|
GCC Middle and Back End API Reference
|
Data Structures | |
| struct | access |
| struct | assign_link |
| struct | uid_decl_hasher |
Typedefs | |
| typedef struct access * | access_p |
Enumerations | |
| enum | sra_mode { SRA_MODE_EARLY_IPA, SRA_MODE_EARLY_INTRA, SRA_MODE_INTRA } |
| enum | unscalarized_data_handling { SRA_UDH_NONE, SRA_UDH_RIGHT, SRA_UDH_LEFT } |
| enum | assignment_mod_result { SRA_AM_NONE, SRA_AM_MODIFIED, SRA_AM_REMOVED } |
| enum | ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES, MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES } |
Variables | |
| static enum sra_mode | sra_mode |
| static alloc_pool | access_pool |
| static alloc_pool | link_pool |
| static struct pointer_map_t * | base_access_vec |
| static bitmap | candidate_bitmap |
| static hash_table < uid_decl_hasher > | candidates |
| static bitmap | should_scalarize_away_bitmap |
| static bitmap | cannot_scalarize_away_bitmap |
| static struct obstack | name_obstack |
| static struct access * | work_queue_head |
| static int | func_param_count |
| static bool | encountered_apply_args |
| static bool | encountered_recursive_call |
| static bool | encountered_unchangable_recursive_call |
| static HOST_WIDE_INT * | bb_dereferences |
| static bitmap | final_bbs |
| static struct access | no_accesses_representant |
| struct { | |
| int replacements | |
| int exprs | |
| int subtree_copies | |
| int subreplacements | |
| int deleted | |
| int separate_lhs_rhs_handling | |
| int deleted_unused_parameters | |
| int scalar_by_ref_to_by_val | |
| int aggregate_params_reduced | |
| int param_reductions_created | |
| } | sra_stats |
| enum ipa_splicing_result |
| enum sra_mode |
@verbatim
Scalar Replacement of Aggregates (SRA) converts some structure references into scalar references, exposing them to the scalar optimizers. Copyright (C) 2008-2013 Free Software Foundation, Inc. Contributed by Martin Jambor mjambor@suse.cz
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/.
This file implements Scalar Reduction of Aggregates (SRA). SRA is run
twice, once in the early stages of compilation (early SRA) and once in the
late stages (late SRA). The aim of both is to turn references to scalar
parts of aggregates into uses of independent scalar variables.
The two passes are nearly identical, the only difference is that early SRA
does not scalarize unions which are used as the result in a GIMPLE_RETURN
statement because together with inlining this can lead to weird type
conversions.
Both passes operate in four stages:
1. The declarations that have properties which make them candidates for
scalarization are identified in function find_var_candidates(). The
candidates are stored in candidate_bitmap.
2. The function body is scanned. In the process, declarations which are
used in a manner that prevent their scalarization are removed from the
candidate bitmap. More importantly, for every access into an aggregate,
an access structure (struct access) is created by create_access() and
stored in a vector associated with the aggregate. Among other
information, the aggregate declaration, the offset and size of the access
and its type are stored in the structure.
On a related note, assign_link structures are created for every assign
statement between candidate aggregates and attached to the related
accesses.
3. The vectors of accesses are analyzed. They are first sorted according to
their offset and size and then scanned for partially overlapping accesses
(i.e. those which overlap but one is not entirely within another). Such
an access disqualifies the whole aggregate from being scalarized.
If there is no such inhibiting overlap, a representative access structure
is chosen for every unique combination of offset and size. Afterwards,
the pass builds a set of trees from these structures, in which children
of an access are within their parent (in terms of offset and size).
Then accesses are propagated whenever possible (i.e. in cases when it
does not create a partially overlapping access) across assign_links from
the right hand side to the left hand side.
Then the set of trees for each declaration is traversed again and those
accesses which should be replaced by a scalar are identified.
4. The function is traversed again, and for every reference into an
aggregate that has some component which is about to be scalarized,
statements are amended and new statements are created as necessary.
Finally, if a parameter got scalarized, the scalar replacements are
initialized with values from respective parameter aggregates. Enumeration of all aggregate reductions we can do.
|
inlinestatic |
Return true iff ACC is non-NULL and has subaccesses.
References get_first_repr_for_decl(), access::next_grp, access::offset, and access::size.
Referenced by contains_vce_or_bfcref_p().
|
static |
Return true iff ACC is (partly) covered by at least one replacement.
|
static |
Return true iff this ACCESS precludes IPA-SRA of the parameter it is associated with. REQ_ALIGN is the minimum required alignment.
Avoid issues such as the second simple testcase in PR 42025. The problem
is incompatible assign in a call statement (and possibly even in asm
statements). This can be relaxed by using a new temporary but only for
non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
intraprocedural SRA we deal with this by keeping the old aggregate around,
something we cannot do in IPA-SRA.)
|
static |
Add ACCESS to the work queue (which is actually a stack).
References dump_file, and print_generic_expr().
Referenced by child_would_conflict_in_lacc().
|
static |
Add LINK to the linked list of assign links of RACC.
Referenced by build_access_from_expr_1().
|
static |
@verbatim
Analyze the subtree of accesses rooted in ROOT, scheduling replacements when both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all sorts of access flags appropriately along the way, notably always set grp_read and grp_assign_read according to MARK_READ and grp_write when MARK_WRITE is true.
Creating a replacement for a scalar access is considered beneficial if its grp_hint is set (this means we are either attempting total scalarization or there is more than one direct read access) or according to the following table:
Access written to through a scalar type (once or more times) | | Written to in an assignment statement | | | | Access read as scalar once | | | | | | Read in an assignment statement | | | |
0 0 0 0 No access for the scalar 0 0 0 1 No access for the scalar 0 0 1 0 No Single read - won't help 0 0 1 1 No The same case 0 1 0 0 No access for the scalar 0 1 0 1 No access for the scalar 0 1 1 0 Yes s = *g; return s.i; 0 1 1 1 Yes The same case as above 1 0 0 0 No Won't help 1 0 0 1 Yes s.i = 1; *g = s; 1 0 1 0 Yes s.i = 5; g = s.i; 1 0 1 1 Yes The same case as above 1 1 0 0 No Won't help. 1 1 0 1 Yes s.i = 1; *g = s; 1 1 1 0 Yes s = *g; return s.i; 1 1 1 1 Yes Any of the above yeses
Always create access replacements that cover the whole access.
For integral types this means the precision has to match.
Avoid assumptions based on the integral type kind, too. But leave bitfield accesses alone.
References access::offset, and access::size.
Referenced by get_access_replacement().
|
static |
Analyze all access trees linked by next_grp by the means of analyze_access_subtree.
|
static |
Analyze the collected accesses and produce a plan what to do with the parameters in the form of adjustments, NULL meaning nothing.
If there are any parameters passed by reference which are not modified
directly, we need to check whether they can be modified indirectly.
|
static |
Go through all accesses collected throughout the (intraprocedural) analysis stage, exclude overlapping ones, identify representatives and build trees out of them, making decisions about scalarization on the way. Return true iff there are any to-be-scalarized variables after this stage.
|
static |
Determine what (parts of) parameters passed by reference that are not assigned to are not certainly dereferenced in this function and thus the dereferencing cannot be safely moved to the caller without potentially introducing a segfault. Mark such REPRESENTATIVES as grp_not_necessarilly_dereferenced. The dereferenced maximum "distance," i.e. the offset + size of the accessed part is calculated rather than simple booleans are calculated for each pointer parameter to handle cases when only a fraction of the whole aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for an example). The maximum dereference distances for each pointer parameter and BB are already stored in bb_dereference. This routine simply propagates these values upwards by propagate_dereference_distances and then compares the distances of individual parameters in the ENTRY BB to the equivalent distances of each representative of a (fraction of a) parameter.
References access::base, dump_access(), dump_file, access::next_grp, access::non_addressable, print_generic_expr(), and tree_low_cst().
|
static |
Analyze what representatives (in linked lists accessible from REPRESENTATIVES) can be modified by side effects of statements in the current function.
All accesses are read ones, otherwise grp_maybe_modified would
be trivially set.
Callback of walk_stmt_load_store_addr_ops visit_addr used to determine GIMPLE_ASM operands with memory constrains which cannot be scalarized.
|
static |
Scan expression EXPR and create access structures for all accesses to candidates for scalarization. Return true if any access has been inserted. STMT must be the statement from which the expression is taken, WRITE must be true if the expression is a store and false otherwise.
This means the aggregate is accesses as a whole in a way other than an
assign statement and thus cannot be removed even if we had a scalar
replacement for everything.
References disqualify_candidate(), and get_base_address().
Referenced by disqualify_ops_if_throwing_stmt().
|
staticread |
Scan expression EXPR and create access structures for all accesses to candidates for scalarization. Return the created access or NULL if none is created.
We need to dive through V_C_Es in order to get the size of its parameter
and not the result type. Ada produces such statements. We are also
capable of handling the topmost V_C_E but not any of those buried in other
handled components. fall through
References add_link_to_rhs(), access::base, bitmap_set_bit(), disqualify_ops_if_throwing_stmt(), gimple_assign_lhs(), gimple_assign_rhs1(), gimple_assign_single_p(), gimple_clobber_p(), gimple_has_volatile_ops(), access::grp_assignment_read, access::grp_assignment_write, access::grp_unscalarizable_region, is_gimple_reg_type(), assign_link::lacc, memset(), pool_alloc(), assign_link::racc, access::size, SRA_MODE_EARLY_INTRA, SRA_MODE_INTRA, access::type, and useless_type_conversion_p().
|
static |
Build a subtree of accesses rooted in *ACCESS, and move the pointer in the linked list along the way. Stop when *ACCESS is NULL or the access pointed to it is not "within" the root. Return false iff some accesses partially overlap.
|
static |
Build a tree of access representatives, ACCESS is the pointer to the first one, others are linked in a list by the next_grp field. Return false iff some accesses partially overlap.
|
static |
Scan expressions occurring in STMT, create access structures for all accesses to candidates for scalarization and remove those candidates which occur in statements or expressions that prevent them from being split apart. Return true if any access has been inserted.
Scope clobbers don't influence scalarization.
|
static |
Attempt to build a memory reference that we could but into a gimple debug_bind statement. Similar to build_ref_for_model but punts if it has to create statements and return s NULL instead. This function also ignores alignment issues and so its results should never end up in non-debug statements.
|
static |
Construct a memory reference to a part of an aggregate BASE at the given OFFSET and of the same type as MODEL. In case this is a reference to a bit-field, the function will replicate the last component_ref of model's expr to access it. GSI and INSERT_AFTER have the same meaning as in build_ref_for_offset.
This access represents a bit-field.
Referenced by generate_subtree_copies(), sra_modify_constructor_assign(), and sra_modify_expr().
| tree build_ref_for_offset | ( | location_t | loc, |
| tree | base, | ||
| HOST_WIDE_INT | offset, | ||
| tree | exp_type, | ||
| gimple_stmt_iterator * | gsi, | ||
| bool | insert_after | ||
| ) |
Construct a MEM_REF that would reference a part of aggregate BASE of type EXP_TYPE at the given OFFSET. If BASE is something for which get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used to insert new statements either before or below the current one as specified by INSERT_AFTER. This function is not capable of handling bitfields. BASE must be either a declaration or a memory reference that has correct alignment ifformation embeded in it (e.g. a pre-existing one in SRA).
get_addr_base_and_unit_offset returns NULL for references with a variable
offset such as array[var_index].
|
static |
Construct a memory reference consisting of component_refs and array_refs to a part of an aggregate *RES (which is of type TYPE). The requested part should have type EXP_TYPE at be the given OFFSET. This function might not succeed, it returns true when it does and only then *RES points to something meaningful. This function should be used only to build expressions that we might need to present to user (e.g. in warnings). In all other situations, build_ref_for_model or build_ref_for_offset should be used instead.
|
inlinestatic |
Return true iff callsite CALL has at least as many actual arguments as there are formal parameters of the function currently processed by IPA-SRA.
For a candidate UID return the candidates decl.
References no_accesses_representant.
|
static |
Return true iff a potential new child of LACC at offset OFFSET and with size SIZE would conflict with an already existing one. If exactly such a child already exists in LACC, store a pointer to it in EXACT_MATCH.
References add_access_to_work_queue(), access::base, bitmap_bit_p(), access::first_link, access::group_representative, assign_link::lacc, assign_link::next, pop_access_from_work_queue(), and propagate_subaccesses_across_link().
|
static |
Helper of QSORT function. There are pointers to accesses in the array. An access is considered smaller than another if it has smaller offset or if the offsets are the same but is size is bigger.
Put any non-aggregate type before any aggregate type.
Put any complex or vector type before any other scalar type.
Put the integral type with the bigger precision first.
Put any integral type with non-full precision last.
Stabilize the sort.
We want the bigger accesses first, thus the opposite operator in the next
line:
|
static |
Create total_scalarization accesses for all scalar type fields in DECL that must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE must be the top-most VAR_DECL representing the variable, OFFSET must be the offset of DECL within BASE. REF must be the memory reference expression for the given decl.
Accesses for intraprocedural SRA can have their stmt NULL.
|
static |
Create total_scalarization accesses for all scalar type fields in VAR and for VAR a a whole. VAR must be of a RECORD_TYPE conforming to type_consists_of_records_p.
|
inlinestatic |
Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a bit-field field declaration somewhere in it.
References access_has_children_p(), access::base, access::first_child, force_gimple_operand_gsi(), generate_subtree_copies(), gimple_assign_rhs1(), GSI_SAME_STMT, and sra_stats.
|
inlinestatic |
Return true if REF has an VIEW_CONVERT_EXPR somewhere in it.
|
static |
Convert all callers of NODE to pass parameters as given in ADJUSTMENTS.
References dbg_cnt().
|
static |
Convert all callers of NODE.
|
staticread |
Create and insert access for EXPR. Return created access, or NULL if it is not possible.
|
staticread |
Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in the three fields. Also add it to the vector of accesses corresponding to the base. Finally, return the new access.
|
static |
Create a variable for the given ACCESS which determines the type, name and a few other properties. Return the variable declaration and store it also to ACCESS->replacement.
Get rid of any SSA_NAMEs embedded in debug_expr,
as DECL_DEBUG_EXPR isn't considered when looking for still
used SSA_NAMEs and thus they could be freed. All debug info
generation cares is whether something is constant or variable
and that get_ref_base_and_extent works properly on the
expression. It cannot handle accesses at a non-constant offset
though, so just give up in those cases. FALLTHRU
Referenced by expr_with_var_bounded_array_refs_p(), and load_assign_lhs_subreplacements().
|
staticread |
Create a new child access of PARENT, with all properties just like MODEL except for its offset and with its grp_write false and grp_read true. Return the new access or NULL if it cannot be created. Note that this access is created long after all splicing and sorting, it's not located in any access vector and is automatically a representative of its group.
|
static |
Decide whether parameters with representative accesses given by REPR should be reduced into components.
Taking the address of a non-addressable field is verboten.
Do not decompose a non-BLKmode param in a way that would
create BLKmode params. Especially for by-reference passing
(thus, pointer-type param) this is hardly worthwhile.
|
static |
Hook fed to pointer_map_traverse, deallocate stored vectors.
|
static |
Search the given tree for a declaration by skipping handled components and exclude it from the candidates.
References SRA_MODE_EARLY_INTRA, SRA_MODE_INTRA, stmt_can_throw_internal(), and stmt_ends_bb_p().
|
static |
Remove DECL from candidates for SRA and write REASON to the dump file if there is one.
Referenced by build_access_from_expr(), and mark_parm_dereference().
|
static |
Disqualify LHS and RHS for scalarization if STMT must end its basic block in modes in which it matters, return true iff they have been disqualified. RHS may be NULL, in that case ignore it. If we scalarize an aggregate in intra-SRA we may need to add statements after each statement. This is not possible if a statement unconditionally has to end the basic block.
References bitmap_set_bit(), build_access_from_expr(), gimple_return_retval(), gsi_end_p(), gsi_next(), gsi_start_bb(), gsi_stmt(), basic_block_def::index, and stmt_can_throw_external().
Referenced by build_access_from_expr_1().
|
static |
References access::first_child.
Referenced by analyze_caller_dereference_legality().
|
static |
Dump all access trees for a variable, given the pointer to the first root in ACCESS.
|
static |
Dump a subtree rooted in ACCESS to file F, indent by LEVEL.
References access::first_child, access::next_sibling, access::offset, and access::size.
|
static |
Dump a dereferences TABLE with heading STR to file F.
|
static |
Perform early intraprocedural SRA.
|
static |
Return true if expr contains some ARRAY_REFs into a variable bounded array.
References create_access_replacement(), access::grp_assignment_read, access::grp_scalar_read, access::grp_to_be_debug_replaced, and access::replacement_decl.
|
staticread |
Find an access with required OFFSET and SIZE in a subtree of accesses rooted in ACCESS. Return NULL if it cannot be found.
Referenced by sra_modify_expr().
|
static |
Identify candidates for reduction for IPA-SRA based on their type and mark them in candidate_bitmap. Note that these do not necessarily include parameter which are unused and thus can be removed. Return true iff any such candidate has been found.
|
static |
The very first phase of intraprocedural SRA. It marks in candidate_bitmap those with type which is suitable for scalarization.
|
static |
References count, current_function_decl, is_gimple_reg_type(), is_unused_scalar_param(), and is_va_list_type().
|
static |
Generate statements copying scalar replacements of accesses within a subtree into or out of AGG. ACCESS, all its children, siblings and their children are to be processed. AGG is an aggregate type expression (can be a declaration but does not have to be, it can for example also be a mem_ref or a series of handled components). TOP_OFFSET is the offset of the processed subtree which has to be subtracted from offsets of individual accesses to get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only replacements in the interval <start_offset, start_offset + chunk_size>, otherwise copy all. GSI is a statement iterator used to place the new statements. WRITE should be true when the statements should write from AGG to the replacement and false if vice versa. if INSERT_AFTER is true, new statements will be added after the current statement in GSI, they will be added before the statement otherwise.
References access::base, build_ref_for_model(), access::expr, access::first_child, force_gimple_operand_gsi(), get_access_for_expr(), get_access_replacement(), gimple_location(), gimple_set_location(), access::grp_partial_lhs, access::grp_to_be_debug_replaced, access::grp_to_be_replaced, gsi_insert_after(), gsi_insert_before(), GSI_NEW_STMT, GSI_SAME_STMT, gsi_stmt(), host_integerp(), HOST_WIDE_INT, access::offset, sra_stats, access::stmt, access::type, type(), and useless_type_conversion_p().
Referenced by contains_vce_or_bfcref_p().
|
staticread |
Search for an access representative for the given expression EXPR and return it or NULL if it cannot be found.
FIXME: This should not be necessary but Ada produces V_C_Es with a type of
a different size than the size of its argument and we need the latter
one.
Referenced by generate_subtree_copies().
|
inlinestatic |
Return ACCESS scalar replacement, create it if it does not exist yet.
References analyze_access_subtree(), access::grp_covered, access::grp_total_scalarization, access::grp_unscalarized_data, access::offset, and access::size.
Referenced by generate_subtree_copies(), and sra_modify_expr().
|
staticread |
Find the first adjustment for a particular parameter BASE in a vector of ADJUSTMENTS which is not a copy_param. Return NULL if there is no such adjustment.
Return a vector of pointers to accesses for the variable given in BASE or NULL if there is none.
|
staticread |
Return the first group representative for DECL or NULL if none exists.
References access::grp_queued, access::next_queued, and work_queue_head.
Referenced by access_has_children_p().
|
inlinestatic |
Return the index of BASE in PARMS. Abort if it is not found.
References gimple_assign_lhs(), gimple_call_lhs(), gimple_phi_result(), is_gimple_assign(), and is_gimple_call().
|
static |
Create and return a new suitable default definition SSA_NAME for RACC which is an access describing an uninitialized part of an aggregate that is being loaded.
|
static |
If a parameter replacement identified by ADJ does not yet exist in the form of declaration, create it and record it, otherwise return the previously created one.
|
static |
If T is an SSA_NAME, return NULL if it is not a default def or return its base variable if it is. Return T if it is not an SSA_NAME.
|
staticread |
Find an access representative for the variable BASE and given OFFSET and SIZE. Requires that access trees have already been built. Return NULL if it cannot be found.
References access::grp_queued, access::next_queued, and work_queue_head.
|
static |
Store all replacements in the access tree rooted in TOP_RACC either to their base aggregate if there are unscalarized data or directly to LHS of the statement that is pointed to by GSI otherwise.
Referenced by sra_modify_expr().
|
static |
If NODE has a caller, return true.
|
static |
Assign zero to all scalar replacements in an access subtree. ACCESS is the the root of the subtree to be processed. GSI is the statement iterator used for inserting statements which are added after the current statement if INSERT_AFTER is true or before it otherwise.
References access::offset, and tree_low_cst().
|
static |
Generate statements initializing scalar replacements of parts of function parameters.
References cfun, has_zero_uses(), is_gimple_reg(), and ssa_default_def().
|
static |
Perform early interprocedural SRA.
|
static |
Return if early ipa sra shall be performed.
|
static |
Traverse the function body and all modifications as described in ADJUSTMENTS. Return true iff the CFG has been changed.
Operands must be processed before the lhs.
|
static |
Return false the function is apparently unsuitable for IPA-SRA based on it's attributes, return true otherwise. NODE is the cgraph node of the current function.
|
static |
Return true iff PARM (which must be a parm_decl) is an unused scalar parameter.
References func_param_count, access::next_grp, and visited.
Referenced by gate_intra_sra().
|
inlinestatic |
Return true iff TYPE is stdarg va_list type.
Referenced by gate_intra_sra().
|
static |
Perform "late" intraprocedural SRA.
|
static |
Try to generate statements to load all sub-replacements in an access subtree formed by children of LACC from scalar replacements in the TOP_RACC subtree. If that is not possible, refresh the TOP_RACC base aggregate and load the accesses from it. LEFT_OFFSET is the offset of the left whole subtree being copied. NEW_GSI is stmt iterator used for statement insertions after the original assignment, OLD_GSI is used to insert statements before the assignment. *REFRESHED keeps the information whether we have needed to refresh replacements of the LHS and from which side of the assignments this takes place.
No suitable access on the right hand side, need to load from
the aggregate. See if we have to update it first...
References cfun, create_access_replacement(), get_or_create_ssa_default_def(), access::grp_to_be_debug_replaced, access::grp_to_be_replaced, and access::replacement_decl.
|
static |
Append a name of the declaration to the name obstack. A helper function for make_fancy_name.
|
static |
Create a human readable name for replacement variable of ACCESS.
References build_int_cst(), access::expr, get_addr_base_and_unit_offset(), HOST_WIDE_INT, int_const_binop(), and unshare_expr().
|
static |
Helper for make_fancy_name.
Arrays with only one element may not have a constant as their
index.
| gimple_opt_pass* make_pass_early_ipa_sra | ( | ) |
| gimple_opt_pass* make_pass_sra | ( | ) |
References access::grp_maybe_modified.
| gimple_opt_pass* make_pass_sra_early | ( | ) |
Callback of walk_aliased_vdefs, marks the access passed as DATA as maybe_modified.
|
static |
Mark a dereference of BASE of distance DIST in a basic block tht STMT belongs to, unless the BB has already been marked as a potentially final.
References disqualify_candidate().
|
static |
Return true if VAR is a candidate for SRA.
|
static |
Perform all the modification required in IPA-SRA for NODE to have parameters as given in ADJUSTMENTS. Return true iff the CFG has been changed.
|
inlinestatic |
Predicate to test the special value.
|
static |
Return false iff all callers have at least as many actual arguments as there are formal parameters in the current function.
|
static |
The "main" function of intraprocedural SRA passes. Runs the analysis and if it reveals there are components of some aggregates to be scalarized, it runs the required transformations.
|
staticread |
Pop an access from the work queue, and return it, assuming there is one.
Referenced by child_would_conflict_in_lacc().
|
static |
Propagate all subaccesses across assignment links.
|
static |
Propagate distances in bb_dereferences in the opposite direction than the control flow edges, in each step storing the maximum of the current value and the minimum of all successors. These steps are repeated until the table stabilizes. Note that BBs which might terminate the functions (according to final_bbs bitmap) never updated in this way.
References access::expr, get_object_alignment(), is_gimple_call(), access::stmt, and access::write.
|
static |
Propagate all subaccesses of RACC across an assignment link to LACC. Return true if any new subaccess was created. Additionally, if RACC is a scalar access but LACC is not, change the type of the latter, if possible.
Referenced by child_would_conflict_in_lacc().
|
static |
Scan immediate uses of a default definition SSA name of a parameter PARM and examine whether there are any direct or otherwise infeasible ones. If so, return true, otherwise return false. PARM must be a gimple register with a non-NULL default definition.
Valid uses include dereferences on the lhs and the rhs.
If the number of valid uses does not match the number of
uses in this stmt there is an unhandled use.
|
static |
Print message to dump file why a variable was rejected.
References access::grp_assignment_read, access::grp_assignment_write, access::grp_read, access::grp_scalar_read, access::grp_scalar_write, access::grp_write, is_gimple_reg_type(), access::type, and access::write.
Referenced by process_alt_operands().
|
static |
Move all link structures in their linked list in OLD_RACC to the linked list in NEW_RACC.
|
static |
If the statement STMT defines an SSA_NAME of a parameter which is to be removed because its value is not used, replace the SSA_NAME with a one relating to a created VAR_DECL together all of its uses and return true. ADJUSTMENTS is a pointer to an adjustments vector.
|
static |
Scan function and look for interesting expressions and create access structures for them. Return true iff any access is created.
|
staticread |
Sort all accesses for the given variable, check for partial overlaps and return NULL if there are any. If there are none, pick a representative for each combination of offset and size and create a linked list out of them. Return the pointer to the first representative and make sure it is the first one in the vector of accesses.
Sort by <OFFSET, SIZE>.
If there are both aggregate-type and scalar-type accesses with
this combination of size and offset, the comparison function
should have put the scalars first.
|
static |
Identify representatives of all accesses to all candidate parameters for IPA-SRA. Return result based on what representatives have been found.
|
staticread |
Sort collected accesses for parameter PARM, identify representatives for each accessed region and link them together. Return NULL if there are different but overlapping accesses, return the special ptr value meaning there are no accesses for this parameter if that is the case and return the first representative otherwise. Set *RO_GRP if there is a group of accesses with only read (i.e. no write) accesses.
Access is about to become group representative unless we find some
nasty overlap which would preclude us from breaking this parameter
apart. All or nothing law for parameters.
References UNMODIF_BY_REF_ACCESSES, and unmodified_by_ref_scalar_representative().
|
static |
Deallocate all general structures.
|
static |
Allocate necessary structures.
|
static |
If the statement pointed to by STMT_PTR contains any expressions that need to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any potential type incompatibilities (GSI is used to accommodate conversion statements and must point to the statement). Return true iff the statement was modified.
V_C_Es of constructors can cause trouble (PR 42714).
This can happen when an assignment in between two single field
structures is turned into an assignment in between two pointers to
scalars (PR 42237).
References current_function_decl.
|
static |
If the expression *EXPR should be replaced by a reduction of a parameter, do so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT specifies whether the function should care about type incompatibility the current and new expressions. If it is false, the function will leave incompatibility issues to the caller. Return true iff the expression was modified.
|
static |
Call gimple_debug_bind_reset_value on all debug statements describing gimple register parameters that are being removed or replaced.
All other users must have been removed by
ipa_sra_modify_function_body. Create a VAR_DECL for debug info purposes.
References cgraph_node::callers.
|
static |
Examine both sides of the assignment statement pointed to by STMT, replace them with a scalare replacement if there is one and generate copying of replacements if scalarized aggregates have been used in the assignment. GSI is used to hold generated statements for type conversions and subtree copying.
If we can avoid creating a VIEW_CONVERT_EXPR do so.
??? This should move to fold_stmt which we simply should
call after building a VIEW_CONVERT_EXPR here. From this point on, the function deals with assignments in between
aggregates when at least one has scalar reductions of some of its
components. There are three possible scenarios: Both the LHS and RHS have
to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
In the first case, we would like to load the LHS components from RHS
components whenever possible. If that is not possible, we would like to
read it directly from the RHS (after updating it by storing in it its own
components). If there are some necessary unscalarized data in the LHS,
those will be loaded by the original assignment too. If neither of these
cases happen, the original statement can be removed. Most of this is done
by load_assign_lhs_subreplacements.
In the second case, we would like to store all RHS scalarized components
directly into LHS and if they cover the aggregate completely, remove the
statement too. In the third case, we want the LHS components to be loaded
directly from the RHS (DSE will remove the original statement if it
becomes redundant).
This is a bit complex but manageable when types match and when unions do
not cause confusion in a way that we cannot really load a component of LHS
from the RHS or vice versa (the access representing this level can have
subaccesses that are accessible only through a different union field at a
higher level - different from the one used in the examined expression).
Unions are fun.
Therefore, I specially handle a fourth case, happening when there is a
specific type cast or it is impossible to locate a scalarized subaccess on
the other side of the expression. If that happens, I simply "refresh" the
RHS by storing in it is scalarized components leave the original statement
there to do the copying and then load the scalar replacements of the LHS.
This is what the first branch does. This gimplification must be done after generate_subtree_copies,
lest we insert the subtree copies in the middle of the gimplified
sequence. When an access represents an unscalarizable region, it usually
represents accesses with variable offset and thus must not be used
to generate new memory accesses. Restore the aggregate RHS from its components so the
prevailing aggregate copy does the right thing. Re-load the components of the aggregate copy destination.
But use the RHS aggregate to load from to expose more
optimization opportunities.
|
static |
Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer to the assignment and GSI is the statement iterator pointing at it. Returns the same values as sra_modify_assign.
Remove clobbers of fully scalarized variables, otherwise
do nothing. I have never seen this code path trigger but if it can happen the
following should handle it gracefully.
References build_ref_for_model(), and gimple_assign_set_lhs().
|
static |
Replace the expression EXPR with a scalar replacement if there is one and generate other statements to do type conversion or subtree copying if necessary. GSI is used to place newly created statements, WRITE is true if the expression is being written to (it is on a LHS of a statement or output in an assembly statement).
If we replace a non-register typed access simply use the original
access expression to extract the scalar component afterwards.
This happens if scalarizing a function return value or parameter
like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
gcc.c-torture/compile/20011217-1.c.
We also want to use this when accessing a complex or vector which can
be accessed as a different type too, potentially creating a need for
type conversion (see PR42196) and when scalarized unions are involved
in assembler statements (see PR42398).
References access::base, build_ref_for_model(), find_access_in_subtree(), force_gimple_operand_gsi(), get_access_replacement(), gimple_set_location(), access::grp_partial_lhs, access::grp_to_be_replaced, gsi_insert_after(), GSI_NEW_STMT, GSI_SAME_STMT, handle_unscalarized_data_in_subtree(), access::offset, access::size, sra_stats, SRA_UDH_LEFT, SRA_UDH_NONE, access::stmt, access::type, update_stmt(), and useless_type_conversion_p().
|
static |
Traverse the function body and all modifications as decided in analyze_all_variable_accesses. Return true iff the CFG has been changed.
Operands must be processed before the lhs.
References dbg_cnt().
|
static |
Convert the decisions made at the representative level into compact parameter adjustments. REPRESENTATIVES are pointers to first representatives of each param accesses, ADJUSTMENTS_COUNT is the expected final number of adjustments.
|
static |
Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple register types or (recursively) records with only these two kinds of fields. It also returns false if any of these records contains a bit-field.
|
static |
Return true iff the type contains a field or an element which does not allow scalarization.
|
staticread |
Return the representative access for the parameter declaration PARM if it is a scalar passed by reference which is not written to and the pointer value is not used directly. Thus, if it is legal to dereference it in the caller and we can rule out modifications through aliases, such parameter should be turned into one passed by value. Return NULL otherwise.
Referenced by splice_param_accesses().
|
static |
Alloc pool for allocating access structures.
| int aggregate_params_reduced |
Number of aggregate parameters that were replaced by one or more of their
components.
|
static |
Base (tree) -> Vector (vec<access_p> *) map.
|
static |
This is a table in which for each basic block and parameter there is a distance (offset + size) in that parameter which is dereferenced and accessed in that BB.
|
static |
Set of candidates.
|
static |
Referenced by build_and_add_sum(), and reorder_basic_blocks().
|
static |
| int deleted |
Number of times sra_modify_assign has deleted a statement.
Referenced by dse_step3().
| int deleted_unused_parameters |
Number of parameters that were removed because they were unused.
|
static |
scan_function sets the following to true if it encounters a call to __builtin_apply_args.
|
static |
Set by scan_function when it finds a recursive call.
|
static |
Set by scan_function when it finds a recursive call with less actual arguments than formal parameters..
| int exprs |
Number of times sra_modify_expr or sra_modify_assign themselves changed an
expression.
|
static |
Bitmap of BBs that can cause the function to "stop" progressing by returning, throwing externally, looping infinitely or calling a function which might abort etc..
|
static |
Number of parameters of the analyzed function when doing early ipa SRA.
Referenced by is_unused_scalar_param().
|
static |
Alloc pool for allocating assign link structures.
|
static |
Obstack for creation of fancy names.
|
static |
Representative of no accesses at all.
Referenced by candidate().
| int param_reductions_created |
Numbber of components created when splitting aggregate parameters.
| int replacements |
Number of processed aggregates is readily available in
analyze_all_variable_accesses and so is not stored here. Number of created scalar replacements.
| int scalar_by_ref_to_by_val |
Number of scalars passed as parameters by reference that have been
converted to be passed by value.
| int separate_lhs_rhs_handling |
Number of times sra_modify_assign has to deal with subaccesses of LHS and
RHS reparately due to type conversions or nonexistent matching
references.
|
static |
Bitmap of candidates which we should try to entirely scalarize away and those which cannot be (because they are and need be used as a whole).
Global variable describing which aggregate reduction we are performing at the moment.
| struct { ... } sra_stats |
Dump contents of ACCESS to file F in a human friendly way. If GRP is true, representative fields are dumped, otherwise those which only describe the individual access are.
Referenced by contains_vce_or_bfcref_p(), generate_subtree_copies(), and sra_modify_expr().
| int subreplacements |
Number of statements created by load_assign_lhs_subreplacements.
| int subtree_copies |
Number of statements created by generate_subtree_copies.
|
static |
Head of a linked list of accesses that need to have its subaccesses propagated to their assignment counterparts.
Referenced by get_first_repr_for_decl(), and get_var_base_offset_size_access().
1.8.1.1