- Generated by
1.8.1.1
|
GCC Middle and Back End API Reference
|
Data Structures | |
| struct | expr |
| struct | occr |
| struct | hash_table_d |
| struct | ls_expr |
| struct | pre_ldst_expr_hasher |
| struct | modify_pair_s |
| struct | bb_data |
| struct | reg_avail_info |
| struct | mem_conflict_info |
Typedefs | |
| typedef struct occr * | occr_t |
| typedef struct modify_pair_s | modify_pair |
Functions | |
| static int | expr_equiv_p (const_rtx, const_rtx) |
| static void | compute_can_copy (void) |
| static void * | gmalloc (static void *gcalloc(size_t) |
| bool | can_copy_p () |
| static void * | gmalloc () |
| static void * | gcalloc () |
| static void * | gcse_alloc () |
| static void | alloc_gcse_mem () |
| static void | free_gcse_mem () |
| static void | compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc, struct hash_table_d *table) |
| static int | want_to_gcse_p () |
| bool | can_assign_to_reg_without_clobbers_p () |
| static int | oprs_unchanged_p () |
| static void | mems_conflict_for_gcse_p (rtx dest, const_rtx setter, void *data) |
| static int | load_killed_in_block_p (const_basic_block bb, int uid_limit, const_rtx x, int avail_p) |
| static int | oprs_anticipatable_p () |
| static int | oprs_available_p () |
| static unsigned int | hash_expr (const_rtx x, enum machine_mode mode, int *do_not_record_p, int hash_table_size) |
| static int | expr_equiv_p () |
| static void | insert_expr_in_table (rtx x, enum machine_mode mode, rtx insn, int antic_p, int avail_p, int max_distance, struct hash_table_d *table) |
| static void | hash_scan_set () |
| static void | hash_scan_clobber (rtx x, rtx insn, struct hash_table_d *table) |
| static void | hash_scan_call (rtx x, rtx insn, struct hash_table_d *table) |
| static void | hash_scan_insn () |
| static void | dump_hash_table () |
| static void | record_last_reg_set_info () |
| static void | canon_list_insert (rtx dest, const_rtx x, void *v_insn) |
| static void | record_last_mem_set_info () |
| static void | record_last_set_info () |
| static void | compute_hash_table_work () |
| static void | alloc_hash_table () |
| static void | free_hash_table () |
| static void | compute_hash_table () |
| static void | clear_modify_mem_tables () |
| static void | free_modify_mem_tables () |
| static void | compute_transp () |
| static void | alloc_pre_mem () |
| static void | free_pre_mem () |
| static void | prune_expressions () |
| static void | prune_insertions_deletions () |
| static struct edge_list * | compute_pre_data () |
| static int | pre_expr_reaches_here_p_work (basic_block occr_bb, struct expr *expr, basic_block bb, char *visited) |
| static int | pre_expr_reaches_here_p () |
| static rtx | process_insert_insn () |
| static void | insert_insn_end_basic_block () |
| static int | pre_edge_insert () |
| static void | pre_insert_copy_insn () |
| static void | pre_insert_copies () |
| static rtx | gcse_emit_move_after () |
| static int | pre_delete () |
| static int | pre_gcse () |
| static int | one_pre_gcse_pass () |
| static void | add_label_notes () |
| static void | alloc_code_hoist_mem () |
| static void | free_code_hoist_mem () |
| static void | compute_code_hoist_vbeinout () |
| static void | compute_code_hoist_data () |
| static int | update_bb_reg_pressure () |
| static int | should_hoist_expr_to_dom (basic_block expr_bb, struct expr *expr, basic_block bb, sbitmap visited, int distance, int *bb_size, enum reg_class pressure_class, int *nregs, bitmap hoisted_bbs, rtx from) |
| static struct occr * | find_occr_in_bb () |
| static int | hoist_code () |
| static enum reg_class | get_regno_pressure_class () |
| static enum reg_class | get_pressure_class_and_nregs () |
| static void | change_pressure () |
| static void | calculate_bb_reg_pressure () |
| static int | one_code_hoisting_pass () |
| static struct ls_expr * | ldst_entry () |
| static void | free_ldst_entry () |
| static void | free_ld_motion_mems () |
| static void | print_ldst_list () |
| static struct ls_expr * | find_rtx_in_ldst () |
| static int | simple_mem () |
| static void | invalidate_any_buried_refs () |
| static void | compute_ld_motion_mems () |
| static void | trim_ld_motion_mems () |
| static void | update_ld_motion_stores () |
| static bool | is_too_expensive () |
| static bool | gate_rtl_pre () |
| static unsigned int | execute_rtl_pre () |
| static bool | gate_rtl_hoist () |
| static unsigned int | execute_rtl_hoist () |
| rtl_opt_pass * | make_pass_rtl_pre () |
| rtl_opt_pass * | make_pass_rtl_hoist () |
| typedef struct modify_pair_s modify_pair |
|
static |
If X contains any LABEL_REF's, add REG_LABEL_OPERAND notes for them to INSN. If such notes are added to an insn which references a CODE_LABEL, the LABEL_NUSES count is incremented. We have to add that note, because the following loop optimization pass requires them.
??? If there was a jump optimization pass after gcse and before loop, then we would not need to do this here, because jump would add the necessary REG_LABEL_OPERAND and REG_LABEL_TARGET notes.
References add_reg_note().
Referenced by insert_insn_end_basic_block().
|
static |
??? We could compute post dominators and run this algorithm in reverse to perform tail merging, doing so would probably be more effective than the tail merging code in jump.c. It's unclear if tail merging could be run in parallel with code hoisting. It would be nice.
Allocate vars used for code hoisting analysis.
References sbitmap_vector_alloc().
Referenced by one_code_hoisting_pass().
|
static |
Allocate memory for the reg/memory set tracking tables. This is called at the start of each pass.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
Allocate space for the set/expr hash TABLE. It is used to determine the number of buckets to use.
References get_max_insn_count(), hash_table_d::size, and hash_table_d::table.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
Allocate vars used for PRE analysis.
References sbitmap_vector_alloc().
Referenced by one_pre_gcse_pass().
|
static |
Calculate register pressure for each basic block by walking insns from last to first.
References bitmap_clear_bit(), bitmap_copy(), bitmap_set_bit(), change_pressure(), curr_reg_pressure, curr_regs_live, df_get_live_in(), df_get_live_out(), DF_REF_CONDITIONAL, DF_REF_PARTIAL, dump_file, basic_block_def::index, ira_setup_eliminable_regset(), reg_class_names, and reg_obstack.
Referenced by one_code_hoisting_pass().
| bool can_assign_to_reg_without_clobbers_p | ( | ) |
Return true if we can assign X to a pseudo register such that the resulting insn does not result in clobbering a hard register as a side-effect. Additionally, if the target requires it, check that the resulting insn can be copied. If it cannot, this means that X is special and probably has hidden side-effects we don't want to mess with. This function is typically used by code motion passes, to verify that it is safe to insert an insn without worrying about clobbering maybe live hard regs.
References added_clobbers_hard_reg_p(), gen_rtx_REG(), general_operand(), make_insn_raw(), recog(), targetm, and word_mode.
Referenced by compute_ld_motion_mems(), find_moveable_store(), and want_to_gcse_p().
| bool can_copy_p | ( | ) |
Returns whether the mode supports reg/reg copy operations.
References compute_can_copy().
Referenced by hash_scan_set(), and may_assign_reg_p().
Record all of the canonicalized MEMs of record_last_mem_set_info's insn. Note we store a pair of elements in the list, so they have to be taken off pairwise.
References canon_rtx(), modify_pair_s::dest, modify_pair_s::dest_addr, and get_addr().
Referenced by record_last_mem_set_info().
|
static |
Increase (if INCR_P) or decrease current register pressure for register REGNO.
References curr_reg_pressure, and get_regno_pressure_class().
Referenced by calculate_bb_reg_pressure().
|
static |
Expression tracking support.
Clear canon_modify_mem_list and modify_mem_list tables.
References bitmap_clear().
Referenced by compute_hash_table_work(), and free_modify_mem_tables().
|
static |
Referenced by can_copy_p().
|
static |
Top level routine to do the dataflow analysis needed by code hoisting.
References calculate_dominance_info(), CDI_DOMINATORS, compute_code_hoist_vbeinout(), compute_local_properties(), dump_file, expr_hash_table, and prune_expressions().
Referenced by one_code_hoisting_pass().
|
static |
Compute the very busy expressions at entry/exit from each block. An expression is very busy if all paths from a given point compute the expression.
References bitmap_intersection_of_succs(), bitmap_ior(), bitmap_or_and(), bitmap_vector_clear(), changed, dump_bitmap_file(), dump_file, basic_block_def::index, and basic_block_def::next_bb.
Referenced by compute_code_hoist_data().
|
static |
Compute the expression hash table TABLE.
References compute_hash_table_work(), memset(), hash_table_d::n_elems, hash_table_d::size, and hash_table_d::table.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
Top level function to create an expression hash table. Expression entries are placed in the hash table if - they are of the form (set (pseudo-reg) src), - src is something we want to perform GCSE on, - none of the operands are subsequently modified in the block Currently src must be a pseudo-reg or a const_int. TABLE is the table computed.
References clear_modify_mem_tables(), free(), hash_scan_insn(), last_bb, max_reg_num(), note_stores(), record_last_mem_set_info(), record_last_reg_set_info(), and record_last_set_info().
Referenced by compute_hash_table().
|
static |
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.
References alloc_INSN_LIST(), can_assign_to_reg_without_clobbers_p(), hash_table< Descriptor, Allocator >::create(), find_reg_equal_equiv_note(), ls_expr::invalid, invalidate_any_buried_refs(), ldst_entry(), ls_expr::loads, SET, simple_mem(), and ls_expr::stores.
Referenced by one_pre_gcse_pass().
|
static |
Compute the local properties of each recorded expression. Local properties are those that are defined by the block, irrespective of other blocks. An expression is transparent in a block if its operands are not modified in the block. An expression is computed (locally available) in a block if it is computed at least once and expression would contain the same value if the computation was moved to the end of the block. An expression is locally anticipatable in a block if it is computed at least once and expression would contain the same value if the computation was moved to the beginning of the block. We call this routine for pre and code hoisting. They all compute basically the same information and thus can easily share this code. TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local properties. If NULL, then it is not necessary to compute or record that particular property. TABLE controls which hash table to look at.
References expr::antic_occr, expr::avail_occr, expr::bitmap_index, bitmap_set_bit(), bitmap_vector_clear(), bitmap_vector_ones(), compute_transp(), occr::copied_p, occr::deleted_p, expr::expr, occr::insn, occr::next, expr::next_same_hash, expr::reaching_reg, hash_table_d::size, and hash_table_d::table.
Referenced by compute_code_hoist_data(), and compute_pre_data().
|
staticread |
Top level routine to do the dataflow analysis needed by PRE.
References bitmap_ior(), bitmap_not(), bitmap_vector_clear(), compute_local_properties(), edge_list, expr_hash_table, basic_block_def::index, hash_table_d::n_elems, pre_edge_lcm(), prune_expressions(), prune_insertions_deletions(), and sbitmap_vector_free().
Referenced by one_pre_gcse_pass().
|
static |
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, reset the INDX bit in BMAP.
References bitmap_clear_bit(), canon_rtx(), canon_true_dependence(), expr::dest, modify_pair_s::dest, modify_pair_s::dest_addr, and get_addr().
Referenced by compute_local_properties().
|
static |
Dump the hash table TABLE to file FILE under the name NAME.
References expr::bitmap_index, expr::expr, free(), expr::max_distance, hash_table_d::n_elems, expr::next_same_hash, print_rtl(), hash_table_d::size, and hash_table_d::table.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
|
static |
|
static |
Return nonzero if exp1 is equivalent to exp2.
References exp_equiv_p().
|
staticread |
|
staticread |
Returns 1 if X is in the list of ldst only expressions.
References hash_table< Descriptor, Allocator >::find_slot(), hash_table< Descriptor, Allocator >::is_created(), and ls_expr::pattern.
Referenced by mems_conflict_for_gcse_p(), and update_ld_motion_stores().
|
static |
Free vars used for code hoisting analysis.
References CDI_DOMINATORS, free_dominance_info(), and sbitmap_vector_free().
Referenced by one_code_hoisting_pass().
|
static |
Free memory allocated by alloc_gcse_mem.
References free_modify_mem_tables().
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
Free things allocated by alloc_hash_table.
References free(), and hash_table_d::table.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
Free up all memory associated with the ldst list.
References hash_table< Descriptor, Allocator >::dispose(), free_ldst_entry(), hash_table< Descriptor, Allocator >::is_created(), ls_expr::next, and pre_ldst_mems.
Referenced by one_pre_gcse_pass().
|
static |
Free up an individual ldst entry.
References free(), free_INSN_LIST_list(), ls_expr::loads, and ls_expr::stores.
Referenced by free_ld_motion_mems(), and trim_ld_motion_mems().
|
static |
Release memory used by modify_mem_list_set.
References clear_modify_mem_tables(), and free().
Referenced by free_gcse_mem().
|
static |
Free vars used for PRE analysis.
References sbitmap_vector_free().
Referenced by one_pre_gcse_pass().
|
static |
References function::calls_setjmp, cfun, dbg_cnt(), and optimize_function_for_size_p().
|
static |
All the passes implemented in this file. Each pass has its own gate and execute function, and at the end of the file a pass definition for passes.c. We do not construct an accurate cfg in functions which call setjmp, so none of these passes runs if the function calls setjmp. FIXME: Should just handle setjmp via REG_SETJMP notes.
References function::calls_setjmp, cfun, dbg_cnt(), and optimize_function_for_speed_p().
|
static |
Cover function to xcalloc to record bytes allocated.
References bytes_used.
|
static |
Cover function to obstack_alloc.
References bytes_used, and gcse_obstack.
|
static |
Emit move from SRC to DEST noting the equivalence with expression computed in INSN.
References copy_insn_1(), emit_insn_after(), find_reg_equal_equiv_note(), gen_move_insn(), reg_mentioned_p(), rtx_equal_p(), and set_unique_reg_note().
Referenced by hoist_code(), and pre_delete().
|
static |
Return pressure class and number of hard registers (through *NREGS) for destination of INSN.
References reg_allocno_class().
Referenced by hoist_code().
|
static |
Return pressure class and number of needed hard registers (through *NREGS) of register REGNO.
References eliminable_regset, and reg_allocno_class().
Referenced by change_pressure(), and update_bb_reg_pressure().
|
static |
Misc. utilities.
Compute which modes support reg/reg copy operations.
References emit_insn(), end_sequence(), gen_rtx_REG(), memset(), recog(), and start_sequence().
|
static |
Cover function to xmalloc to record bytes allocated.
References bytes_used.
|
static |
Hash expression X. 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.
References hash_rtx().
Referenced by insert_expr_in_table(), and lookup_expr_in_table().
|
static |
Referenced by hash_scan_insn(), and hash_scan_set().
|
static |
Referenced by hash_scan_insn().
|
static |
Process INSN and add hash table entries as appropriate.
References hash_scan_call(), hash_scan_clobber(), hash_scan_set(), and SET.
Referenced by compute_hash_table_work().
|
static |
Scan SET present in INSN and add an entry to the hash TABLE.
References can_copy_p(), can_throw_internal(), find_reg_equal_equiv_note(), find_reg_note(), hash_scan_call(), insert_expr_in_table(), multiple_sets(), oprs_anticipatable_p(), oprs_available_p(), set_noop_p(), and want_to_gcse_p().
Referenced by hash_scan_insn().
|
static |
Actually perform code hoisting.
The code hoisting pass can hoist multiple computations of the same
expression along dominated path to a dominating basic block, like
from b2/b3 to b1 as depicted below:
b1 ------
/\ |
/ \ |
bx by distance
/ \ |
/ \ |
b2 b3 ------
Unfortunately code hoisting generally extends the live range of an
output pseudo register, which increases register pressure and hurts
register allocation. To address this issue, an attribute MAX_DISTANCE
is computed and attached to each expression. The attribute is computed
from rtx cost of the corresponding expression and it's used to control
how long the expression can be hoisted up in flow graph. As the
expression is hoisted up in flow graph, GCC decreases its DISTANCE
and stops the hoist if DISTANCE reaches 0. Code hoisting can decrease
register pressure if live ranges of inputs are shrunk.
Option "-fira-hoist-pressure" implements register pressure directed
hoist based on upper method. The rationale is:
1. Calculate register pressure for each basic block by reusing IRA
facility.
2. When expression is hoisted through one basic block, GCC checks
the change of live ranges for inputs/output. The basic block's
register pressure will be increased because of extended live
range of output. However, register pressure will be decreased
if the live ranges of inputs are shrunk.
3. After knowing how hoisting affects register pressure, GCC prefers
to hoist the expression if it can decrease register pressure, by
increasing DISTANCE of the corresponding expression.
4. If hoisting the expression increases register pressure, GCC checks
register pressure of the basic block and decrease DISTANCE only if
the register pressure is high. In other words, expression will be
hoisted through at no cost if the basic block has low register
pressure.
5. Update register pressure information for basic blocks through
which expression is hoisted.
References expr::antic_occr, bb_data::backup, bitmap_bit_p(), bitmap_clear(), bitmap_copy(), expr::bitmap_index, bitmap_set_bit(), CDI_DOMINATORS, changed, dbg_cnt(), delete_insn(), occr::deleted_p, expr::expr, expr_hash_table, find_occr_in_bb(), free(), gcse_emit_move_after(), gcse_subst_count, gen_reg_rtx_and_attrs(), get_all_dominated_blocks(), get_dominated_to_depth(), get_max_uid(), get_pressure_class_and_nregs(), basic_block_def::index, insert_insn_end_basic_block(), occr::insn, bb_data::live_in, expr::max_distance, bb_data::max_reg_pressure, hash_table_d::n_elems, nearest_common_dominator_for_set(), expr::next_same_hash, bb_data::old_pressure, expr::reaching_reg, should_hoist_expr_to_dom(), hash_table_d::size, hash_table_d::table, and vNULL.
Referenced by one_code_hoisting_pass().
|
static |
Insert expression X in INSN in the hash TABLE. If it is already present, record it as the last occurrence in INSN's basic block. MODE is the mode of the value X is being stored into. It is only used if X is a CONST_INT. ANTIC_P is nonzero if X is an anticipatable expression. AVAIL_P is nonzero if X is an available expression. MAX_DISTANCE is the maximum distance in instructions this expression can be moved.
References expr::antic_occr, expr::avail_occr, expr::bitmap_index, bytes_used, occr::deleted_p, expr::expr, expr_equiv_p(), expr::hash, hash_expr(), occr::insn, expr::max_distance, hash_table_d::n_elems, occr::next, expr::next_same_hash, hash_table_d::size, and hash_table_d::table.
Referenced by hash_scan_set().
|
static |
Add EXPR to the end of basic block BB. This is used by both the PRE and code hoisting.
References add_label_notes(), expr::bitmap_index, dump_file, emit_insn_after_noloc(), emit_insn_before_noloc(), find_first_parameter_load(), find_reg_note(), gcse_create_count, basic_block_def::index, prev_nonnote_insn(), process_insert_insn(), expr::reaching_reg, sets_cc0_p(), single_succ_edge(), and single_succ_p().
Referenced by hoist_code(), and pre_edge_insert().
|
static |
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.
References ls_expr::invalid, ldst_entry(), and simple_mem().
Referenced by compute_ld_motion_mems().
|
static |
Return true if the graph is too expensive to optimize. PASS is the optimization about to be performed.
References max_reg_num(), and warning().
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
|
staticread |
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 killed 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.
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 will search the ldst list for a matching expression. If it doesn't find one, we create one and initialize it.
References ls_expr::expr, hash_table< Descriptor, Allocator >::find_slot_with_hash(), ls_expr::hash_index, hash_rtx(), ls_expr::index, ls_expr::invalid, ls_expr::loads, ls_expr::next, ls_expr::pattern, ls_expr::pattern_regs, pre_ldst_mems, ls_expr::reaching_reg, and ls_expr::stores.
Referenced by compute_ld_motion_mems(), and invalidate_any_buried_refs().
|
static |
Return nonzero if the expression in X (a memory reference) is killed in block BB before or after the insn with the LUID in UID_LIMIT. AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills before UID_LIMIT. To check the entire block, set UID_LIMIT to max_uid + 1 and AVAIL_P to 0.
References mem_conflict_info::conflict, basic_block_def::index, mem_conflict_info::mem, mems_conflict_for_gcse_p(), and note_stores().
Referenced by hash_scan_set(), and oprs_unchanged_p().
| rtl_opt_pass* make_pass_rtl_hoist | ( | ) |
| rtl_opt_pass* make_pass_rtl_pre | ( | ) |
DEST is the output of an instruction. If it is a memory reference and possibly conflicts with the load found in DATA, then communicate this information back through DATA.
References mem_conflict_info::conflict, expr_equiv_p(), find_rtx_in_ldst(), mem_conflict_info::mem, and true_dependence().
Referenced by load_killed_in_block_p().
|
static |
Top level routine to perform one code hoisting (aka unification) pass Return nonzero if a change was made.
References alloc_aux_for_blocks(), alloc_code_hoist_mem(), alloc_gcse_mem(), alloc_hash_table(), bytes_used, calculate_bb_reg_pressure(), changed, compute_code_hoist_data(), compute_hash_table(), current_function_name(), doing_code_hoisting_p, dump_file, dump_hash_table(), end_alias_analysis(), expr_hash_table, free_aux_for_blocks(), free_code_hoist_mem(), free_gcse_mem(), free_hash_table(), free_reg_info(), gcse_create_count, gcse_obstack, gcse_subst_count, hoist_code(), init_alias_analysis(), ira_set_pseudo_classes(), is_too_expensive(), hash_table_d::n_elems, regstat_free_n_sets_and_refs(), and regstat_init_n_sets_and_refs().
Referenced by execute_rtl_hoist().
|
static |
Top level routine to perform one PRE GCSE pass. Return nonzero if a change was made.
References add_noreturn_fake_exit_edges(), alloc_gcse_mem(), alloc_hash_table(), alloc_pre_mem(), bytes_used, changed, compute_hash_table(), compute_ld_motion_mems(), compute_pre_data(), current_function_name(), dump_file, dump_hash_table(), edge_list, end_alias_analysis(), expr_hash_table, free_edge_list(), free_gcse_mem(), free_hash_table(), free_ld_motion_mems(), free_pre_mem(), gcse_create_count, gcse_obstack, gcse_subst_count, init_alias_analysis(), is_too_expensive(), hash_table_d::n_elems, pre_gcse(), remove_fake_exit_edges(), and trim_ld_motion_mems().
Referenced by execute_rtl_pre().
|
static |
Return nonzero if the operands of expression X are unchanged from the start of INSN's basic block up to but not including INSN.
References oprs_unchanged_p().
Referenced by hash_scan_set().
|
static |
Return nonzero if the operands of expression X are unchanged from INSN to the end of INSN's basic block.
References oprs_unchanged_p().
Referenced by hash_scan_set().
|
static |
Return nonzero if the operands of expression X are unchanged from the start of INSN's basic block up to but not including INSN (if AVAIL_P == 0), or from INSN to the end of INSN's basic block (if AVAIL_P != 0).
References reg_avail_info::first_set, reg_avail_info::last_bb, reg_avail_info::last_set, and load_killed_in_block_p().
Referenced by eliminate_partially_redundant_loads(), hash_scan_set(), oprs_anticipatable_p(), oprs_available_p(), and oprs_unchanged_p().
|
static |
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. Return nonzero if a change is made.
References expr::antic_occr, bitmap_bit_p(), expr::bitmap_index, changed, dbg_cnt(), delete_insn(), occr::deleted_p, dump_file, expr::expr, expr_hash_table, gcse_emit_move_after(), gcse_subst_count, gen_reg_rtx_and_attrs(), basic_block_def::index, occr::insn, occr::next, expr::next_same_hash, expr::reaching_reg, hash_table_d::size, and hash_table_d::table.
Referenced by pre_gcse().
|
static |
Insert partially redundant expressions on edges in the CFG to make the expressions fully redundant.
References expr::antic_occr, bitmap_bit_p(), expr::bitmap_index, bitmap_set_bit(), bitmap_vector_clear(), occr::deleted_p, dump_file, simple_bitmap_def::elms, expr_hash_table, edge_def::flags, gcse_create_count, basic_block_def::index, insert(), insert_insn_end_basic_block(), insert_insn_on_edge(), inserted, occr::insn, hash_table_d::n_elems, occr::next, edge_list::num_edges, process_insert_insn(), expr::reaching_reg, sbitmap_vector_alloc(), sbitmap_vector_free(), simple_bitmap_def::size, and update_ld_motion_stores().
Referenced by pre_gcse().
|
static |
The wrapper for pre_expr_reaches_here_work that ensures that any memory allocated for that function is returned.
References free(), pre_expr_reaches_here_p_work(), and visited.
Referenced by pre_insert_copies().
|
static |
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.
References bitmap_bit_p(), expr::bitmap_index, basic_block_def::index, basic_block_def::preds, and edge_def::src.
Referenced by pre_expr_reaches_here_p().
|
static |
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.
References expr::bitmap_index, changed, commit_edge_insertions(), expr::expr, expr_hash_table, free(), hash_table_d::n_elems, expr::next_same_hash, pre_delete(), pre_edge_insert(), pre_insert_copies(), hash_table_d::size, and hash_table_d::table.
Referenced by one_pre_gcse_pass().
|
static |
Copy available expressions that reach the redundant expression to `reaching_reg'.
References expr::antic_occr, expr::avail_occr, occr::copied_p, occr::deleted_p, expr::expr, expr_hash_table, occr::insn, occr::next, expr::next_same_hash, pre_expr_reaches_here_p(), pre_insert_copy_insn(), expr::reaching_reg, hash_table_d::size, hash_table_d::table, and update_ld_motion_stores().
Referenced by pre_gcse().
|
static |
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.
References expr::bitmap_index, dump_file, emit_insn_after(), emit_insn_before(), expr::expr, expr_equiv_p(), gcse_create_count, gen_move_insn(), expr::reaching_reg, SET, and validate_change().
Referenced by pre_insert_copies().
|
static |
Dump debugging info about the ldst list.
References ls_expr::index, ls_expr::loads, ls_expr::next, ls_expr::pattern, print_rtl(), and ls_expr::stores.
Referenced by trim_ld_motion_mems().
|
static |
Generate RTL to copy an EXPR to its `reaching_reg' and return it.
References copy_rtx(), emit_insn(), emit_move_insn(), end_sequence(), exp(), expr::expr, general_operand(), get_insns(), insn_invalid_p(), expr::reaching_reg, and start_sequence().
Referenced by insert_insn_end_basic_block(), and pre_edge_insert().
|
static |
Remove certain expressions from anticipatable and transparent sets of basic blocks that have incoming abnormal edge. For PRE remove potentially trapping expressions to avoid placing them on abnormal edges. For hoisting remove memory references that can be clobbered by calls.
References bitmap_and_compl(), bitmap_clear(), expr::bitmap_index, bitmap_set_bit(), expr::expr, expr_hash_table, edge_def::flags, basic_block_def::index, may_trap_p(), hash_table_d::n_elems, expr::next_same_hash, basic_block_def::preds, sbitmap_alloc(), sbitmap_free(), hash_table_d::size, edge_def::src, hash_table_d::table, and ui.
Referenced by compute_code_hoist_data(), and compute_pre_data().
|
static |
It may be necessary to insert a large number of insns on edges to make the existing occurrences of expressions fully redundant. This routine examines the set of insertions and deletions and if the ratio of insertions to deletions is too high for a particular expression, then the expression is removed from the insertion/deletion sets. N_ELEMS is the number of elements in the hash table.
References bitmap_clear(), bitmap_clear_bit(), bitmap_set_bit(), free(), insertions, sbitmap_alloc(), and sbitmap_free().
Referenced by compute_pre_data().
|
static |
Record memory modification information for INSN. We do not actually care about the memory location(s) that are set, or even how they are set (consider a CALL_INSN). We merely need to record which insns modify memory.
References bitmap_set_bit(), canon_list_insert(), and note_stores().
Referenced by compute_hash_table_work(), record_last_set_info(), and record_opr_changes().
|
static |
Record register first/last/block set information for REGNO in INSN. first_set records the first place in the block where the register is set and is used to compute "anticipatability". last_set records the last place in the block where the register is set and is used to compute "availability". last_bb records the block for which first_set and last_set are valid, as a quick test to invalidate them.
References current_bb, reg_avail_info::first_set, reg_avail_info::last_bb, and reg_avail_info::last_set.
Referenced by compute_hash_table_work(), record_last_set_info(), and record_opr_changes().
|
static |
Called from compute_hash_table via note_stores to handle one SET or CLOBBER in an insn. DATA is really the instruction in which the SET is taking place.
References push_operand(), record_last_mem_set_info(), and record_last_reg_set_info().
Referenced by compute_hash_table_work(), and record_opr_changes().
|
static |
Determine if the expression EXPR should be hoisted to EXPR_BB up in flow graph, if it can reach BB unimpared. Stop the search if the expression would need to be moved more than DISTANCE instructions. DISTANCE is the number of instructions through which EXPR can be hoisted up in flow graph. BB_SIZE points to an array which contains the number of instructions for each basic block. PRESSURE_CLASS and NREGS are register class and number of hard registers for storing EXPR. HOISTED_BBS points to a bitmap indicating basic blocks through which EXPR is hoisted. FROM is the instruction from which EXPR is hoisted. It's unclear exactly what Muchnick meant by "unimpared". It seems to me that the expression must either be computed or transparent in *every* block in the path(s) from EXPR_BB to BB. Any other definition would allow the expression to be hoisted out of loops, even if the expression wasn't a loop invariant. Contrast this to reachability for PRE where an expression is considered reachable if *any* path reaches instead of *all* paths.
References bb_data::backup, bitmap_bit_p(), bitmap_clear(), bitmap_copy(), expr::bitmap_index, bitmap_set_bit(), expr::expr, basic_block_def::index, bb_data::live_in, bb_data::max_reg_pressure, bb_data::old_pressure, basic_block_def::preds, sbitmap_alloc(), sbitmap_free(), edge_def::src, and update_bb_reg_pressure().
Referenced by hoist_code().
|
static |
Load Motion for loads which only kill themselves.
Return true if x, a MEM, is a simple access with no 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.
References function::can_throw_non_call_exceptions, cfun, may_trap_p(), reg_mentioned_p(), and side_effects_p().
Referenced by compute_ld_motion_mems(), and invalidate_any_buried_refs().
|
static |
Remove any references that have been either invalidated or are not in the expression list for pre gcse.
References dump_file, expr::expr, ls_expr::expr, expr_equiv_p(), expr_hash_table, free_ldst_entry(), expr::hash, ls_expr::hash_index, ls_expr::invalid, last, ls_expr::next, expr::next_same_hash, ls_expr::pattern, pre_ldst_mems, print_ldst_list(), hash_table< Descriptor, Allocator >::remove_elt_with_hash(), hash_table_d::size, and hash_table_d::table.
Referenced by one_pre_gcse_pass().
|
static |
Update register pressure for BB when hoisting an expression from instruction FROM, if live ranges of inputs are shrunk. Also maintain live_in information if live range of register referred in FROM is shrunk. Return 0 if register pressure doesn't change, otherwise return the number by which register pressure is decreased. NOTE: Register pressure won't be increased in this function.
References bitmap_bit_p(), bitmap_clear_bit(), edge_def::dest, get_regno_pressure_class(), and basic_block_def::succs.
Referenced by should_hoist_expr_to_dom().
|
static |
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.
References copy(), copy_rtx(), df_insn_rescan(), dump_file, emit_insn_before(), expr::expr, find_rtx_in_ldst(), gcse_create_count, gen_move_insn(), print_inline_rtx(), print_rtl(), expr::reaching_reg, and ls_expr::stores.
Referenced by pre_edge_insert(), and pre_insert_copies().
|
static |
See whether X, the source of a set, is something we want to consider for GCSE.
References avoid_constant_pool_reference(), can_assign_to_reg_without_clobbers_p(), cfun, doing_code_hoisting_p, optimize_function_for_size_p(), optimize_function_for_speed_p(), and set_src_cost().
Referenced by hash_scan_set().
|
static |
For available exprs
|
static |
Nonzero for expressions that are locally anticipatable in the block.
|
static |
Bitmap indexed by block numbers to record which blocks contain function calls.
|
static |
Various variables for statistics gathering.
Memory used in a pass. This isn't intended to be absolutely precise. Its intent is only to keep an eye on memory usage.
Referenced by gcalloc(), gcse_alloc(), gmalloc(), insert_expr_in_table(), one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
This array parallels modify_mem_list, except that it stores MEMs being set and their canonicalized memory addresses.
|
static |
Nonzero for expressions that are computed (available) in the block.
Referenced by determine_roots(), determine_use_iv_cost_condition(), dump_components(), filter_suitable_components(), find_opt(), get_computation_cost_at(), graphds_dfs(), graphds_scc(), graphite_translate_clast_equation(), invert_exp_1(), iv_elimination_compare_lt(), op_iter_init_phidef(), op_iter_init_phiuse(), remove_unused_ivs(), rewrite_use_compare(), rewrite_use_nonlinear_expr(), rtl_lv_add_condition_to_bb(), set_component_ssa_name(), set_use_iv_cost(), split_data_refs_to_components(), TB_get_command(), TB_get_tree_code(), update_phi_components(), and validate_all_switches().
|
static |
Referenced by rtl_verify_bb_layout(), and schedule_region().
|
static |
Current register pressure for each pressure class.
Referenced by calculate_bb_reg_pressure(), and change_pressure().
|
static |
Referenced by record_last_reg_set_info().
| struct target_gcse default_target_gcse |
@verbatim Partial redundancy elimination / Hoisting for RTL.
Copyright (C) 1997-2013 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see http://www.gnu.org/licenses/.
TODO
- reordering of memory allocation and freeing to be more space efficient
- calc rough register pressure information and use the info to drive all
kinds of code motion (including code hoisting) in a unified way. References searched while implementing this.
Compilers Principles, Techniques and Tools
Aho, Sethi, Ullman
Addison-Wesley, 1988
Global Optimization by Suppression of Partial Redundancies
E. Morel, C. Renvoise
communications of the acm, Vol. 22, Num. 2, Feb. 1979
A Portable Machine-Independent Global Optimizer - Design and Measurements
Frederick Chow
Stanford Ph.D. thesis, Dec. 1983
A Fast Algorithm for Code Movement Optimization
D.M. Dhamdhere
SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988
A Solution to a Problem with Morel and Renvoise's
Global Optimization by Suppression of Partial Redundancies
K-H Drechsler, M.P. Stadel
ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988
Practical Adaptation of the Global Optimization
Algorithm of Morel and Renvoise
D.M. Dhamdhere
ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991
Efficiently Computing Static Single Assignment Form and the Control
Dependence Graph
R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck
ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991
Lazy Code Motion
J. Knoop, O. Ruthing, B. Steffen
ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
What's In a Region? Or Computing Control Dependence Regions in Near-Linear
Time for Reducible Flow Control
Thomas Ball
ACM Letters on Programming Languages and Systems,
Vol. 2, Num. 1-4, Mar-Dec 1993
An Efficient Representation for Sparse Sets
Preston Briggs, Linda Torczon
ACM Letters on Programming Languages and Systems,
Vol. 2, Num. 1-4, Mar-Dec 1993
A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion
K-H Drechsler, M.P. Stadel
ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993
Partial Dead Code Elimination
J. Knoop, O. Ruthing, B. Steffen
ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
Effective Partial Redundancy Elimination
P. Briggs, K.D. Cooper
ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
The Program Structure Tree: Computing Control Regions in Linear Time
R. Johnson, D. Pearson, K. Pingali
ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
Optimal Code Motion: Theory and Practice
J. Knoop, O. Ruthing, B. Steffen
ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994
The power of assignment motion
J. Knoop, O. Ruthing, B. Steffen
ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
Global code motion / global value numbering
C. Click
ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
Value Driven Redundancy Elimination
L.T. Simpson
Rice University Ph.D. thesis, Apr. 1996
Value Numbering
L.T. Simpson
Massively Scalar Compiler Project, Rice University, Sep. 1996
High Performance Compilers for Parallel Computing
Michael Wolfe
Addison-Wesley, 1996
Advanced Compiler Design and Implementation
Steven Muchnick
Morgan Kaufmann, 1997
Building an Optimizing Compiler
Robert Morgan
Digital Press, 1998
People wishing to speed up the code here should read:
Elimination Algorithms for Data Flow Analysis
B.G. Ryder, M.C. Paull
ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986
How to Analyze Large Programs Efficiently and Informatively
D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck
ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
People wishing to do something different can find various possibilities
in the above papers and elsewhere. We support GCSE via Partial Redundancy Elimination. PRE optimizations
are a superset of those done by classic GCSE.
Two passes of copy/constant propagation are done around PRE or hoisting
because the first one enables more GCSE and the second one helps to clean
up the copies that PRE and HOIST create. This is needed more for PRE than
for HOIST because code hoisting will try to use an existing register
containing the common subexpression rather than create a new one. This is
harder to do for PRE because of the code motion (which HOIST doesn't do).
Expressions we are interested in GCSE-ing are of the form
(set (pseudo-reg) (expression)).
Function want_to_gcse_p says what these are.
In addition, expressions in REG_EQUAL notes are candidates for GCSE-ing.
This allows PRE to hoist expressions that are expressed in multiple insns,
such as complex address calculations (e.g. for PIC code, or loads with a
high part and a low part).
PRE handles moving invariant expressions out of loops (by treating them as
partially redundant).
**********************
We used to support multiple passes but there are diminishing returns in
doing so. The first pass usually makes 90% of the changes that are doable.
A second pass can make a few more changes made possible by the first pass.
Experiments show any further passes don't make enough changes to justify
the expense.
A study of spec92 using an unlimited number of passes:
[1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
[6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
[12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
It was found doing copy propagation between each pass enables further
substitutions.
This study was done before expressions in REG_EQUAL notes were added as
candidate expressions for optimization, and before the GIMPLE optimizers
were added. Probably, multiple passes is even less efficient now than
at the time when the study was conducted.
PRE is quite expensive in complicated functions because the DFA can take
a while to converge. Hence we only perform one pass.
**********************
The steps for PRE are:
1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
2) Perform the data flow analysis for PRE.
3) Delete the redundant instructions
4) Insert the required copies [if any] that make the partially
redundant instructions fully redundant.
5) For other reaching expressions, insert an instruction to copy the value
to a newly created pseudo that will reach the redundant instruction.
The deletion is done first so that when we do insertions we
know which pseudo reg to use.
Various papers have argued that PRE DFA is expensive (O(n^2)) and others
argue it is not. The number of iterations for the algorithm to converge
is typically 2-4 so I don't view it as that expensive (relatively speaking).
PRE GCSE depends heavily on the second CPROP pass to clean up the copies
we create. To make an expression reach the place where it's redundant,
the result of the expression is copied to a new register, and the redundant
expression is deleted by replacing it with this new register. Classic GCSE
doesn't have this problem as much as it computes the reaching defs of
each register in each block and thus can try to use an existing
register. GCSE global vars.
|
static |
Doing code hoisting.
Referenced by one_code_hoisting_pass(), and want_to_gcse_p().
|
static |
Expression hash table.
Referenced by compute_code_hoist_data(), compute_pre_data(), hoist_code(), one_code_hoisting_pass(), one_pre_gcse_pass(), pre_delete(), pre_edge_insert(), pre_gcse(), pre_insert_copies(), prune_expressions(), and trim_ld_motion_mems().
| int flag_rerun_cse_after_global_opts |
Set to non-zero if CSE should run after all GCSE optimizations are done.
Referenced by execute_rtl_cprop(), execute_rtl_hoist(), execute_rtl_pre(), execute_rtl_store_motion(), gate_handle_cse_after_global_opts(), and rest_of_clean_state().
|
static |
Number of copy instructions created.
Referenced by insert_insn_end_basic_block(), one_code_hoisting_pass(), one_pre_gcse_pass(), pre_edge_insert(), pre_insert_copy_insn(), and update_ld_motion_stores().
|
static |
An obstack for our working variables.
Referenced by gcse_alloc(), one_code_hoisting_pass(), and one_pre_gcse_pass().
|
static |
GCSE substitutions made.
Referenced by hoist_code(), one_code_hoisting_pass(), one_pre_gcse_pass(), and pre_delete().
|
static |
Code Hoisting variables and subroutines.
Very busy expressions.
|
static |
Array, indexed by basic block number for a list of insns which modify memory within that block.
|
static |
|
static |
Nonzero for expressions which should be deleted in a specific block.
|
static |
Nonzero for expressions which should be inserted on a specific edge.
|
static |
Head of the list of load/store memory refs.
Referenced by free_ld_motion_mems(), ldst_entry(), and trim_ld_motion_mems().
|
static |
Hashtable for the load/store memory refs.
|
static |
Nonzero for expressions where this block is an optimal computation point.
|
static |
Nonzero for expressions which are redundant in a particular block.
|
static |
|
static |
Bitmap containing one bit for each register in the program. Used when performing GCSE to track which registers have been set since the start of the basic block.
|
static |
Used internally by can_assign_to_reg_without_clobbers_p.
Referenced by can_reload_into().
| struct target_gcse* this_target_gcse = &default_target_gcse |
|
static |
Compute PRE+LCM working variables.
Local properties of expressions.
Nonzero for expressions that are transparent in the block.
Referenced by find_moveable_pseudos().
1.8.1.1