- Generated by
1.8.1.1
|
GCC Middle and Back End API Reference
|
Data Structures | |
| struct | store_info |
| struct | read_info |
| struct | insn_info |
| struct | bb_info |
| struct | group_info |
| struct | deferred_change |
| struct | clear_alias_mode_holder |
| struct | invariant_group_base_hasher |
| struct | note_add_store_info |
Typedefs | |
| typedef struct store_info * | store_info_t |
| typedef struct read_info * | read_info_t |
| typedef struct insn_info * | insn_info_t |
| typedef struct bb_info * | bb_info_t |
| typedef struct group_info * | group_info_t |
| typedef struct group_info * | const_group_info_t |
| typedef struct deferred_change * | deferred_change_t |
Variables | |
| static bitmap_obstack | dse_bitmap_obstack |
| static struct obstack | dse_obstack |
| static bitmap | scratch = NULL |
| static alloc_pool | cse_store_info_pool |
| static alloc_pool | rtx_store_info_pool |
| static alloc_pool | read_info_pool |
| static alloc_pool | insn_info_pool |
| static insn_info_t | active_local_stores |
| static int | active_local_stores_len |
| static alloc_pool | bb_info_pool |
| static bb_info_t * | bb_table |
| static alloc_pool | rtx_group_info_pool |
| static int | rtx_group_next_id |
| static vec< group_info_t > | rtx_group_vec |
| static alloc_pool | deferred_change_pool |
| static deferred_change_t | deferred_change_list = NULL |
| static group_info_t | clear_alias_group |
| static htab_t | clear_alias_mode_table |
| static bool | stores_off_frame_dead_at_return |
| static int | globally_deleted |
| static int | locally_deleted |
| static int | spill_deleted |
| static bitmap | all_blocks |
| static bitmap | kill_on_calls |
| static unsigned int | current_position |
| static hash_table < invariant_group_base_hasher > | rtx_group_table |
| typedef struct group_info* const_group_info_t |
| typedef struct deferred_change* deferred_change_t |
| typedef struct group_info* group_info_t |
| typedef struct insn_info* insn_info_t |
| typedef struct read_info* read_info_t |
| typedef struct store_info* store_info_t |
|
static |
Set the BB_INFO so that the last insn is marked as a wild read of non-frame locations.
|
static |
Set the BB_INFO so that the last insn is marked as a wild read.
Referenced by replace_read(), and set_position_unneeded().
|
inlinestatic |
Return TRUE if all bytes START through START+WIDTH-1 from S_INFO store are needed.
Referenced by check_mem_read_rtx().
|
inlinestatic |
Return TRUE if any bytes from S_INFO store are needed.
|
static |
Return whether EXPR can possibly escape the current function scope.
References free_read_records(), bb_info::last_insn, reset_active_stores(), and insn_info::wild_read.
|
static |
Take all reasonable action to put the address of MEM into the form that we can do analysis on. The gold standard is to get the address into the form: address + OFFSET where address is something that rtx_varies_p considers a constant. When we can get the address in this form, we can do global analysis on it. Note that for constant bases, address is not actually returned, only the group_id. The address can be obtained from that. If that fails, we try cselib to get a value we can at least use locally. If that fails we return false. The GROUP_ID is set to -1 for cselib bases and the index of the group for non_varying bases. FOR_READ is true if this is a mem read and false if not.
First see if just canon_rtx (mem_address) is const or frame,
if not, try cselib_expand_value_rtx and call canon_rtx on that. Use cselib to replace all of the reg references with the full
expression. This will take care of the case where we have
r_x = base + offset;
val = *r_x;
by making it into
val = *(base + offset); If this fails, just go with the address from first
iteration. Split the address into canonical BASE + OFFSET terms.
| bool check_for_inc_dec | ( | ) |
Entry point for postreload. If you work on reload_cse, or you need this anywhere else, consider if you can provide register liveness information and add a parameter to this function so that it can be passed down in insn_info.fixed_regs_live.
|
static |
Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it is there, is split into a separate insn. Return true on success (or if there was nothing to do), false on failure.
Referenced by dse_step3().
|
static |
A for_each_rtx callback in which DATA is the bb_info. Check to see if LOC is a mem and if it is look at the address and kill any appropriate stores that may be active.
If it is reading readonly mem, then there can be no conflict with
another write. For alias_set != 0 canon_true_dependence should be never called.
We ignore the clobbers in store_info. The is mildly aggressive,
but there really should not be a clobber followed by a read. Skip the clobbers.
This is the restricted case where the base is a constant or
the frame pointer and offset is a constant. Skip the clobbers.
There are three cases here.
We have a cselib store followed by a read from a
const base. This is a block mode load. We may get lucky and
canon_true_dependence may save the day. If this read is just reading back something that we just
stored, rewrite the read. The bases are the same, just see if the offsets
overlap. else
The else case that is missing here is that the
bases are constant but different. There is nothing
to do here because there is no overlap. Skip the clobbers.
If this read is just reading back something that we just
stored, rewrite the read.
References all_positions_needed_p(), store_info::begin, bb_info::regs_live, replace_read(), and store_info::rhs.
|
static |
A for_each_rtx callback in which DATA points the INSN_INFO for as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns true for any part of *LOC.
References store_info::is_set, store_info::next, and insn_info::store_rec.
|
staticread |
Find the entry associated with ALIAS_SET.
|
static |
Clear the rhs field from the active_local_stores array.
Skip the clobbers.
Referenced by set_all_positions_unneeded().
|
static |
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.
Note that we have to test for the actual rtx used for the frame
and arg pointers and not just the register number in case we have
eliminated the frame and/or arg pointer and are using it
for pseudos. The arg pointer varies if it is not a fixed register.
|
static |
Return a bitmap of the fixed registers contained in IN.
|
static |
Delete the insn and free all of the fields inside INSN_INFO.
References bitmap_set_bit(), group_info::escaped_n, group_info::escaped_p, group_info::offset_map_size_n, group_info::offset_map_size_p, group_info::store1_n, group_info::store1_p, group_info::store2_n, and group_info::store2_p.
Referenced by dse_transfer_function().
|
static |
Confluence function for blocks with no successors. Create an out set from the gen set of the exit block. This block logically has the exit block as a successor.
References bitmap_clear(), dump_file, and dump_flags.
|
static |
Propagate the information from the in set of the dest of E to the out set of the src of E. If the various in or out sets are not there, that means they are all ones.
|
static |
Initialization of data structures.
Referenced by dse_step4().
|
static |
Do all of step 1.
Scan the insns.
This is something of a hack, because the global algorithm
is supposed to take care of the case where stores go dead
at the end of the function. However, the global
algorithm must take a more conservative view of block
mode reads than the local alg does. So to get the case
where you have a store to the frame followed by a non
overlapping block more read, we look at the active local
stores at the end of the function and delete all of the
frame and spill based ones. Skip the clobbers.
Get rid of the loads that were discovered in
replace_read. Cselib is finished with this block. There is no reason to validate this change. That was
done earlier. Get rid of all of the cselib based store_infos in this
block and mark the containing insns as not being
deletable. Free at least positions_needed bitmaps.
Referenced by dse_step4().
|
static |
For all non stack related bases, we only consider a store to
be deletable if there are two or more stores for that
position. This is because it takes one store to make the
other store redundant. However, for the stores that are
stack related, we consider them if there is only one store
for the position. We do this because the stack related
stores can be deleted if their is no read between them and
the end of the function.
To make this work in the current framework, we take the stack
related bases add all of the bits from store1 into store2.
This has the effect of making the eligible even if there is
only one store.
References bitmap_and_compl_into(), bitmap_ior_into(), group_info::frame_related, group_info::group_kill, and group_info::process_globally.
Referenced by dse_step4().
|
static |
Init the offset tables for the normal case.
Position 0 is unused because 0 is used in the maps to mean
unused.
References bitmap_and_compl_into(), bitmap_ior_into(), and group_info::group_kill.
Referenced by dse_step4().
|
static |
Build the transfer functions for the function.
If this is the second time dataflow is run, delete the old
sets. For any block in an infinite loop, we must initialize the out set
to all ones. This could be expensive, but almost never occurs in
practice. However, it is common in regression tests.
References store_info::alias_set, store_info::begin, bitmap_bit_p(), bitmap_empty_p(), bitmap_print(), insn_info::cannot_delete, check_for_inc_dec_1(), dbg_cnt(), delete_insn(), deleted, dump_file, dump_flags, get_bitmap_index(), globally_deleted, store_info::group_id, HOST_WIDE_INT, basic_block_def::index, insn_info::insn, store_info::is_set, bb_info::last_insn, store_info::next, bb_info::out, scan_stores_nospill(), and insn_info::store_rec.
Referenced by dse_step4().
|
static |
Set the gen set of the exit block, and also any block with no successors that does not have a wild read.
The gen set is all 0's for the exit block except for the
frame_pointer_group.
|
static |
Scan the insns in BB_INFO starting at PTR and going to the top of the block in order to build the gen and kill sets for the block. We start at ptr which may be the last insn in the block or may be the first insn with a wild read. In the latter case we are able to skip the rest of the block because it just does not matter: anything that happens is hidden by the wild read.
There are no wild reads in the spill case.
In the spill case or in the no_spill case if there is no wild
read in the block, we will need a kill set. There may have been code deleted by the dce pass run before
this phase. Process the read(s) last.
|
static |
Solve the dataflow equations.
References df_analyze(), DF_DEFER_INSN_RESCAN, DF_LR_RUN_DCE, df_note_add_problem(), df_set_flags(), dse_step0(), dse_step1(), dse_step2_init(), dse_step2_nospill(), dse_step3(), dse_step5_nospill(), dse_step6(), dse_step7(), dump_file, dump_flags, globally_deleted, locally_deleted, and spill_deleted.
|
static |
There may have been code deleted by the dce pass run before
this phase. Try to delete the current insn.
Skip the clobbers.
We do want to process the local info if the insn was
deleted. For instance, if the insn did a wild read, we
no longer need to trash the info.
References dbg_cnt().
Referenced by dse_step4().
|
static |
There may have been code deleted by the dce pass run before
this phase.
Referenced by dse_step4().
|
static |
Referenced by dse_step4().
|
static |
@verbatim
Propagate the info from the out to the in set of BB_INDEX's basic block. There are three cases:
1) The block has no kill set. In this case the kill set is all ones. It does not matter what the out set of the block is, none of the info can reach the top. The only thing that reaches the top is the gen set and we just copy the set.
2) There is a kill set but no out set and bb has successors. In this case we just return. Eventually an out set will be created and it is better to wait than to create a set of ones.
3) There is both a kill and out set. We apply the obvious transfer function.
Case 3 above.
Case 2 above.
Case 1 above. If there is already an in set, nothing
happens.
References delete_dead_store_insn(), dump_file, dump_flags, insn_info::insn, and store_info::redundant_reason.
|
static |
References bitmap_set_range(), and regs_set.
|
static |
Callback for for_each_inc_dec that emits an INSN that sets DEST to SRC + SRCOFF before insn ARG.
We can reuse all operands without copying, because we are about
to delete the insn that contained it. If a failure was flagged above, return 1 so that for_each_inc_dec will
return it immediately, communicating the failure to its caller.
|
static |
Return the insn in BB_INFO before the first wild read or if there are no wild reads in the block, return the last insn.
Block starts with wild read.
References bitmap_copy(), bb_info::gen, basic_block_def::index, and bb_info::out.
Referenced by scan_stores_nospill().
|
static |
If the modes are different and the value's source and target do not line up, we need to extract the value from lower part of the rhs of the store, shift it, and then put it into a form that can be shoved into the read_insn. This function generates a right SHIFT of a value that is at least ACCESS_SIZE bytes wide of READ_MODE. The shift sequence is returned or NULL if we failed to find a shift.
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
shift insns that shift values within 32 or 64 bit registers.
This loop tries to find the smallest shift insn that will right
justify the value we want to read but is available in one insn on
the machine. If a constant was stored into memory, try to simplify it here,
otherwise the cost of the shift might preclude this optimization
e.g. at -Os, even when no actual shift will be needed. Try a wider mode if truncating the store mode to NEW_MODE
requires a real instruction. Also try a wider mode if the necessary punning is either not
desirable or not possible. In theory we could also check for an ashr. Ian Taylor knows
of one dsp where the cost of these two was not the same. But
this really is a rare case anyway. 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. 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.
References store_info::begin, store_info::const_rhs, copy_rtx(), store_info::end, extract_low_bits(), gen_int_mode(), HOST_BITS_PER_WIDE_INT, HOST_WIDE_INT, int_mode_for_mode(), store_info::mem, optimize_bb_for_speed_p(), store_info::rhs, and shift.
|
static |
Free all READ_REC of the LAST_INSN of BB_INFO.
Referenced by can_escape().
|
static |
Delete all of the store_info recs from INSN_INFO.
|
static |
|
static |
|
static |
Look up the bitmap index for OFFSET in GROUP_INFO. If it is not there, return 0.
References insn_info::prev_insn, and insn_info::wild_read.
Referenced by dse_step3().
|
static |
Get arguments passed to CALL_INSN. Return TRUE if successful. So far it only handles arguments passed in registers.
|
static |
Get the GROUP for BASE. Add a new group if it is not there.
Find the store_base_info structure for BASE, creating a new one
if necessary.
|
static |
Helper function for replace_read and record_store. Attempt to return a value stored in STORE_INFO, from READ_BEGIN to one before READ_END bytes read in READ_MODE. Return NULL if not successful. If REQUIRE_CST is true, return always constant.
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.
The store is a memset (addr, const_val, const_size).
References bitmap_and_into(), bitmap_empty_p(), df_print_regset(), dump_file, dump_flags, look_for_hardregs(), note_stores(), reg_obstack, and regs_set.
|
static |
Return whether DECL, a local variable, can possibly escape the current function scope.
If this is a partitioned variable, we need to consider all the variables
in the partition. This is necessary because a store into one of them can
be replaced with a store into another and this may not change the outcome
of the escape analysis.
|
static |
Call back for note_stores to find the hard regs set or clobbered by insn. Data is a bitmap of the hardregs set so far.
Referenced by get_stored_val().
|
static |
Return a bitmask with the first N low bits set.
| rtl_opt_pass* make_pass_rtl_dse1 | ( | ) |
| rtl_opt_pass* make_pass_rtl_dse2 | ( | ) |
|
static |
Find all of the blocks that are not backwards reachable from the exit block or any block with no successors (BB). These are the infinite loops or infinite self loops. These blocks will still have their bits set in UNREACHABLE_BLOCKS.
|
static |
Callback for emit_inc_dec_insn_before via note_stores. Check if a register is clobbered which is live afterwards.
If this register is referenced by the current or an earlier insn,
that's OK. E.g. this applies to the register that is being incremented
with this addition. If we come here, we have a clobber of a register that's only OK
if that register is not live. If we don't have liveness information
available, fail now. Now check if this is a live fixed register.
|
static |
BODY is an instruction pattern that belongs to INSN. Return 1 if there is a candidate store, after adding it to the appropriate local store group if so.
If this is not used, then this cannot be used to keep the insn
from being deleted. On the other hand, it does provide something
that can be used to prove that another store is dead. Check whether that value is a suitable memory location.
If the set or clobber is unused, then it does not effect our
ability to get rid of the entire insn. At this point we know mem is a mem.
Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0))
as memset (addr, 0, 36); If the set or clobber is unused, then it does not effect our
ability to get rid of the entire insn. We can still process a volatile mem, we just cannot delete it.
In the restrictive case where the base is a constant or the
frame pointer we can do global analysis. No place to keep the value after ra.
Sometimes the store and reload is used for truncation and
rounding. Check to see if this stores causes some other stores to be
dead. For alias_set != 0 canon_true_dependence should be never called.
Skip the clobbers. We delete the active insn if this insn
shadows the set. To have been put on the active list, it
has exactly on set. Generally, spills cannot be processed if and of the
references to the slot have a different mode. But if
we are in the same block and mode is exactly the same
between this store and one before in the same block,
we can still delete it. Even if PTR won't be eliminated as unneeded, if both
PTR and this insn store the same constant value, we might
eliminate this insn instead. Need to see if it is possible for this store to overwrite
the value of store_info. If it is, set the rhs to NULL to
keep it from being used to remove a load. An insn can be deleted if every position of every one of
its s_infos is zero. Finish filling in the store_info.
If this is a clobber, we return 0. We will only be able to
delete this insn if there is only one store USED store, but we
can use the clobber to delete other stores earlier.
References insn_info::contains_cselib_groups, dump_file, dump_flags, may_be_sp_based_p(), offset, pool_alloc(), and insn_info::stack_pointer_based.
|
static |
Remove BASE from the set of active_local_stores. This is a callback from cselib that is used to get rid of the stores in active_local_stores.
If ANY of the store_infos match the cselib group that is
being deleted, then the insn can not be deleted.
|
static |
Take a sequence of:
A <- r1
...
... <- A
and change it into
r2 <- r1
A <- r1
...
... <- r2
or
r3 <- extract (r1)
r3 <- r3 >> shift
r2 <- extract (r3)
... <- r2
or
r2 <- extract (r1)
... <- r2
Depending on the alignment and the mode of the store and
subsequent load.
The STORE_INFO and STORE_INSN are for the store and READ_INFO
and READ_INSN are for the read. Return true if the replacement
went ok. 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. Force the value into a new register so that it won't be clobbered
between the store and the load. Now we have to scan the set of new instructions to see if the
sequence contains and sets of hardregs that happened to be
live at this point. For instance, this can happen if one of
the insns sets the CC and the CC happened to be live at that
point. This does occasionally happen, see PR 37922. Insert this right before the store insn where it will be safe
from later insns that might change it before the read. And now for the kludge part: cselib croaks if you just
return at this point. There are two reasons for this:
1) Cselib has an idea of how many pseudos there are and
that does not include the new ones we just added.
2) Cselib does not know about the move insn we added
above the store_info, and there is no way to tell it
about it, because it has "moved on".
Problem (1) is fixable with a certain amount of engineering.
Problem (2) is requires starting the bb from scratch. This
could be expensive.
So we are just going to have to lie. The move/extraction
insns are not really an issue, cselib did not see them. But
the use of the new pseudo read_insn is a real problem because
cselib has not scanned this insn. The way that we solve this
problem is that we are just going to put the mem back for now
and when we are finished with the block, we undo this. We
keep a table of mems to get rid of. At the end of the basic
block we can put them back. Get rid of the read_info, from the point of view of the
rest of dse, play like this read never happened.
References add_wild_read(), insn_info::cannot_delete, dump_file, and dump_flags.
Referenced by check_mem_read_rtx().
|
static |
References get_address_mode().
Referenced by can_escape().
|
static |
------------------------------------------------------------------------- DSE -------------------------------------------------------------------------
Callback for running pass_rtl_dse.
Need the notes since we must track live hardregs in the forwards
direction.
|
static |
Apply record_store to all candidate stores in INSN. Mark INSN if some part of it is not a candidate store and assigns to a non-register target.
Cselib clears the table for this case, so we have to essentially
do the same. Look at all of the uses in the insn.
Const functions cannot do anything bad i.e. read memory,
however, they can read their parameters which may have
been pushed onto the stack.
memset and bzero don't read memory either. See the head comment of the frame_read field.
Loop over the active stores and remove those which are
killed by the const function call. The stack pointer based stores are always killed.
If the frame is read, the frame related stores are killed.
Skip the clobbers.
Every other call, including pure functions, may read any memory
that is not relative to the frame. Assuming that there are sets in these insns, we cannot delete
them. If we found some sets of mems, add it into the active_local_stores so
that it can be locally deleted if found dead or used for
replace_read and redundant constant store elimination. Otherwise mark
it as cannot delete. This simplifies the processing later.
|
static |
Process the READ_INFOs into the bitmaps into GEN and KILL. KILL may be NULL.
If this insn reads the frame, kill all the frame related stores.
Kill all non-frame related stores. Kill all stores of variables that
escape. Begin > end for block mode reads.
The groups are the same, just process the
offsets. The groups are different, if the alias sets
conflict, clear the entire group. We only need
to apply this test if the read_info is a cselib
read. Anything with a constant base cannot alias
something else with a different constant
base.
References bitmap_ior_into(), group_info::frame_related, bb_info::gen, group_info::group_kill, group_info::process_globally, and stores_off_frame_dead_at_return.
Referenced by scan_stores_spill().
|
static |
Process the READ_INFOs into the bitmaps into GEN and KILL. KILL may be NULL.
References bitmap_copy(), and bb_info::out.
Referenced by scan_stores_spill().
|
static |
Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL may be NULL.
References bitmap_clear(), find_insn_before_first_wild_read(), basic_block_def::index, bb_info::kill, and bb_info::last_insn.
Referenced by dse_step3(), and scan_stores_spill().
|
static |
Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL may be NULL.
References bb_info::gen, insn_info::insn, bb_info::kill, insn_info::prev_insn, insn_info::read_rec, scan_reads_nospill(), scan_reads_spill(), scan_stores_nospill(), and insn_info::store_rec.
|
inlinestatic |
Mark the whole store S_INFO as unneeded.
References insn_info::cannot_delete, and clear_rhs_from_active_local_stores().
|
inlinestatic |
Mark byte POS bytes from the beginning of store S_INFO as unneeded.
References add_wild_read(), insn_info::cannot_delete, dump_file, and dump_flags.
|
static |
Set the store* bitmaps offset_map_size* fields in GROUP based on OFFSET and WIDTH.
|
static |
The linked list of stores that are under consideration in this basic block.
|
static |
|
static |
Referenced by setjmp_args_warning(), and setjmp_vars_warning().
|
static |
|
static |
Table to hold all bb_infos.
|
static |
The group that holds all of the clear_alias_sets.
|
static |
The modes of the clear_alias_sets.
|
static |
|
static |
The number of bits used in the global bitmaps.
|
static |
|
static |
|
static |
Obstack for the DSE dataflow bitmaps. We don't want to put these on the default obstack because these bitmaps can grow quite large (~2GB for the small (!) test case of PR54146) and we'll hold on to all that memory until the end of the compiler run. As a bonus, delete_tree_live_info can destroy all the bitmaps by just releasing the whole obstack.
|
static |
Obstack for other data. As for above: Kinda nice to be able to throw it all away at the end in one big sweep.
|
static |
Counter for stats.
Referenced by dse_step3(), and dse_step4().
|
static |
|
static |
Locations that are killed by calls in the global phase.
|
static |
Referenced by dse_step4().
|
static |
|
static |
|
static |
Index into the rtx_group_vec.
|
static |
Tables of group_info structures, hashed by base value.
|
static |
|
static |
|
static |
Scratch bitmap for cselib's cselib_expand_value_rtx.
Referenced by return_insn_p().
|
static |
Referenced by dse_step4().
|
static |
This is true except if cfun->stdarg -- i.e. we cannot do this for vararg functions because they play games with the frame.
Referenced by scan_reads_nospill().
1.8.1.1