)
+ {
+ int indx;
+ basic_block bb = INDEX_EDGE_PRED_BB (edge_list, e);
-static int
-bypass_block (basic_block bb, rtx setcc, rtx jump)
-{
- rtx insn, note;
- edge e, enext, edest;
- int i, change;
- int may_be_loop_header;
+ for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
+ {
+ SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
- insn = (setcc != NULL) ? setcc : jump;
+ 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;
- /* 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);
+ /* Now look at each deleted occurrence of this expression. */
+ for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
+ {
+ if (! occr->deleted_p)
+ continue;
- 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;
- }
-
- change = 0;
- for (e = bb->pred; e; e = enext)
- {
- 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;
-
- 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;
-
- 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)));
+ /* 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);
- new = simplify_replace_rtx (src, reg_used->reg_rtx,
- SET_SRC (set->expr));
+ /* 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. */
- /* 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 (eg->flags & 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 (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;
+ 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);
+ }
- /* Avoid unification of the edge with other edges from original
- branch. We would end up emitting the instruction on "both"
- edges. */
-
- if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc))))
- {
- edge e2;
- for (e2 = e->src->succ; e2; e2 = e2->succ_next)
- if (e2->dest == dest)
- {
- dest = NULL;
- break;
+ update_ld_motion_stores (expr);
+ SET_BIT (inserted[e], j);
+ did_insert = 1;
+ gcse_create_count++;
+ }
}
- }
-
- 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 (bb);
- insn != NULL && insn != NEXT_INSN (BB_END (bb));
- 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;
+ sbitmap_vector_free (inserted);
+ return did_insert;
}
-\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. */
+/* 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
-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 (bb);
- rtx new_insn;
- rtx reg = expr->reaching_reg;
- int regno = REGNO (reg);
- rtx pat, pat_end;
-
- pat = process_insert_insn (expr);
- if (pat == NULL_RTX || ! INSN_P (pat))
- abort ();
-
- 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 (bb));
-
- /* 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)
- {
- 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 EXPR->EXPR generated by INSN to EXPR->REACHING_REG.
- Given "old_reg <- expr" (INSN), instead of adding after it
- reaching_reg <- old_reg
- it's better to do the following:
- reaching_reg <- expr
- old_reg <- reaching_reg
- because this way copy propagation can discover additional PRE
- opportunities. But if this fails, we try the old way.
- When "expr" is a store, i.e.
- given "MEM <- old_reg", instead of adding after it
- reaching_reg <- old_reg
- it's better to add it before as follows:
- reaching_reg <- old_reg
- MEM <- reaching_reg. */
-
-static void
-pre_insert_copy_insn (struct expr *expr, rtx insn)
+pre_insert_copy_insn (struct expr *expr, rtx insn)
{
rtx reg = expr->reaching_reg;
int regno = REGNO (reg);
rtx pat = PATTERN (insn);
rtx set, new_insn;
rtx old_reg;
- int i;
-
- /* This block matches the logic in hash_scan_insn. */
- if (GET_CODE (pat) == SET)
- set = pat;
- else if (GET_CODE (pat) == PARALLEL)
- {
- /* Search through the parallel looking for the set whose
- source was the expression that we're interested in. */
- set = NULL_RTX;
- for (i = 0; i < XVECLEN (pat, 0); i++)
- {
- rtx x = XVECEXP (pat, 0, i);
- if (GET_CODE (x) == SET
- && expr_equiv_p (SET_SRC (x), expr->expr))
- {
- set = x;
- break;
- }
- }
- }
- else
- abort ();
-
- if (GET_CODE (SET_DEST (set)) == REG)
- {
- 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);
-
- /* Keep register set table up to date. */
- replace_one_set (REGNO (old_reg), insn, new_insn);
- record_one_set (regno, insn);
- }
- else
- {
- new_insn = gen_move_insn (reg, old_reg);
- new_insn = emit_insn_after (new_insn, insn);
-
- /* Keep register set table up to date. */
- record_one_set (regno, new_insn);
- }
- }
- else /* This is possible only in case of a store to memory. */
- {
- old_reg = SET_SRC (set);
- new_insn = gen_move_insn (reg, old_reg);
-
- /* Check if we can modify the set source in the original insn. */
- if (validate_change (insn, &SET_SRC (set), reg, 0))
- new_insn = emit_insn_before (new_insn, insn);
- else
- new_insn = emit_insn_after (new_insn, insn);
-
- /* Keep register set table up to date. */
- record_one_set (regno, new_insn);
- }
-
- 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);
-}
-
-/* Copy available expressions that reach the redundant expression
- to `reaching_reg'. */
-
-static void
-pre_insert_copies (void)
-{
- unsigned int i, added_copy;
- 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;
-
- /* Set when we add a copy for that expression. */
- added_copy = 0;
-
- for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
- {
- if (! occr->deleted_p)
- continue;
-
- 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;
-
- added_copy = 1;
-
- /* Copy the result of avail to reaching_reg. */
- pre_insert_copy_insn (expr, insn);
- avail->copied_p = 1;
- }
- }
-
- if (added_copy)
- update_ld_motion_stores (expr);
- }
-}
-
-/* Emit move from SRC to DEST noting the equivalence with expression computed
- in INSN. */
-static rtx
-gcse_emit_move_after (rtx src, rtx dest, rtx insn)
-{
- rtx new;
- rtx 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);
-
- /* We only delete insns that have a single_set. */
- if (TEST_BIT (pre_delete_map[bb->index], indx)
- && (set = single_set (insn)) != 0)
- {
- /* 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 (bb)) != 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. */
+ int i;
-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)
+ /* This block matches the logic in hash_scan_insn. */
+ switch (GET_CODE (pat))
{
- rtx insn, stop_insn;
-
- /* Set the current block for invalidate_nonnull_info. */
- npi->current_block = current_block;
+ case SET:
+ set = pat;
+ break;
- /* Scan each insn in the basic block looking for memory references and
- register sets. */
- stop_insn = NEXT_INSN (BB_END (current_block));
- for (insn = BB_HEAD (current_block);
- insn != stop_insn;
- insn = NEXT_INSN (insn))
+ case PARALLEL:
+ /* Search through the parallel looking for the set whose
+ source was the expression that we're interested in. */
+ set = NULL_RTX;
+ for (i = 0; i < XVECLEN (pat, 0); i++)
{
- rtx 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)
+ rtx x = XVECEXP (pat, 0, i);
+ if (GET_CODE (x) == SET
+ && expr_equiv_p (SET_SRC (x), expr->expr))
{
- note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
- continue;
+ set = x;
+ break;
}
-
- /* 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);
}
- }
+ break;
- /* 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);
+ default:
+ gcc_unreachable ();
+ }
- /* Now look at each bb and see if it ends with a compare of a value
- against zero. */
- FOR_EACH_BB (bb)
+ if (REG_P (SET_DEST (set)))
{
- rtx last_insn = BB_END (bb);
- rtx condition, earliest;
- int compare_and_branch;
-
- /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
- since BLOCK_REG[BB] is zero if this block did not end with a
- comparison against zero, this condition works. */
- if (block_reg[bb->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, false);
-
- /* 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;
+ 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);
- /* 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;
+ /* Keep register set table up to date. */
+ record_one_set (regno, insn);
+ }
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);
-#ifdef HAVE_cc0
- if (compare_and_branch == 2)
- delete_insn (earliest);
-#endif
- purge_dead_edges (bb);
+ {
+ new_insn = gen_move_insn (reg, old_reg);
+ new_insn = emit_insn_after (new_insn, insn);
- /* Don't check this block again. (Note that BB_END is
- invalid here; we deleted the last instruction in the
- block.) */
- block_reg[bb->index] = 0;
+ /* Keep register set table up to date. */
+ record_one_set (regno, new_insn);
+ }
}
-
- 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 = max_reg_num ();
- struct null_pointer_info npi;
- int something_changed = 0;
-
- /* If we have only a single block, or it is too expensive, give up. */
- if (n_basic_blocks <= 1
- || is_too_expensive (_ ("NULL pointer checks disabled")))
- return 0;
-
- /* We need four bitmaps, each with a bit for each register in each
- basic block. */
- 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)
+ else /* This is possible only in case of a store to memory. */
{
- rtx last_insn = BB_END (bb);
- rtx condition, earliest, reg;
-
- /* We only want conditional branches. */
- if (GET_CODE (last_insn) != JUMP_INSN
- || !any_condjump_p (last_insn)
- || !onlyjump_p (last_insn))
- continue;
-
- /* LAST_INSN is a conditional jump. Get its condition. */
- condition = get_condition (last_insn, &earliest, false);
-
- /* 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;
+ old_reg = SET_SRC (set);
+ new_insn = gen_move_insn (reg, old_reg);
- block_reg[bb->index] = REGNO (reg);
- }
+ /* Check if we can modify the set source in the original insn. */
+ if (validate_change (insn, &SET_SRC (set), reg, 0))
+ new_insn = emit_insn_before (new_insn, insn);
+ else
+ new_insn = emit_insn_after (new_insn, insn);
- /* 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);
+ /* Keep register set table up to date. */
+ record_one_set (regno, new_insn);
}
- /* 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);
+ gcse_create_count++;
- return something_changed;
+ 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);
}
-/* Code Hoisting variables and subroutines. */
+/* Copy available expressions that reach the redundant expression
+ to `reaching_reg'. */
-/* Very busy expressions. */
-static sbitmap *hoist_vbein;
-static sbitmap *hoist_vbeout;
+static void
+pre_insert_copies (void)
+{
+ unsigned int i, added_copy;
+ struct expr *expr;
+ struct occr *occr;
+ struct occr *avail;
-/* Hoistable expressions. */
-static sbitmap *hoist_exprs;
+ /* For each available expression in the table, copy the result to
+ `reaching_reg' if the expression reaches a deleted one.
-/* ??? 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.
+ ??? The current algorithm is rather brute force.
+ Need to do some profiling. */
- It's unclear if tail merging could be run in parallel with
- code hoisting. It would be nice. */
+ 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;
-/* Allocate vars used for code hoisting analysis. */
+ /* Set when we add a copy for that expression. */
+ added_copy = 0;
-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);
+ for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
+ {
+ if (! occr->deleted_p)
+ continue;
- 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);
-}
+ for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
+ {
+ rtx insn = avail->insn;
-/* Free vars used for code hoisting analysis. */
+ /* No need to handle this one if handled already. */
+ if (avail->copied_p)
+ continue;
-static void
-free_code_hoist_mem (void)
-{
- sbitmap_vector_free (antloc);
- sbitmap_vector_free (transp);
- sbitmap_vector_free (comp);
+ /* Don't handle this one if it's a redundant one. */
+ if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
+ continue;
- sbitmap_vector_free (hoist_vbein);
- sbitmap_vector_free (hoist_vbeout);
- sbitmap_vector_free (hoist_exprs);
- sbitmap_vector_free (transpout);
+ /* 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;
- free_dominance_info (CDI_DOMINATORS);
-}
+ added_copy = 1;
-/* Compute the very busy expressions at entry/exit from each block.
+ /* Copy the result of avail to reaching_reg. */
+ pre_insert_copy_insn (expr, insn);
+ avail->copied_p = 1;
+ }
+ }
- An expression is very busy if all paths from a given point
- compute the expression. */
+ if (added_copy)
+ update_ld_motion_stores (expr);
+ }
+}
-static void
-compute_code_hoist_vbeinout (void)
+/* 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)
{
- 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;
+ rtx new;
+ rtx set = single_set (insn), set2;
+ rtx note;
+ rtx eqv;
- /* 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);
- }
+ /* This should never fail since we're creating a reg->reg copy
+ we've verified to be valid. */
- passes++;
- }
+ new = emit_insn_after (gen_move_insn (dest, src), insn);
- if (gcse_file)
- fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
-}
+ /* 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);
-/* Top level routine to do the dataflow analysis needed by code hoisting. */
+ set_unique_reg_note (new, REG_EQUAL, copy_insn_1 (eqv));
-static void
-compute_code_hoist_data (void)
-{
- compute_local_properties (transp, comp, antloc, &expr_hash_table);
- compute_transpout ();
- compute_code_hoist_vbeinout ();
- calculate_dominance_info (CDI_DOMINATORS);
- if (gcse_file)
- fprintf (gcse_file, "\n");
+ return new;
}
-/* 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.
+/* 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.
- Contrast this to reachability for PRE where an expression is
- considered reachable if *any* path reaches instead of *all*
- paths. */
+ Returns nonzero if a change is made. */
static int
-hoist_expr_reaches_here_p (basic_block expr_bb, int expr_index, basic_block bb, char *visited)
+pre_delete (void)
{
- edge pred;
- int visited_allocated_locally = 0;
+ 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;
- if (visited == NULL)
- {
- visited_allocated_locally = 1;
- visited = xcalloc (last_basic_block, 1);
- }
+ /* We only need to search antic_occr since we require
+ ANTLOC != 0. */
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
- {
- basic_block pred_bb = pred->src;
+ for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
+ {
+ rtx insn = occr->insn;
+ rtx set;
+ basic_block bb = BLOCK_FOR_INSN (insn);
- if (pred->src == ENTRY_BLOCK_PTR)
- break;
- else if (pred_bb == expr_bb)
- continue;
- else if (visited[pred_bb->index])
- continue;
+ /* We only delete insns that have a single_set. */
+ if (TEST_BIT (pre_delete_map[bb->index], indx)
+ && (set = single_set (insn)) != 0)
+ {
+ /* 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)));
- /* 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;
+ 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++;
- /* 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);
+ 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 (pred == NULL);
+ return changed;
}
-\f
-/* Actually perform code hoisting. */
-static void
-hoist_code (void)
+/* 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)
{
- basic_block bb, dominated;
- basic_block *domby;
- unsigned int domby_len;
- unsigned int i,j;
+ unsigned int i;
+ int did_insert, changed;
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. */
for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
index_map[expr->bitmap_index] = expr;
- /* Walk over each basic block looking for potentially hoistable
- expressions, nothing gets hoisted from the entry block. */
- FOR_EACH_BB (bb)
- {
- int found = 0;
- int insn_inserted_p;
-
- domby_len = get_dominated_by (CDI_DOMINATORS, bb, &domby);
- /* Examine each expression that is very busy at the exit of this
- block. These are the potentially hoistable expressions. */
- for (i = 0; i < hoist_vbeout[bb->index]->n_bits; i++)
- {
- int hoistable = 0;
-
- 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; 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 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;
-
- 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;
- if (!insn_inserted_p)
- {
- insert_insn_end_bb (index_map[i], bb, 0);
- insn_inserted_p = 1;
- }
- }
- }
- }
- }
- free (domby);
+ /* 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 code hoisting (aka unification) pass
+/* Top level routine to perform one PRE GCSE pass.
Return nonzero if a change was made. */
static int
-one_code_hoisting_pass (void)
+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, "Code Hosting Expressions", &expr_hash_table);
+ dump_hash_table (gcse_file, "Expression", &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 ();
+ 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);
+ 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
-/* 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.
-
- int i;
- float a[10];
-
- void foo(float scale)
- {
- for (i=0; i<10; i++)
- a[i] *= scale;
- }
-
- 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
- the load out since its live around the loop, and stored at the bottom
- of the loop.
-
- The 'Load 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.
+/* 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. */
- 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. */
+/* ??? This is very similar to the loop.c add_label_notes function. We
+ could probably share code here. */
-/* This will search the ldst list for a matching expression. If it
- doesn't find one, we create one and initialize it. */
+/* ??? 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 struct ls_expr *
-ldst_entry (rtx x)
+static void
+add_label_notes (rtx x, rtx insn)
{
- int do_not_record_p = 0;
- struct ls_expr * ptr;
- unsigned int hash;
-
- hash = hash_expr_1 (x, GET_MODE (x), & do_not_record_p);
+ enum rtx_code code = GET_CODE (x);
+ int i, j;
+ const char *fmt;
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- if (ptr->hash_index == hash && expr_equiv_p (ptr->pattern, x))
- return ptr;
+ 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.
- ptr = xmalloc (sizeof (struct ls_expr));
+ We no longer ignore such label references (see LABEL_REF handling in
+ mark_jump_label for additional information). */
- 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;
+ 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;
+ }
- return ptr;
+ 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);
+ }
}
-/* Free up an individual ldst entry. */
-
-static void
-free_ldst_entry (struct ls_expr * ptr)
-{
- free_INSN_LIST_list (& ptr->loads);
- free_INSN_LIST_list (& ptr->stores);
+/* Compute transparent outgoing information for each block.
- free (ptr);
-}
+ 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.
-/* Free up all memory associated with the ldst list. */
+ 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
-free_ldst_mems (void)
+compute_transpout (void)
{
- while (pre_ldst_mems)
+ basic_block bb;
+ unsigned int i;
+ struct expr *expr;
+
+ sbitmap_vector_ones (transpout, last_basic_block);
+
+ FOR_EACH_BB (bb)
{
- struct ls_expr * tmp = pre_ldst_mems;
+ /* Note that flow inserted a nop a the end of basic blocks that
+ end in call instructions for reasons other than abnormal
+ control flow. */
+ if (! CALL_P (BB_END (bb)))
+ continue;
- pre_ldst_mems = pre_ldst_mems->next;
+ 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;
- free_ldst_entry (tmp);
+ /* ??? 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);
+ }
}
-
- pre_ldst_mems = NULL;
}
-/* Dump debugging info about the ldst list. */
+/* Code Hoisting variables and subroutines. */
-static void
-print_ldst_list (FILE * file)
-{
- struct ls_expr * ptr;
+/* Very busy expressions. */
+static sbitmap *hoist_vbein;
+static sbitmap *hoist_vbeout;
- fprintf (file, "LDST list: \n");
+/* Hoistable expressions. */
+static sbitmap *hoist_exprs;
- for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
- {
- fprintf (file, " Pattern (%3d): ", ptr->index);
+/* ??? 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.
- print_rtl (file, ptr->pattern);
+ It's unclear if tail merging could be run in parallel with
+ code hoisting. It would be nice. */
- fprintf (file, "\n Loads : ");
+/* Allocate vars used for code hoisting analysis. */
- if (ptr->loads)
- print_rtl (file, ptr->loads);
- else
- fprintf (file, "(nil)");
+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);
- fprintf (file, "\n Stores : ");
+ 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);
+}
- if (ptr->stores)
- print_rtl (file, ptr->stores);
- else
- fprintf (file, "(nil)");
+/* Free vars used for code hoisting analysis. */
- fprintf (file, "\n\n");
- }
+static void
+free_code_hoist_mem (void)
+{
+ sbitmap_vector_free (antloc);
+ sbitmap_vector_free (transp);
+ sbitmap_vector_free (comp);
- fprintf (file, "\n");
+ sbitmap_vector_free (hoist_vbein);
+ sbitmap_vector_free (hoist_vbeout);
+ sbitmap_vector_free (hoist_exprs);
+ sbitmap_vector_free (transpout);
+
+ free_dominance_info (CDI_DOMINATORS);
}
-/* Returns 1 if X is in the list of ldst only expressions. */
+/* Compute the very busy expressions at entry/exit from each block.
-static struct ls_expr *
-find_rtx_in_ldst (rtx x)
+ An expression is very busy if all paths from a given point
+ compute the expression. */
+
+static void
+compute_code_hoist_vbeinout (void)
{
- struct ls_expr * ptr;
+ int changed, passes;
+ basic_block bb;
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- if (expr_equiv_p (ptr->pattern, x) && ! ptr->invalid)
- return ptr;
+ sbitmap_vector_zero (hoist_vbeout, last_basic_block);
+ sbitmap_vector_zero (hoist_vbein, last_basic_block);
- return NULL;
-}
+ passes = 0;
+ changed = 1;
-/* Assign each element of the list of mems a monotonically increasing value. */
+ while (changed)
+ {
+ changed = 0;
-static int
-enumerate_ldsts (void)
-{
- struct ls_expr * ptr;
- int n = 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);
+ }
- for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
- ptr->index = n++;
+ passes++;
+ }
- return n;
+ if (gcse_file)
+ fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
}
-/* Return first item in the list. */
+/* Top level routine to do the dataflow analysis needed by code hoisting. */
-static inline struct ls_expr *
-first_ls_expr (void)
+static void
+compute_code_hoist_data (void)
{
- return pre_ldst_mems;
+ compute_local_properties (transp, comp, antloc, &expr_hash_table);
+ compute_transpout ();
+ compute_code_hoist_vbeinout ();
+ calculate_dominance_info (CDI_DOMINATORS);
+ if (gcse_file)
+ fprintf (gcse_file, "\n");
}
-/* Return the next item in the list after the specified one. */
+/* Determine if the expression identified by EXPR_INDEX would
+ reach BB unimpared if it was placed at the end of EXPR_BB.
-static inline struct ls_expr *
-next_ls_expr (struct ls_expr * ptr)
-{
- return ptr->next;
-}
-\f
-/* Load Motion for loads which only kill themselves. */
+ 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.
-/* 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. */
+ Contrast this to reachability for PRE where an expression is
+ considered reachable if *any* path reaches instead of *all*
+ paths. */
static int
-simple_mem (rtx x)
+hoist_expr_reaches_here_p (basic_block expr_bb, int expr_index, basic_block bb, char *visited)
{
- if (GET_CODE (x) != MEM)
- return 0;
+ edge pred;
+ edge_iterator ei;
+ int visited_allocated_locally = 0;
- if (MEM_VOLATILE_P (x))
- return 0;
- if (GET_MODE (x) == BLKmode)
- return 0;
+ if (visited == NULL)
+ {
+ visited_allocated_locally = 1;
+ visited = xcalloc (last_basic_block, 1);
+ }
- /* 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_EDGE (pred, ei, bb->preds)
+ {
+ basic_block pred_bb = pred->src;
- if (side_effects_p (x))
- return 0;
+ if (pred->src == ENTRY_BLOCK_PTR)
+ break;
+ else if (pred_bb == expr_bb)
+ continue;
+ else if (visited[pred_bb->index])
+ continue;
- /* Do not consider function arguments passed on stack. */
- if (reg_mentioned_p (stack_pointer_rtx, x))
- return 0;
+ /* 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 (flag_float_store && FLOAT_MODE_P (GET_MODE (x)))
- return 0;
+ /* 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);
- return 1;
+ return (pred == NULL);
}
-
-/* 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. */
+\f
+/* Actually perform code hoisting. */
static void
-invalidate_any_buried_refs (rtx x)
+hoist_code (void)
{
- const char * fmt;
- int i, j;
- struct ls_expr * ptr;
-
- /* Invalidate it in the list. */
- if (GET_CODE (x) == MEM && simple_mem (x))
- {
- ptr = ldst_entry (x);
- ptr->invalid = 1;
- }
-
- /* Recursively process the insn. */
- fmt = GET_RTX_FORMAT (GET_CODE (x));
-
- for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
- {
- if (fmt[i] == 'e')
- 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));
- }
-}
+ basic_block bb, dominated;
+ basic_block *domby;
+ unsigned int domby_len;
+ unsigned int i,j;
+ struct expr **index_map;
+ struct expr *expr;
-/* 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. */
+ sbitmap_vector_zero (hoist_exprs, last_basic_block);
-static void
-compute_ld_motion_mems (void)
-{
- struct ls_expr * ptr;
- basic_block bb;
- rtx insn;
+ /* Compute a mapping from expression number (`bitmap_index') to
+ hash table entry. */
- pre_ldst_mems = NULL;
+ 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;
+ /* Walk over each basic block looking for potentially hoistable
+ expressions, nothing gets hoisted from the entry block. */
FOR_EACH_BB (bb)
{
- for (insn = BB_HEAD (bb);
- insn && insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
+ int found = 0;
+ int insn_inserted_p;
+
+ domby_len = get_dominated_by (CDI_DOMINATORS, bb, &domby);
+ /* Examine each expression that is very busy at the exit of this
+ block. These are the potentially hoistable expressions. */
+ for (i = 0; i < hoist_vbeout[bb->index]->n_bits; i++)
{
- if (INSN_P (insn))
+ int hoistable = 0;
+
+ if (TEST_BIT (hoist_vbeout[bb->index], i)
+ && TEST_BIT (transpout[bb->index], i))
{
- if (GET_CODE (PATTERN (insn)) == 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++)
{
- rtx src = SET_SRC (PATTERN (insn));
- rtx dest = SET_DEST (PATTERN (insn));
+ 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;
- /* 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 if the expression would reach the dominated block
+ unimpared if it was placed at the end of BB.
- /* 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);
+ 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 (GET_CODE (src) != MEM
- && GET_CODE (src) != ASM_OPERANDS
- /* Check for REG manually since want_to_gcse_p
- returns 0 for all REGs. */
- && can_assign_to_reg_p (src))
- ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
- else
- ptr->invalid = 1;
- }
+ /* 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;
}
- else
- invalidate_any_buried_refs (PATTERN (insn));
}
}
- }
-}
+ /* If we found nothing to hoist, then quit now. */
+ if (! found)
+ {
+ free (domby);
+ continue;
+ }
-/* Remove any references that have been either invalidated or are not in the
- expression list for pre gcse. */
+ /* 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;
-static void
-trim_ld_motion_mems (void)
-{
- struct ls_expr * * last = & pre_ldst_mems;
- struct ls_expr * ptr = pre_ldst_mems;
+ /* 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;
- while (ptr != NULL)
- {
- struct expr * expr;
+ /* 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;
- /* 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;
+ /* 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;
- 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;
+ /* Find the right occurrence of this expression. */
+ while (BLOCK_FOR_INSN (occr->insn) != dominated && occr)
+ occr = occr->next;
- if (expr)
- {
- /* Set the expression field if we are keeping it. */
- ptr->expr = expr;
- last = & ptr->next;
- ptr = ptr->next;
- }
- else
- {
- *last = ptr->next;
- free_ldst_entry (ptr);
- ptr = * last;
+ gcc_assert (occr);
+ insn = occr->insn;
+ set = single_set (insn);
+ gcc_assert (set);
+
+ /* Create a pseudo-reg to store the result of reaching
+ expressions into. Get the mode for the new pseudo
+ from the mode of the original destination pseudo. */
+ if (expr->reaching_reg == NULL)
+ expr->reaching_reg
+ = gen_reg_rtx (GET_MODE (SET_DEST (set)));
+
+ 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);
}
- /* Show the world what we've found. */
- if (gcse_file && pre_ldst_mems != NULL)
- print_ldst_list (gcse_file);
+ free (index_map);
}
-/* 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. */
+/* Top level routine to perform one code hoisting (aka unification) pass
-static void
-update_ld_motion_stores (struct expr * expr)
+ Return nonzero if a change was made. */
+
+static int
+one_code_hoisting_pass (void)
{
- struct ls_expr * mem_ptr;
+ int changed = 0;
- if ((mem_ptr = find_rtx_in_ldst (expr->expr)))
+ 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)
{
- /* 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. */
+ alloc_code_hoist_mem (last_basic_block, expr_hash_table.n_elems);
+ compute_code_hoist_data ();
+ hoist_code ();
+ free_code_hoist_mem ();
+ }
- /* 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_hash_table (&expr_hash_table);
- 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;
+ 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.
- /* If we've already copied it, continue. */
- if (expr->reaching_reg == src)
- continue;
+ int i;
+ float a[10];
- if (gcse_file)
+ void foo(float scale)
{
- 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");
+ for (i=0; i<10; i++)
+ a[i] *= scale;
}
- 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;
-
- /* un-recognize this pattern since it's probably different now. */
- INSN_CODE (insn) = -1;
- gcse_create_count++;
- }
- }
-}
-\f
-/* Store motion code. */
+ '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.
-#define ANTIC_STORE_LIST(x) ((x)->loads)
-#define AVAIL_STORE_LIST(x) ((x)->stores)
-#define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
+ 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.
-/* 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;
+ 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. */
-/* And current insn, for the same routine. */
-static rtx compute_store_table_current_insn;
+/* This will search the ldst list for a matching expression. If it
+ doesn't find one, we create one and initialize it. */
-/* Used in computing the reverse edge graph bit vectors. */
-static sbitmap * st_antloc;
+static struct ls_expr *
+ldst_entry (rtx x)
+{
+ int do_not_record_p = 0;
+ struct ls_expr * ptr;
+ unsigned int hash;
-/* Global holding the number of store expressions we are dealing with. */
-static int num_stores;
+ hash = hash_rtx (x, GET_MODE (x), &do_not_record_p,
+ NULL, /*have_reg_qty=*/false);
-/* Checks to set if we need to mark a register set. Called from
- note_stores. */
+ for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
+ if (ptr->hash_index == hash && expr_equiv_p (ptr->pattern, x))
+ return ptr;
-static void
-reg_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED,
- void *data)
-{
- sbitmap bb_reg = data;
+ ptr = xmalloc (sizeof (struct ls_expr));
- if (GET_CODE (dest) == SUBREG)
- dest = SUBREG_REG (dest);
+ 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;
- if (GET_CODE (dest) == REG)
- {
- regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
- if (bb_reg)
- SET_BIT (bb_reg, REGNO (dest));
- }
+ return ptr;
}
-/* Clear any mark that says that this insn sets dest. Called from
- note_stores. */
+/* Free up an individual ldst entry. */
static void
-reg_clear_last_set (rtx dest, rtx setter ATTRIBUTE_UNUSED,
- void *data)
-{
- int *dead_vec = data;
-
- if (GET_CODE (dest) == SUBREG)
- dest = SUBREG_REG (dest);
+free_ldst_entry (struct ls_expr * ptr)
+{
+ free_INSN_LIST_list (& ptr->loads);
+ free_INSN_LIST_list (& ptr->stores);
- if (GET_CODE (dest) == REG &&
- dead_vec[REGNO (dest)] == INSN_UID (compute_store_table_current_insn))
- dead_vec[REGNO (dest)] = 0;
+ free (ptr);
}
-/* Return zero if some of the registers in list X are killed
- due to set of registers in bitmap REGS_SET. */
+/* Free up all memory associated with the ldst list. */
-static bool
-store_ops_ok (rtx x, int *regs_set)
+static void
+free_ldst_mems (void)
{
- rtx reg;
-
- for (; x; x = XEXP (x, 1))
+ while (pre_ldst_mems)
{
- reg = XEXP (x, 0);
- if (regs_set[REGNO(reg)])
- return false;
+ struct ls_expr * tmp = pre_ldst_mems;
+
+ pre_ldst_mems = pre_ldst_mems->next;
+
+ free_ldst_entry (tmp);
}
- return true;
+ pre_ldst_mems = NULL;
}
-/* Returns a list of registers mentioned in X. */
-static rtx
-extract_mentioned_regs (rtx x)
-{
- return extract_mentioned_regs_helper (x, NULL_RTX);
-}
+/* Dump debugging info about the ldst list. */
-/* Helper for extract_mentioned_regs; ACCUM is used to accumulate used
- registers. */
-static rtx
-extract_mentioned_regs_helper (rtx x, rtx accum)
+static void
+print_ldst_list (FILE * file)
{
- int i;
- enum rtx_code code;
- const char * fmt;
-
- /* Repeat is used to turn tail-recursion into iteration. */
- repeat:
+ struct ls_expr * ptr;
- if (x == 0)
- return accum;
+ fprintf (file, "LDST list: \n");
- code = GET_CODE (x);
- switch (code)
+ for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
{
- case REG:
- return alloc_EXPR_LIST (0, x, accum);
+ fprintf (file, " Pattern (%3d): ", ptr->index);
- case MEM:
- x = XEXP (x, 0);
- goto repeat;
+ print_rtl (file, ptr->pattern);
- case PRE_DEC:
- case PRE_INC:
- case POST_DEC:
- case POST_INC:
- /* We do not run this function with arguments having side effects. */
- abort ();
+ fprintf (file, "\n Loads : ");
- 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;
+ if (ptr->loads)
+ print_rtl (file, ptr->loads);
+ else
+ fprintf (file, "(nil)");
- default:
- break;
- }
+ fprintf (file, "\n Stores : ");
- i = GET_RTX_LENGTH (code) - 1;
- fmt = GET_RTX_FORMAT (code);
+ if (ptr->stores)
+ print_rtl (file, ptr->stores);
+ else
+ fprintf (file, "(nil)");
- for (; i >= 0; i--)
- {
- if (fmt[i] == 'e')
- {
- rtx tem = XEXP (x, i);
+ fprintf (file, "\n\n");
+ }
- /* 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;
- }
+ fprintf (file, "\n");
+}
- accum = extract_mentioned_regs_helper (tem, accum);
- }
- else if (fmt[i] == 'E')
- {
- int j;
+/* Returns 1 if X is in the list of ldst only expressions. */
- for (j = 0; j < XVECLEN (x, i); j++)
- accum = extract_mentioned_regs_helper (XVECEXP (x, i, j), accum);
- }
- }
+static struct ls_expr *
+find_rtx_in_ldst (rtx x)
+{
+ struct ls_expr * ptr;
- return accum;
+ for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
+ if (expr_equiv_p (ptr->pattern, x) && ! ptr->invalid)
+ return ptr;
+
+ return NULL;
}
-/* 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.
+/* Assign each element of the list of mems a monotonically increasing value. */
- The results are stored this way:
+static int
+enumerate_ldsts (void)
+{
+ struct ls_expr * ptr;
+ int n = 0;
- -- 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.
+ for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
+ ptr->index = n++;
- 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.
- */
+ return n;
+}
-static void
-find_moveable_store (rtx insn, int *regs_set_before, int *regs_set_after)
+/* Return first item in the list. */
+
+static inline struct ls_expr *
+first_ls_expr (void)
{
- struct ls_expr * ptr;
- rtx dest, set, tmp;
- int check_anticipatable, check_available;
- basic_block bb = BLOCK_FOR_INSN (insn);
+ return pre_ldst_mems;
+}
- set = single_set (insn);
- if (!set)
- return;
+/* Return the next item in the list after the specified one. */
- dest = SET_DEST (set);
+static inline struct ls_expr *
+next_ls_expr (struct ls_expr * ptr)
+{
+ return ptr->next;
+}
+\f
+/* Load Motion for loads which only kill themselves. */
- if (GET_CODE (dest) != MEM || MEM_VOLATILE_P (dest)
- || GET_MODE (dest) == BLKmode)
- return;
+/* 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. */
- if (side_effects_p (dest))
- return;
+static int
+simple_mem (rtx x)
+{
+ if (! MEM_P (x))
+ return 0;
+
+ if (MEM_VOLATILE_P (x))
+ return 0;
+
+ if (GET_MODE (x) == BLKmode)
+ 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 (dest))
- return;
+ if (flag_non_call_exceptions && may_trap_p (x))
+ return 0;
- /* Even if the destination cannot trap, the source may. In this case we'd
- need to handle updating the REG_EH_REGION note. */
- if (find_reg_note (insn, REG_EH_REGION, NULL_RTX))
- return;
+ if (side_effects_p (x))
+ return 0;
- ptr = ldst_entry (dest);
- if (!ptr->pattern_regs)
- ptr->pattern_regs = extract_mentioned_regs (dest);
+ /* Do not consider function arguments passed on stack. */
+ if (reg_mentioned_p (stack_pointer_rtx, x))
+ return 0;
- /* 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));
- }
+ if (flag_float_store && FLOAT_MODE_P (GET_MODE (x)))
+ return 0;
- /* 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
+ return 1;
+}
+
+/* Make sure there isn't a buried reference in this pattern anywhere.
+ If there is, invalidate the entry for it since we're not capable
+ of fixing it up just yet.. We have to be sure we know about ALL
+ 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
+invalidate_any_buried_refs (rtx x)
+{
+ const char * fmt;
+ int i, j;
+ struct ls_expr * ptr;
+
+ /* Invalidate it in the list. */
+ if (MEM_P (x) && simple_mem (x))
{
- tmp = XEXP (AVAIL_STORE_LIST (ptr), 0);
- if (BLOCK_FOR_INSN (tmp) != bb)
- check_available = 1;
+ ptr = ldst_entry (x);
+ ptr->invalid = 1;
}
- if (check_available)
+
+ /* Recursively process the insn. */
+ fmt = GET_RTX_FORMAT (GET_CODE (x));
+
+ for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
{
- /* Check that we have already reached the insn at that the check
- failed last time. */
- if (LAST_AVAIL_CHECK_FAILURE (ptr))
- {
- for (tmp = BB_END (bb);
- tmp != insn && tmp != LAST_AVAIL_CHECK_FAILURE (ptr);
- tmp = PREV_INSN (tmp))
- continue;
- if (tmp == insn)
- check_available = 0;
- }
- else
- check_available = store_killed_after (dest, ptr->pattern_regs, insn,
- bb, regs_set_after,
- &LAST_AVAIL_CHECK_FAILURE (ptr));
+ if (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));
}
- if (!check_available)
- AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn, AVAIL_STORE_LIST (ptr));
}
-/* Find available and anticipatable stores. */
+/* 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
-compute_store_table (void)
+static void
+compute_ld_motion_mems (void)
{
- int ret;
+ struct ls_expr * ptr;
basic_block bb;
- unsigned regno;
- rtx insn, pat, tmp;
- int *last_set_in, *already_set;
- struct ls_expr * ptr, **prev_next_ptr_ptr;
-
- max_gcse_regno = max_reg_num ();
+ rtx insn;
- reg_set_in_block = sbitmap_vector_alloc (last_basic_block,
- max_gcse_regno);
- sbitmap_vector_zero (reg_set_in_block, last_basic_block);
- pre_ldst_mems = 0;
- last_set_in = xcalloc (max_gcse_regno, sizeof (int));
- already_set = xmalloc (sizeof (int) * max_gcse_regno);
+ pre_ldst_mems = NULL;
- /* Find all the stores we care about. */
FOR_EACH_BB (bb)
{
- /* First compute the registers set in this block. */
- regvec = last_set_in;
-
for (insn = BB_HEAD (bb);
- insn != NEXT_INSN (BB_END (bb));
+ insn && insn != NEXT_INSN (BB_END (bb));
insn = NEXT_INSN (insn))
{
- if (! INSN_P (insn))
- continue;
-
- if (GET_CODE (insn) == CALL_INSN)
+ if (INSN_P (insn))
{
- bool clobbers_all = false;
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
-
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- {
- last_set_in[regno] = INSN_UID (insn);
- SET_BIT (reg_set_in_block[bb->index], regno);
- }
- }
-
- pat = PATTERN (insn);
- compute_store_table_current_insn = insn;
- note_stores (pat, reg_set_info, reg_set_in_block[bb->index]);
- }
+ if (GET_CODE (PATTERN (insn)) == SET)
+ {
+ rtx src = SET_SRC (PATTERN (insn));
+ rtx dest = SET_DEST (PATTERN (insn));
- /* Now find the stores. */
- memset (already_set, 0, sizeof (int) * max_gcse_regno);
- regvec = already_set;
- for (insn = BB_HEAD (bb);
- insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
- {
- if (! INSN_P (insn))
- continue;
+ /* 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);
+ }
- 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
+ /* 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);
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
- already_set[regno] = 1;
+ if (! MEM_P (src)
+ && GET_CODE (src) != ASM_OPERANDS
+ /* Check for REG manually since want_to_gcse_p
+ returns 0 for all REGs. */
+ && can_assign_to_reg_p (src))
+ ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
+ else
+ ptr->invalid = 1;
+ }
+ }
+ else
+ invalidate_any_buried_refs (PATTERN (insn));
}
+ }
+ }
+}
- pat = PATTERN (insn);
- note_stores (pat, reg_set_info, NULL);
+/* Remove any references that have been either invalidated or are not in the
+ expression list for pre gcse. */
- /* Now that we've marked regs, look for stores. */
- find_moveable_store (insn, already_set, last_set_in);
+static void
+trim_ld_motion_mems (void)
+{
+ struct ls_expr * * last = & pre_ldst_mems;
+ struct ls_expr * ptr = pre_ldst_mems;
- /* Unmark regs that are no longer set. */
- compute_store_table_current_insn = insn;
- note_stores (pat, reg_clear_last_set, last_set_in);
- 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
+ while (ptr != NULL)
+ {
+ struct expr * expr;
- 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))
- last_set_in[regno] = 0;
- }
- }
+ /* 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;
-#ifdef ENABLE_CHECKING
- /* last_set_in should now be all-zero. */
- for (regno = 0; regno < max_gcse_regno; regno++)
- if (last_set_in[regno] != 0)
- abort ();
-#endif
+ 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;
- /* Clear temporary marks. */
- for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
+ if (expr)
{
- 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);
+ /* Set the expression field if we are keeping it. */
+ ptr->expr = expr;
+ last = & ptr->next;
+ ptr = ptr->next;
}
- }
-
- /* 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))
+ else
{
- *prev_next_ptr_ptr = ptr->next;
+ *last = ptr->next;
free_ldst_entry (ptr);
+ ptr = * last;
}
- else
- prev_next_ptr_ptr = &ptr->next;
- }
-
- ret = enumerate_ldsts ();
-
- if (gcse_file)
- {
- fprintf (gcse_file, "ST_avail and ST_antic (shown under loads..)\n");
- print_ldst_list (gcse_file);
}
- free (last_set_in);
- free (already_set);
- return ret;
+ /* Show the world what we've found. */
+ if (gcse_file && pre_ldst_mems != NULL)
+ print_ldst_list (gcse_file);
}
-/* 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. */
+/* 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 bool
-load_kills_store (rtx x, rtx store_pattern, int after)
+static void
+update_ld_motion_stores (struct expr * expr)
{
- if (after)
- return anti_dependence (x, store_pattern);
- else
- return true_dependence (store_pattern, GET_MODE (store_pattern), x,
- rtx_addr_varies_p);
-}
+ struct ls_expr * mem_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. */
+ 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. */
-static bool
-find_loads (rtx x, rtx store_pattern, int after)
-{
- const char * fmt;
- int i, j;
- int ret = false;
+ /* 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;
- if (!x)
- return false;
+ 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;
- if (GET_CODE (x) == SET)
- x = SET_SRC (x);
+ /* If we've already copied it, continue. */
+ if (expr->reaching_reg == src)
+ continue;
- if (GET_CODE (x) == MEM)
- {
- if (load_kills_store (x, store_pattern, after))
- return true;
- }
+ 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");
+ }
- /* Recursively process the insn. */
- fmt = GET_RTX_FORMAT (GET_CODE (x));
+ 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;
- 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);
+ /* un-recognize this pattern since it's probably different now. */
+ INSN_CODE (insn) = -1;
+ gcse_create_count++;
+ }
}
- return ret;
}
+\f
+/* Store motion code. */
-/* 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. */
+#define ANTIC_STORE_LIST(x) ((x)->loads)
+#define AVAIL_STORE_LIST(x) ((x)->stores)
+#define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
-static bool
-store_killed_in_insn (rtx x, rtx x_regs, rtx insn, int after)
-{
- rtx reg, base, note;
+/* 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;
- if (!INSN_P (insn))
- return false;
+/* And current insn, for the same routine. */
+static rtx compute_store_table_current_insn;
- if (GET_CODE (insn) == CALL_INSN)
- {
- /* A normal or pure call might read from pattern,
- but a const call will not. */
- if (! CONST_OR_PURE_CALL_P (insn) || pure_call_p (insn))
- return true;
+/* Used in computing the reverse edge graph bit vectors. */
+static sbitmap * st_antloc;
- /* 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;
- }
+/* Global holding the number of store expressions we are dealing with. */
+static int num_stores;
- return false;
- }
+/* Checks to set if we need to mark a register set. Called from
+ note_stores. */
- if (GET_CODE (PATTERN (insn)) == SET)
- {
- rtx pat = PATTERN (insn);
- rtx dest = SET_DEST (pat);
+static void
+reg_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED,
+ void *data)
+{
+ sbitmap bb_reg = data;
- if (GET_CODE (dest) == SIGN_EXTRACT
- || GET_CODE (dest) == ZERO_EXTRACT)
- dest = XEXP (dest, 0);
+ if (GET_CODE (dest) == SUBREG)
+ dest = SUBREG_REG (dest);
- /* 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;
- }
- }
- if (find_loads (SET_SRC (pat), x, after))
- return true;
+ if (REG_P (dest))
+ {
+ regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
+ if (bb_reg)
+ SET_BIT (bb_reg, REGNO (dest));
}
- else if (find_loads (PATTERN (insn), x, after))
- return true;
+}
- /* If this insn has a REG_EQUAL or REG_EQUIV note referencing a memory
- location aliased with X, then this insn kills X. */
- note = find_reg_equal_equiv_note (insn);
- if (! note)
- return false;
- note = XEXP (note, 0);
+/* Clear any mark that says that this insn sets dest. Called from
+ note_stores. */
- /* However, if the note represents a must alias rather than a may
- alias relationship, then it does not kill X. */
- if (expr_equiv_p (note, x))
- return false;
+static void
+reg_clear_last_set (rtx dest, rtx setter ATTRIBUTE_UNUSED,
+ void *data)
+{
+ int *dead_vec = data;
- /* See if there are any aliased loads in the note. */
- return find_loads (note, x, after);
+ if (GET_CODE (dest) == SUBREG)
+ dest = SUBREG_REG (dest);
+
+ if (REG_P (dest) &&
+ dead_vec[REGNO (dest)] == INSN_UID (compute_store_table_current_insn))
+ dead_vec[REGNO (dest)] = 0;
}
-/* 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. */
+/* Return zero if some of the registers in list X are killed
+ due to set of registers in bitmap REGS_SET. */
static bool
-store_killed_after (rtx x, rtx x_regs, rtx insn, basic_block bb,
- int *regs_set_after, rtx *fail_insn)
+store_ops_ok (rtx x, int *regs_set)
{
- rtx last = BB_END (bb), act;
+ rtx reg;
- if (!store_ops_ok (x_regs, regs_set_after))
+ for (; x; x = XEXP (x, 1))
{
- /* We do not know where it will happen. */
- if (fail_insn)
- *fail_insn = NULL_RTX;
- return true;
+ reg = XEXP (x, 0);
+ if (regs_set[REGNO(reg)])
+ return false;
}
- /* 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 true;
+}
- return false;
+/* Returns a list of registers mentioned in X. */
+static rtx
+extract_mentioned_regs (rtx x)
+{
+ return extract_mentioned_regs_helper (x, NULL_RTX);
}
-/* 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)
+/* Helper for extract_mentioned_regs; ACCUM is used to accumulate used
+ registers. */
+static rtx
+extract_mentioned_regs_helper (rtx x, rtx accum)
{
- rtx first = BB_HEAD (bb);
+ int i;
+ enum rtx_code code;
+ const char * fmt;
- if (!store_ops_ok (x_regs, regs_set_before))
- return true;
+ /* Repeat is used to turn tail-recursion into iteration. */
+ repeat:
- for ( ; insn != PREV_INSN (first); insn = PREV_INSN (insn))
- if (store_killed_in_insn (x, x_regs, insn, true))
- return true;
+ if (x == 0)
+ return accum;
- return false;
-}
+ code = GET_CODE (x);
+ switch (code)
+ {
+ case REG:
+ return alloc_EXPR_LIST (0, x, accum);
-/* 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;
+ case MEM:
+ x = XEXP (x, 0);
+ goto repeat;
- /* 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);
+ case PRE_DEC:
+ case PRE_INC:
+ case POST_DEC:
+ case POST_INC:
+ /* We do not run this function with arguments having side effects. */
+ gcc_unreachable ();
- st_antloc = sbitmap_vector_alloc (last_basic_block, num_stores);
- sbitmap_vector_zero (st_antloc, last_basic_block);
+ case PC:
+ case CC0: /*FIXME*/
+ case CONST:
+ case CONST_INT:
+ case CONST_DOUBLE:
+ case CONST_VECTOR:
+ case SYMBOL_REF:
+ case LABEL_REF:
+ case ADDR_VEC:
+ case ADDR_DIFF_VEC:
+ return accum;
+
+ default:
+ break;
+ }
+
+ i = GET_RTX_LENGTH (code) - 1;
+ fmt = GET_RTX_FORMAT (code);
- for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
+ for (; i >= 0; i--)
{
- for (st = AVAIL_STORE_LIST (ptr); st != NULL; st = XEXP (st, 1))
+ if (fmt[i] == 'e')
{
- insn = XEXP (st, 0);
- bb = BLOCK_FOR_INSN (insn);
+ rtx tem = XEXP (x, i);
- /* If we've already seen an available expression in this block,
- we can delete this one (It occurs earlier in the block). We'll
- copy the SRC expression to an unused register in case there
- are any side effects. */
- if (TEST_BIT (ae_gen[bb->index], ptr->index))
+ /* If we are about to do the last recursive call
+ needed at this level, change it into iteration. */
+ if (i == 0)
{
- rtx r = gen_reg_rtx (GET_MODE (ptr->pattern));
- if (gcse_file)
- fprintf (gcse_file, "Removing redundant store:\n");
- replace_store_insn (r, XEXP (st, 0), bb, ptr);
- continue;
+ x = tem;
+ goto repeat;
}
- SET_BIT (ae_gen[bb->index], ptr->index);
- }
- for (st = ANTIC_STORE_LIST (ptr); st != NULL; st = XEXP (st, 1))
+ accum = extract_mentioned_regs_helper (tem, accum);
+ }
+ else if (fmt[i] == 'E')
{
- insn = XEXP (st, 0);
- bb = BLOCK_FOR_INSN (insn);
- SET_BIT (st_antloc[bb->index], ptr->index);
+ int j;
+
+ for (j = 0; j < XVECLEN (x, i); j++)
+ accum = extract_mentioned_regs_helper (XVECEXP (x, i, j), accum);
}
}
- ae_kill = sbitmap_vector_alloc (last_basic_block, num_stores);
- sbitmap_vector_zero (ae_kill, last_basic_block);
-
- transp = sbitmap_vector_alloc (last_basic_block, num_stores);
- sbitmap_vector_zero (transp, last_basic_block);
- regs_set_in_block = xmalloc (sizeof (int) * max_gcse_regno);
-
- FOR_EACH_BB (bb)
- {
- for (regno = 0; regno < max_gcse_regno; regno++)
- regs_set_in_block[regno] = TEST_BIT (reg_set_in_block[bb->index], regno);
+ return accum;
+}
- for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
- {
- if (store_killed_after (ptr->pattern, ptr->pattern_regs, BB_HEAD (bb),
- bb, regs_set_in_block, NULL))
- {
- /* It should not be necessary to consider the expression
- killed if it is both anticipatable and available. */
- if (!TEST_BIT (st_antloc[bb->index], ptr->index)
- || !TEST_BIT (ae_gen[bb->index], ptr->index))
- SET_BIT (ae_kill[bb->index], ptr->index);
- }
- else
- SET_BIT (transp[bb->index], ptr->index);
- }
- }
+/* 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.
- free (regs_set_in_block);
+ The results are stored this way:
- 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);
- }
-}
+ -- 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.
-/* Insert an instruction at the beginning of a basic block, and update
- the BB_HEAD if needed. */
+ The things are complicated a bit by fact that there already may be stores
+ to the same MEM from other blocks; also caller must take care of the
+ necessary cleanup of the temporary markers after end of the basic block.
+ */
static void
-insert_insn_start_bb (rtx insn, basic_block bb)
+find_moveable_store (rtx insn, int *regs_set_before, int *regs_set_after)
{
- /* Insert at start of successor block. */
- rtx prev = PREV_INSN (BB_HEAD (bb));
- rtx before = BB_HEAD (bb);
- while (before != 0)
- {
- if (GET_CODE (before) != CODE_LABEL
- && (GET_CODE (before) != NOTE
- || NOTE_LINE_NUMBER (before) != NOTE_INSN_BASIC_BLOCK))
- break;
- prev = before;
- if (prev == BB_END (bb))
- break;
- before = NEXT_INSN (before);
- }
-
- insn = emit_insn_after (insn, prev);
+ struct ls_expr * ptr;
+ rtx dest, set, tmp;
+ int check_anticipatable, check_available;
+ basic_block bb = BLOCK_FOR_INSN (insn);
- if (gcse_file)
- {
- fprintf (gcse_file, "STORE_MOTION insert store at start of BB %d:\n",
- bb->index);
- print_inline_rtx (gcse_file, insn, 6);
- fprintf (gcse_file, "\n");
- }
-}
+ set = single_set (insn);
+ if (!set)
+ return;
-/* 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. */
+ dest = SET_DEST (set);
-static int
-insert_store (struct ls_expr * expr, edge e)
-{
- rtx reg, insn;
- basic_block bb;
- edge tmp;
+ if (! MEM_P (dest) || MEM_VOLATILE_P (dest)
+ || GET_MODE (dest) == BLKmode)
+ return;
- /* We did all the deleted before this insert, so if we didn't delete a
- store, then we haven't set the reaching reg yet either. */
- if (expr->reaching_reg == NULL_RTX)
- return 0;
+ if (side_effects_p (dest))
+ return;
- if (e->flags & EDGE_FAKE)
- 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 (dest))
+ return;
- reg = expr->reaching_reg;
- insn = gen_move_insn (copy_rtx (expr->pattern), reg);
+ /* Even if the destination cannot trap, the source may. In this case we'd
+ need to handle updating the REG_EH_REGION note. */
+ if (find_reg_note (insn, REG_EH_REGION, NULL_RTX))
+ return;
- /* 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)
- if (!(tmp->flags & EDGE_FAKE))
- {
- 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;
- }
+ ptr = ldst_entry (dest);
+ if (!ptr->pattern_regs)
+ ptr->pattern_regs = extract_mentioned_regs (dest);
- /* 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)
+ /* 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
{
- for (tmp = e->dest->pred; tmp ; tmp = tmp->pred_next)
- {
- 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;
+ tmp = XEXP (ANTIC_STORE_LIST (ptr), 0);
+ if (tmp != NULL_RTX
+ && BLOCK_FOR_INSN (tmp) != bb)
+ check_anticipatable = 1;
}
-
- /* 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)
+ if (check_anticipatable)
{
- insert_insn_start_bb (insn, bb);
- return 0;
+ 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));
}
- insert_insn_on_edge (insn, e);
-
- if (gcse_file)
+ /* 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
{
- 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");
+ tmp = XEXP (AVAIL_STORE_LIST (ptr), 0);
+ if (BLOCK_FOR_INSN (tmp) != bb)
+ check_available = 1;
}
-
- return 1;
-}
-
-/* Remove any REG_EQUAL or REG_EQUIV notes containing a reference to the
- memory location in SMEXPR set in basic block BB.
-
- This could be rather expensive. */
-
-static void
-remove_reachable_equiv_notes (basic_block bb, struct ls_expr *smexpr)
-{
- edge *stack = xmalloc (sizeof (edge) * n_basic_blocks), act;
- sbitmap visited = sbitmap_alloc (last_basic_block);
- int stack_top = 0;
- rtx last, insn, note;
- rtx mem = smexpr->pattern;
-
- sbitmap_zero (visited);
- act = bb->succ;
-
- while (1)
+ if (check_available)
{
- if (!act)
- {
- if (!stack_top)
- {
- free (stack);
- sbitmap_free (visited);
- return;
- }
- act = stack[--stack_top];
- }
- bb = act->dest;
-
- if (bb == EXIT_BLOCK_PTR
- || TEST_BIT (visited, bb->index))
- {
- act = act->succ_next;
- continue;
- }
- SET_BIT (visited, bb->index);
-
- if (TEST_BIT (st_antloc[bb->index], smexpr->index))
- {
- for (last = ANTIC_STORE_LIST (smexpr);
- BLOCK_FOR_INSN (XEXP (last, 0)) != bb;
- last = XEXP (last, 1))
+ /* Check that we have already reached the insn at that the check
+ failed last time. */
+ if (LAST_AVAIL_CHECK_FAILURE (ptr))
+ {
+ for (tmp = BB_END (bb);
+ tmp != insn && tmp != LAST_AVAIL_CHECK_FAILURE (ptr);
+ tmp = PREV_INSN (tmp))
continue;
- last = XEXP (last, 0);
+ if (tmp == insn)
+ check_available = 0;
}
else
- last = NEXT_INSN (BB_END (bb));
-
- for (insn = BB_HEAD (bb); insn != last; insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- {
- note = find_reg_equal_equiv_note (insn);
- if (!note || !expr_equiv_p (XEXP (note, 0), mem))
- continue;
-
- if (gcse_file)
- fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
- INSN_UID (insn));
- remove_note (insn, note);
- }
- act = act->succ_next;
- if (bb->succ)
- {
- if (act)
- stack[stack_top++] = act;
- act = bb->succ;
- }
+ check_available = store_killed_after (dest, ptr->pattern_regs, insn,
+ bb, regs_set_after,
+ &LAST_AVAIL_CHECK_FAILURE (ptr));
}
+ if (!check_available)
+ AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn, AVAIL_STORE_LIST (ptr));
}
-/* This routine will replace a store with a SET to a specified register. */
+/* Find available and anticipatable stores. */
-static void
-replace_store_insn (rtx reg, rtx del, basic_block bb, struct ls_expr *smexpr)
+static int
+compute_store_table (void)
{
- rtx insn, mem, note, set, ptr;
-
- mem = smexpr->pattern;
- insn = gen_move_insn (reg, SET_SRC (single_set (del)));
- insn = emit_insn_after (insn, del);
-
- if (gcse_file)
- {
- fprintf (gcse_file,
- "STORE_MOTION delete insn in BB %d:\n ", bb->index);
- print_inline_rtx (gcse_file, del, 6);
- fprintf (gcse_file, "\nSTORE MOTION replaced with insn:\n ");
- print_inline_rtx (gcse_file, insn, 6);
- fprintf (gcse_file, "\n");
- }
-
- for (ptr = ANTIC_STORE_LIST (smexpr); ptr; ptr = XEXP (ptr, 1))
- if (XEXP (ptr, 0) == del)
- {
- XEXP (ptr, 0) = insn;
- break;
- }
- delete_insn (del);
-
- /* Now we must handle REG_EQUAL notes whose contents is equal to the mem;
- they are no longer accurate provided that they are reached by this
- definition, so drop them. */
- for (; insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- {
- set = single_set (insn);
- if (!set)
- continue;
- if (expr_equiv_p (SET_DEST (set), mem))
- return;
- note = find_reg_equal_equiv_note (insn);
- if (!note || !expr_equiv_p (XEXP (note, 0), mem))
- continue;
-
- if (gcse_file)
- fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
- INSN_UID (insn));
- remove_note (insn, note);
- }
- remove_reachable_equiv_notes (bb, smexpr);
-}
+ int ret;
+ basic_block bb;
+ unsigned regno;
+ rtx insn, pat, tmp;
+ int *last_set_in, *already_set;
+ struct ls_expr * ptr, **prev_next_ptr_ptr;
+ max_gcse_regno = max_reg_num ();
-/* Delete a store, but copy the value that would have been stored into
- the reaching_reg for later storing. */
+ reg_set_in_block = sbitmap_vector_alloc (last_basic_block,
+ max_gcse_regno);
+ sbitmap_vector_zero (reg_set_in_block, last_basic_block);
+ pre_ldst_mems = 0;
+ last_set_in = xcalloc (max_gcse_regno, sizeof (int));
+ already_set = xmalloc (sizeof (int) * max_gcse_regno);
-static void
-delete_store (struct ls_expr * expr, basic_block bb)
-{
- rtx reg, i, del;
+ /* Find all the stores we care about. */
+ FOR_EACH_BB (bb)
+ {
+ /* First compute the registers set in this block. */
+ regvec = last_set_in;
- if (expr->reaching_reg == NULL_RTX)
- expr->reaching_reg = gen_reg_rtx (GET_MODE (expr->pattern));
+ for (insn = BB_HEAD (bb);
+ insn != NEXT_INSN (BB_END (bb));
+ insn = NEXT_INSN (insn))
+ {
+ if (! INSN_P (insn))
+ continue;
- reg = expr->reaching_reg;
+ if (CALL_P (insn))
+ {
+ for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
+ {
+ last_set_in[regno] = INSN_UID (insn);
+ SET_BIT (reg_set_in_block[bb->index], regno);
+ }
+ }
- for (i = AVAIL_STORE_LIST (expr); i; i = XEXP (i, 1))
- {
- del = XEXP (i, 0);
- if (BLOCK_FOR_INSN (del) == bb)
- {
- /* We know there is only one since we deleted redundant
- ones during the available computation. */
- replace_store_insn (reg, del, bb, expr);
- break;
+ pat = PATTERN (insn);
+ compute_store_table_current_insn = insn;
+ note_stores (pat, reg_set_info, reg_set_in_block[bb->index]);
}
- }
-}
-/* Free memory used by store motion. */
+ /* Now find the stores. */
+ memset (already_set, 0, sizeof (int) * max_gcse_regno);
+ regvec = already_set;
+ for (insn = BB_HEAD (bb);
+ insn != NEXT_INSN (BB_END (bb));
+ insn = NEXT_INSN (insn))
+ {
+ if (! INSN_P (insn))
+ continue;
-static void
-free_store_memory (void)
-{
- free_ldst_mems ();
+ if (CALL_P (insn))
+ {
+ for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
+ already_set[regno] = 1;
+ }
- 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);
+ pat = PATTERN (insn);
+ note_stores (pat, reg_set_info, NULL);
- ae_gen = ae_kill = transp = st_antloc = NULL;
- pre_insert_map = pre_delete_map = reg_set_in_block = NULL;
-}
+ /* Now that we've marked regs, look for stores. */
+ find_moveable_store (insn, already_set, last_set_in);
-/* Perform store motion. Much like gcse, except we move expressions the
- other way by looking at the flowgraph in reverse. */
+ /* Unmark regs that are no longer set. */
+ compute_store_table_current_insn = insn;
+ note_stores (pat, reg_clear_last_set, last_set_in);
+ if (CALL_P (insn))
+ {
+ for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
+ if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno)
+ && last_set_in[regno] == INSN_UID (insn))
+ last_set_in[regno] = 0;
+ }
+ }
-static void
-store_motion (void)
-{
- basic_block bb;
- int x;
- struct ls_expr * ptr;
- int update_flow = 0;
+#ifdef ENABLE_CHECKING
+ /* last_set_in should now be all-zero. */
+ for (regno = 0; regno < max_gcse_regno; regno++)
+ gcc_assert (!last_set_in[regno]);
+#endif
- if (gcse_file)
- {
- fprintf (gcse_file, "before store motion\n");
- print_rtl (gcse_file, get_insns ());
+ /* 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);
+ }
}
- init_alias_analysis ();
-
- /* Find all the available and anticipatable stores. */
- num_stores = compute_store_table ();
- if (num_stores == 0)
+ /* 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)
{
- sbitmap_vector_free (reg_set_in_block);
- end_alias_analysis ();
- return;
+ if (!AVAIL_STORE_LIST (ptr))
+ {
+ *prev_next_ptr_ptr = ptr->next;
+ free_ldst_entry (ptr);
+ }
+ else
+ prev_next_ptr_ptr = &ptr->next;
}
- /* Now compute kill & transp vectors. */
- build_store_vectors ();
- add_noreturn_fake_exit_edges ();
- connect_infinite_loops_to_exit ();
-
- edge_list = pre_edge_rev_lcm (gcse_file, num_stores, transp, ae_gen,
- st_antloc, ae_kill, &pre_insert_map,
- &pre_delete_map);
+ ret = enumerate_ldsts ();
- /* 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 (gcse_file)
{
- FOR_EACH_BB (bb)
- if (TEST_BIT (pre_delete_map[bb->index], ptr->index))
- delete_store (ptr, bb);
-
- for (x = 0; x < NUM_EDGES (edge_list); x++)
- if (TEST_BIT (pre_insert_map[x], ptr->index))
- update_flow |= insert_store (ptr, INDEX_EDGE (edge_list, x));
+ fprintf (gcse_file, "ST_avail and ST_antic (shown under loads..)\n");
+ print_ldst_list (gcse_file);
}
- if (update_flow)
- commit_edge_insertions ();
+ free (last_set_in);
+ free (already_set);
+ return ret;
+}
- free_store_memory ();
- free_edge_list (edge_list);
- remove_fake_edges ();
- end_alias_analysis ();
+/* Check to see if the load X is aliased with STORE_PATTERN.
+ AFTER is true if we are checking the case when STORE_PATTERN occurs
+ after the X. */
+
+static bool
+load_kills_store (rtx x, rtx store_pattern, int after)
+{
+ if (after)
+ return anti_dependence (x, store_pattern);
+ else
+ return true_dependence (store_pattern, GET_MODE (store_pattern), x,
+ rtx_addr_varies_p);
}
-\f
-/* Entry point for jump bypassing optimization pass. */
+/* 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. */
-int
-bypass_jumps (FILE *file)
+static bool
+find_loads (rtx x, rtx store_pattern, int after)
{
- int changed;
+ const char * fmt;
+ int i, j;
+ int ret = false;
- /* 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;
+ if (!x)
+ return false;
- /* For calling dump_foo fns from gdb. */
- debug_stderr = stderr;
- gcse_file = file;
+ if (GET_CODE (x) == SET)
+ x = SET_SRC (x);
- /* Identify the basic block information for this function, including
- successors and predecessors. */
- max_gcse_regno = max_reg_num ();
+ if (MEM_P (x))
+ {
+ if (load_kills_store (x, store_pattern, after))
+ return true;
+ }
- if (file)
- dump_flow_info (file);
+ /* Recursively process the insn. */
+ fmt = GET_RTX_FORMAT (GET_CODE (x));
- /* Return if there's nothing to do, or it is too expensive. */
- if (n_basic_blocks <= 1 || is_too_expensive (_ ("jump bypassing disabled")))
- return 0;
+ for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0 && !ret; i--)
+ {
+ if (fmt[i] == 'e')
+ ret |= find_loads (XEXP (x, i), store_pattern, after);
+ else if (fmt[i] == 'E')
+ for (j = XVECLEN (x, i) - 1; j >= 0; j--)
+ ret |= find_loads (XVECEXP (x, i, j), store_pattern, after);
+ }
+ return ret;
+}
- gcc_obstack_init (&gcse_obstack);
- bytes_used = 0;
+/* Check if INSN kills the store pattern X (is aliased with it).
+ AFTER is true if we are checking the case when store X occurs
+ after the insn. Return true if it does. */
- /* We need alias. */
- init_alias_analysis ();
+static bool
+store_killed_in_insn (rtx x, rtx x_regs, rtx insn, int after)
+{
+ rtx reg, base, note;
- /* 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.
+ if (!INSN_P (insn))
+ return false;
- 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. */
+ if (CALL_P (insn))
+ {
+ /* A normal or pure call might read from pattern,
+ but a const call will not. */
+ if (! CONST_OR_PURE_CALL_P (insn) || pure_call_p (insn))
+ return true;
- alloc_reg_set_mem (max_gcse_regno);
- compute_sets (get_insns ());
+ /* 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;
+ }
- max_gcse_regno = max_reg_num ();
- alloc_gcse_mem (get_insns ());
- changed = one_cprop_pass (1, 1, 1);
- free_gcse_mem ();
+ return false;
+ }
- if (file)
+ if (GET_CODE (PATTERN (insn)) == SET)
{
- fprintf (file, "BYPASS of %s: %d basic blocks, ",
- current_function_name (), n_basic_blocks);
- fprintf (file, "%d bytes\n\n", bytes_used);
+ rtx pat = PATTERN (insn);
+ rtx dest = SET_DEST (pat);
+
+ if (GET_CODE (dest) == ZERO_EXTRACT)
+ dest = XEXP (dest, 0);
+
+ /* Check for memory stores to aliased objects. */
+ if (MEM_P (dest)
+ && !expr_equiv_p (dest, x))
+ {
+ if (after)
+ {
+ if (output_dependence (dest, x))
+ return true;
+ }
+ else
+ {
+ if (output_dependence (x, dest))
+ return true;
+ }
+ }
+ if (find_loads (SET_SRC (pat), x, after))
+ return true;
}
+ else if (find_loads (PATTERN (insn), x, after))
+ return true;
- obstack_free (&gcse_obstack, NULL);
- free_reg_set_mem ();
+ /* If this insn has a REG_EQUAL or REG_EQUIV note referencing a memory
+ location aliased with X, then this insn kills X. */
+ note = find_reg_equal_equiv_note (insn);
+ if (! note)
+ return false;
+ note = XEXP (note, 0);
- /* We are finished with alias. */
- end_alias_analysis ();
- allocate_reg_info (max_reg_num (), FALSE, FALSE);
+ /* However, if the note represents a must alias rather than a may
+ alias relationship, then it does not kill X. */
+ if (expr_equiv_p (note, x))
+ return false;
- return changed;
+ /* See if there are any aliased loads in the note. */
+ return find_loads (note, x, after);
}
-/* Return true if the graph is too expensive to optimize. PASS is the
- optimization about to be performed. */
+/* Returns true if the expression X is loaded or clobbered on or after INSN
+ within basic block BB. REGS_SET_AFTER is bitmap of registers set in
+ or after the insn. X_REGS is list of registers mentioned in X. If the store
+ is killed, return the last insn in that it occurs in FAIL_INSN. */
static bool
-is_too_expensive (const char *pass)
+store_killed_after (rtx x, rtx x_regs, rtx insn, basic_block bb,
+ int *regs_set_after, rtx *fail_insn)
{
- /* Trying to perform global optimizations on flow graphs which have
- a high connectivity will take a long time and is unlikely to be
- particularly useful.
-
- In normal circumstances a cfg should have about twice as many
- edges as blocks. But we do not want to punish small functions
- which have a couple switch statements. Rather than simply
- threshold the number of blocks, uses something with a more
- graceful degradation. */
- if (n_edges > 20000 + n_basic_blocks * 4)
- {
- if (warn_disabled_optimization)
- warning ("%s: %d basic blocks and %d edges/basic block",
- pass, n_basic_blocks, n_edges / n_basic_blocks);
-
- return true;
- }
+ rtx last = BB_END (bb), act;
- /* If allocating memory for the cprop bitmap would take up too much
- storage it's better just to disable the optimization. */
- if ((n_basic_blocks
- * SBITMAP_SET_SIZE (max_reg_num ())
- * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
+ if (!store_ops_ok (x_regs, regs_set_after))
{
- if (warn_disabled_optimization)
- warning ("%s: %d basic blocks and %d registers",
- pass, n_basic_blocks, max_reg_num ());
-
+ /* We do not know where it will happen. */
+ if (fail_insn)
+ *fail_insn = NULL_RTX;
return true;
}
+ /* Scan from the end, so that fail_insn is determined correctly. */
+ for (act = last; act != PREV_INSN (insn); act = PREV_INSN (act))
+ if (store_killed_in_insn (x, x_regs, act, false))
+ {
+ if (fail_insn)
+ *fail_insn = act;
+ return true;
+ }
+
return false;
}
-/* The following code implements gcse after reload, the purpose of this
- pass is to cleanup redundant loads generated by reload and other
- optimizations that come after gcse. It searches for simple inter-block
- redundancies and tries to eliminate them by adding moves and loads
- in cold places. */
-
-/* The following structure holds the information about the occurrences of
- the redundant instructions. */
-struct unoccr
-{
- struct unoccr *next;
- edge pred;
- rtx insn;
-};
-
-static bool reg_used_on_edge (rtx, edge);
-static rtx reg_set_between_after_reload_p (rtx, rtx, rtx);
-static rtx reg_used_between_after_reload_p (rtx, rtx, rtx);
-static rtx get_avail_load_store_reg (rtx);
-static bool is_jump_table_basic_block (basic_block);
-static bool bb_has_well_behaved_predecessors (basic_block);
-static struct occr* get_bb_avail_insn (basic_block, struct occr *);
-static void hash_scan_set_after_reload (rtx, rtx, struct hash_table *);
-static void compute_hash_table_after_reload (struct hash_table *);
-static void eliminate_partially_redundant_loads (basic_block,
- rtx,
- struct expr *);
-static void gcse_after_reload (void);
-static struct occr* get_bb_avail_insn (basic_block, struct occr *);
-void gcse_after_reload_main (rtx, FILE *);
-
-
-/* Check if register REG is used in any insn waiting to be inserted on E.
- Assumes no such insn can be a CALL_INSN; if so call reg_used_between_p
- with PREV(insn),NEXT(insn) instead of calling
- reg_overlap_mentioned_p. */
-
+/* 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
-reg_used_on_edge (rtx reg, edge e)
+store_killed_before (rtx x, rtx x_regs, rtx insn, basic_block bb,
+ int *regs_set_before)
{
- rtx insn;
+ rtx first = BB_HEAD (bb);
+
+ if (!store_ops_ok (x_regs, regs_set_before))
+ return true;
- for (insn = e->insns; insn; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn)))
+ for ( ; insn != PREV_INSN (first); insn = PREV_INSN (insn))
+ if (store_killed_in_insn (x, x_regs, insn, true))
return true;
return false;
}
-/* Return the insn that sets register REG or clobbers it in between
- FROM_INSN and TO_INSN (exclusive of those two).
- Just like reg_set_between but for hard registers and not pseudos. */
-
-static rtx
-reg_set_between_after_reload_p (rtx reg, rtx from_insn, rtx to_insn)
+/* 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)
{
- rtx insn;
- int regno;
+ basic_block bb;
+ int *regs_set_in_block;
+ rtx insn, st;
+ struct ls_expr * ptr;
+ unsigned regno;
+
+ /* Build the gen_vector. This is any store in the table which is not killed
+ by aliasing later in its block. */
+ ae_gen = sbitmap_vector_alloc (last_basic_block, num_stores);
+ sbitmap_vector_zero (ae_gen, last_basic_block);
+
+ st_antloc = sbitmap_vector_alloc (last_basic_block, num_stores);
+ sbitmap_vector_zero (st_antloc, last_basic_block);
+
+ for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
+ {
+ for (st = AVAIL_STORE_LIST (ptr); st != NULL; st = XEXP (st, 1))
+ {
+ insn = XEXP (st, 0);
+ bb = BLOCK_FOR_INSN (insn);
+
+ /* If we've already seen an available expression in this block,
+ we can delete this one (It occurs earlier in the block). We'll
+ copy the SRC expression to an unused register in case there
+ are any side effects. */
+ if (TEST_BIT (ae_gen[bb->index], ptr->index))
+ {
+ rtx r = gen_reg_rtx (GET_MODE (ptr->pattern));
+ if (gcse_file)
+ fprintf (gcse_file, "Removing redundant store:\n");
+ replace_store_insn (r, XEXP (st, 0), bb, ptr);
+ continue;
+ }
+ SET_BIT (ae_gen[bb->index], ptr->index);
+ }
- if (GET_CODE (reg) != REG)
- abort ();
- regno = REGNO (reg);
+ 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);
+ }
+ }
- /* We are called after register allocation. */
- if (regno >= FIRST_PSEUDO_REGISTER)
- abort ();
+ ae_kill = sbitmap_vector_alloc (last_basic_block, num_stores);
+ sbitmap_vector_zero (ae_kill, last_basic_block);
- if (from_insn == to_insn)
- return NULL_RTX;
+ transp = sbitmap_vector_alloc (last_basic_block, num_stores);
+ sbitmap_vector_zero (transp, last_basic_block);
+ regs_set_in_block = xmalloc (sizeof (int) * max_gcse_regno);
- for (insn = NEXT_INSN (from_insn);
- insn != to_insn;
- insn = NEXT_INSN (insn))
+ FOR_EACH_BB (bb)
{
- if (INSN_P (insn))
+ for (regno = 0; regno < max_gcse_regno; regno++)
+ regs_set_in_block[regno] = TEST_BIT (reg_set_in_block[bb->index], regno);
+
+ for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
{
- if (FIND_REG_INC_NOTE (insn, reg)
- || (GET_CODE (insn) == CALL_INSN
- && call_used_regs[regno])
- || find_reg_fusage (insn, CLOBBER, reg))
- return insn;
+ if (store_killed_after (ptr->pattern, ptr->pattern_regs, BB_HEAD (bb),
+ bb, regs_set_in_block, NULL))
+ {
+ /* It should not be necessary to consider the expression
+ killed if it is both anticipatable and available. */
+ if (!TEST_BIT (st_antloc[bb->index], ptr->index)
+ || !TEST_BIT (ae_gen[bb->index], ptr->index))
+ SET_BIT (ae_kill[bb->index], ptr->index);
+ }
+ else
+ SET_BIT (transp[bb->index], ptr->index);
}
- if (set_of (reg, insn) != NULL_RTX)
- return insn;
}
- return NULL_RTX;
+
+ free (regs_set_in_block);
+
+ if (gcse_file)
+ {
+ dump_sbitmap_vector (gcse_file, "st_antloc", "", st_antloc, last_basic_block);
+ dump_sbitmap_vector (gcse_file, "st_kill", "", ae_kill, last_basic_block);
+ dump_sbitmap_vector (gcse_file, "Transpt", "", transp, last_basic_block);
+ dump_sbitmap_vector (gcse_file, "st_avloc", "", ae_gen, last_basic_block);
+ }
}
-/* Return the insn that uses register REG in between FROM_INSN and TO_INSN
- (exclusive of those two). Similar to reg_used_between but for hard
- registers and not pseudos. */
+/* Insert an instruction at the beginning of a basic block, and update
+ the BB_HEAD if needed. */
-static rtx
-reg_used_between_after_reload_p (rtx reg, rtx from_insn, rtx to_insn)
+static void
+insert_insn_start_bb (rtx insn, basic_block bb)
{
- rtx insn;
- int regno;
-
- if (GET_CODE (reg) != REG)
- return to_insn;
- regno = REGNO (reg);
+ /* Insert at start of successor block. */
+ rtx prev = PREV_INSN (BB_HEAD (bb));
+ rtx before = BB_HEAD (bb);
+ while (before != 0)
+ {
+ if (! LABEL_P (before)
+ && (! NOTE_P (before)
+ || NOTE_LINE_NUMBER (before) != NOTE_INSN_BASIC_BLOCK))
+ break;
+ prev = before;
+ if (prev == BB_END (bb))
+ break;
+ before = NEXT_INSN (before);
+ }
- /* We are called after register allocation. */
- if (regno >= FIRST_PSEUDO_REGISTER)
- abort ();
- if (from_insn == to_insn)
- return NULL_RTX;
+ insn = emit_insn_after_noloc (insn, prev);
- for (insn = NEXT_INSN (from_insn);
- insn != to_insn;
- insn = NEXT_INSN (insn))
- if (INSN_P (insn)
- && (reg_overlap_mentioned_p (reg, PATTERN (insn))
- || (GET_CODE (insn) == CALL_INSN
- && call_used_regs[regno])
- || find_reg_fusage (insn, USE, reg)
- || find_reg_fusage (insn, CLOBBER, reg)))
- return insn;
- return NULL_RTX;
+ if (gcse_file)
+ {
+ fprintf (gcse_file, "STORE_MOTION insert store at start of BB %d:\n",
+ bb->index);
+ print_inline_rtx (gcse_file, insn, 6);
+ fprintf (gcse_file, "\n");
+ }
}
-/* Return the loaded/stored register of a load/store instruction. */
+/* This routine will insert a store on an edge. EXPR is the ldst entry for
+ the memory reference, and E is the edge to insert it on. Returns nonzero
+ if an edge insertion was performed. */
-static rtx
-get_avail_load_store_reg (rtx insn)
+static int
+insert_store (struct ls_expr * expr, edge e)
{
- if (GET_CODE (SET_DEST (PATTERN (insn))) == REG) /* A load. */
- return SET_DEST(PATTERN(insn));
- if (GET_CODE (SET_SRC (PATTERN (insn))) == REG) /* A store. */
- return SET_SRC (PATTERN (insn));
- abort ();
-}
+ rtx reg, insn;
+ basic_block bb;
+ edge tmp;
+ edge_iterator ei;
-/* Don't handle ABNORMAL edges or jump tables. */
+ /* 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;
-static bool
-is_jump_table_basic_block (basic_block bb)
-{
- rtx insn = BB_END (bb);
+ if (e->flags & EDGE_FAKE)
+ return 0;
- if (GET_CODE (insn) == JUMP_INSN &&
- (GET_CODE (PATTERN (insn)) == ADDR_VEC
- || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC))
- return true;
- return false;
-}
+ reg = expr->reaching_reg;
+ insn = gen_move_insn (copy_rtx (expr->pattern), reg);
-/* Return nonzero if the predecessors of BB are "well behaved". */
+ /* If we are inserting this expression on ALL predecessor edges of a BB,
+ insert it at the start of the BB, and reset the insert bits on the other
+ edges so we don't try to insert it on the other edges. */
+ bb = e->dest;
+ FOR_EACH_EDGE (tmp, ei, e->dest->preds)
+ if (!(tmp->flags & EDGE_FAKE))
+ {
+ int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
+
+ gcc_assert (index != EDGE_INDEX_NO_EDGE);
+ if (! TEST_BIT (pre_insert_map[index], expr->index))
+ break;
+ }
-static bool
-bb_has_well_behaved_predecessors (basic_block bb)
-{
- edge pred;
+ /* If tmp is NULL, we found an insertion on every edge, blank the
+ insertion vector for these edges, and insert at the start of the BB. */
+ if (!tmp && bb != EXIT_BLOCK_PTR)
+ {
+ FOR_EACH_EDGE (tmp, ei, e->dest->preds)
+ {
+ int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
+ RESET_BIT (pre_insert_map[index], expr->index);
+ }
+ insert_insn_start_bb (insn, bb);
+ return 0;
+ }
- if (! bb->pred)
- return false;
- for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
- if (((pred->flags & EDGE_ABNORMAL) && EDGE_CRITICAL_P (pred))
- || is_jump_table_basic_block (pred->src))
- return false;
- return true;
-}
+ /* We can't put stores in the front of blocks pointed to by abnormal
+ edges since that may put a store where one didn't used to be. */
+ gcc_assert (!(e->flags & EDGE_ABNORMAL));
+ insert_insn_on_edge (insn, e);
-/* Search for the occurrences of expression in BB. */
+ 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");
+ }
-static struct occr*
-get_bb_avail_insn (basic_block bb, struct occr *occr)
-{
- for (; occr != NULL; occr = occr->next)
- if (BLOCK_FOR_INSN (occr->insn)->index == bb->index)
- return occr;
- return NULL;
+ return 1;
}
-/* Perform partial GCSE pass after reload, try to eliminate redundant loads
- created by the reload pass. We try to look for a full or partial
- redundant loads fed by one or more loads/stores in predecessor BBs,
- and try adding loads to make them fully redundant. We also check if
- it's worth adding loads to be able to delete the redundant load.
-
- Algorithm:
- 1. Build available expressions hash table:
- For each load/store instruction, if the loaded/stored memory didn't
- change until the end of the basic block add this memory expression to
- the hash table.
- 2. Perform Redundancy elimination:
- For each load instruction do the following:
- perform partial redundancy elimination, check if it's worth adding
- loads to make the load fully redundant. If so add loads and
- register copies and delete the load.
-
- Future enhancement:
- if loaded register is used/defined between load and some store,
- look for some other free register between load and all its stores,
- and replace load with a copy from this register to the loaded
- register. */
-
+/* Remove any REG_EQUAL or REG_EQUIV notes containing a reference to the
+ memory location in SMEXPR set in basic block BB.
-/* This handles the case where several stores feed a partially redundant
- load. It checks if the redundancy elimination is possible and if it's
- worth it. */
+ This could be rather expensive. */
static void
-eliminate_partially_redundant_loads (basic_block bb, rtx insn,
- struct expr *expr)
+remove_reachable_equiv_notes (basic_block bb, struct ls_expr *smexpr)
{
- edge pred;
- rtx avail_insn = NULL_RTX;
- rtx avail_reg;
- rtx dest, pat;
- struct occr *a_occr;
- struct unoccr *occr, *avail_occrs = NULL;
- struct unoccr *unoccr, *unavail_occrs = NULL;
- int npred_ok = 0;
- gcov_type ok_count = 0; /* Redundant load execution count. */
- gcov_type critical_count = 0; /* Execution count of critical edges. */
-
- /* The execution count of the loads to be added to make the
- load fully redundant. */
- gcov_type not_ok_count = 0;
- basic_block pred_bb;
-
- pat = PATTERN (insn);
- dest = SET_DEST (pat);
- /* Check that the loaded register is not used, set, or killed from the
- beginning of the block. */
- if (reg_used_between_after_reload_p (dest,
- PREV_INSN (BB_HEAD (bb)), insn)
- || reg_set_between_after_reload_p (dest,
- PREV_INSN (BB_HEAD (bb)), insn))
- return;
+ edge_iterator *stack, ei;
+ int sp;
+ edge act;
+ sbitmap visited = sbitmap_alloc (last_basic_block);
+ rtx last, insn, note;
+ rtx mem = smexpr->pattern;
- /* Check potential for replacing load with copy for predecessors. */
- for (pred = bb->pred; pred; pred = pred->pred_next)
- {
- rtx next_pred_bb_end;
+ stack = xmalloc (sizeof (edge_iterator) * n_basic_blocks);
+ sp = 0;
+ ei = ei_start (bb->succs);
+
+ sbitmap_zero (visited);
- avail_insn = NULL_RTX;
- pred_bb = pred->src;
- next_pred_bb_end = NEXT_INSN (BB_END (pred_bb));
- for (a_occr = get_bb_avail_insn (pred_bb, expr->avail_occr); a_occr;
- a_occr = get_bb_avail_insn (pred_bb, a_occr->next))
+ act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
+ while (1)
+ {
+ if (!act)
{
- /* Check if the loaded register is not used. */
- avail_insn = a_occr->insn;
- if (! (avail_reg = get_avail_load_store_reg (avail_insn)))
- abort ();
- /* Make sure we can generate a move from register avail_reg to
- dest. */
- extract_insn (gen_move_insn (copy_rtx (dest),
- copy_rtx (avail_reg)));
- if (! constrain_operands (1)
- || reg_killed_on_edge (avail_reg, pred)
- || reg_used_on_edge (dest, pred))
+ if (!sp)
{
- avail_insn = NULL;
- continue;
+ free (stack);
+ sbitmap_free (visited);
+ return;
}
- if (! reg_set_between_after_reload_p (avail_reg, avail_insn,
- next_pred_bb_end))
- /* AVAIL_INSN remains non-null. */
- break;
- else
- avail_insn = NULL;
+ act = ei_edge (stack[--sp]);
}
- if (avail_insn != NULL_RTX)
+ bb = act->dest;
+
+ if (bb == EXIT_BLOCK_PTR
+ || TEST_BIT (visited, bb->index))
{
- npred_ok++;
- ok_count += pred->count;
- if (EDGE_CRITICAL_P (pred))
- critical_count += pred->count;
- occr = gmalloc (sizeof (struct unoccr));
- occr->insn = avail_insn;
- occr->pred = pred;
- occr->next = avail_occrs;
- avail_occrs = occr;
+ if (!ei_end_p (ei))
+ ei_next (&ei);
+ act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
+ continue;
}
- else
+ SET_BIT (visited, bb->index);
+
+ if (TEST_BIT (st_antloc[bb->index], smexpr->index))
{
- not_ok_count += pred->count;
- if (EDGE_CRITICAL_P (pred))
- critical_count += pred->count;
- unoccr = gmalloc (sizeof (struct unoccr));
- unoccr->insn = NULL_RTX;
- unoccr->pred = pred;
- unoccr->next = unavail_occrs;
- unavail_occrs = unoccr;
+ for (last = ANTIC_STORE_LIST (smexpr);
+ BLOCK_FOR_INSN (XEXP (last, 0)) != bb;
+ last = XEXP (last, 1))
+ continue;
+ last = XEXP (last, 0);
}
- }
-
- if (npred_ok == 0 /* No load can be replaced by copy. */
- || (optimize_size && npred_ok > 1)) /* Prevent exploding the code. */
- return;
+ else
+ last = NEXT_INSN (BB_END (bb));
- /* Check if it's worth applying the partial redundancy elimination. */
- if (ok_count < GCSE_AFTER_RELOAD_PARTIAL_FRACTION * not_ok_count)
- return;
+ for (insn = BB_HEAD (bb); insn != last; insn = NEXT_INSN (insn))
+ if (INSN_P (insn))
+ {
+ note = find_reg_equal_equiv_note (insn);
+ if (!note || !expr_equiv_p (XEXP (note, 0), mem))
+ continue;
- if (ok_count < GCSE_AFTER_RELOAD_CRITICAL_FRACTION * critical_count)
- return;
+ if (gcse_file)
+ fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
+ INSN_UID (insn));
+ remove_note (insn, note);
+ }
- /* Generate moves to the loaded register from where
- the memory is available. */
- for (occr = avail_occrs; occr; occr = occr->next)
- {
- avail_insn = occr->insn;
- pred = occr->pred;
- /* Set avail_reg to be the register having the value of the
- memory. */
- avail_reg = get_avail_load_store_reg (avail_insn);
- if (! avail_reg)
- abort ();
-
- insert_insn_on_edge (gen_move_insn (copy_rtx (dest),
- copy_rtx (avail_reg)),
- pred);
-
- if (gcse_file)
- fprintf (gcse_file,
- "GCSE AFTER reload generating move from %d to %d on \
- edge from %d to %d\n",
- REGNO (avail_reg),
- REGNO (dest),
- pred->src->index,
- pred->dest->index);
- }
+ if (!ei_end_p (ei))
+ ei_next (&ei);
+ act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
- /* Regenerate loads where the memory is unavailable. */
- for (unoccr = unavail_occrs; unoccr; unoccr = unoccr->next)
- {
- pred = unoccr->pred;
- insert_insn_on_edge (copy_insn (PATTERN (insn)), pred);
-
- if (gcse_file)
- fprintf (gcse_file,
- "GCSE AFTER reload: generating on edge from %d to %d\
- a copy of load:\n",
- pred->src->index,
- pred->dest->index);
+ if (EDGE_COUNT (bb->succs) > 0)
+ {
+ if (act)
+ stack[sp++] = ei;
+ ei = ei_start (bb->succs);
+ act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
+ }
}
-
- /* Delete the insn if it is not available in this block and mark it
- for deletion if it is available. If insn is available it may help
- discover additional redundancies, so mark it for later deletion.*/
- for (a_occr = get_bb_avail_insn (bb, expr->avail_occr);
- a_occr && (a_occr->insn != insn);
- a_occr = get_bb_avail_insn (bb, a_occr->next));
-
- if (!a_occr)
- delete_insn (insn);
- else
- a_occr->deleted_p = 1;
}
-/* Performing the redundancy elimination as described before. */
+/* This routine will replace a store with a SET to a specified register. */
static void
-gcse_after_reload (void)
+replace_store_insn (rtx reg, rtx del, basic_block bb, struct ls_expr *smexpr)
{
- unsigned int i;
- rtx insn;
- basic_block bb;
- struct expr *expr;
- struct occr *occr;
-
- /* Note we start at block 1. */
+ rtx insn, mem, note, set, ptr, pair;
- if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
- return;
+ mem = smexpr->pattern;
+ insn = gen_move_insn (reg, SET_SRC (single_set (del)));
+ insn = emit_insn_after (insn, del);
- FOR_BB_BETWEEN (bb,
- ENTRY_BLOCK_PTR->next_bb->next_bb,
- EXIT_BLOCK_PTR,
- next_bb)
+ if (gcse_file)
{
- if (! bb_has_well_behaved_predecessors (bb))
- continue;
-
- /* Do not try this optimization on cold basic blocks. */
- if (probably_cold_bb_p (bb))
- continue;
+ 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");
+ }
- reset_opr_set_tables ();
+ for (ptr = ANTIC_STORE_LIST (smexpr); ptr; ptr = XEXP (ptr, 1))
+ if (XEXP (ptr, 0) == del)
+ {
+ XEXP (ptr, 0) = insn;
+ break;
+ }
- for (insn = BB_HEAD (bb);
- insn != NULL
- && insn != NEXT_INSN (BB_END (bb));
- insn = NEXT_INSN (insn))
- {
- /* Is it a load - of the form (set (reg) (mem))? */
- if (GET_CODE (insn) == INSN
- && GET_CODE (PATTERN (insn)) == SET
- && GET_CODE (SET_DEST (PATTERN (insn))) == REG
- && GET_CODE (SET_SRC (PATTERN (insn))) == MEM)
- {
- rtx pat = PATTERN (insn);
- rtx src = SET_SRC (pat);
- struct expr *expr;
-
- if (general_operand (src, GET_MODE (src))
- /* Is the expression recorded? */
- && (expr = lookup_expr (src, &expr_hash_table)) != NULL
- /* Are the operands unchanged since the start of the
- block? */
- && oprs_not_set_p (src, insn)
- && ! MEM_VOLATILE_P (src)
- && GET_MODE (src) != BLKmode
- && !(flag_non_call_exceptions && may_trap_p (src))
- && !side_effects_p (src))
- {
- /* We now have a load (insn) and an available memory at
- its BB start (expr). Try to remove the loads if it is
- redundant. */
- eliminate_partially_redundant_loads (bb, insn, expr);
- }
- }
+ /* Move the notes from the deleted insn to its replacement, and patch
+ up the LIBCALL notes. */
+ REG_NOTES (insn) = REG_NOTES (del);
- /* Keep track of everything modified by this insn. */
- if (INSN_P (insn))
- mark_oprs_set (insn);
- }
+ note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
+ if (note)
+ {
+ pair = XEXP (note, 0);
+ note = find_reg_note (pair, REG_LIBCALL, NULL_RTX);
+ XEXP (note, 0) = insn;
+ }
+ note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
+ if (note)
+ {
+ pair = XEXP (note, 0);
+ note = find_reg_note (pair, REG_RETVAL, NULL_RTX);
+ XEXP (note, 0) = insn;
}
- commit_edge_insertions ();
+ delete_insn (del);
+
+ /* Now we must handle REG_EQUAL notes whose contents is equal to the mem;
+ they are no longer accurate provided that they are reached by this
+ definition, so drop them. */
+ for (; insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn))
+ if (INSN_P (insn))
+ {
+ set = single_set (insn);
+ if (!set)
+ continue;
+ if (expr_equiv_p (SET_DEST (set), mem))
+ return;
+ note = find_reg_equal_equiv_note (insn);
+ if (!note || !expr_equiv_p (XEXP (note, 0), mem))
+ continue;
- /* Go over the expression hash table and delete insns that were
- marked for later deletion. */
- for (i = 0; i < expr_hash_table.size; i++)
- {
- for (expr = expr_hash_table.table[i];
- expr != NULL;
- expr = expr->next_same_hash)
- for (occr = expr->avail_occr; occr; occr = occr->next)
- if (occr->deleted_p)
- delete_insn (occr->insn);
- }
+ if (gcse_file)
+ fprintf (gcse_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
+ INSN_UID (insn));
+ remove_note (insn, note);
+ }
+ remove_reachable_equiv_notes (bb, smexpr);
}
-/* Scan pattern PAT of INSN and add an entry to the hash TABLE.
- After reload we are interested in loads/stores only. */
+
+/* Delete a store, but copy the value that would have been stored into
+ the reaching_reg for later storing. */
static void
-hash_scan_set_after_reload (rtx pat, rtx insn, struct hash_table *table)
+delete_store (struct ls_expr * expr, basic_block bb)
{
- rtx src = SET_SRC (pat);
- rtx dest = SET_DEST (pat);
+ rtx reg, i, del;
- if (GET_CODE (src) != MEM && GET_CODE (dest) != MEM)
- return;
+ if (expr->reaching_reg == NULL_RTX)
+ expr->reaching_reg = gen_reg_rtx (GET_MODE (expr->pattern));
- if (GET_CODE (dest) == REG)
- {
- if (/* 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)
- /* Is SET_SRC something we want to gcse? */
- && general_operand (src, GET_MODE (src))
- /* Don't CSE a nop. */
- && ! set_noop_p (pat)
- && ! JUMP_P (insn))
- {
- /* An expression is not available if its operands are
- subsequently modified, including this insn. */
- if (oprs_available_p (src, insn))
- insert_expr_in_table (src, GET_MODE (dest), insn, 0, 1, table);
- }
- }
- else if ((GET_CODE (src) == REG))
+ reg = expr->reaching_reg;
+
+ for (i = AVAIL_STORE_LIST (expr); i; i = XEXP (i, 1))
{
- /* Only record sets of pseudo-regs in the hash table. */
- if (/* Don't GCSE something if we can't do a reg/reg copy. */
- can_copy_p (GET_MODE (src))
- /* GCSE commonly inserts instruction after the insn. We can't
- do that easily for EH_REGION notes so disable GCSE on these
- for now. */
- && ! find_reg_note (insn, REG_EH_REGION, NULL_RTX)
- /* Is SET_DEST something we want to gcse? */
- && general_operand (dest, GET_MODE (dest))
- /* Don't CSE a nop. */
- && ! set_noop_p (pat)
- &&! JUMP_P (insn)
- && ! (flag_float_store && FLOAT_MODE_P (GET_MODE (dest)))
- /* Check if the memory expression is killed after insn. */
- && ! load_killed_in_block_p (BLOCK_FOR_INSN (insn),
- INSN_CUID (insn) + 1,
- dest,
- 1)
- && oprs_unchanged_p (XEXP (dest, 0), insn, 1))
+ del = XEXP (i, 0);
+ if (BLOCK_FOR_INSN (del) == bb)
{
- insert_expr_in_table (dest, GET_MODE (dest), insn, 0, 1, table);
+ /* We know there is only one since we deleted redundant
+ ones during the available computation. */
+ replace_store_insn (reg, del, bb, expr);
+ break;
}
}
}
-
-/* Create hash table of memory expressions available at end of basic
- blocks. */
+/* Free memory used by store motion. */
static void
-compute_hash_table_after_reload (struct hash_table *table)
+free_store_memory (void)
{
- unsigned int i;
-
- table->set_p = 0;
-
- /* Initialize count of number of entries in hash table. */
- table->n_elems = 0;
- memset ((char *) table->table, 0,
- table->size * sizeof (struct expr *));
+ free_ldst_mems ();
- /* While we compute the hash table we also compute a bit array of which
- registers are set in which blocks. */
- sbitmap_vector_zero (reg_set_in_block, last_basic_block);
+ 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);
- /* Re-cache any INSN_LIST nodes we have allocated. */
- clear_modify_mem_tables ();
+ ae_gen = ae_kill = transp = st_antloc = NULL;
+ pre_insert_map = pre_delete_map = reg_set_in_block = NULL;
+}
- /* 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));
+/* Perform store motion. Much like gcse, except we move expressions the
+ other way by looking at the flowgraph in reverse. */
- for (i = 0; i < max_gcse_regno; ++i)
- reg_avail_info[i].last_bb = NULL;
+static void
+store_motion (void)
+{
+ basic_block bb;
+ int x;
+ struct ls_expr * ptr;
+ int update_flow = 0;
- FOR_EACH_BB (current_bb)
+ if (gcse_file)
{
- rtx insn;
- unsigned int regno;
-
- /* First pass over the instructions records information used to
- determine when registers and memory are first and last set. */
- for (insn = BB_HEAD (current_bb);
- insn && insn != NEXT_INSN (BB_END (current_bb));
- insn = NEXT_INSN (insn))
- {
- if (! INSN_P (insn))
- continue;
+ fprintf (gcse_file, "before store motion\n");
+ print_rtl (gcse_file, get_insns ());
+ }
- if (GET_CODE (insn) == CALL_INSN)
- {
- bool clobbers_all = false;
+ init_alias_analysis ();
-#ifdef NON_SAVING_SETJMP
- if (NON_SAVING_SETJMP
- && find_reg_note (insn, REG_SETJMP, NULL_RTX))
- clobbers_all = true;
-#endif
+ /* 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;
+ }
- for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
- if (clobbers_all
- || TEST_HARD_REG_BIT (regs_invalidated_by_call,
- regno))
- record_last_reg_set_info (insn, regno);
+ /* Now compute kill & transp vectors. */
+ build_store_vectors ();
+ add_noreturn_fake_exit_edges ();
+ connect_infinite_loops_to_exit ();
- mark_call (insn);
- }
+ edge_list = pre_edge_rev_lcm (gcse_file, num_stores, transp, ae_gen,
+ st_antloc, ae_kill, &pre_insert_map,
+ &pre_delete_map);
- note_stores (PATTERN (insn), record_last_set_info, insn);
+ /* Now we want to insert the new stores which are going to be needed. */
+ for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
+ {
+ /* If any of the edges we have above are abnormal, we can't move this
+ store. */
+ for (x = NUM_EDGES (edge_list) - 1; x >= 0; x--)
+ if (TEST_BIT (pre_insert_map[x], ptr->index)
+ && (INDEX_EDGE (edge_list, x)->flags & EDGE_ABNORMAL))
+ break;
- if (GET_CODE (PATTERN (insn)) == SET)
- {
- rtx src, dest;
+ if (x >= 0)
+ {
+ if (gcse_file != NULL)
+ fprintf (gcse_file,
+ "Can't replace store %d: abnormal edge from %d to %d\n",
+ ptr->index, INDEX_EDGE (edge_list, x)->src->index,
+ INDEX_EDGE (edge_list, x)->dest->index);
+ continue;
+ }
+
+ /* Now we want to insert the new stores which are going to be needed. */
- src = SET_SRC (PATTERN (insn));
- dest = SET_DEST (PATTERN (insn));
- if (GET_CODE (src) == MEM && auto_inc_p (XEXP (src, 0)))
- {
- regno = REGNO (XEXP (XEXP (src, 0), 0));
- record_last_reg_set_info (insn, regno);
- }
- if (GET_CODE (dest) == MEM && auto_inc_p (XEXP (dest, 0)))
- {
- regno = REGNO (XEXP (XEXP (dest, 0), 0));
- record_last_reg_set_info (insn, regno);
- }
- }
- }
+ FOR_EACH_BB (bb)
+ if (TEST_BIT (pre_delete_map[bb->index], ptr->index))
+ delete_store (ptr, bb);
- /* The next pass builds the hash table. */
- for (insn = BB_HEAD (current_bb);
- insn && insn != NEXT_INSN (BB_END (current_bb));
- insn = NEXT_INSN (insn))
- if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == SET)
- if (! find_reg_note (insn, REG_LIBCALL, NULL_RTX))
- hash_scan_set_after_reload (PATTERN (insn), insn, table);
+ 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));
}
- free (reg_avail_info);
- reg_avail_info = NULL;
-}
+ if (update_flow)
+ commit_edge_insertions ();
+ free_store_memory ();
+ free_edge_list (edge_list);
+ remove_fake_exit_edges ();
+ end_alias_analysis ();
+}
-/* Main entry point of the GCSE after reload - clean some redundant loads
- due to spilling. */
+\f
+/* Entry point for jump bypassing optimization pass. */
-void
-gcse_after_reload_main (rtx f, FILE* file)
+int
+bypass_jumps (FILE *file)
{
- gcse_subst_count = 0;
- gcse_create_count = 0;
+ int changed;
+
+ /* We do not construct an accurate cfg in functions which call
+ setjmp, so just punt to be safe. */
+ if (current_function_calls_setjmp)
+ return 0;
+ /* For calling dump_foo fns from gdb. */
+ debug_stderr = stderr;
gcse_file = file;
+ /* Identify the basic block information for this function, including
+ successors and predecessors. */
+ max_gcse_regno = max_reg_num ();
+
+ if (file)
+ dump_flow_info (file);
+
+ /* Return if there's nothing to do, or it is too expensive. */
+ if (n_basic_blocks <= 1 || is_too_expensive (_ ("jump bypassing disabled")))
+ return 0;
+
gcc_obstack_init (&gcse_obstack);
bytes_used = 0;
/* We need alias. */
init_alias_analysis ();
- max_gcse_regno = max_reg_num ();
+ /* 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.
- alloc_reg_set_mem (max_gcse_regno);
- alloc_gcse_mem (f);
- alloc_hash_table (max_cuid, &expr_hash_table, 0);
- compute_hash_table_after_reload (&expr_hash_table);
+ 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. */
- if (gcse_file)
- dump_hash_table (gcse_file, "Expression", &expr_hash_table);
+ alloc_reg_set_mem (max_gcse_regno);
+ compute_sets (get_insns ());
- if (expr_hash_table.n_elems > 0)
- gcse_after_reload ();
+ max_gcse_regno = max_reg_num ();
+ alloc_gcse_mem (get_insns ());
+ changed = one_cprop_pass (MAX_GCSE_PASSES + 2, 1, 1);
+ free_gcse_mem ();
- free_hash_table (&expr_hash_table);
+ if (file)
+ {
+ fprintf (file, "BYPASS of %s: %d basic blocks, ",
+ current_function_name (), n_basic_blocks);
+ fprintf (file, "%d bytes\n\n", bytes_used);
+ }
- free_gcse_mem ();
+ 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);
- obstack_free (&gcse_obstack, NULL);
+ return changed;
+}
+
+/* Return true if the graph is too expensive to optimize. PASS is the
+ optimization about to be performed. */
+
+static bool
+is_too_expensive (const char *pass)
+{
+ /* Trying to perform global optimizations on flow graphs which have
+ a high connectivity will take a long time and is unlikely to be
+ particularly useful.
+
+ In normal circumstances a cfg should have about twice as many
+ edges as blocks. But we do not want to punish small functions
+ which have a couple switch statements. Rather than simply
+ threshold the number of blocks, uses something with a more
+ graceful degradation. */
+ if (n_edges > 20000 + n_basic_blocks * 4)
+ {
+ if (warn_disabled_optimization)
+ warning ("%s: %d basic blocks and %d edges/basic block",
+ pass, n_basic_blocks, n_edges / n_basic_blocks);
+
+ return true;
+ }
+
+ /* If allocating memory for the cprop bitmap would take up too much
+ storage it's better just to disable the optimization. */
+ if ((n_basic_blocks
+ * SBITMAP_SET_SIZE (max_reg_num ())
+ * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
+ {
+ if (warn_disabled_optimization)
+ warning ("%s: %d basic blocks and %d registers",
+ pass, n_basic_blocks, max_reg_num ());
+
+ return true;
+ }
+
+ return false;
}
#include "gt-gcse.h"
OSDN Git Service
