#include "config.h"#include "system.h"#include "coretypes.h"#include "backend.h"#include "tree.h"#include "gimple.h"#include "cfghooks.h"#include "tree-pass.h"#include "ssa.h"#include "fold-const.h"#include "cfganal.h"#include "gimplify.h"#include "gimple-iterator.h"#include "gimplify-me.h"#include "tree-cfg.h"#include "tree-ssa-loop-manip.h"#include "tree-into-ssa.h"#include "tree-ssa.h"#include "cfgloop.h"#include "tree-scalar-evolution.h"#include "tree-vectorizer.h"#include "tree-ssa-loop-ivopts.h"#include "gimple-fold.h"#include "tree-ssa-loop-niter.h"#include "internal-fn.h"#include "stor-layout.h"#include "optabs-query.h"#include "vec-perm-indices.h"#include "insn-config.h"#include "rtl.h"#include "recog.h"#include "langhooks.h"#include "tree-vector-builder.h"#include "optabs-tree.h"
Data Structures | |
| struct | adjust_info |
Variables | |
| static vec< adjust_info, va_heap > | adjust_vec |
Function Documentation
◆ add_header_seq()
|
static |
Add SEQ to the beginning of LOOP's header block.
References gsi_after_labels(), gsi_insert_seq_before(), GSI_SAME_STMT, and loop::header.
Referenced by vect_set_loop_condition_partial_vectors().
◆ add_preheader_seq()
|
static |
Add SEQ to the end of LOOP's preheader block.
References gcc_assert, gsi_insert_seq_on_edge_immediate(), and loop_preheader_edge().
Referenced by vect_set_loop_condition_partial_vectors(), and vect_set_loop_condition_partial_vectors_avx512().
◆ adjust_debug_stmts()
|
static |
Adjust any debug stmts that referenced FROM values to use the loop-closed TO, if the references are dominated by BB and not by the definition of FROM. If adjust_vec is non-NULL, adjustments will be postponed until adjust_vec_debug_stmts is called.
References adjust_debug_stmts_now(), adjust_vec, adjust_info::bb, adjust_info::from, MAY_HAVE_DEBUG_BIND_STMTS, SSA_NAME_IS_DEFAULT_DEF, adjust_info::to, TREE_CODE, and virtual_operand_p().
Referenced by adjust_phi_and_debug_stmts(), and slpeel_tree_duplicate_loop_to_edge_cfg().
◆ adjust_debug_stmts_now()
|
static |
Adjust any debug stmts that referenced AI->from values to use the loop-closed AI->to, if the references are dominated by AI->bb and not by the definition of AI->from.
References adjust_info::bb, CDI_DOMINATORS, dom_info_available_p(), dominated_by_p(), FOR_EACH_IMM_USE_ON_STMT, FOR_EACH_IMM_USE_STMT, adjust_info::from, gcc_assert, gimple_bb(), gimple_debug_bind_p(), gimple_debug_bind_reset_value(), is_gimple_debug(), SET_USE, SSA_NAME_DEF_STMT, adjust_info::to, and update_stmt().
Referenced by adjust_debug_stmts(), and adjust_vec_debug_stmts().
◆ adjust_phi_and_debug_stmts()
Change E's phi arg in UPDATE_PHI to NEW_DEF, and record information to adjust any debug stmts that referenced the old phi arg, presumably non-loop-closed references left over from other transformations.
References adjust_debug_stmts(), gcc_assert, gimple_bb(), MAY_HAVE_DEBUG_BIND_STMTS, PHI_ARG_DEF_FROM_EDGE, PHI_RESULT, SET_PHI_ARG_DEF, and TREE_CODE.
Referenced by slpeel_tree_duplicate_loop_to_edge_cfg(), slpeel_update_phi_nodes_for_guard1(), vect_loop_versioning(), and vect_update_ivs_after_vectorizer().
◆ adjust_vec_debug_stmts()
|
static |
Adjust debug stmts as scheduled before.
References adjust_debug_stmts_now(), adjust_vec, gcc_assert, and MAY_HAVE_DEBUG_BIND_STMTS.
Referenced by vect_do_peeling().
◆ chain_cond_expr()
Set *COND_EXPR to a tree that is true when both the original *COND_EXPR and PART_COND_EXPR are true. Treat a null *COND_EXPR as "true".
References boolean_type_node, and fold_build2.
Referenced by vect_create_cond_for_align_checks(), vect_create_cond_for_lower_bounds(), and vect_create_cond_for_unequal_addrs().
◆ find_loop_location()
| dump_user_location_t find_loop_location | ( | class loop * | loop | ) |
Function find_loop_location. Extract the location of the loop in the source code. If the loop is not well formed for vectorization, an estimated location is calculated. Return the loop location if succeed and NULL if not.
References BUILTINS_LOCATION, cfun, dump_user_location_t::from_function_decl(), get_loop_exit_condition(), get_loop_exit_edges(), gimple_location(), gsi_end_p(), gsi_next(), gsi_start_bb(), gsi_stmt(), loop::header, LOCATION_LOCUS, loop_outer(), LOOPS_HAVE_RECORDED_EXITS, loops_state_satisfies_p(), NULL, and si.
Referenced by canonicalize_loop_induction_variables(), loop_distribution::execute(), find_loop_guard(), hoist_guard(), tree_loop_interchange::interchange(), optimize_mask_stores(), parallelize_loops(), tree_loop_unroll_and_jam(), tree_ssa_iv_optimize_loop(), tree_ssa_unswitch_loops(), tree_unswitch_outer_loop(), and try_vectorize_loop_1().
◆ get_live_virtual_operand_on_edge()
Get the virtual operand live on E. The precondition on this is valid immediate dominators and an actual virtual definition dominating E.
??? Costly band-aid. For the use in question we can populate a live-on-exit/end-of-BB virtual operand when copying stmts.
References CDI_DOMINATORS, get_immediate_dominator(), get_virtual_phi(), gimple_phi_result(), gimple_vdef(), gimple_vuse(), gsi_end_p(), gsi_last_bb(), gsi_prev(), and gsi_stmt().
Referenced by slpeel_tree_duplicate_loop_to_edge_cfg().
◆ get_misalign_in_elems()
|
static |
Return a gimple value containing the misalignment (measured in vector elements) for the loop described by LOOP_VINFO, i.e. how many elements it is away from a perfectly aligned address. Add any new statements to SEQ.
References build_int_cst(), dr_info::dr, DR_STEP, DR_TARGET_ALIGNMENT, exact_log2(), fold_build2, fold_convert, int_cst_value(), LOOP_VINFO_UNALIGNED_DR, offset, size_int, size_zero_node, STMT_VINFO_VECTYPE, target_align(), tree_int_cst_compare(), TREE_INT_CST_LOW, TREE_TYPE, TYPE_SIZE_UNIT, TYPE_VECTOR_SUBPARTS(), unsigned_type_for(), and vect_create_addr_base_for_vector_ref().
Referenced by vect_gen_prolog_loop_niters(), and vect_prepare_for_masked_peels().
◆ interleave_supported_p()
|
static |
Return true if the target can interleave elements of two vectors. OFFSET is 0 if the first half of the vectors should be interleaved or 1 if the second half should. When returning true, store the associated permutation in INDICES.
References can_vec_perm_const_p(), i, offset, TYPE_MODE, and TYPE_VECTOR_SUBPARTS().
Referenced by vect_maybe_permute_loop_masks().
◆ iv_phi_p()
|
static |
Return true if the phi described by STMT_INFO defines an IV of the loop to be vectorized.
References as_a(), PHI_RESULT, STMT_VINFO_DEF_TYPE, vect_double_reduction_def, vect_reduction_def, and virtual_operand_p().
Referenced by vect_can_advance_ivs_p(), and vect_update_ivs_after_vectorizer().
◆ rename_use_op()
|
static |
Vectorizer Specific Loop Manipulations Copyright (C) 2003-2024 Free Software Foundation, Inc. Contributed by Dorit Naishlos <dorit@il.ibm.com> and Ira Rosen <irar@il.ibm.com> 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/>.
Renames the use *OP_P.
References get_current_def(), SET_USE, TREE_CODE, and USE_FROM_PTR.
Referenced by rename_variables_in_bb(), and slpeel_tree_duplicate_loop_to_edge_cfg().
◆ rename_variables_in_bb()
|
static |
Renames the variables in basic block BB. Allow renaming of PHI arguments on edges incoming from outer-block header if RENAME_FROM_OUTER_LOOP is true.
References flow_bb_inside_loop_p(), FOR_EACH_EDGE, FOR_EACH_SSA_USE_OPERAND, gcc_assert, gsi_end_p(), gsi_next(), gsi_start_bb(), gsi_start_phis(), gsi_stmt(), loop::header, loop::inner, basic_block_def::loop_father, loop_outer(), loop::next, NULL, PHI_ARG_DEF_PTR_FROM_EDGE, basic_block_def::preds, rename_use_op(), single_pred(), single_pred_p(), and SSA_OP_ALL_USES.
Referenced by slpeel_tree_duplicate_loop_to_edge_cfg().
◆ slpeel_add_loop_guard()
|
static |
Given the condition expression COND, put it as the last statement of GUARD_BB; set both edges' probability; set dominator of GUARD_TO to DOM_BB; return the skip edge. GUARD_TO is the target basic block to skip the loop. PROBABILITY is the skip edge's probability. Mark the new edge as irreducible if IRREDUCIBLE_P is true.
References CDI_DOMINATORS, EDGE_SUCC, force_gimple_operand_1(), gimple_build_cond_from_tree(), gsi_insert_after(), gsi_insert_seq_after(), gsi_last_bb(), GSI_NEW_STMT, profile_probability::invert(), is_gimple_condexpr_for_cond(), make_edge(), NULL, NULL_TREE, set_immediate_dominator(), and split_edge().
Referenced by vect_do_peeling().
◆ slpeel_can_duplicate_loop_p()
| bool slpeel_can_duplicate_loop_p | ( | const class loop * | loop, |
| const_edge | exit_e, | ||
| const_edge | e ) |
This function verifies that the following restrictions apply to LOOP:
(1) it consists of exactly 2 basic blocks - header, and an empty latch
for innermost loop and 5 basic blocks for outer-loop.
(2) it is single entry, single exit
(3) its exit condition is the last stmt in the header
(4) E is the entry/exit edge of LOOP.
References can_copy_bbs_p(), empty_block_p(), free(), get_loop_body_with_size(), get_loop_exit_condition(), gsi_last_bb(), gsi_stmt(), loop::latch, loop_outer(), loop_preheader_edge(), and loop::num_nodes.
Referenced by vect_analyze_loop_2(), vect_do_peeling(), and vect_enhance_data_refs_alignment().
◆ slpeel_tree_duplicate_loop_to_edge_cfg()
| class loop * slpeel_tree_duplicate_loop_to_edge_cfg | ( | class loop * | loop, |
| edge | loop_exit, | ||
| class loop * | scalar_loop, | ||
| edge | scalar_exit, | ||
| edge | e, | ||
| edge * | new_e, | ||
| bool | flow_loops, | ||
| vec< basic_block > * | updated_doms ) |
Given LOOP this function generates a new copy of it and puts it on E which is either the entry or exit of LOOP. If SCALAR_LOOP is non-NULL, assume LOOP and SCALAR_LOOP are equivalent and copy the basic blocks from SCALAR_LOOP instead of LOOP, but to either the entry or exit of LOOP. If FLOW_LOOPS then connect LOOP to SCALAR_LOOP as a continuation. This is correct for cases where one loop continues from the other like in the vectorizer, but not true for uses in e.g. loop distribution where the contents of the loop body are split but the iteration space of both copies remains the same. If UPDATED_DOMS is not NULL it is update with the list of basic blocks whoms dominators were updated during the peeling. When doing early break vectorization then LOOP_VINFO needs to be provided and is used to keep track of any newly created memory references that need to be updated should we decide to vectorize.
References add_phi_arg(), add_phi_args_after_copy(), adjust_debug_stmts(), adjust_phi_and_debug_stmts(), CDI_DOMINATORS, checking_verify_dominators(), copy_bbs(), copy_ssa_name(), create_phi_node(), delete_basic_block(), duplicate_loop(), duplicate_subloops(), EDGE_COUNT, EDGE_PRED, first_dom_son(), flow_bb_inside_loop_p(), flush_pending_stmts(), FOR_EACH_EDGE, free(), gcc_assert, hash_map< KeyId, Value, Traits >::get(), get_all_dominated_blocks(), get_bb_copy(), get_immediate_dominator(), get_live_virtual_operand_on_edge(), get_loop_body_with_size(), get_loop_copy(), get_loop_exit_edges(), get_virtual_phi(), gimple_phi_arg_def_from_edge(), gimple_phi_num_args(), gimple_phi_result(), gsi_end_p(), gsi_for_stmt(), gsi_next(), gsi_start_phis(), gsi_stmt(), loop::header, i, loop::inner, iterate_fix_dominators(), loop::latch, loop_latch_edge(), loop_outer(), loop_preheader_edge(), MAY_HAVE_DEBUG_BIND_STMTS, next_dom_son(), NULL, NULL_TREE, loop::num_nodes, PHI_ARG_DEF_FROM_EDGE, PHI_ARG_DEF_PTR_FROM_EDGE, PHI_RESULT, basic_block_def::preds, hash_map< KeyId, Value, Traits >::put(), queue, redirect_edge_and_branch(), redirect_edge_and_branch_force(), redirect_edge_pred(), redirect_edge_var_map_clear(), remove_phi_node(), rename_use_op(), rename_variables_in_bb(), set_immediate_dominator(), SET_PHI_ARG_DEF, SET_PHI_ARG_DEF_ON_EDGE, single_pred(), single_pred_edge(), single_succ_edge(), single_succ_p(), split_edge(), TREE_CODE, true, UNKNOWN_LOCATION, and virtual_operand_p().
Referenced by copy_loop_before(), and vect_do_peeling().
◆ slpeel_update_phi_nodes_for_guard1()
|
static |
Function slpeel_add_loop_guard adds guard skipping from the beginning
of SKIP_LOOP to the beginning of UPDATE_LOOP. GUARD_EDGE and MERGE_EDGE
are two pred edges of the merge point before UPDATE_LOOP. The two loops
appear like below:
guard_bb:
if (cond)
goto merge_bb;
else
goto skip_loop;
skip_loop:
header_a:
i_1 = PHI<i_0, i_2>;
...
i_2 = i_1 + 1;
if (cond_a)
goto latch_a;
else
goto exit_a;
latch_a:
goto header_a;
exit_a:
i_5 = PHI<i_2>;
merge_bb:
;; PHI (i_x = PHI<i_0, i_5>) to be created at merge point.
update_loop:
header_b:
i_3 = PHI<i_5, i_4>; ;; Use of i_5 to be replaced with i_x.
...
i_4 = i_3 + 1;
if (cond_b)
goto latch_b;
else
goto exit_bb;
latch_b:
goto header_b;
exit_bb:
This function creates PHI nodes at merge_bb and replaces the use of i_5
in the update_loop's PHI node with the result of new PHI result.
References add_phi_arg(), adjust_phi_and_debug_stmts(), copy_ssa_name(), create_phi_node(), gimple_phi_arg_location_from_edge(), gsi_end_p(), gsi_next(), gsi_start_phis(), loop::header, loop_preheader_edge(), gphi_iterator::phi(), PHI_ARG_DEF_FROM_EDGE, and PHI_RESULT.
Referenced by vect_do_peeling().
◆ vect_adjust_loop_lens_control()
|
static |
Populate DEST_RGM->controls, given that they should add up to STEP.
STEP = MIN_EXPR <ivtmp_34, VF>;
First length (MIN (X, VF/N)):
loop_len_15 = MIN_EXPR <STEP, VF/N>;
Second length:
tmp = STEP - loop_len_15;
loop_len_16 = MIN (tmp, VF/N);
Third length:
tmp2 = tmp - loop_len_16;
loop_len_17 = MIN (tmp2, VF/N);
Last length:
loop_len_18 = tmp2 - loop_len_17;
References build_int_cst(), rgroup_controls::controls, rgroup_controls::factor, gimple_build(), gimple_build_assign(), gimple_seq_add_stmt(), i, rgroup_controls::type, and TYPE_VECTOR_SUBPARTS().
Referenced by vect_set_loop_condition_partial_vectors().
◆ vect_build_loop_niters()
| tree vect_build_loop_niters | ( | loop_vec_info | loop_vinfo, |
| bool * | new_var_p ) |
This function builds ni_name = number of iterations. Statements are emitted on the loop preheader edge. If NEW_VAR_P is not NULL, set it to TRUE if new ssa_var is generated.
References create_tmp_var, force_gimple_operand(), gsi_insert_seq_on_edge_immediate(), loop_preheader_edge(), LOOP_VINFO_LOOP, LOOP_VINFO_NITERS, NULL, TREE_CODE, TREE_TYPE, and unshare_expr().
Referenced by vect_do_peeling(), and vect_transform_loop().
◆ vect_can_advance_ivs_p()
| bool vect_can_advance_ivs_p | ( | loop_vec_info | loop_vinfo | ) |
Function vect_can_advance_ivs_p In case the number of iterations that LOOP iterates is unknown at compile time, an epilog loop will be generated, and the loop induction variables (IVs) will be "advanced" to the value they are supposed to take just before the epilog loop. Here we check that the access function of the loop IVs and the expression that represents the loop bound are simple enough. These restrictions will be relaxed in the future.
References dump_enabled_p(), dump_printf(), dump_printf_loc(), expr_invariant_in_loop_p(), gsi_end_p(), gsi_next(), gsi_start_phis(), loop::header, iv_phi_p(), vec_info::lookup_stmt(), LOOP_VINFO_LOOP, MSG_MISSED_OPTIMIZATION, MSG_NOTE, NULL_TREE, gphi_iterator::phi(), STMT_VINFO_LOOP_PHI_EVOLUTION_PART, STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE, tree_is_chrec(), vect_can_peel_nonlinear_iv_p(), vect_location, and vect_step_op_add.
Referenced by vect_analyze_loop_2(), vect_do_peeling(), and vect_enhance_data_refs_alignment().
◆ vect_can_peel_nonlinear_iv_p()
|
static |
Return true if vectorizer can peel for nonlinear iv.
References dump_enabled_p(), dump_printf_loc(), wi::exact_log2(), LOOP_VINFO_INT_NITERS, LOOP_VINFO_MASK_SKIP_NITERS, LOOP_VINFO_NITERS_KNOWN_P, LOOP_VINFO_PEELING_FOR_ALIGNMENT, LOOP_VINFO_USING_PARTIAL_VECTORS_P, LOOP_VINFO_VECT_FACTOR, MSG_MISSED_OPTIMIZATION, NULL_TREE, STMT_VINFO_LOOP_PHI_EVOLUTION_PART, STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE, wi::to_wide(), TREE_CODE, TREE_INT_CST_LOW, TREE_TYPE, TYPE_PRECISION, vect_location, vect_step_op_mul, vect_step_op_neg, and vect_use_loop_mask_for_alignment_p().
Referenced by vect_can_advance_ivs_p().
◆ vect_create_cond_for_alias_checks()
| void vect_create_cond_for_alias_checks | ( | loop_vec_info | loop_vinfo, |
| tree * | cond_expr ) |
Function vect_create_cond_for_alias_checks.
Create a conditional expression that represents the run-time checks for
overlapping of address ranges represented by a list of data references
relations passed as input.
Input:
COND_EXPR - input conditional expression. New conditions will be chained
with logical AND operation. If it is NULL, then the function
is used to return the number of alias checks.
LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
to be checked.
Output:
COND_EXPR - conditional expression.
The returned COND_EXPR is the conditional expression to be used in the if
statement that controls which version of the loop gets executed at runtime.
References create_runtime_alias_checks(), dump_enabled_p(), dump_printf_loc(), LOOP_VINFO_COMP_ALIAS_DDRS, LOOP_VINFO_LOOP, MSG_NOTE, and vect_location.
Referenced by vect_loop_versioning().
◆ vect_create_cond_for_align_checks()
|
static |
Function vect_create_cond_for_align_checks.
Create a conditional expression that represents the alignment checks for
all of data references (array element references) whose alignment must be
checked at runtime.
Input:
COND_EXPR - input conditional expression. New conditions will be chained
with logical AND operation.
LOOP_VINFO - two fields of the loop information are used.
LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
Output:
COND_EXPR_STMT_LIST - statements needed to construct the conditional
expression.
The returned value is the conditional expression to be used in the if
statement that controls which version of the loop gets executed at runtime.
The algorithm makes two assumptions:
1) The number of bytes "n" in a vector is a power of 2.
2) An address "a" is aligned if a%n is zero and that this
test can be done as a&(n-1) == 0. For example, for 16
byte vectors the test is a&0xf == 0.
References boolean_type_node, build_int_cst(), chain_cond_expr(), DR_STEP, fold_build2, FOR_EACH_VEC_ELT, gcc_assert, gimple_build_assign(), gimple_seq_add_seq(), gimple_seq_add_stmt(), i, LOOP_VINFO_MAY_MISALIGN_STMTS, LOOP_VINFO_PTR_MASK, make_temp_ssa_name(), NULL, NULL_TREE, offset, ptr_type_node, signed_type_for(), size_int, size_zero_node, STMT_VINFO_DATA_REF, STMT_VINFO_VECTYPE, tree_int_cst_compare(), TREE_INT_CST_LOW, TREE_TYPE, TYPE_SIZE_UNIT, TYPE_VECTOR_SUBPARTS(), and vect_create_addr_base_for_vector_ref().
Referenced by vect_loop_versioning().
◆ vect_create_cond_for_lower_bounds()
|
static |
Create an expression that is true when all lower-bound conditions for the vectorized loop are met. Chain this condition with *COND_EXPR.
References boolean_type_node, build_int_cstu(), chain_cond_expr(), fold_build2, fold_convert, i, LOOP_VINFO_LOWER_BOUNDS, TREE_TYPE, and unsigned_type_for().
Referenced by vect_loop_versioning().
◆ vect_create_cond_for_niters_checks()
|
static |
Function vect_create_cond_for_niters_checks.
Create a conditional expression that represents the run-time checks for
loop's niter. The loop is guaranteed to terminate if the run-time
checks hold.
Input:
COND_EXPR - input conditional expression. New conditions will be chained
with logical AND operation. If it is NULL, then the function
is used to return the number of alias checks.
LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
to be checked.
Output:
COND_EXPR - conditional expression.
The returned COND_EXPR is the conditional expression to be used in the
if statement that controls which version of the loop gets executed at
runtime.
References boolean_type_node, fold_build2, and LOOP_VINFO_NITERS_ASSUMPTIONS.
Referenced by vect_loop_versioning().
◆ vect_create_cond_for_unequal_addrs()
|
static |
If LOOP_VINFO_CHECK_UNEQUAL_ADDRS contains <A1, B1>, ..., <An, Bn>, create a tree representation of: (&A1 != &B1) && ... && (&An != &Bn). Set *COND_EXPR to a tree that is true when both the original *COND_EXPR and this new condition are true. Treat a null *COND_EXPR as "true".
References boolean_type_node, build_fold_addr_expr, chain_cond_expr(), fold_build2, FOR_EACH_VEC_ELT, i, and LOOP_VINFO_CHECK_UNEQUAL_ADDRS.
Referenced by vect_loop_versioning().
◆ vect_do_peeling()
| class loop * vect_do_peeling | ( | loop_vec_info | loop_vinfo, |
| tree | niters, | ||
| tree | nitersm1, | ||
| tree * | niters_vector, | ||
| tree * | step_vector, | ||
| tree * | niters_vector_mult_vf_var, | ||
| int | th, | ||
| bool | check_profitability, | ||
| bool | niters_no_overflow, | ||
| tree * | advance ) |
Function vect_do_peeling.
Input:
- LOOP_VINFO: Represent a loop to be vectorized, which looks like:
preheader:
LOOP:
header_bb:
loop_body
if (exit_loop_cond) goto exit_bb
else goto header_bb
exit_bb:
- NITERS: The number of iterations of the loop.
- NITERSM1: The number of iterations of the loop's latch.
- NITERS_NO_OVERFLOW: No overflow in computing NITERS.
- TH, CHECK_PROFITABILITY: Threshold of niters to vectorize loop if
CHECK_PROFITABILITY is true.
Output:
- *NITERS_VECTOR and *STEP_VECTOR describe how the main loop should
iterate after vectorization; see vect_set_loop_condition for details.
- *NITERS_VECTOR_MULT_VF_VAR is either null or an SSA name that
should be set to the number of scalar iterations handled by the
vector loop. The SSA name is only used on exit from the loop.
This function peels prolog and epilog from the loop, adds guards skipping
PROLOG and EPILOG for various conditions. As a result, the changed CFG
would look like:
guard_bb_1:
if (prefer_scalar_loop) goto merge_bb_1
else goto guard_bb_2
guard_bb_2:
if (skip_prolog) goto merge_bb_2
else goto prolog_preheader
prolog_preheader:
PROLOG:
prolog_header_bb:
prolog_body
if (exit_prolog_cond) goto prolog_exit_bb
else goto prolog_header_bb
prolog_exit_bb:
merge_bb_2:
vector_preheader:
VECTOR LOOP:
vector_header_bb:
vector_body
if (exit_vector_cond) goto vector_exit_bb
else goto vector_header_bb
vector_exit_bb:
guard_bb_3:
if (skip_epilog) goto merge_bb_3
else goto epilog_preheader
merge_bb_1:
epilog_preheader:
EPILOG:
epilog_header_bb:
epilog_body
if (exit_epilog_cond) goto merge_bb_3
else goto epilog_header_bb
merge_bb_3:
Note this function peels prolog and epilog only if it's necessary,
as well as guards.
This function returns the epilogue loop if a decision was made to vectorize
it, otherwise NULL.
The analysis resulting in this epilogue loop's loop_vec_info was performed
in the same vect_analyze_loop call as the main loop's. At that time
vect_analyze_loop constructs a list of accepted loop_vec_info's for lower
vectorization factors than the main loop. This list is chained in the
loop's loop_vec_info in the 'epilogue_vinfo' member. When we decide to
vectorize the epilogue loop for a lower vectorization factor, the
loop_vec_info in epilogue_vinfo is updated and linked to the epilogue loop.
This is later used to vectorize the epilogue.
The reason the loop_vec_info needs updating is that it was
constructed based on the original main loop, and the epilogue loop is a
copy of this loop, so all links pointing to statements in the original loop
need updating. Furthermore, these loop_vec_infos share the
data_reference's records, which will also need to be updated.
TODO: Guard for prefer_scalar_loop should be emitted along with
versioning conditions if loop versioning is needed.
References add_phi_arg(), adjust_vec, adjust_vec_debug_stmts(), advance(), profile_probability::always(), loop::any_estimate, loop::any_likely_upper_bound, loop::any_upper_bound, profile_probability::apply_scale(), loop::aux, boolean_type_node, build_int_cst(), build_one_cst(), build_zero_cst(), CDI_DOMINATORS, cfun, basic_block_def::count, create_phi_node(), DEF_FROM_PTR, delete_update_ssa(), DR_TARGET_ALIGNMENT, EDGE_PRED, _loop_vec_info::epilogue_vinfo, first_dom_son(), flow_bb_inside_loop_p(), flow_loop_nested_p(), fold_build2, FOR_EACH_IMM_USE_STMT, FOR_EACH_SSA_DEF_OPERAND, loop::force_vectorize, free(), free_original_copy_tables(), gcc_assert, gcc_checking_assert, get_bb_original(), get_immediate_dominator(), get_loop_body(), get_loop_copy(), get_loop_exit_edges(), gimple_bb(), gimple_build_assign(), gimple_build_nop(), gimple_debug_bind_p(), gimple_debug_bind_reset_value(), gimple_phi_arg_def_from_edge(), gimple_phi_result(), gsi_end_p(), gsi_for_stmt(), gsi_insert_after(), gsi_insert_before(), gsi_last_bb(), GSI_NEW_STMT, gsi_next(), gsi_start_bb(), gsi_start_phis(), gsi_stmt(), profile_probability::guessed_always(), i, initialize_original_copy_tables(), profile_probability::initialized_p(), integer_onep(), profile_probability::invert(), poly_int< N, C >::is_constant(), iterate_fix_dominators(), LOOP_C_INFINITE, loop_constraint_clear(), loop_preheader_edge(), LOOP_REQUIRES_VERSIONING, LOOP_VINFO_BBS, LOOP_VINFO_EARLY_BREAKS, LOOP_VINFO_EARLY_BREAKS_VECT_PEELED, LOOP_VINFO_EPILOGUE_IV_EXIT, LOOP_VINFO_INT_NITERS, LOOP_VINFO_IV_EXIT, LOOP_VINFO_LOOP, LOOP_VINFO_NITERS, LOOP_VINFO_NITERS_KNOWN_P, LOOP_VINFO_NITERSM1, LOOP_VINFO_PEELING_FOR_ALIGNMENT, LOOP_VINFO_PEELING_FOR_GAPS, LOOP_VINFO_PEELING_FOR_NITER, LOOP_VINFO_SCALAR_IV_EXIT, LOOP_VINFO_SCALAR_LOOP, LOOP_VINFO_UNALIGNED_DR, LOOP_VINFO_USING_PARTIAL_VECTORS_P, LOOP_VINFO_VECT_FACTOR, _loop_vec_info::main_loop_edge, make_ssa_name(), MAY_HAVE_DEBUG_BIND_STMTS, loop::nb_iterations_estimate, loop::nb_iterations_likely_upper_bound, loop::nb_iterations_upper_bound, need_ssa_update_p(), next_dom_son(), profile_count::nonzero_p(), NULL, NULL_TREE, loop::num_nodes, PHI_RESULT, profile_count::probability_in(), queue, record_niter_bound(), reset_original_copy_tables(), scale_bbs_frequencies(), scale_loop_profile(), scev_reset(), set_immediate_dominator(), set_range_info(), single_pred_edge(), single_pred_p(), single_succ_edge(), _loop_vec_info::skip_main_loop_edge, _loop_vec_info::skip_this_loop_edge, slpeel_add_loop_guard(), slpeel_can_duplicate_loop_p(), slpeel_tree_duplicate_loop_to_edge_cfg(), slpeel_update_phi_nodes_for_guard1(), split_edge(), SSA_NAME_DEF_STMT, SSA_OP_DEF, poly_int< N, C >::to_constant(), wi::to_wide(), TREE_CODE, TREE_TYPE, TYPE_MAX_VALUE, ui, UNKNOWN_LOCATION, update_stmt(), vect_build_loop_niters(), vect_can_advance_ivs_p(), vect_determine_partial_vectors_and_peeling(), vect_gen_prolog_loop_niters(), vect_gen_scalar_loop_niters(), vect_gen_vector_loop_niters(), vect_gen_vector_loop_niters_mult_vf(), vect_set_loop_condition(), vect_update_inits_of_drs(), vect_update_ivs_after_vectorizer(), vect_use_loop_mask_for_alignment_p(), vect_vf_for_cost(), and virtual_operand_p().
Referenced by vect_transform_loop().
◆ vect_gen_prolog_loop_niters()
|
static |
Function vect_gen_prolog_loop_niters Generate the number of iterations which should be peeled as prolog for the loop represented by LOOP_VINFO. It is calculated as the misalignment of DR - the data reference recorded in LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of this loop, the data reference DR will refer to an aligned location. The following computation is generated: If the misalignment of DR is known at compile time: addr_mis = int mis = DR_MISALIGNMENT (dr); Else, compute address misalignment in bytes: addr_mis = addr & (target_align - 1) prolog_niters = ((VF - addr_mis/elem_size)&(VF-1))/step (elem_size = element type size; an element is the scalar element whose type is the inner type of the vectype) The computations will be emitted at the end of BB. We also compute and store upper bound (included) of the result in BOUND. When the step of the data-ref in the loop is not 1 (as in interleaved data and SLP), the number of iterations of the prolog must be divided by the step (which is equal to the size of interleaved group). The above formulas assume that VF == number of elements in the vector. This may not hold when there are multiple-types in the loop. In this case, for some data-references in the loop the VF does not represent the number of elements that fit in the vector. Therefore, instead of VF we use TYPE_VECTOR_SUBPARTS.
References build_int_cst(), create_tmp_var, dr_info::dr, DR_STEP, DR_TARGET_ALIGNMENT, dump_enabled_p(), dump_printf_loc(), fold_build2, fold_convert, force_gimple_operand(), gcc_assert, get_misalign_in_elems(), gimple_seq_add_seq(), gsi_end_p(), gsi_insert_seq_after(), gsi_insert_seq_before(), gsi_last_bb(), GSI_SAME_STMT, int_cst_value(), poly_int< N, C >::is_constant(), LOOP_VINFO_NITERS, LOOP_VINFO_PEELING_FOR_ALIGNMENT, LOOP_VINFO_UNALIGNED_DR, MSG_NOTE, NULL, single_succ_p(), size_zero_node, STMT_VINFO_VECTYPE, target_align(), tree_int_cst_compare(), TREE_TYPE, TYPE_SIZE_UNIT, and vect_location.
Referenced by vect_do_peeling().
◆ vect_gen_scalar_loop_niters()
|
static |
Calculate the number of iterations above which vectorized loop will be preferred than scalar loop. NITERS_PROLOG is the number of iterations of prolog loop. If it's integer const, the integer number is also passed in INT_NITERS_PROLOG. BOUND_PROLOG is the upper bound (inclusive) of the number of iterations of the prolog loop. BOUND_EPILOG is the corresponding value for the epilog loop. If CHECK_PROFITABILITY is true, TH is the threshold below which the scalar (rather than vectorized) loop will be executed. This function stores the upper bound (inclusive) of the result in BOUND_SCALAR.
References build_int_cst(), fold_build2, known_ge, maybe_gt, and TREE_TYPE.
Referenced by vect_do_peeling().
◆ vect_gen_vector_loop_niters()
| void vect_gen_vector_loop_niters | ( | loop_vec_info | loop_vinfo, |
| tree | niters, | ||
| tree * | niters_vector_ptr, | ||
| tree * | step_vector_ptr, | ||
| bool | niters_no_overflow ) |
NITERS is the number of times that the original scalar loop executes
after peeling. Work out the maximum number of iterations N that can
be handled by the vectorized form of the loop and then either:
a) set *STEP_VECTOR_PTR to the vectorization factor and generate:
niters_vector = N
b) set *STEP_VECTOR_PTR to one and generate:
niters_vector = N / vf
In both cases, store niters_vector in *NITERS_VECTOR_PTR and add
any new statements on the loop preheader edge. NITERS_NO_OVERFLOW
is true if NITERS doesn't overflow (i.e. if NITERS is always nonzero).
References build_int_cst(), build_one_cst(), create_tmp_var, exact_log2(), fold_build2, force_gimple_operand(), gsi_insert_seq_on_edge_immediate(), poly_int< N, C >::is_constant(), is_gimple_val(), loop_preheader_edge(), LOOP_VINFO_LOOP, LOOP_VINFO_PEELING_FOR_GAPS, LOOP_VINFO_USING_PARTIAL_VECTORS_P, LOOP_VINFO_VECT_FACTOR, wi::max_value(), NULL, NULL_TREE, wi::one(), wi::rshift(), set_range_info(), TREE_TYPE, TYPE_PRECISION, and TYPE_SIGN.
Referenced by vect_do_peeling(), and vect_transform_loop().
◆ vect_gen_vector_loop_niters_mult_vf()
|
static |
Given NITERS_VECTOR which is the number of iterations for vectorized loop specified by LOOP_VINFO after vectorization, compute the number of iterations before vectorization (niters_vector * vf) and store it to NITERS_VECTOR_MULT_VF_PTR.
References build_int_cst(), create_tmp_var, exact_log2(), fold_build2, force_gimple_operand(), gcc_assert, gsi_insert_seq_before(), GSI_SAME_STMT, gsi_start_bb(), is_gimple_val(), LOOP_VINFO_EARLY_BREAKS_VECT_PEELED, LOOP_VINFO_IV_EXIT, LOOP_VINFO_VECT_FACTOR, NULL, and TREE_TYPE.
Referenced by vect_do_peeling().
◆ vect_get_main_loop_result()
| tree vect_get_main_loop_result | ( | loop_vec_info | loop_vinfo, |
| tree | main_loop_value, | ||
| tree | skip_value ) |
LOOP_VINFO is an epilogue loop whose corresponding main loop can be skipped. Return a value that equals: - MAIN_LOOP_VALUE when LOOP_VINFO is entered from the main loop and - SKIP_VALUE when the main loop is skipped.
References add_phi_arg(), create_phi_node(), gcc_assert, _loop_vec_info::main_loop_edge, make_ssa_name(), _loop_vec_info::skip_main_loop_edge, TREE_TYPE, and UNKNOWN_LOCATION.
Referenced by vect_transform_cycle_phi().
◆ vect_iv_increment_position()
| void vect_iv_increment_position | ( | edge | loop_exit, |
| gimple_stmt_iterator * | bsi, | ||
| bool * | insert_after ) |
Stores the standard position for induction variable increment in belonging to LOOP_EXIT (just before the exit condition of the given exit to BSI. INSERT_AFTER is set to true if the increment should be inserted after *BSI.
References gsi_last_bb().
Referenced by vect_create_epilog_for_reduction(), vect_set_loop_condition_normal(), vect_set_loop_condition_partial_vectors_avx512(), and vect_set_loop_controls_directly().
◆ vect_loop_versioning()
| class loop * vect_loop_versioning | ( | loop_vec_info | loop_vinfo, |
| gimple * | loop_vectorized_call ) |
Function vect_loop_versioning. If the loop has data references that may or may not be aligned or/and has data reference relations whose independence was not proven then two versions of the loop need to be generated, one which is vectorized and one which isn't. A test is then generated to control which of the loops is executed. The test checks for the alignment of all of the data references that may or may not be aligned. An additional sequence of runtime tests is generated for each pairs of DDRs whose independence was not proven. The vectorized version of loop is executed only if both alias and alignment tests are passed. The test generated to check which version of loop is executed is modified to also check for profitability as indicated by the cost model threshold TH. The versioning precondition(s) are placed in *COND_EXPR and *COND_EXPR_STMT_LIST.
References add_phi_arg(), adjust_phi_and_debug_stmts(), profile_probability::always(), profile_count::apply_probability(), as_a(), bb_loop_depth(), boolean_false_node, boolean_true_node, boolean_type_node, build_int_cst(), build_zero_cst(), CDI_DOMINATORS, copy_ssa_name(), basic_block_def::count, create_phi_node(), dom_info_available_p(), dominated_by_p(), dump_enabled_p(), dump_printf_loc(), EDGE_COMPLEX, EDGE_COUNT, EDGE_PRED, EDGE_SUCC, expr, EXPR_P, extract_true_false_edges_from_block(), first_dom_son(), flow_bb_inside_loop_p(), flow_loop_nested_p(), flush_pending_stmts(), fold_build2, fold_loop_internal_call(), FOR_EACH_SSA_USE_OPERAND, force_gimple_operand_1(), free_original_copy_tables(), gcc_assert, get_current_def(), dump_user_location_t::get_location_t(), get_loop_copy(), gimple_bb(), gimple_build_cond(), gimple_call_lhs(), gimple_cond_set_condition_from_tree(), gimple_phi_arg_location_from_edge(), gimple_seq_add_seq(), gsi_end_p(), gsi_for_stmt(), gsi_insert_after(), gsi_insert_seq_before(), gsi_last_bb(), GSI_NEW_STMT, gsi_next(), GSI_SAME_STMT, gsi_start(), gsi_start_phis(), gsi_stmt(), loop::header, initialize_original_copy_tables(), loop::inner, profile_probability::invert(), is_gimple_condexpr_for_cond(), is_gimple_val(), profile_probability::likely(), LOCATION_LOCUS, LOOP_C_FINITE, loop_constraint_set_p(), loop_outer(), loop_preheader_edge(), LOOP_REQUIRES_VERSIONING_FOR_ALIAS, LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT, LOOP_REQUIRES_VERSIONING_FOR_NITERS, LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND, loop_version(), LOOP_VINFO_COST_MODEL_THRESHOLD, LOOP_VINFO_IV_EXIT, LOOP_VINFO_LOOP, LOOP_VINFO_NITERSM1, LOOP_VINFO_SCALAR_LOOP, LOOP_VINFO_SCALAR_LOOP_SCALING, LOOP_VINFO_VERSIONING_THRESHOLD, make_edge(), MSG_NOTE, MSG_OPTIMIZED_LOCATIONS, MSG_PRIORITY_USER_FACING, loop::next, next_dom_son(), NULL, NULL_TREE, loop::num, outermost_invariant_loop_for_expr(), gphi_iterator::phi(), PHI_ARG_DEF_FROM_EDGE, PHI_RESULT, basic_block_def::preds, replace_uses_by(), vec_info::resync_stmt_addr(), scale_loop_frequencies(), set_immediate_dominator(), single_exit(), split_block(), split_edge(), profile_probability::sqrt(), SSA_NAME_DEF_STMT, SSA_OP_USE, ssa_redirect_edge(), superloop_at_depth(), TODO_update_ssa_no_phi, TREE_CODE, TREE_TYPE, UNKNOWN_LOCATION, unshare_expr(), update_ssa(), update_stmt(), USE_FROM_PTR, vect_apply_runtime_profitability_check_p(), vect_create_cond_for_alias_checks(), vect_create_cond_for_align_checks(), vect_create_cond_for_lower_bounds(), vect_create_cond_for_niters_checks(), vect_create_cond_for_unequal_addrs(), vect_free_loop_info_assumptions(), vect_location, and vect_loop_vectorized_call().
Referenced by vect_transform_loop().
◆ vect_maybe_permute_loop_masks()
|
static |
Try to use permutes to define the masks in DEST_RGM using the masks in SRC_RGM, given that the former has twice as many masks as the latter. Return true on success, adding any new statements to SEQ.
References build1(), rgroup_controls::controls, gcc_assert, gimple_build_assign(), gimple_seq_add_stmt(), i, insn_data, interleave_supported_p(), make_ssa_name(), rgroup_controls::max_nscalars_per_iter, insn_operand_data::mode, insn_data_d::operand, optab_handler(), rgroup_controls::type, TYPE_MODE, vect_gen_perm_mask_checked(), and vect_halve_mask_nunits().
Referenced by vect_set_loop_condition_partial_vectors().
◆ vect_prepare_for_masked_peels()
| void vect_prepare_for_masked_peels | ( | loop_vec_info | loop_vinfo | ) |
For the information recorded in LOOP_VINFO prepare the loop for peeling by masking. This involves calculating the number of iterations to be peeled and then aligning all memory references appropriately.
References build_int_cst(), dump_enabled_p(), dump_printf_loc(), fold_convert, force_gimple_operand(), gcc_assert, get_misalign_in_elems(), gimple_seq_add_seq(), gsi_insert_seq_on_edge_immediate(), loop_preheader_edge(), LOOP_VINFO_LOOP, LOOP_VINFO_MASK_SKIP_NITERS, LOOP_VINFO_NITERS, LOOP_VINFO_PEELING_FOR_ALIGNMENT, LOOP_VINFO_VECT_FACTOR, MSG_NOTE, NULL, NULL_TREE, TREE_TYPE, vect_location, vect_update_inits_of_drs(), and vect_use_loop_mask_for_alignment_p().
Referenced by vect_transform_loop().
◆ vect_set_loop_condition()
| void vect_set_loop_condition | ( | class loop * | loop, |
| edge | loop_e, | ||
| loop_vec_info | loop_vinfo, | ||
| tree | niters, | ||
| tree | step, | ||
| tree | final_iv, | ||
| bool | niters_maybe_zero ) |
If we're using fully-masked loops, make LOOP iterate: N == (NITERS - 1) / STEP + 1 times. When NITERS is zero, this is equivalent to making the loop execute (1 << M) / STEP times, where M is the precision of NITERS. NITERS_MAYBE_ZERO is true if this last case might occur. If we're not using fully-masked loops, make LOOP iterate: N == (NITERS - STEP) / STEP + 1 times, where NITERS is known to be outside the range [1, STEP - 1]. This is equivalent to making the loop execute NITERS / STEP times when NITERS is nonzero and (1 << M) / STEP times otherwise. NITERS_MAYBE_ZERO again indicates whether this last case might occur. If FINAL_IV is nonnull, it is an SSA name that should be set to N * STEP on exit from the loop. Assumption: the exit-condition of LOOP is the last stmt in the loop.
References dump_enabled_p(), dump_printf_loc(), get_loop_exit_condition(), gsi_for_stmt(), gsi_remove(), vec_info::lookup_stmt(), LOOP_VINFO_PARTIAL_VECTORS_STYLE, LOOP_VINFO_USING_PARTIAL_VECTORS_P, MSG_NOTE, vec_info::remove_stmt(), vect_location, vect_partial_vectors_avx512, vect_set_loop_condition_normal(), vect_set_loop_condition_partial_vectors(), and vect_set_loop_condition_partial_vectors_avx512().
Referenced by vect_do_peeling(), and vect_transform_loop().
◆ vect_set_loop_condition_normal()
|
static |
Like vect_set_loop_condition, but handle the case in which the vector loop handles exactly VF scalars per iteration.
References add_phi_arg(), build_all_ones_cst(), build_int_cst(), build_zero_cst(), copy_ssa_name(), create_iv(), create_phi_node(), fold_build2, force_gimple_operand(), force_gimple_operand_gsi(), gcc_assert, get_loop_exit_condition(), gimple_build(), gimple_build_assign(), gimple_build_cond(), gsi_after_labels(), gsi_for_stmt(), gsi_insert_before(), gsi_insert_seq_on_edge_immediate(), GSI_SAME_STMT, integer_onep(), integer_zerop(), loop_preheader_edge(), loop::nb_iterations, NULL, NULL_TREE, single_pred_p(), TREE_TYPE, UNKNOWN_LOCATION, and vect_iv_increment_position().
Referenced by vect_set_loop_condition().
◆ vect_set_loop_condition_partial_vectors()
|
static |
Set up the iteration condition and rgroup controls for LOOP, given that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is the number of iterations of the original scalar loop that should be handled by the vector loop. NITERS_MAYBE_ZERO and FINAL_IV are as for vect_set_loop_condition. Insert the branch-back condition before LOOP_COND_GSI and return the final gcond.
References add_header_seq(), add_preheader_seq(), build_int_cst(), build_minus_one_cst(), build_one_cst(), build_zero_cst(), rgroup_controls::controls, rgroup_controls::factor, fold_build2, FOR_EACH_VEC_ELT, gcc_assert, gimple_build(), gimple_build_assign(), gimple_build_cond(), gimple_convert(), gsi_insert_before(), gsi_insert_on_edge_immediate(), GSI_SAME_STMT, i, LOOP_VINFO_EARLY_BREAKS_VECT_PEELED, LOOP_VINFO_FULLY_MASKED_P, LOOP_VINFO_LENS, LOOP_VINFO_MASK_SKIP_NITERS, LOOP_VINFO_MASKS, LOOP_VINFO_RGROUP_COMPARE_TYPE, LOOP_VINFO_RGROUP_IV_TYPE, LOOP_VINFO_USING_DECREMENTING_IV_P, LOOP_VINFO_USING_SELECT_VL_P, LOOP_VINFO_VECT_FACTOR, rgroup_controls::max_nscalars_per_iter, loop::nb_iterations, NULL, NULL_TREE, TREE_TYPE, TYPE_PRECISION, TYPE_UNSIGNED, vect_adjust_loop_lens_control(), vect_maybe_permute_loop_masks(), vect_rgroup_iv_might_wrap_p(), and vect_set_loop_controls_directly().
Referenced by vect_set_loop_condition().
◆ vect_set_loop_condition_partial_vectors_avx512()
|
static |
Set up the iteration condition and rgroup controls for LOOP in AVX512 style, given that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is the number of iterations of the original scalar loop that should be handled by the vector loop. NITERS_MAYBE_ZERO and FINAL_IV are as for vect_set_loop_condition. Insert the branch-back condition before LOOP_COND_GSI and return the final gcond.
References add_preheader_seq(), tree_vector_builder::build(), build_int_cst(), build_minus_one_cst(), build_one_cst(), create_iv(), fold_build2, fold_convert, FOR_EACH_VEC_ELT_REVERSE, gcc_assert, GET_MODE_CLASS, gimple_build(), gimple_build_assign(), gimple_build_cond(), gimple_build_vector_from_val(), gimple_convert(), GSI_CONTINUE_LINKING, gsi_insert_before(), gsi_insert_on_edge_immediate(), GSI_SAME_STMT, i, known_ge, known_le, LOOP_VINFO_EARLY_BREAKS_VECT_PEELED, LOOP_VINFO_MASK_SKIP_NITERS, LOOP_VINFO_MASKS, LOOP_VINFO_RGROUP_IV_TYPE, LOOP_VINFO_VECT_FACTOR, loop::nb_iterations, NULL, NULL_TREE, poly_int_tree_p(), poly_int< N, C >::to_constant(), TREE_TYPE, lang_hooks_for_types::type_for_mode, TYPE_MODE, TYPE_PRECISION, TYPE_UNSIGNED, TYPE_VECTOR_SUBPARTS(), lang_hooks::types, UNKNOWN_LOCATION, vect_iv_increment_position(), and vect_set_loop_control().
Referenced by vect_set_loop_condition().
◆ vect_set_loop_control()
|
static |
Define one loop rgroup control CTRL from loop LOOP. INIT_CTRL is the value that the control should have during the first iteration and NEXT_CTRL is the value that it should have on subsequent iterations.
References add_phi_arg(), create_phi_node(), loop::header, loop_latch_edge(), loop_preheader_edge(), and UNKNOWN_LOCATION.
Referenced by vect_set_loop_condition_partial_vectors_avx512(), and vect_set_loop_controls_directly().
◆ vect_set_loop_controls_directly()
|
static |
Helper for vect_set_loop_condition_partial_vectors. Generate definitions for all the rgroup controls in RGC and return a control that is nonzero when the loop needs to iterate. Add any new preheader statements to PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond. RGC belongs to loop LOOP. The loop originally iterated NITERS times and has been vectorized according to LOOP_VINFO. If NITERS_SKIP is nonnull, the first iteration of the vectorized loop starts with NITERS_SKIP dummy iterations of the scalar loop before the real work starts. The mask elements for these dummy iterations must be 0, to ensure that the extra iterations do not have an effect. It is known that: NITERS * RGC->max_nscalars_per_iter * RGC->factor does not overflow. However, MIGHT_WRAP_P says whether an induction variable that starts at 0 and has step: VF * RGC->max_nscalars_per_iter * RGC->factor might overflow before hitting a value above: (NITERS + NITERS_SKIP) * RGC->max_nscalars_per_iter * RGC->factor This means that we cannot guarantee that such an induction variable would ever hit a value that produces a set of all-false masks or zero lengths for RGC. Note: the cost of the code generated by this function is modeled by vect_estimate_min_profitable_iters, so changes here may need corresponding changes there.
References rgroup_controls::bias_adjusted_ctrl, build_int_cst(), build_minus_one_cst(), rgroup_controls::controls, create_iv(), end(), rgroup_controls::factor, FOR_EACH_VEC_ELT_REVERSE, gcc_assert, gimple_build(), gimple_build_assign(), gimple_convert(), gimple_seq_add_seq(), gimple_seq_add_stmt(), gsi_insert_seq_before(), GSI_SAME_STMT, i, known_ge, known_le, LOOP_VINFO_FULLY_MASKED_P, LOOP_VINFO_IV_EXIT, LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS, LOOP_VINFO_RGROUP_COMPARE_TYPE, LOOP_VINFO_RGROUP_IV_TYPE, LOOP_VINFO_USING_DECREMENTING_IV_P, LOOP_VINFO_USING_SELECT_VL_P, LOOP_VINFO_VECT_FACTOR, make_ssa_name(), make_temp_ssa_name(), rgroup_controls::max_nscalars_per_iter, NULL, NULL_TREE, poly_int_tree_p(), TREE_TYPE, rgroup_controls::type, TYPE_VECTOR_SUBPARTS(), vect_gen_len(), vect_gen_while(), vect_gen_while_not(), vect_iv_increment_position(), and vect_set_loop_control().
Referenced by vect_set_loop_condition_partial_vectors().
◆ vect_update_init_of_dr()
|
static |
Function vect_update_init_of_dr If CODE is PLUS, the vector loop starts NITERS iterations after the scalar one, otherwise CODE is MINUS and the vector loop starts NITERS iterations before the scalar one (using masking to skip inactive elements). This function updates the information recorded in DR to account for the difference. Specifically, it updates the OFFSET field of DR_INFO.
References build_zero_cst(), dr_info::dr, DR_STEP, fold_build2, fold_convert, offset, and sizetype.
Referenced by vect_update_inits_of_drs().
◆ vect_update_inits_of_drs()
| void vect_update_inits_of_drs | ( | loop_vec_info | loop_vinfo, |
| tree | niters, | ||
| tree_code | code ) |
Function vect_update_inits_of_drs Apply vect_update_inits_of_dr to all accesses in LOOP_VINFO. CODE and NITERS are as for vect_update_inits_of_dr.
References DUMP_VECT_SCOPE, fold_convert, FOR_EACH_VEC_ELT, i, vec_info::lookup_dr(), LOOP_VINFO_DATAREFS, sizetype, STMT_VINFO_GATHER_SCATTER_P, STMT_VINFO_SIMD_LANE_ACCESS_P, TREE_TYPE, types_compatible_p(), and vect_update_init_of_dr().
Referenced by update_epilogue_loop_vinfo(), vect_do_peeling(), and vect_prepare_for_masked_peels().
◆ vect_update_ivs_after_vectorizer()
|
static |
Function vect_update_ivs_after_vectorizer.
"Advance" the induction variables of LOOP to the value they should take
after the execution of LOOP. This is currently necessary because the
vectorizer does not handle induction variables that are used after the
loop. Such a situation occurs when the last iterations of LOOP are
peeled, because:
1. We introduced new uses after LOOP for IVs that were not originally used
after LOOP: the IVs of LOOP are now used by an epilog loop.
2. LOOP is going to be vectorized; this means that it will iterate N/VF
times, whereas the loop IVs should be bumped N times.
Input:
- LOOP - a loop that is going to be vectorized. The last few iterations
of LOOP were peeled.
- NITERS - the number of iterations that LOOP executes (before it is
vectorized). i.e, the number of times the ivs should be bumped.
- UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
coming out from LOOP on which there are uses of the LOOP ivs
(this is the path from LOOP->exit to epilog_loop->preheader).
The new definitions of the ivs are placed in LOOP->exit.
The phi args associated with the edge UPDATE_E in the bb
UPDATE_E->dest are updated accordingly.
Assumption 1: Like the rest of the vectorizer, this function assumes
a single loop exit that has a single predecessor.
Assumption 2: The phi nodes in the LOOP header and in update_bb are
organized in the same order.
Assumption 3: The access function of the ivs is simple enough (see
vect_can_advance_ivs_p). This assumption will be relaxed in the future.
Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
coming out of LOOP on which the ivs of LOOP are used (this is the path
that leads to the epilog loop; other paths skip the epilog loop). This
path starts with the edge UPDATE_E, and its destination (denoted update_bb)
needs to have its phis updated.
References adjust_phi_and_debug_stmts(), create_tmp_var, dump_enabled_p(), dump_printf_loc(), fold_build2, fold_build_pointer_plus, fold_convert, force_gimple_operand(), gcc_assert, gimple_phi_result(), gsi_end_p(), gsi_insert_seq_after(), gsi_insert_seq_before(), gsi_last_bb(), gsi_next(), GSI_SAME_STMT, gsi_start_phis(), loop::header, init_expr(), iv_phi_p(), vec_info::lookup_stmt(), loop_preheader_edge(), LOOP_VINFO_IV_EXIT, LOOP_VINFO_LOOP, MSG_NOTE, NULL, gphi_iterator::phi(), PHI_ARG_DEF_FROM_EDGE, POINTER_TYPE_P, STMT_VINFO_LOOP_PHI_EVOLUTION_PART, STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE, tree_is_chrec(), TREE_TYPE, type(), unshare_expr(), vect_location, vect_peel_nonlinear_iv_init(), vect_step_op_add, and vect_step_op_neg.
Referenced by vect_do_peeling().
Variable Documentation
◆ adjust_vec
|
static |
A stack of values to be adjusted in debug stmts. We have to process them LIFO, so that the closest substitution applies. If we processed them FIFO, without the stack, we might substitute uses with a PHI DEF that would soon become non-dominant, and when we got to the suitable one, it wouldn't have anything to substitute any more.
Referenced by adjust_debug_stmts(), adjust_vec_debug_stmts(), and vect_do_peeling().
