#include "config.h"#include "system.h"#include "coretypes.h"#include "backend.h"#include "target.h"#include "rtl.h"#include "tree.h"#include "predict.h"#include "df.h"#include "memmodel.h"#include "tm_p.h"#include "insn-config.h"#include "print-rtl.h"#include "regs.h"#include "ira.h"#include "recog.h"#include "diagnostic-core.h"#include "cfgrtl.h"#include "cfganal.h"#include "lcm.h"#include "cfgcleanup.h"#include "expr.h"#include "intl.h"#include "tree-pass.h"#include "dbgcnt.h"#include "gcse.h"#include "gcse-common.h"#include "function-abi.h"#include "gt-gcse.h"
Data Structures | |
| struct | gcse_expr |
| struct | gcse_occr |
| struct | gcse_hash_table_d |
| struct | ls_expr |
| struct | pre_ldst_expr_hasher |
| struct | bb_data |
| struct | reg_avail_info |
| struct | mem_conflict_info |
| struct | set_data |
Macros | |
| #define | BB_DATA(bb) |
| #define | GNEW(T) |
| #define | GCNEW(T) |
| #define | GNEWVEC(T, N) |
| #define | GCNEWVEC(T, N) |
| #define | GNEWVAR(T, S) |
| #define | GCNEWVAR(T, S) |
| #define | GOBNEW(T) |
| #define | GOBNEWVAR(T, S) |
| #define | can_copy (this_target_gcse->x_can_copy) |
| #define | can_copy_init_p (this_target_gcse->x_can_copy_init_p) |
Typedefs | |
| typedef struct gcse_occr * | occr_t |
Macro Definition Documentation
◆ BB_DATA
| #define BB_DATA | ( | bb | ) |
Referenced by calculate_bb_reg_pressure(), change_pressure(), hoist_code(), should_hoist_expr_to_dom(), and update_bb_reg_pressure().
◆ can_copy
| #define can_copy (this_target_gcse->x_can_copy) |
Misc. utilities.
Referenced by can_copy_p(), and compute_can_copy().
◆ can_copy_init_p
| #define can_copy_init_p (this_target_gcse->x_can_copy_init_p) |
Referenced by can_copy_p().
◆ GCNEW
| #define GCNEW | ( | T | ) |
◆ GCNEWVAR
◆ GCNEWVEC
Referenced by alloc_gcse_mem(), and prune_insertions_deletions().
◆ GNEW
| #define GNEW | ( | T | ) |
◆ GNEWVAR
Referenced by alloc_hash_table().
◆ GNEWVEC
Referenced by compute_hash_table_work().
◆ GOBNEW
| #define GOBNEW | ( | T | ) |
Referenced by insert_expr_in_table().
◆ GOBNEWVAR
Typedef Documentation
◆ occr_t
Function Documentation
◆ add_label_notes()
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_label_notes(), add_reg_note(), gcc_assert, GET_CODE, GET_RTX_FORMAT, GET_RTX_LENGTH, i, JUMP_P, LABEL_NUSES, LABEL_P, label_ref_label(), LABEL_REF_NONLOCAL_P, XEXP, XVECEXP, and XVECLEN.
Referenced by add_label_notes(), and insert_insn_end_basic_block().
◆ alloc_code_hoist_mem()
|
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.cc. 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 antloc, comp, hoist_vbein, hoist_vbeout, sbitmap_vector_alloc(), and transp.
Referenced by one_code_hoisting_pass().
◆ alloc_gcse_mem()
|
static |
Allocate memory for the reg/memory set tracking tables. This is called at the start of each pass.
References ALLOC_REG_SET, BITMAP_ALLOC, blocks_with_calls, canon_modify_mem_list, cfun, GCNEWVEC, last_basic_block_for_fn, modify_mem_list, modify_mem_list_set, NULL, and reg_set_bitmap.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
◆ alloc_hash_table()
|
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(), GNEWVAR, and table.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
◆ alloc_pre_mem()
|
static |
Allocate vars used for PRE analysis.
References ae_kill, antloc, comp, NULL, pre_delete_map, pre_insert_map, pre_optimal, pre_redundant, sbitmap_vector_alloc(), and transp.
Referenced by one_pre_gcse_pass().
◆ calculate_bb_reg_pressure()
|
static |
Calculate register pressure for each basic block by walking insns from last to first.
References BB_DATA, BITMAP_ALLOC, bitmap_clear_bit(), bitmap_copy(), BITMAP_FREE, bitmap_set_bit, cfun, change_pressure(), curr_bb, curr_reg_pressure, curr_regs_live, df_get_live_in(), df_get_live_out(), DF_REF_CONDITIONAL, DF_REF_FLAGS, DF_REF_PARTIAL, DF_REF_REAL_REG, dump_file, EXECUTE_IF_SET_IN_BITMAP, FOR_BB_INSNS_REVERSE, FOR_EACH_BB_FN, FOR_EACH_INSN_DEF, FOR_EACH_INSN_USE, gcc_assert, i, basic_block_def::index, ira_pressure_classes, ira_pressure_classes_num, ira_setup_eliminable_regset(), NONDEBUG_INSN_P, NULL, reg_class_names, reg_obstack, REG_P, and REGNO.
Referenced by one_code_hoisting_pass().
◆ can_assign_to_reg_without_clobbers_p()
Return true if we can assign X to a pseudo register of mode MODE 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(), const0_rtx, gen_rtx_REG(), general_operand(), GET_MODE, INSN_LOCATION(), make_insn_raw(), NULL_RTX, PATTERN(), PUT_MODE(), recog(), SET_DEST, SET_NEXT_INSN(), SET_PREV_INSN(), SET_SRC, targetm, test_insn, UNKNOWN_LOCATION, and word_mode.
Referenced by compute_ld_motion_mems(), find_moveable_store(), and want_to_gcse_p().
◆ can_copy_p()
| bool can_copy_p | ( | machine_mode | mode | ) |
Returns whether the mode supports reg/reg copy operations.
References can_copy, can_copy_init_p, and compute_can_copy().
Referenced by hash_scan_set(), hash_scan_set(), hash_scan_set(), and may_assign_reg_p().
◆ change_pressure()
|
static |
Increase (if INCR_P) or decrease current register pressure for register REGNO.
References BB_DATA, curr_bb, curr_reg_pressure, and get_regno_pressure_class().
Referenced by calculate_bb_reg_pressure().
◆ clear_modify_mem_tables()
|
static |
Expression tracking support.
Clear canon_modify_mem_list and modify_mem_list tables.
References bitmap_clear(), blocks_with_calls, canon_modify_mem_list, EXECUTE_IF_SET_IN_BITMAP, i, modify_mem_list, and modify_mem_list_set.
Referenced by compute_hash_table_work(), and free_modify_mem_tables().
◆ compute_can_copy()
|
static |
Compute which modes support reg/reg copy operations.
References can_copy, emit_insn(), end_sequence(), gen_rtx_REG(), GET_MODE_CLASS, i, LAST_VIRTUAL_REGISTER, NULL, PATTERN(), recog(), and start_sequence().
Referenced by can_copy_p().
◆ compute_code_hoist_data()
|
static |
Top level routine to do the dataflow analysis needed by code hoisting.
References antloc, calculate_dominance_info(), CDI_DOMINATORS, comp, compute_code_hoist_vbeinout(), compute_local_properties(), dump_file, expr_hash_table, prune_expressions(), and transp.
Referenced by one_code_hoisting_pass().
◆ compute_code_hoist_vbeinout()
|
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 antloc, bitmap_intersection_of_succs(), bitmap_ior(), bitmap_or_and(), bitmap_vector_clear(), cfun, changed, comp, dump_bitmap_file(), dump_file, EXIT_BLOCK_PTR_FOR_FN, FOR_EACH_BB_FN, FOR_EACH_BB_REVERSE_FN, hoist_vbein, hoist_vbeout, basic_block_def::index, last_basic_block_for_fn, basic_block_def::next_bb, and transp.
Referenced by compute_code_hoist_data().
◆ compute_hash_table()
|
static |
Compute the expression hash table TABLE.
References compute_hash_table_work(), and table.
◆ compute_hash_table_work()
|
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 CALL_P, can_throw_external(), cfun, clear_modify_mem_tables(), current_bb, EXECUTE_IF_SET_IN_HARD_REG_SET, FOR_BB_INSNS, FOR_EACH_BB_FN, free(), function_abi::full_and_partial_reg_clobbers(), GNEWVEC, hash_scan_insn(), i, insn_callee_abi(), reg_avail_info::last_bb, max_reg_num(), NONDEBUG_INSN_P, note_stores(), NULL, record_last_mem_set_info(), record_last_reg_set_info(), record_last_set_info(), RTL_CONST_OR_PURE_CALL_P, RTL_LOOPING_CONST_OR_PURE_CALL_P, and table.
Referenced by compute_hash_table().
◆ compute_ld_motion_mems()
|
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 can_assign_to_reg_without_clobbers_p(), cfun, find_reg_equal_equiv_note(), FOR_BB_INSNS, FOR_EACH_BB_FN, GET_CODE, GET_MODE, ls_expr::invalid, invalidate_any_buried_refs(), ldst_entry(), MEM_P, NONDEBUG_INSN_P, NULL, PATTERN(), pre_ldst_mems, pre_ldst_table, REG_NOTE_KIND, REG_P, SET, SET_DEST, SET_SRC, simple_mem(), ls_expr::stores, and XEXP.
Referenced by one_pre_gcse_pass().
◆ compute_local_properties()
|
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 antloc, expr::avail_occr, expr::bitmap_index, bitmap_set_bit, bitmap_vector_clear(), bitmap_vector_ones(), BLOCK_FOR_INSN(), blocks_with_calls, canon_modify_mem_list, cfun, comp, compute_transp(), occr::deleted_p, expr, expr::expr, i, basic_block_def::index, occr::insn, last_basic_block_for_fn, modify_mem_list_set, occr::next, NULL, table, and transp.
Referenced by compute_code_hoist_data(), and compute_pre_data().
◆ compute_pre_data()
|
static |
Top level routine to do the dataflow analysis needed by PRE.
References ae_kill, antloc, bitmap_ior(), bitmap_not(), bitmap_vector_clear(), cfun, comp, compute_local_properties(), edge_list, expr_hash_table, FOR_EACH_BB_FN, basic_block_def::index, last_basic_block_for_fn, gcse_hash_table_d::n_elems, NULL, pre_delete_map, pre_edge_lcm(), pre_insert_map, prune_expressions(), prune_insertions_deletions(), sbitmap_vector_free(), and transp.
Referenced by one_pre_gcse_pass().
◆ dump_hash_table()
|
static |
Dump the hash table TABLE to file FILE under the name NAME.
References expr::bitmap_index, expr, expr::expr, free(), HOST_WIDE_INT_PRINT_DEC, i, NULL, print_rtl(), and table.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
◆ execute_rtl_hoist()
|
static |
◆ execute_rtl_pre()
|
static |
◆ expr_equiv_p()
Return true if exp1 is equivalent to exp2.
References exp_equiv_p(), and y.
Referenced by pre_ldst_expr_hasher::equal(), insert_expr_in_table(), mems_conflict_for_gcse_p(), pre_insert_copy_insn(), and trim_ld_motion_mems().
◆ find_occr_in_bb()
|
static |
Find occurrence in BB.
References BLOCK_FOR_INSN(), occr::insn, and occr::next.
Referenced by hoist_code().
◆ find_rtx_in_ldst()
Returns 1 if X is in the list of ldst only expressions.
References NULL, ls_expr::pattern, and pre_ldst_table.
Referenced by mems_conflict_for_gcse_p(), and update_ld_motion_stores().
◆ free_code_hoist_mem()
|
static |
Free vars used for code hoisting analysis.
References antloc, CDI_DOMINATORS, comp, free_dominance_info(), hoist_vbein, hoist_vbeout, sbitmap_vector_free(), and transp.
Referenced by one_code_hoisting_pass().
◆ free_gcse_mem()
|
static |
Free memory allocated by alloc_gcse_mem.
References BITMAP_FREE, blocks_with_calls, free_modify_mem_tables(), FREE_REG_SET, modify_mem_list_set, and reg_set_bitmap.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
◆ free_hash_table()
|
static |
Free things allocated by alloc_hash_table.
Referenced by one_code_hoisting_pass(), and one_pre_gcse_pass().
◆ free_ld_motion_mems()
|
static |
Free up all memory associated with the ldst list.
References free_ldst_entry(), ls_expr::next, NULL, pre_ldst_mems, and pre_ldst_table.
Referenced by one_pre_gcse_pass().
◆ free_ldst_entry()
|
static |
Free up an individual ldst entry.
References free(), and ls_expr::stores.
Referenced by free_ld_motion_mems(), and trim_ld_motion_mems().
◆ free_modify_mem_tables()
|
static |
Release memory used by modify_mem_list_set.
References canon_modify_mem_list, clear_modify_mem_tables(), free(), and modify_mem_list.
Referenced by free_gcse_mem().
◆ free_pre_mem()
|
static |
Free vars used for PRE analysis.
References comp, NULL, pre_delete_map, pre_insert_map, pre_optimal, pre_redundant, sbitmap_vector_free(), and transp.
Referenced by one_pre_gcse_pass().
◆ gcalloc()
|
static |
Cover function to xcalloc to record bytes allocated.
References bytes_used.
◆ gcse_alloc()
|
static |
Cover function to obstack_alloc.
References bytes_used, and gcse_obstack.
◆ gcse_cc_finalize()
| void gcse_cc_finalize | ( | void | ) |
Reset all state within gcse.cc so that we can rerun the compiler within the same process. For use by toplev::finalize.
References NULL, and test_insn.
Referenced by toplev::finalize().
◆ gcse_emit_move_after()
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(), set_data::insn, NULL_RTX, reg_mentioned_p(), REG_P, rtx_equal_p(), SET_DEST, SET_SRC, set_unique_reg_note(), single_set(), single_set_gcse(), and XEXP.
Referenced by hoist_code(), and pre_delete().
◆ gcse_or_cprop_is_too_expensive()
| bool gcse_or_cprop_is_too_expensive | ( | const char * | pass | ) |
Return true if the graph is too expensive to optimize. PASS is the optimization about to be performed.
References cfun, max_reg_num(), n_basic_blocks_for_fn, n_edges_for_fn, SBITMAP_ELT_TYPE, SBITMAP_SET_SIZE, and warning().
Referenced by gcse_after_reload_main(), one_code_hoisting_pass(), one_cprop_pass(), and one_pre_gcse_pass().
◆ get_pressure_class_and_nregs()
|
static |
Return pressure class and number of hard registers (through *NREGS) for destination of INSN.
References gcc_assert, GET_CODE, GET_MODE, ira_pressure_class_translate, ira_reg_class_max_nregs, MEM_P, reg_allocno_class(), REG_P, REGNO, SET_DEST, SET_SRC, single_set_gcse(), and SUBREG_REG.
Referenced by hoist_code().
◆ get_regno_pressure_class()
|
static |
Return pressure class and number of needed hard registers (through *NREGS) of register REGNO.
References eliminable_regset, ira_no_alloc_regs, ira_pressure_class_translate, ira_reg_class_max_nregs, PSEUDO_REGNO_MODE, reg_allocno_class(), and TEST_HARD_REG_BIT.
Referenced by change_pressure(), and update_bb_reg_pressure().
◆ gmalloc()
|
static |
Cover function to xmalloc to record bytes allocated.
References bytes_used.
◆ hash_expr()
|
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(), and NULL.
Referenced by insert_expr_in_table().
◆ hash_scan_call()
|
static |
Referenced by hash_scan_insn(), and hash_scan_set().
◆ hash_scan_clobber()
|
static |
Referenced by hash_scan_insn().
◆ hash_scan_insn()
|
static |
Process INSN and add hash table entries as appropriate.
References GET_CODE, hash_scan_call(), hash_scan_clobber(), hash_scan_set(), i, gcse_occr::insn, PATTERN(), SET, table, XVECEXP, and XVECLEN.
Referenced by compute_hash_table_work().
◆ hash_scan_set()
|
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(), GET_CODE, GET_MODE, hash_scan_call(), insert_expr_in_table(), gcse_occr::insn, JUMP_P, MEM_P, multiple_sets(), NULL, NULL_RTX, oprs_anticipatable_p(), oprs_available_p(), REG_NOTE_KIND, REG_P, REGNO, SET_DEST, set_noop_p(), SET_SRC, table, want_to_gcse_p(), and XEXP.
Referenced by hash_scan_insn().
◆ hoist_code()
|
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 antloc, BASIC_BLOCK_FOR_FN, BB_DATA, BITMAP_ALLOC, bitmap_bit_p, bitmap_clear(), bitmap_copy(), BITMAP_FREE, expr::bitmap_index, bitmap_set_bit, CDI_DOMINATORS, cfun, changed, comp, dbg_cnt(), delete_insn(), occr::deleted_p, EDGE_COUNT, EDGE_SUCC, ENTRY_BLOCK_PTR_FOR_FN, EXECUTE_IF_SET_IN_BITMAP, expr, expr_hash_table, find_occr_in_bb(), FOR_BB_INSNS, FOR_EACH_BB_FN, FOR_EACH_VEC_ELT, free(), gcc_assert, 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(), hoist_vbeout, i, basic_block_def::index, insert_insn_end_basic_block(), occr::insn, INSN_UID(), last_basic_block_for_fn, gcse_expr::max_distance, gcse_hash_table_d::n_elems, nearest_common_dominator_for_set(), NONDEBUG_INSN_P, NULL, SBITMAP_SIZE, SET_DEST, should_hoist_expr_to_dom(), single_set_gcse(), gcse_hash_table_d::size, gcse_hash_table_d::table, and vNULL.
Referenced by one_code_hoisting_pass().
◆ insert_expr_in_table()
|
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 true if X is an anticipatable expression. AVAIL_P is true if X is an available expression. MAX_DISTANCE is the maximum distance in instructions this expression can be moved.
References gcse_expr::antic_occr, gcse_expr::avail_occr, gcse_expr::bitmap_index, BLOCK_FOR_INSN(), bytes_used, gcse_occr::deleted_p, gcse_expr::expr, expr_equiv_p(), gcc_assert, GOBNEW, hash_expr(), gcse_occr::insn, gcse_expr::max_distance, gcse_occr::next, gcse_expr::next_same_hash, NULL, and table.
Referenced by hash_scan_set().
◆ insert_insn_end_basic_block() [1/2]
| rtx_insn * insert_insn_end_basic_block | ( | rtx_insn * | pat, |
| basic_block | bb ) |
Return the INSN which is added at the end of the block BB with same instruction pattern with PAT.
References add_label_notes(), BB_END, BB_HEAD, CALL_P, emit_insn_after_noloc(), emit_insn_before_noloc(), find_first_parameter_load(), gcc_assert, INSN_P, JUMP_P, LABEL_P, NEXT_INSN(), NONJUMP_INSN_P, NOTE_INSN_BASIC_BLOCK_P, NULL_RTX, PATTERN(), single_succ_edge(), and single_succ_p().
◆ insert_insn_end_basic_block() [2/2]
|
static |
Add EXPR to the end of basic block BB. This is used by both the PRE and code hoisting.
References expr::bitmap_index, dump_file, gcse_create_count, basic_block_def::index, insert_insn_end_basic_block(), INSN_UID(), process_insert_insn(), and REGNO.
Referenced by hoist_code(), insert_insn_end_basic_block(), and pre_edge_insert().
◆ invalidate_any_buried_refs()
|
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 GET_CODE, GET_RTX_FORMAT, GET_RTX_LENGTH, i, ls_expr::invalid, invalidate_any_buried_refs(), ldst_entry(), MEM_P, simple_mem(), XEXP, XVECEXP, and XVECLEN.
Referenced by compute_ld_motion_mems(), and invalidate_any_buried_refs().
◆ ldst_entry()
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, GET_MODE, ls_expr::hash_index, hash_rtx(), ls_expr::index, ls_expr::invalid, ls_expr::next, NULL, NULL_RTX, ls_expr::pattern, ls_expr::pattern_regs, pre_ldst_mems, pre_ldst_table, ls_expr::reaching_reg, and ls_expr::stores.
Referenced by compute_ld_motion_mems(), and invalidate_any_buried_refs().
◆ load_killed_in_block_p()
|
static |
Return true 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 true 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 false.
References CALL_P, mem_conflict_info::conflict, DF_INSN_LUID, FOR_EACH_VEC_ELT_REVERSE, basic_block_def::index, mem_conflict_info::mem, MEM_READONLY_P, mems_conflict_for_gcse_p(), modify_mem_list, and note_stores().
Referenced by oprs_unchanged_p().
◆ make_pass_rtl_hoist()
| rtl_opt_pass * make_pass_rtl_hoist | ( | gcc::context * | ctxt | ) |
◆ make_pass_rtl_pre()
| rtl_opt_pass * make_pass_rtl_pre | ( | gcc::context * | ctxt | ) |
◆ mems_conflict_for_gcse_p()
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(), GET_CODE, GET_MODE, mem_conflict_info::mem, MEM_P, NULL, pre_ldst_mems, true_dependence(), and XEXP.
Referenced by load_killed_in_block_p().
◆ one_code_hoisting_pass()
|
static |
Top level routine to perform one code hoisting (aka unification) pass Return true 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(), cfun, 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(), gcc_obstack_init, gcse_create_count, gcse_obstack, gcse_or_cprop_is_too_expensive(), gcse_subst_count, hoist_code(), init_alias_analysis(), ira_set_pseudo_classes(), last_basic_block_for_fn, n_basic_blocks_for_fn, gcse_hash_table_d::n_elems, NULL, NUM_FIXED_BLOCKS, regstat_free_n_sets_and_refs(), and regstat_init_n_sets_and_refs().
Referenced by execute_rtl_hoist().
◆ one_pre_gcse_pass()
|
static |
Top level routine to perform one PRE GCSE pass. Return true if a change was made.
References _, add_noreturn_fake_exit_edges(), alloc_gcse_mem(), alloc_hash_table(), alloc_pre_mem(), bytes_used, cfun, 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(), gcc_obstack_init, gcse_create_count, gcse_obstack, gcse_or_cprop_is_too_expensive(), gcse_subst_count, init_alias_analysis(), last_basic_block_for_fn, n_basic_blocks_for_fn, gcse_hash_table_d::n_elems, NULL, NUM_FIXED_BLOCKS, pre_gcse(), remove_fake_exit_edges(), and trim_ld_motion_mems().
Referenced by execute_rtl_pre().
◆ oprs_anticipatable_p()
Return true 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().
◆ oprs_available_p()
Return true 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().
◆ oprs_unchanged_p()
Return true 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 == false), or from INSN to the end of INSN's basic block (if AVAIL_P == true).
References CASE_CONST_ANY, current_bb, DF_INSN_LUID, reg_avail_info::first_set, GET_CODE, GET_RTX_FORMAT, GET_RTX_LENGTH, i, reg_avail_info::last_bb, reg_avail_info::last_set, load_killed_in_block_p(), oprs_unchanged_p(), REGNO, XEXP, XVECEXP, and XVECLEN.
Referenced by oprs_anticipatable_p(), oprs_available_p(), and oprs_unchanged_p().
◆ pre_delete()
|
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 to propagate the copy or eliminate it. Return true if a change is made.
References bitmap_bit_p, expr::bitmap_index, BLOCK_FOR_INSN(), changed, dbg_cnt(), delete_insn(), occr::deleted_p, dump_file, expr, expr_hash_table, gcse_emit_move_after(), gcse_subst_count, gen_reg_rtx_and_attrs(), i, basic_block_def::index, gcse_occr::insn, occr::insn, INSN_UID(), occr::next, NULL, pre_delete_map, REGNO, SET_DEST, single_set(), gcse_hash_table_d::size, and gcse_hash_table_d::table.
Referenced by pre_gcse().
◆ pre_edge_insert()
Insert partially redundant expressions on edges in the CFG to make the expressions fully redundant.
References bitmap_bit_p, expr::bitmap_index, bitmap_set_bit, bitmap_vector_clear(), occr::deleted_p, dump_file, simple_bitmap_def::elms, expr_hash_table, gcse_create_count, i, basic_block_def::index, INDEX_EDGE, INDEX_EDGE_PRED_BB, INDEX_EDGE_SUCC_BB, insert(), insert_insn_end_basic_block(), insert_insn_on_edge(), gcse_occr::insn, gcse_hash_table_d::n_elems, occr::next, NULL, NULL_RTX, NUM_EDGES, edge_list::num_edges, pre_insert_map, process_insert_insn(), SBITMAP_ELT_BITS, SBITMAP_ELT_TYPE, sbitmap_vector_alloc(), sbitmap_vector_free(), simple_bitmap_def::size, and update_ld_motion_stores().
Referenced by pre_gcse().
◆ pre_expr_reaches_here_p()
|
static |
The wrapper for pre_expr_reaches_here_work that ensures that any memory allocated for that function is returned.
References cfun, free(), last_basic_block_for_fn, pre_expr_reaches_here_p_work(), and visited.
Referenced by pre_insert_copies().
◆ pre_expr_reaches_here_p_work()
|
static |
PRE utilities
Return true 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, cfun, comp, ENTRY_BLOCK_PTR_FOR_FN, FOR_EACH_EDGE, basic_block_def::index, pre_expr_reaches_here_p_work(), basic_block_def::preds, transp, and visited.
Referenced by pre_expr_reaches_here_p(), and pre_expr_reaches_here_p_work().
◆ pre_gcse()
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_hash_table, free(), i, gcse_hash_table_d::n_elems, pre_delete(), pre_edge_insert(), pre_insert_copies(), gcse_hash_table_d::size, and gcse_hash_table_d::table.
Referenced by one_pre_gcse_pass().
◆ pre_insert_copies()
|
static |
Copy available expressions that reach the redundant expression to `reaching_reg'.
References expr::avail_occr, BLOCK_FOR_INSN(), gcse_occr::copied_p, rtx_insn::deleted(), occr::deleted_p, expr, expr_hash_table, i, gcse_occr::insn, occr::insn, gcse_occr::next, occr::next, NULL, pre_expr_reaches_here_p(), pre_insert_copy_insn(), gcse_hash_table_d::size, gcse_hash_table_d::table, and update_ld_motion_stores().
Referenced by pre_gcse().
◆ pre_insert_copy_insn()
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, BLOCK_FOR_INSN(), dump_file, emit_insn_after(), emit_insn_before(), expr::expr, expr_equiv_p(), gcc_assert, gcc_unreachable, gcse_create_count, gen_move_insn(), GET_CODE, i, gcse_occr::insn, INSN_UID(), NULL_RTX, PATTERN(), REG_P, REGNO, SET, SET_DEST, SET_SRC, validate_change(), XVECEXP, and XVECLEN.
Referenced by pre_insert_copies().
◆ prepare_copy_insn()
Generate RTL to copy an EXP to REG and return it.
References emit_insn(), emit_move_insn(), end_sequence(), exp(), gcc_unreachable, general_operand(), get_insns(), GET_MODE, insn_invalid_p(), and start_sequence().
Referenced by process_insert_insn(), and replace_store_insn().
◆ print_ldst_list()
|
static |
Dump debugging info about the ldst list.
References ls_expr::index, ls_expr::next, NULL, ls_expr::pattern, pre_ldst_mems, print_rtl(), print_rtx_insn_vec(), and ls_expr::stores.
Referenced by trim_ld_motion_mems().
◆ process_insert_insn()
Generate RTL to copy an EXPR to its `reaching_reg' and return it.
References copy_rtx(), exp(), expr::expr, and prepare_copy_insn().
Referenced by insert_insn_end_basic_block(), and pre_edge_insert().
◆ prune_expressions()
|
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 antloc, BB_END, bitmap_and_compl(), bitmap_clear(), expr::bitmap_index, bitmap_set_bit, CALL_P, cfun, CONSTANT_POOL_ADDRESS_P, contains_mem_rtx_p(), expr, expr::expr, expr_hash_table, FOR_EACH_BB_FN, FOR_EACH_EDGE, GET_CODE, basic_block_def::index, may_trap_p(), MEM_NOTRAP_P, MEM_P, MEM_READONLY_P, MEM_VOLATILE_P, gcse_hash_table_d::n_elems, basic_block_def::preds, gcse_hash_table_d::size, gcse_hash_table_d::table, transp, ui, and XEXP.
Referenced by compute_code_hoist_data(), and compute_pre_data().
◆ prune_insertions_deletions()
|
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, cfun, EXECUTE_IF_SET_IN_BITMAP, free(), GCNEWVEC, i, insertions, last_basic_block_for_fn, n_edges_for_fn, pre_delete_map, and pre_insert_map.
Referenced by compute_pre_data().
◆ record_last_mem_set_info()
|
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 blocks_with_calls, canon_modify_mem_list, modify_mem_list, modify_mem_list_set, and record_last_mem_set_info_common().
Referenced by compute_hash_table_work(), and record_last_set_info().
◆ record_last_reg_set_info()
|
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, DF_INSN_LUID, reg_avail_info::first_set, reg_avail_info::last_bb, and reg_avail_info::last_set.
Referenced by compute_hash_table_work(), and record_last_set_info().
◆ record_last_set_info()
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 GET_CODE, GET_MODE, MEM_P, push_operand(), record_last_mem_set_info(), record_last_reg_set_info(), REG_P, REGNO, and SUBREG_REG.
Referenced by compute_hash_table_work().
◆ record_set_data()
Increment number of sets and record set in DATA.
References find_reg_note(), GET_CODE, set_data::insn, set_data::nsets, SET, set_data::set, SET_DEST, and side_effects_p().
Referenced by single_set_gcse().
◆ should_hoist_expr_to_dom()
|
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, bitmap_bit_p, bitmap_clear(), bitmap_copy(), expr::bitmap_index, bitmap_set_bit, cfun, CONST_INT_P, ENTRY_BLOCK_PTR_FOR_FN, EXECUTE_IF_SET_IN_BITMAP, expr::expr, FOR_EACH_EDGE, gcc_assert, i, basic_block_def::index, ira_class_hard_regs_num, last_basic_block_for_fn, bb_data::max_reg_pressure, NULL, basic_block_def::preds, sbitmap_alloc(), sbitmap_free(), should_hoist_expr_to_dom(), transp, update_bb_reg_pressure(), and visited.
Referenced by hoist_code(), and should_hoist_expr_to_dom().
◆ simple_mem()
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 cfun, FLOAT_MODE_P, GET_MODE, may_trap_p(), MEM_VOLATILE_P, reg_mentioned_p(), side_effects_p(), and stack_pointer_rtx.
Referenced by compute_ld_motion_mems(), and invalidate_any_buried_refs().
◆ single_set_gcse()
References gcc_assert, GET_CODE, set_data::insn, INSN_P, note_pattern_stores(), set_data::nsets, PATTERN(), record_set_data(), SET, and set_data::set.
Referenced by gcse_emit_move_after(), get_pressure_class_and_nregs(), and hoist_code().
◆ trim_ld_motion_mems()
|
static |
Remove any references that have been either invalidated or are not in the expression list for pre gcse.
References dump_file, expr, expr::expr, ls_expr::expr, expr_equiv_p(), expr_hash_table, free_ldst_entry(), ls_expr::hash_index, ls_expr::invalid, last, ls_expr::next, NULL, ls_expr::pattern, pre_ldst_mems, pre_ldst_table, print_ldst_list(), gcse_hash_table_d::size, and gcse_hash_table_d::table.
Referenced by one_pre_gcse_pass().
◆ update_bb_reg_pressure()
|
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 BB_DATA, bitmap_bit_p, bitmap_clear_bit(), BLOCK_FOR_INSN(), cfun, DF_REF_INSN, DF_REF_INSN_INFO, DF_REF_NEXT_REG, DF_REF_REAL_REG, DF_REG_USE_CHAIN, EXIT_BLOCK_PTR_FOR_FN, FOR_EACH_EDGE, FOR_EACH_INSN_USE, get_regno_pressure_class(), NONDEBUG_INSN_P, NULL, REGNO, and basic_block_def::succs.
Referenced by should_hoist_expr_to_dom().
◆ update_ld_motion_stores()
|
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_rtx(), df_insn_rescan(), dump_file, emit_insn_before(), expr::expr, find_rtx_in_ldst(), FOR_EACH_VEC_ELT_REVERSE, gcse_create_count, gen_move_insn(), i, INSN_CODE, PATTERN(), print_inline_rtx(), print_rtl(), SET_SRC, and ls_expr::stores.
Referenced by pre_edge_insert(), and pre_insert_copies().
◆ want_to_gcse_p()
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(), CASE_CONST_ANY, cfun, COSTS_N_INSNS, doing_code_hoisting_p, gcc_assert, GET_CODE, GET_MODE, optimize_function_for_size_p(), optimize_function_for_speed_p(), and set_src_cost().
Referenced by hash_scan_set().
Variable Documentation
◆ ae_kill
|
static |
For available exprs
Referenced by alloc_pre_mem(), and compute_pre_data().
◆ antloc
|
static |
Nonzero for expressions that are locally anticipatable in the block.
Referenced by alloc_code_hoist_mem(), alloc_pre_mem(), compute_antinout_edge(), compute_code_hoist_data(), compute_code_hoist_vbeinout(), compute_insert_delete(), compute_laterin(), compute_local_properties(), compute_pre_data(), free_code_hoist_mem(), hoist_code(), pre_edge_lcm(), pre_edge_lcm_avs(), and prune_expressions().
◆ blocks_with_calls
|
static |
Bitmap indexed by block numbers to record which blocks contain function calls.
Referenced by alloc_gcse_mem(), clear_modify_mem_tables(), compute_local_properties(), compute_transp(), free_gcse_mem(), record_last_mem_set_info(), and record_last_mem_set_info_common().
◆ bytes_used
|
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().
◆ canon_modify_mem_list
|
static |
This array parallels modify_mem_list, except that it stores MEMs being set and their canonicalized memory addresses.
Referenced by alloc_gcse_mem(), clear_modify_mem_tables(), compute_local_properties(), compute_transp(), free_modify_mem_tables(), record_last_mem_set_info(), and record_last_mem_set_info_common().
◆ comp
|
static |
Nonzero for expressions that are computed (available) in the block.
Referenced by alloc_code_hoist_mem(), alloc_pre_mem(), build_poly_int_cst(), compare_and_jump_seq(), compute_code_hoist_data(), compute_code_hoist_vbeinout(), compute_local_properties(), compute_local_properties(), compute_pre_data(), determine_group_iv_cost_cond(), pcom_worker::determine_roots(), pcom_worker::determine_roots_comp(), dump_component(), dump_components(), emit_conditional_move(), expand_ifn_atomic_op_fetch_cmp_0(), pcom_worker::filter_suitable_components(), find_opt(), free_code_hoist_mem(), free_pre_mem(), graphds_dfs(), graphds_scc(), hoist_code(), invert_exp_1(), iv_elimination_compare_lt(), make_invariant_chain(), may_eliminate_iv(), op_iter_init_phidef(), op_iter_init_phiuse(), pre_expr_reaches_here_p_work(), remove_unused_ivs(), rewrite_use_compare(), rewrite_use_nonlinear_expr(), rtl_lv_add_condition_to_bb(), set_component_ssa_name(), set_group_iv_cost(), typed_splay_tree< KEY_TYPE, VALUE_TYPE >::splay_tree_lookup(), typed_splay_tree< KEY_TYPE, VALUE_TYPE >::splay_tree_remove(), pcom_worker::split_data_refs_to_components(), pcom_worker::suitable_component_p(), typed_splay_tree< KEY_TYPE, VALUE_TYPE >::typed_splay_tree(), update_phi_components(), and validate_all_switches().
◆ curr_bb
|
static |
◆ curr_reg_pressure
|
static |
Current register pressure for each pressure class.
Referenced by calculate_bb_reg_pressure(), and change_pressure().
◆ current_bb
|
static |
Referenced by compute_hash_table_work(), oprs_unchanged_p(), and record_last_reg_set_info().
◆ default_target_gcse
| struct target_gcse default_target_gcse |
Partial redundancy elimination / Hoisting for RTL. Copyright (C) 1997-2024 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.
◆ doing_code_hoisting_p
Doing code hoisting.
Referenced by one_code_hoisting_pass(), and want_to_gcse_p().
◆ expr_hash_table
|
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().
◆ flag_rerun_cse_after_global_opts
| 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(), and rest_of_clean_state().
◆ gcse_create_count
|
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().
◆ gcse_obstack
|
static |
An obstack for our working variables.
Referenced by gcse_alloc(), one_code_hoisting_pass(), and one_pre_gcse_pass().
◆ gcse_subst_count
|
static |
GCSE substitutions made.
Referenced by hoist_code(), one_code_hoisting_pass(), one_pre_gcse_pass(), and pre_delete().
◆ hoist_vbein
|
static |
Code Hoisting variables and subroutines.
Very busy expressions.
Referenced by alloc_code_hoist_mem(), compute_code_hoist_vbeinout(), and free_code_hoist_mem().
◆ hoist_vbeout
|
static |
Referenced by alloc_code_hoist_mem(), compute_code_hoist_vbeinout(), free_code_hoist_mem(), and hoist_code().
◆ modify_mem_list
Array, indexed by basic block number for a list of insns which modify memory within that block.
Referenced by alloc_gcse_mem(), clear_modify_mem_tables(), free_modify_mem_tables(), load_killed_in_block_p(), record_last_mem_set_info(), and record_last_mem_set_info_common().
◆ modify_mem_list_set
|
static |
◆ pre_delete_map
|
static |
Nonzero for expressions which should be deleted in a specific block.
Referenced by alloc_pre_mem(), compute_pre_data(), free_pre_mem(), pre_delete(), and prune_insertions_deletions().
◆ pre_insert_map
|
static |
Nonzero for expressions which should be inserted on a specific edge.
Referenced by alloc_pre_mem(), compute_pre_data(), free_pre_mem(), pre_edge_insert(), and prune_insertions_deletions().
◆ pre_ldst_mems
Head of the list of load/store memory refs.
Referenced by compute_ld_motion_mems(), free_ld_motion_mems(), ldst_entry(), mems_conflict_for_gcse_p(), print_ldst_list(), and trim_ld_motion_mems().
◆ pre_ldst_table
|
static |
Hashtable for the load/store memory refs.
Referenced by compute_ld_motion_mems(), find_rtx_in_ldst(), free_ld_motion_mems(), ldst_entry(), and trim_ld_motion_mems().
◆ pre_optimal
|
static |
Nonzero for expressions where this block is an optimal computation point.
Referenced by alloc_pre_mem(), and free_pre_mem().
◆ pre_redundant
|
static |
Nonzero for expressions which are redundant in a particular block.
Referenced by alloc_pre_mem(), and free_pre_mem().
◆ reg_avail_info
|
static |
◆ reg_set_bitmap
|
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.
Referenced by alloc_gcse_mem(), and free_gcse_mem().
◆ test_insn
|
static |
Used internally by can_assign_to_reg_without_clobbers_p.
Referenced by can_assign_to_reg_without_clobbers_p(), can_reload_into(), and gcse_cc_finalize().
◆ transp
|
static |
Compute PRE+LCM working variables.
Local properties of expressions.
Nonzero for expressions that are transparent in the block.
Referenced by alloc_code_hoist_mem(), alloc_pre_mem(), compute_antinout_edge(), compute_code_hoist_data(), compute_code_hoist_vbeinout(), compute_local_properties(), compute_pre_data(), find_moveable_pseudos(), free_code_hoist_mem(), free_pre_mem(), pre_edge_lcm(), pre_edge_lcm_avs(), pre_edge_rev_lcm(), pre_expr_reaches_here_p_work(), prune_expressions(), and should_hoist_expr_to_dom().
