#include "tm.h"
#include "rtl.h"
#include "tree.h"
+#include "tm_p.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "flags.h"
/* An bitmask as wide as the number of bytes in the word that
contains a 1 if the byte may be needed. The store is unused if
all of the bits are 0. */
- long positions_needed;
+ unsigned HOST_WIDE_INT positions_needed;
/* The next store info for this insn. */
struct store_info *next;
rtx rhs;
};
+/* Return a bitmask with the first N low bits set. */
+
+static unsigned HOST_WIDE_INT
+lowpart_bitmask (int n)
+{
+ unsigned HOST_WIDE_INT mask = ~(unsigned HOST_WIDE_INT) 0;
+ return mask >> (HOST_BITS_PER_WIDE_INT - n);
+}
+
typedef struct store_info *store_info_t;
static alloc_pool cse_store_info_pool;
static alloc_pool rtx_store_info_pool;
contains a wild read, the use_rec will be null. */
bool wild_read;
- /* This field is set for const function calls. Const functions
- cannot read memory, but they can read the stack because that is
- where they may get their parms. So having this set is less
- severe than a wild read, it just means that all of the stores to
- the stack are killed rather than all stores. */
- bool stack_read;
+ /* This field is only used for the processing of const functions.
+ These functions cannot read memory, but they can read the stack
+ because that is where they may get their parms. We need to be
+ this conservative because, like the store motion pass, we don't
+ consider CALL_INSN_FUNCTION_USAGE when processing call insns.
+ Moreover, we need to distinguish two cases:
+ 1. Before reload (register elimination), the stores related to
+ outgoing arguments are stack pointer based and thus deemed
+ of non-constant base in this pass. This requires special
+ handling but also means that the frame pointer based stores
+ need not be killed upon encountering a const function call.
+ 2. After reload, the stores related to outgoing arguments can be
+ either stack pointer or hard frame pointer based. This means
+ that we have no other choice than also killing all the frame
+ pointer based stores upon encountering a const function call.
+ This field is set after reload for const function calls. Having
+ this set is less severe than a wild read, it just means that all
+ the frame related stores are killed rather than all the stores. */
+ bool frame_read;
+
+ /* This field is only used for the processing of const functions.
+ It is set if the insn may contain a stack pointer based store. */
+ bool stack_pointer_based;
/* This is true if any of the sets within the store contains a
cselib base. Such stores can only be deleted by the local
static alloc_pool clear_alias_mode_pool;
-/* This is true except for two cases:
- (1) current_function_stdarg -- i.e. we cannot do this
- for vararg functions because they play games with the frame.
- (2) In ada, it is sometimes not safe to do assume that any stores
- based off the stack frame go dead at the exit to a function. */
+/* This is true except if current_function_stdarg -- i.e. we cannot do
+ this for vararg functions because they play games with the frame. */
static bool stores_off_frame_dead_at_return;
/* Counter for stats. */
static bool gate_dse (void);
+static bool gate_dse1 (void);
+static bool gate_dse2 (void);
\f
/*----------------------------------------------------------------------------
bb_table = XCNEWVEC (bb_info_t, last_basic_block);
rtx_group_next_id = 0;
- stores_off_frame_dead_at_return =
- (!(TREE_CODE (TREE_TYPE (current_function_decl)) == FUNCTION_TYPE
- && (TYPE_RETURNS_STACK_DEPRESSED (TREE_TYPE (current_function_decl)))))
- && (!current_function_stdarg);
+ stores_off_frame_dead_at_return = !current_function_stdarg;
init_alias_analysis ();
}
-/* Return true if X is a constant or one of the registers that behaves
- as a constant over the life of a function. */
+/* Return true if X is a constant or one of the registers that behave
+ as a constant over the life of a function. This is equivalent to
+ !rtx_varies_p for memory addresses. */
static bool
const_or_frame_p (rtx x)
}
else
{
- store_info = pool_alloc (cse_store_info_pool);
+ rtx base_term = find_base_term (XEXP (mem, 0));
+ if (!base_term
+ || (GET_CODE (base_term) == ADDRESS
+ && GET_MODE (base_term) == Pmode
+ && XEXP (base_term, 0) == stack_pointer_rtx))
+ insn_info->stack_pointer_based = true;
insn_info->contains_cselib_groups = true;
+
+ store_info = pool_alloc (cse_store_info_pool);
group_id = -1;
if (dump_file)
&& (GET_MODE (mem) == entry->mode))
{
delete = true;
- s_info->positions_needed = 0;
+ s_info->positions_needed = (unsigned HOST_WIDE_INT) 0;
}
if (dump_file)
fprintf (dump_file, " trying spill store in insn=%d alias_set=%d\n",
(int)s_info->begin, (int)s_info->end);
for (i = offset; i < offset+width; i++)
if (i >= s_info->begin && i < s_info->end)
- s_info->positions_needed &= ~(1L << (i - s_info->begin));
+ s_info->positions_needed
+ &= ~(((unsigned HOST_WIDE_INT) 1) << (i - s_info->begin));
}
else if (s_info->rhs)
/* Need to see if it is possible for this store to overwrite
/* An insn can be deleted if every position of every one of
its s_infos is zero. */
- if (s_info->positions_needed != 0)
+ if (s_info->positions_needed != (unsigned HOST_WIDE_INT) 0)
delete = false;
if (delete)
ptr = next;
}
- gcc_assert ((unsigned) width < sizeof (store_info->positions_needed) * CHAR_BIT);
+ gcc_assert ((unsigned) width <= HOST_BITS_PER_WIDE_INT);
/* Finish filling in the store_info. */
store_info->next = insn_info->store_rec;
store_info->alias_set = spill_alias_set;
store_info->mem_addr = get_addr (XEXP (mem, 0));
store_info->cse_base = base;
- store_info->positions_needed = (1L << width) - 1;
+ store_info->positions_needed = lowpart_bitmask (width);
store_info->group_id = group_id;
store_info->begin = offset;
store_info->end = offset + width;
if (store_info->is_set
/* No place to keep the value after ra. */
&& !reload_completed
- /* The careful reviewer may wish to comment my checking that the
- rhs of a store is always a reg. */
- && REG_P (SET_SRC (body))
+ && (REG_P (SET_SRC (body))
+ || GET_CODE (SET_SRC (body)) == SUBREG
+ || CONSTANT_P (SET_SRC (body)))
/* Sometimes the store and reload is used for truncation and
rounding. */
&& !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store)))
shift. */
static rtx
-find_shift_sequence (rtx read_reg,
- int access_size,
+find_shift_sequence (int access_size,
store_info_t store_info,
read_info_t read_info,
int shift)
{
enum machine_mode store_mode = GET_MODE (store_info->mem);
enum machine_mode read_mode = GET_MODE (read_info->mem);
+ enum machine_mode new_mode;
+ rtx read_reg = NULL;
/* Some machines like the x86 have shift insns for each size of
operand. Other machines like the ppc or the ia-64 may only have
justify the value we want to read but is available in one insn on
the machine. */
- for (; access_size <= UNITS_PER_WORD; access_size *= 2)
+ for (new_mode = smallest_mode_for_size (access_size * BITS_PER_UNIT,
+ MODE_INT);
+ GET_MODE_BITSIZE (new_mode) <= BITS_PER_WORD;
+ new_mode = GET_MODE_WIDER_MODE (new_mode))
{
- rtx target, new_reg;
- enum machine_mode new_mode;
+ rtx target, new_reg, shift_seq, insn, new_lhs;
+ int cost;
- /* Try a wider mode if truncating the store mode to ACCESS_SIZE
- bytes requires a real instruction. */
- if (access_size < GET_MODE_SIZE (store_mode)
- && !TRULY_NOOP_TRUNCATION (access_size * BITS_PER_UNIT,
+ /* Try a wider mode if truncating the store mode to NEW_MODE
+ requires a real instruction. */
+ if (GET_MODE_BITSIZE (new_mode) < GET_MODE_BITSIZE (store_mode)
+ && !TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (new_mode),
GET_MODE_BITSIZE (store_mode)))
continue;
- new_mode = smallest_mode_for_size (access_size * BITS_PER_UNIT,
- GET_MODE_CLASS (read_mode));
+ /* Also try a wider mode if the necessary punning is either not
+ desirable or not possible. */
+ if (!CONSTANT_P (store_info->rhs)
+ && !MODES_TIEABLE_P (new_mode, store_mode))
+ continue;
+ new_lhs = simplify_gen_subreg (new_mode, copy_rtx (store_info->rhs),
+ store_mode, 0);
+ if (new_lhs == NULL_RTX)
+ continue;
+
new_reg = gen_reg_rtx (new_mode);
start_sequence ();
target = expand_binop (new_mode, lshr_optab, new_reg,
GEN_INT (shift), new_reg, 1, OPTAB_DIRECT);
- if (target == new_reg)
- {
- rtx shift_seq = get_insns ();
- end_sequence ();
+ shift_seq = get_insns ();
+ end_sequence ();
- /* If cost is too great, set target to NULL and
- let the iteration happen. */
- if (shift_seq != NULL)
- {
- int cost = 0;
- rtx insn;
-
- for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- cost += insn_rtx_cost (PATTERN (insn));
-
- /* The computation up to here is essentially independent
- of the arguments and could be precomputed. It may
- not be worth doing so. We could precompute if
- worthwhile or at least cache the results. The result
- technically depends on SHIFT, ACCESS_SIZE, and
- GET_MODE_CLASS (READ_MODE). But in practice the
- answer will depend only on ACCESS_SIZE. */
-
- if (cost <= COSTS_N_INSNS (1))
- {
- /* We found an acceptable shift. Generate a move to
- take the value from the store and put it into the
- shift pseudo, then shift it, then generate another
- move to put in into the target of the read. */
- start_sequence ();
- emit_move_insn (new_reg, gen_lowpart (new_mode, store_info->rhs));
- emit_insn (shift_seq);
- convert_move (read_reg, new_reg, 1);
-
- if (dump_file)
- {
- fprintf (dump_file, " -- adding extract insn r%d:%s = r%d:%s\n",
- REGNO (new_reg), GET_MODE_NAME (new_mode),
- REGNO (store_info->rhs), GET_MODE_NAME (store_mode));
-
- fprintf (dump_file, " -- with shift of r%d by %d\n",
- REGNO(new_reg), shift);
- fprintf (dump_file, " -- and second extract insn r%d:%s = r%d:%s\n",
- REGNO (read_reg), GET_MODE_NAME (read_mode),
- REGNO (new_reg), GET_MODE_NAME (new_mode));
- }
-
- /* Get the three insn sequence and return it. */
- shift_seq = get_insns ();
- end_sequence ();
- return shift_seq;
- }
- }
- }
- else
- /* End the sequence. */
- end_sequence ();
+ if (target != new_reg || shift_seq == NULL)
+ continue;
+
+ cost = 0;
+ for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn))
+ if (INSN_P (insn))
+ cost += insn_rtx_cost (PATTERN (insn));
+
+ /* The computation up to here is essentially independent
+ of the arguments and could be precomputed. It may
+ not be worth doing so. We could precompute if
+ worthwhile or at least cache the results. The result
+ technically depends on both SHIFT and ACCESS_SIZE,
+ but in practice the answer will depend only on ACCESS_SIZE. */
+
+ if (cost > COSTS_N_INSNS (1))
+ continue;
+
+ /* We found an acceptable shift. Generate a move to
+ take the value from the store and put it into the
+ shift pseudo, then shift it, then generate another
+ move to put in into the target of the read. */
+ emit_move_insn (new_reg, new_lhs);
+ emit_insn (shift_seq);
+ read_reg = extract_low_bits (read_mode, new_mode, new_reg);
+ break;
}
- return NULL;
+ return read_reg;
}
enum machine_mode read_mode = GET_MODE (read_info->mem);
int shift;
int access_size; /* In bytes. */
- rtx read_reg = gen_reg_rtx (read_mode);
- rtx shift_seq = NULL;
+ rtx insns, read_reg;
if (!dbg_cnt (dse))
return false;
- if (GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))
- return false;
-
/* To get here the read is within the boundaries of the write so
shift will never be negative. Start out with the shift being in
bytes. */
/* From now on it is bits. */
shift *= BITS_PER_UNIT;
- /* We need to keep this in perspective. We are replacing a read
+ /* Create a sequence of instructions to set up the read register.
+ This sequence goes immediately before the store and its result
+ is read by the load.
+
+ We need to keep this in perspective. We are replacing a read
with a sequence of insns, but the read will almost certainly be
in cache, so it is not going to be an expensive one. Thus, we
are not willing to do a multi insn shift or worse a subroutine
call to get rid of the read. */
+ if (dump_file)
+ fprintf (dump_file, "trying to replace %smode load in insn %d"
+ " from %smode store in insn %d\n",
+ GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn),
+ GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn));
+ start_sequence ();
if (shift)
+ read_reg = find_shift_sequence (access_size, store_info, read_info, shift);
+ else
+ read_reg = extract_low_bits (read_mode, store_mode,
+ copy_rtx (store_info->rhs));
+ if (read_reg == NULL_RTX)
{
- if (access_size > UNITS_PER_WORD || FLOAT_MODE_P (store_mode))
- return false;
-
- shift_seq = find_shift_sequence (read_reg, access_size, store_info,
- read_info, shift);
- if (!shift_seq)
- return false;
+ end_sequence ();
+ if (dump_file)
+ fprintf (dump_file, " -- could not extract bits of stored value\n");
+ return false;
}
-
- if (dump_file)
- fprintf (dump_file, "replacing load at %d from store at %d\n",
- INSN_UID (read_insn->insn), INSN_UID (store_insn->insn));
+ /* Force the value into a new register so that it won't be clobbered
+ between the store and the load. */
+ read_reg = copy_to_mode_reg (read_mode, read_reg);
+ insns = get_insns ();
+ end_sequence ();
if (validate_change (read_insn->insn, loc, read_reg, 0))
{
- rtx insns;
deferred_change_t deferred_change = pool_alloc (deferred_change_pool);
- if (read_mode == store_mode)
- {
- start_sequence ();
-
- /* The modes are the same and everything lines up. Just
- generate a simple move. */
- emit_move_insn (read_reg, store_info->rhs);
- if (dump_file)
- fprintf (dump_file, " -- adding move insn r%d = r%d\n",
- REGNO (read_reg), REGNO (store_info->rhs));
- insns = get_insns ();
- end_sequence ();
- }
- else if (shift)
- insns = shift_seq;
- else
- {
- /* The modes are different but the lsb are in the same
- place, we need to extract the value in the right from the
- rhs of the store. */
- start_sequence ();
- convert_move (read_reg, store_info->rhs, 1);
-
- if (dump_file)
- fprintf (dump_file, " -- adding extract insn r%d:%s = r%d:%s\n",
- REGNO (read_reg), GET_MODE_NAME (read_mode),
- REGNO (store_info->rhs), GET_MODE_NAME (store_mode));
- insns = get_insns ();
- end_sequence ();
- }
-
/* Insert this right before the store insn where it will be safe
from later insns that might change it before the read. */
emit_insn_before (insns, store_insn->insn);
rest of dse, play like this read never happened. */
read_insn->read_rec = read_info->next;
pool_free (read_info_pool, read_info);
+ if (dump_file)
+ {
+ fprintf (dump_file, " -- replaced the loaded MEM with ");
+ print_simple_rtl (dump_file, read_reg);
+ fprintf (dump_file, "\n");
+ }
return true;
}
else
{
if (dump_file)
- fprintf (dump_file, " -- validation failure\n");
+ {
+ fprintf (dump_file, " -- replacing the loaded MEM with ");
+ print_simple_rtl (dump_file, read_reg);
+ fprintf (dump_file, " led to an invalid instruction\n");
+ }
return false;
}
}
&& (offset >= store_info->begin)
&& (offset + width <= store_info->end))
{
- int mask = ((1L << width) - 1) << (offset - store_info->begin);
-
+ unsigned HOST_WIDE_INT mask
+ = (lowpart_bitmask (width)
+ << (offset - store_info->begin));
+
if ((store_info->positions_needed & mask) == mask
&& replace_read (store_info, i_ptr,
read_info, insn_info, loc))
&& (offset >= store_info->begin)
&& (offset + width <= store_info->end))
{
- int mask = ((1L << width) - 1) << (offset - store_info->begin);
-
+ unsigned HOST_WIDE_INT mask
+ = (lowpart_bitmask (width)
+ << (offset - store_info->begin));
+
if ((store_info->positions_needed & mask) == mask
&& replace_read (store_info, i_ptr,
read_info, insn_info, loc))
if (CALL_P (insn))
{
insn_info->cannot_delete = true;
+
/* Const functions cannot do anything bad i.e. read memory,
- however, they can read their parameters which may have been
- pushed onto the stack. */
+ however, they can read their parameters which may have
+ been pushed onto the stack. */
if (CONST_OR_PURE_CALL_P (insn) && !pure_call_p (insn))
{
insn_info_t i_ptr = active_local_stores;
if (dump_file)
fprintf (dump_file, "const call %d\n", INSN_UID (insn));
+ /* See the head comment of the frame_read field. */
+ if (reload_completed)
+ insn_info->frame_read = true;
+
+ /* Loop over the active stores and remove those which are
+ killed by the const function call. */
while (i_ptr)
{
- store_info_t store_info = i_ptr->store_rec;
+ bool remove_store = false;
- /* Skip the clobbers. */
- while (!store_info->is_set)
- store_info = store_info->next;
+ /* The stack pointer based stores are always killed. */
+ if (i_ptr->stack_pointer_based)
+ remove_store = true;
+
+ /* If the frame is read, the frame related stores are killed. */
+ else if (insn_info->frame_read)
+ {
+ store_info_t store_info = i_ptr->store_rec;
+
+ /* Skip the clobbers. */
+ while (!store_info->is_set)
+ store_info = store_info->next;
- /* Remove the frame related stores. */
- if (store_info->group_id >= 0
- && VEC_index (group_info_t, rtx_group_vec, store_info->group_id)->frame_related)
+ if (store_info->group_id >= 0
+ && VEC_index (group_info_t, rtx_group_vec,
+ store_info->group_id)->frame_related)
+ remove_store = true;
+ }
+
+ if (remove_store)
{
if (dump_file)
dump_insn_info ("removing from active", i_ptr);
}
else
last = i_ptr;
+
i_ptr = i_ptr->next_local_store;
}
-
- insn_info->stack_read = true;
-
- return;
}
- /* Every other call, including pure functions may read memory. */
- add_wild_read (bb_info);
+ else
+ /* Every other call, including pure functions, may read memory. */
+ add_wild_read (bb_info);
+
return;
}
/* Assuming that there are sets in these insns, we cannot delete
them. */
if ((GET_CODE (PATTERN (insn)) == CLOBBER)
- || volatile_insn_p (PATTERN (insn))
+ || volatile_refs_p (PATTERN (insn))
|| (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
|| (RTX_FRAME_RELATED_P (insn))
|| find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX))
int i;
group_info_t group;
- /* For const function calls kill the stack related stores. */
- if (insn_info->stack_read)
+ /* If this insn reads the frame, kill all the frame related stores. */
+ if (insn_info->frame_read)
{
for (i = 0; VEC_iterate (group_info_t, rtx_group_vec, i, group); i++)
if (group->process_globally && group->frame_related)
static bool
gate_dse (void)
{
- return optimize > 0 && flag_dse;
+ return gate_dse1 () || gate_dse2 ();
+}
+
+static bool
+gate_dse1 (void)
+{
+ return optimize > 0 && flag_dse
+ && dbg_cnt (dse1);
+}
+
+static bool
+gate_dse2 (void)
+{
+ return optimize > 0 && flag_dse
+ && dbg_cnt (dse2);
}
-struct tree_opt_pass pass_rtl_dse1 =
+struct rtl_opt_pass pass_rtl_dse1 =
{
+ {
+ RTL_PASS,
"dse1", /* name */
- gate_dse, /* gate */
+ gate_dse1, /* gate */
rest_of_handle_dse, /* execute */
NULL, /* sub */
NULL, /* next */
0, /* todo_flags_start */
TODO_dump_func |
TODO_df_finish | TODO_verify_rtl_sharing |
- TODO_ggc_collect, /* todo_flags_finish */
- 'w' /* letter */
+ TODO_ggc_collect /* todo_flags_finish */
+ }
};
-struct tree_opt_pass pass_rtl_dse2 =
+struct rtl_opt_pass pass_rtl_dse2 =
{
+ {
+ RTL_PASS,
"dse2", /* name */
- gate_dse, /* gate */
+ gate_dse2, /* gate */
rest_of_handle_dse, /* execute */
NULL, /* sub */
NULL, /* next */
0, /* todo_flags_start */
TODO_dump_func |
TODO_df_finish | TODO_verify_rtl_sharing |
- TODO_ggc_collect, /* todo_flags_finish */
- 'w' /* letter */
+ TODO_ggc_collect /* todo_flags_finish */
+ }
};