GCC Middle and Back End API Reference
|
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "hash-table.h"
#include "tm.h"
#include "tree.h"
#include "flags.h"
#include "tm_p.h"
#include "basic-block.h"
#include "cfgloop.h"
#include "function.h"
#include "gimple-pretty-print.h"
#include "gimple.h"
#include "gimple-ssa.h"
#include "tree-cfg.h"
#include "tree-phinodes.h"
#include "ssa-iterators.h"
#include "tree-ssanames.h"
#include "tree-into-ssa.h"
#include "domwalk.h"
#include "tree-pass.h"
#include "tree-ssa-propagate.h"
#include "tree-ssa-threadupdate.h"
#include "langhooks.h"
#include "params.h"
#include "tree-ssa-threadedge.h"
#include "tree-ssa-dom.h"
Data Structures | |
struct | hashable_expr |
struct | cond_equivalence_s |
struct | edge_info |
struct | expr_hash_elt |
struct | expr_elt_hasher |
struct | opt_stats_d |
class | dom_opt_dom_walker |
Typedefs | |
typedef struct cond_equivalence_s | cond_equivalence |
typedef struct expr_hash_elt * | expr_hash_elt_t |
Enumerations | |
enum | expr_kind { EXPR_SINGLE, EXPR_UNARY, EXPR_BINARY, EXPR_TERNARY, EXPR_CALL, EXPR_PHI } |
Variables | |
static vec< expr_hash_elt_t > | avail_exprs_stack |
static hash_table < expr_elt_hasher > | avail_exprs |
static vec< tree > | const_and_copies_stack |
static bool | cfg_altered |
static bitmap | need_eh_cleanup |
static struct opt_stats_d | opt_stats |
typedef struct cond_equivalence_s cond_equivalence |
Structure for recording known values of a conditional expression at the exits from its block.
typedef struct expr_hash_elt* expr_hash_elt_t |
Stack of available expressions in AVAIL_EXPRs. Each block pushes any expressions it enters into the hash table along with a marker entry (null). When we finish processing the block, we pop off entries and remove the expressions from the global hash table until we hit the marker.
enum expr_kind |
SSA Dominator optimizations for trees Copyright (C) 2001-2013 Free Software Foundation, Inc. Contributed by Diego Novillo dnovi llo@ redha t.co m
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 optimizations on the dominator tree. Representation of a "naked" right-hand-side expression, to be used in recording available expressions in the expression hash table.
|
staticread |
Allocate an EDGE_INFO for edge E and attach it to E. Return the new EDGE_INFO structure.
Referenced by cprop_into_successor_phis().
|
static |
Referenced by initialize_expr_from_cond().
|
static |
Hashing and equality functions for AVAIL_EXPRS. We compute a value number for expressions using the code of the expression and the SSA numbers of its operands.
If the hash table entry is not associated with a statement, then we can just hash the expression and not worry about virtual operands and such.
Add the SSA version numbers of the vuse operand. This is important because compound variables like arrays are not renamed in the operands. Rather, the rename is done on the virtual variable representing all the elements of the array.
|
static |
Build a cond_equivalence record indicating that the comparison CODE holds between operands OP0 and OP1 and push it to **P.
Referenced by debug_dominator_optimization_stats(), and record_cond().
|
static |
Given a conditional statement CONDSTMT, convert the condition to a canonical form.
If it would be profitable to swap the operands, then do so to canonicalize the statement, enabling better optimization.
By placing canonicalization of such expressions here we transparently keep statements in canonical form, even when the statement is modified.
For relationals we need to swap the operands and change the code.
References dump_file, and print_expr_hash_elt().
|
static |
CONST_AND_COPIES is a table which maps an SSA_NAME to the current known value for that SSA_NAME (or NULL if no value is known).
Propagate values from CONST_AND_COPIES into the uses, vuses and vdef_ops of STMT.
|
static |
CONST_AND_COPIES is a table which maps an SSA_NAME to the current known value for that SSA_NAME (or NULL if no value is known).
Propagate values from CONST_AND_COPIES into the PHI nodes of the successors of BB.
If this is an abnormal edge, then we do not want to copy propagate into the PHI alternative associated with this edge.
We may have an equivalence associated with this edge. While we can not propagate it into non-dominated blocks, we can propagate them into PHIs in non-dominated blocks.
Push the unwind marker so we can reset the const and copies table back to its original state after processing this edge.
Extract and record any simple NAME = VALUE equivalences. Don't bother with [01] = COND equivalences, they're not useful here.
The alternative may be associated with a constant, so verify it is an SSA_NAME before doing anything with it.
If we have *ORIG_P in our constant/copy table, then replace ORIG_P with its value in our constant/copy table.
References allocate_edge_info(), CASE_LOW, edge_def::dest, error_mark_node, fold_convert_loc(), basic_block_def::index, edge_info::lhs, NULL, edge_info::rhs, and TREE_TYPE.
|
static |
Replace *OP_P in STMT with any known equivalent value for *OP_P from CONST_AND_COPIES.
If the operand has a known constant value or it is known to be a copy of some other variable, use the value or copy stored in CONST_AND_COPIES.
Do not replace hard register operands in asm statements.
Certain operands are not allowed to be copy propagated due to their interaction with exception handling and some GCC extensions.
Do not propagate addresses that point to volatiles into memory stmts without volatile operands.
Do not propagate copies if the propagated value is at a deeper loop depth than the propagatee. Otherwise, this may move loop variant variables outside of their loops and prevent coalescing opportunities. If the value was loop invariant, it will be hoisted by LICM and exposed for copy propagation.
Do not propagate copies into simple IV increment statements. See PR23821 for how this can disturb IV analysis.
Dump details.
And note that we modified this statement. This is now safe, even if we changed virtual operands since we will rescan the statement and rewrite its operands again.
DEBUG_FUNCTION void debug_dominator_optimization_stats | ( | void | ) |
Dump SSA statistics on stderr.
References build_and_record_new_cond(), and edge_info::cond_equivalences.
void dump_dominator_optimization_stats | ( | ) |
Dump SSA statistics on FILE.
References COMPARISON_CLASS_P, TREE_CODE, and TREE_OPERAND.
Referenced by record_equivalences_from_incoming_edge().
|
static |
STMT is either a PHI node (potentially a degenerate PHI node) or a statement that is a trivial copy or constant initialization.
Attempt to eliminate T by propagating its RHS into all uses of its LHS. This may in turn set new bits in INTERESTING_NAMES for nodes we want to revisit later.
All exit paths should clear INTERESTING_NAMES for the result of STMT.
If the LHS of this statement or PHI has no uses, then we can just eliminate it. This can occur if, for example, the PHI was created by block duplication due to threading and its only use was in the conditional at the end of the block which was deleted.
Get the RHS of the assignment or PHI node if the PHI is a degenerate.
For virtual operands we have to propagate into all uses as otherwise we will create overlapping life-ranges.
Note that STMT may well have been deleted by now, so do not access it, instead use the saved version # to clear T's entry in the worklist.
References eliminate_degenerate_phis(), execute(), and gate_dominator().
|
static |
A very simple pass to eliminate degenerate PHI nodes from the IL. This is meant to be fast enough to be able to be run several times in the optimization pipeline.
Certain optimizations, particularly those which duplicate blocks or remove edges from the CFG can create or expose PHIs which are trivial copies or constant initializations.
While we could pick up these optimizations in DOM or with the combination of copy-prop and CCP, those solutions are far too heavy-weight for our needs.
This implementation has two phases so that we can efficiently eliminate the first order degenerate PHIs and second order degenerate PHIs.
The first phase performs a dominator walk to identify and eliminate the vast majority of the degenerate PHIs. When a degenerate PHI is identified and eliminated any affected statements or PHIs are put on a worklist.
The second phase eliminates degenerate PHIs and trivial copies or constant initializations using the worklist. This is how we pick up the secondary optimization opportunities with minimal cost.
Bitmap of blocks which need EH information updated. We can not update it on-the-fly as doing so invalidates the dominator tree.
INTERESTING_NAMES is effectively our worklist, indexed by SSA_NAME_VERSION. A set bit indicates that the statement or PHI node which defines the SSA_NAME should be (re)examined to determine if it has become a degenerate PHI or trivial const/copy propagation opportunity. Experiments have show we generally get better compilation time behavior with bitmaps rather than sbitmaps.
First phase. Eliminate degenerate PHIs via a dominator walk of the CFG. Experiments have indicated that we generally get better compile-time behavior by visiting blocks in the first phase in dominator order. Presumably this is because walking in dominator order leaves fewer PHIs for later examination by the worklist phase.
Second phase. Eliminate second order degenerate PHIs as well as trivial copies or constant initializations identified by the first phase or this phase. Basically we keep iterating until our set of INTERESTING_NAMEs is empty.
EXECUTE_IF_SET_IN_BITMAP does not like its bitmap changed during the loop. Copy it to another bitmap and use that.
Ignore SSA_NAMEs that have been released because their defining statement was deleted (unreachable).
If we changed the CFG schedule loops for fixup by cfgcleanup.
Propagation of const and copies may make some EH edges dead. Purge such edges from the CFG as needed.
Referenced by eliminate_const_or_copy().
|
static |
The first phase in degenerate PHI elimination.
Eliminate the degenerate PHIs in BB, then recurse on the dominator children of BB.
Recurse into the dominator children of BB.
|
static |
|
static |
Search for redundant computations in STMT. If any are found, then replace them with the variable holding the result of the computation.
If safe, record this expression into the available expression hash table.
Certain expressions on the RHS can be optimized away, but can not themselves be entered into the hash tables.
Do not record equivalences for increments of ivs. This would create overlapping live ranges for a very questionable gain.
Check if the expression has been computed before.
Get the type of the expression we are trying to optimize.
We can't propagate into a phi, so the logic below doesn't apply. Instead record an equivalence between the cached LHS and the PHI result of this statement, provided they are in the same block. This should be sufficient to kill the redundant phi.
It is safe to ignore types here since we have already done type checking in the hashing and equality routines. In fact type checking here merely gets in the way of constant propagation. Also, make sure that it is safe to propagate CACHED_LHS into the expression in STMT.
Since it is always necessary to mark the result as modified, perhaps we should move this into propagate_tree_value_into_stmt itself.
References dump_file, dump_flags, print_generic_expr(), and set_ssa_name_value().
|
static |
Free all EDGE_INFO structures associated with edges in the CFG. If a particular edge can be threaded, copy the redirection target from the EDGE_INFO structure into the edge's AUX field as required by code to update the CFG and SSA graph for jump threading.
References BITMAP_ALLOC, calculate_dominance_info(), CDI_DOMINATORS, hash_table< Descriptor, Allocator >::create(), loop_optimizer_init(), LOOPS_HAVE_SIMPLE_LATCHES, NULL, and opt_stats.
|
static |
|
static |
Delete an expr_hash_elt and reclaim its storage.
References edge_def::aux, edge_info::cond_equivalences, and NULL.
|
static |
Delete variable sized pieces of the expr_hash_elt ELEMENT.
References edge_def::aux, FOR_EACH_BB, FOR_EACH_EDGE, and basic_block_def::preds.
|
static |
Referenced by eliminate_const_or_copy().
|
static |
Given a statement STMT, which is either a PHI node or an assignment, return the "lhs" of the node.
|
static |
Given a statement STMT, which is either a PHI node or an assignment, return the "rhs" of the node, in the case of a non-degenerate phi, NULL is returned.
|
static |
Hashtable helpers.
Compare two hashable_expr structures for equivalence. They are considered equivalent when the the expressions they denote must necessarily be equal. The logic is intended to follow that of operand_equal_p in fold-const.c
If either type is NULL, there is nothing to check.
If both types don't have the same signedness, precision, and mode, then we can't consider them equal.
For commutative ops, allow the other order.
For commutative ops, allow the other order.
If the calls are to different functions, then they clearly cannot be equal.
|
static |
|
static |
Dump statistics for the hash table HTAB.
|
static |
Given a conditional expression COND as a tree, initialize a hashable_expr expression EXPR. The conditional must be a comparison or logical negation. A constant or a variable is not permitted.
References avail_expr_hash(), expr_hash_elt::expr, expr_hash_elt::hash, expr_hash_elt::lhs, NULL, expr_hash_elt::stamp, and expr_hash_elt::stmt.
|
static |
Given a statement STMT, initialize the hash table element pointed to by ELEMENT.
|
static |
Given a hashable_expr expression EXPR and an LHS, initialize the hash table element pointed to by ELEMENT.
References EXPR_SINGLE, EXPR_UNARY, operand_equal_p(), hashable_expr::ops, and hashable_expr::single.
|
static |
Compute a hash value for a hashable_expr value EXPR and a previously accumulated hash value VAL. If two hashable_expr values compare equal with hashable_expr_equal_p, they must hash to the same value, given an identical value of VAL. The logic is intended to follow iterative_hash_expr in tree.c.
Make sure to include signedness in the hash computation. Don't hash the type, that can lead to having nodes which compare equal according to operand_equal_p, but which have different hash codes.
|
static |
Search for an existing instance of STMT in the AVAIL_EXPRS table. If found, return its LHS. Otherwise insert STMT in the table and return NULL_TREE.
Also, when an expression is first inserted in the table, it is also is also added to AVAIL_EXPRS_STACK, so that it can be removed when we finish processing this block and its children.
Get LHS of phi, assignment, or call; else NULL_TREE.
Don't bother remembering constant assignments and copy operations. Constants and copy operations are handled by the constant/copy propagator in optimize_stmt.
Finally try to find the expression in the main expression hash table.
Extract the LHS of the assignment so that it can be used as the current definition of another variable.
See if the LHS appears in the CONST_AND_COPIES table. If it does, then use the value from the const_and_copies table.
References dump_file, dump_flags, FOR_EACH_IMM_USE_ON_STMT, FOR_EACH_IMM_USE_STMT, gimple_debug_bind_p(), loop_depth_of_name(), may_propagate_copy(), may_propagate_copy_into_asm(), print_generic_expr(), print_gimple_stmt(), propagate_value(), SSA_NAME_DEF_STMT, SSA_NAME_OCCURS_IN_ABNORMAL_PHI, and TREE_CODE.
int loop_depth_of_name | ( | ) |
Return the loop depth of the basic block of the defining statement of X. This number should not be treated as absolutely correct because the loop information may not be completely up-to-date when dom runs. However, it will be relatively correct, and as more passes are taught to keep loop info up to date, the result will become more and more accurate.
If it's not an SSA_NAME, we have no clue where the definition is.
Otherwise return the loop depth of the defining statement's bb. Note that there may not actually be a bb for this statement, if the ssa_name is live on entry.
Referenced by lookup_avail_expr().
gimple_opt_pass* make_pass_dominator | ( | ) |
References NULL.
gimple_opt_pass* make_pass_phi_only_cprop | ( | ) |
|
static |
Local functions.
|
static |
Optimize the statement pointed to by iterator SI.
We try to perform some simplistic global redundancy elimination and constant propagation:
1- To detect global redundancy, we keep track of expressions that have been computed in this block and its dominators. If we find that the same expression is computed more than once, we eliminate repeated computations by using the target of the first one.
2- Constant values and copy assignments. This is used to do very simplistic constant and copy propagation. When a constant or copy assignment is found, we map the value on the RHS of the assignment to the variable in the LHS in the CONST_AND_COPIES table.
Const/copy propagate into USES, VUSES and the RHS of VDEFs.
If the statement has been modified with constant replacements, fold its RHS before checking for redundant computations.
Try to fold the statement making sure that STMT is kept up to date.
We only need to consider cases that can yield a gimple operand.
This should never be an ADDR_EXPR.
Indicate that maybe_clean_or_replace_eh_stmt needs to be called, even if fold_stmt updated the stmt already and thus cleared gimple_modified_p flag on it.
Check for redundant computations. Do this optimization only for assignments that have no volatile ops and conditionals.
Resolve __builtin_constant_p. If it hasn't been folded to integer_one_node by now, it's fairly certain that the value simply isn't constant.
Perform simple redundant store elimination.
Build a new statement with the RHS and LHS exchanged.
Record any additional equivalences created by this statement.
If STMT is a COND_EXPR and it was modified, then we may know where it goes. If that is the case, then mark the CFG as altered. This will cause us to later call remove_unreachable_blocks and cleanup_tree_cfg when it is safe to do so. It is not safe to clean things up here since removal of edges and such can trigger the removal of PHI nodes, which in turn can release SSA_NAMEs to the manager. That's all fine and good, except that once SSA_NAMEs are released to the manager, we must not call create_ssa_name until all references to released SSA_NAMEs have been eliminated. All references to the deleted SSA_NAMEs can not be eliminated until we remove unreachable blocks. We can not remove unreachable blocks until after we have completed any queued jump threading. We can not complete any queued jump threads until we have taken appropriate variables out of SSA form. Taking variables out of SSA form can call create_ssa_name and thus we lose. Ultimately I suspect we're going to need to change the interface into the SSA_NAME manager.
If we simplified a statement in such a way as to be shown that it cannot trap, update the eh information and the cfg to match.
References bitmap_set_bit, boolean_type_node, dump_file, dump_flags, find_taken_edge(), fold_binary_loc(), gimple_cond_code(), gimple_cond_lhs(), gimple_cond_rhs(), gimple_location(), gimple_switch_index(), basic_block_def::index, maybe_clean_or_replace_eh_stmt(), NULL, TREE_CODE, and update_stmt_if_modified().
|
static |
Print a diagnostic dump of an expression hash table entry.
References hashable_expr::call, expr_hash_elt::expr, hashable_expr::fn_from, gimple_call_fn(), gimple_call_internal_fn(), gimple_call_internal_p(), internal_fn_name(), hashable_expr::nargs, hashable_expr::ops, and print_generic_expr().
Referenced by canonicalize_comparison().
|
static |
Propagate RHS into all uses of LHS (when possible).
RHS and LHS are derived from STMT, which is passed in solely so that we can remove it if propagation is successful.
When propagating into a PHI node or into a statement which turns into a trivial copy or constant initialization, set the appropriate bit in INTERESTING_NAMEs so that we will visit those nodes as well in an effort to pick up secondary optimization opportunities.
First verify that propagation is valid and isn't going to move a loop variant variable outside its loop.
Dump details.
Walk over every use of LHS and try to replace the use with RHS. At this point the only reason why such a propagation would not be successful would be if the use occurs in an ASM_EXPR.
Leave debug stmts alone. If we succeed in propagating all non-debug uses, we'll drop the DEF, and propagation into debug stmts will occur then.
It's not always safe to propagate into an ASM_EXPR.
It's not ok to propagate into the definition stmt of RHS. <bb 9>: # prephitmp.12_36 = PHI <g_67.1_6(9)> g_67.1_6 = prephitmp.12_36; goto <bb 9>; While this is strictly all dead code we do not want to deal with this here.
Dump details.
Propagate the RHS into this use of the LHS.
Special cases to avoid useless calls into the folding routines, operand scanning, etc. Propagation into a PHI may cause the PHI to become a degenerate, so mark the PHI as interesting. No other actions are necessary.
Dump details.
From this point onward we are propagating into a real statement. Folding may (or may not) be possible, we may expose new operands, expose dead EH edges, etc.
NOTE tuples. In the tuples world, fold_stmt_inplace cannot fold a call that simplifies to a constant, because the GIMPLE_CALL must be replaced by a GIMPLE_ASSIGN, and there is no way to effect such a transformation in-place. We might want to consider using the more general fold_stmt here.
Sometimes propagation can expose new operands to the renamer.
Dump details.
If we replaced a variable index with a constant, then we would need to update the invariant flag for ADDR_EXPRs.
If we cleaned up EH information from the statement, mark its containing block as needing EH cleanups.
Propagation may expose new trivial copy/constant propagation opportunities.
Propagation into these nodes may make certain edges in the CFG unexecutable. We want to identify them as PHI nodes at the destination of those unexecutable edges may become degenerates.
Remove all outgoing edges except TE.
Mark all the PHI nodes at the destination of the unexecutable edge as interesting.
And fixup the flags on the single remaining edge.
Ensure there is nothing else to do.
If we were able to propagate away all uses of LHS, then we can remove STMT.
References bitmap_set_bit, gimple_phi_result(), gsi_stmt(), and SSA_NAME_VERSION.
|
static |
|
static |
Enter condition equivalence into the expression hash table. This indicates that a conditional expression has a known boolean value.
References build_and_record_new_cond(), and edge_info::cond_equivalences.
|
static |
Record that COND is true and INVERTED is false into the edge information structure. Also record that any conditions dominated by COND are true as well.
For example, if a < b is true, then a <= b must also be true.
Now store the original true and false conditions into the first two slots.
It is possible for INVERTED to be the negation of a comparison, and not a valid RHS or GIMPLE_COND condition. This happens because invert_truthvalue may return such an expression when asked to invert a floating-point comparison. These comparisons are not assumed to obey the trichotomy law.
Referenced by record_equality().
|
static |
Record that X is equal to Y in const_and_copies. Record undo information in the block-local vector.
References edge_def::dest, edge_def::flags, FOR_EACH_EDGE, gsi_start_phis(), and basic_block_def::succs.
|
static |
A helper function for record_const_or_copy and record_equality. Do the work of recording the value and undo info.
|
static |
We have finished optimizing BB, record any information implied by taking a specific outgoing edge from BB.
A COND_EXPR may create equivalences too.
Special case comparing booleans against a constant as we know the value of OP0 on both arms of the branch. i.e., we can record an equivalence for OP0 rather than COND.
??? TRUTH_NOT_EXPR can create an equivalence too.
References edge_info::lhs, and edge_info::rhs.
Referenced by single_incoming_edge_ignoring_loop_edges().
|
static |
Similarly, but assume that X and Y are the two operands of an EQ_EXPR. This constrains the cases in which we may treat this as assignment.
If one of the previous values is invariant, or invariant in more loops (by depth), then use that. Otherwise it doesn't matter which value we choose, just so long as we canonicalize on one value.
After the swapping, we must have one SSA_NAME.
For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a variable compared against zero. If we're honoring signed zeros, then we cannot record this value unless we know that the value is nonzero.
References edge_info::lhs, record_const_or_copy(), edge_info::rhs, and TREE_CODE.
|
static |
|
static |
Record any equivalences created by the incoming edge to BB. If BB has more than one incoming edge, then no equivalence is created.
If our parent block ended with a control statement, then we may be able to record some equivalences based on which outgoing edge from the parent was followed.
If we had a single incoming edge from our parent block, then enter any data associated with the edge into our tables.
If LHS is an SSA_NAME and RHS is a constant integer and LHS was set via a widening type conversion, then we may be able to record additional equivalences.
If the conversion widens the original value and the constant is in the range of the type of OLD_RHS, then convert the constant and record the equivalence. Note that int_fits_type_p does not check the precision if the upper and lower bounds are OK.
References dump_dominator_optimization_stats().
|
static |
|
static |
PHI nodes can create equivalences too.
Ignoring any alternatives which are the same as the result, if all the alternatives are equal, then the PHI node creates an equivalence.
Ignore alternatives which are the same as our LHS. Since LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we can simply compare pointers.
If we have not processed an alternative yet, then set RHS to this alternative.
If we have processed an alternative (stored in RHS), then see if it is equal to this one. If it isn't, then stop the search.
If we had no interesting alternatives, then all the RHS alternatives must have been the same as LHS.
If we managed to iterate through each PHI alternative without breaking out of the loop, then we have a PHI which may create a useful equivalence. We do not need to record unwind data for this, since this is a true assignment and not an equivalence inferred from a comparison. All uses of this ssa name are dominated by this assignment, so unwinding just costs time and space.
|
static |
|
static |
STMT, a GIMPLE_ASSIGN, may create certain equivalences, in either the available expressions table or the const_and_copies table. Detect and record those equivalences. We handle only very simple copy equivalences here. The heavy lifing is done by eliminate_redundant_computations.
If the RHS of the assignment is a constant or another variable that may be propagated, register it in the CONST_AND_COPIES table. We do not need to record unwind data for this, since this is a true assignment and not an equivalence inferred from a comparison. All uses of this ssa name are dominated by this assignment, so unwinding just costs time and space.
A memory store, even an aliased store, creates a useful equivalence. By exchanging the LHS and RHS, creating suitable vops and recording the result in the available expression table, we may be able to expose more redundant loads. Build a new statement with the RHS and LHS exchanged.
NOTE tuples. The call to gimple_build_assign below replaced a call to build_gimple_modify_stmt, which did not set the SSA_NAME_DEF_STMT on the LHS of the assignment. Doing so may cause an SSA validation failure, as the LHS may be a default-initialized name and should have no definition. I'm a bit dubious of this, as the artificial statement that we generate here may in fact be ill-formed, but it is simply used as an internal device in this pass, and never becomes part of the CFG.
Finally enter the statement into the available expression table.
|
static |
Record into the equivalence tables any equivalences implied by traversing edge E (which are cached in E->aux).
Callers are responsible for managing the unwinding markers.
If we have info associated with this edge, record it into our equivalence tables.
If we have a simple NAME = VALUE equivalence, record it.
If we have 0 = COND or 1 = COND equivalences, record them into our expression hash tables.
References gimple_phi_arg_def(), gimple_phi_num_args(), gimple_phi_result(), gsi_end_p(), gsi_next(), gsi_start_phis(), gsi_stmt(), NULL, and operand_equal_for_phi_arg_p().
Referenced by restore_vars_to_original_value().
|
static |
Initialize local stacks for this optimizer and record equivalences upon entry to BB. Equivalences can come from the edge traversed to reach BB or they may come from PHI nodes at the start of BB. Remove all the expressions in LOCALS from TABLE, stopping when there are LIMIT entries left in LOCALs.
Remove all the expressions made available in this block.
This must precede the actual removal from the hash table, as ELEMENT and the table entry may share a call argument vector which will be freed during removal.
|
static |
PHI-ONLY copy and constant propagation. This pass is meant to clean up degenerate PHIs created by or exposed by jump threading. Given a statement STMT, which is either a PHI node or an assignment, remove it from the IL.
References bitmap_set_bit, dump_file, and dump_flags.
|
static |
Use the source/dest pairs in CONST_AND_COPIES_STACK to restore CONST_AND_COPIES to its original state, stopping when we hit a NULL marker.
References gimple_build_cond(), integer_zero_node, dom_opt_dom_walker::m_dummy_cond, NULL, NULL_TREE, record_temporary_equivalences(), simplify_stmt_for_jump_threading(), and dom_opt_dom_walker::thread_across_edge().
bool simple_iv_increment_p | ( | ) |
Returns true when STMT is a simple iv increment. It detects the following situation:
i_1 = phi (..., i_2) i_2 = i_1 +/- ...
A trivial wrapper so that we can present the generic jump threading code with a simple API for simplifying statements.
Referenced by restore_vars_to_original_value().
|
static |
|
static |
Ignoring loop backedges, if BB has precisely one incoming edge then return that edge. Otherwise return NULL.
A loop back edge can be identified by the destination of the edge dominating the source of the edge.
If we have already seen a non-loop edge, then we must have multiple incoming non-loop edges and thus we return NULL.
This is the first non-loop incoming edge we have found. Record it.
References fold_convert, record_equality(), and TREE_TYPE.
|
static |
Jump threading, redundancy elimination and const/copy propagation.
This pass may expose new symbols that need to be renamed into SSA. For every new symbol exposed, its corresponding bit will be set in VARS_TO_RENAME.
Create our hash tables.
We need to know loop structures in order to avoid destroying them in jump threading. Note that we still can e.g. thread through loop headers to an exit edge, or through loop header to the loop body, assuming that we update the loop info.
Initialize the value-handle array.
We need accurate information regarding back edges in the CFG for jump threading; this may include back edges that are not part of a single loop.
Recursively walk the dominator tree optimizing statements.
If we exposed any new variables, go ahead and put them into SSA form now, before we handle jump threading. This simplifies interactions between rewriting of _DECL nodes into SSA form and rewriting SSA_NAME nodes into SSA form after block duplication and CFG manipulation.
Thread jumps, creating duplicate blocks as needed.
Removal of statements may make some EH edges dead. Purge such edges from the CFG as needed.
Jump threading may have created forwarder blocks from blocks needing EH cleanup; the new successor of these blocks, which has inherited from the original block, needs the cleanup. Don't clear bits in the bitmap, as that can break the bitmap iterator.
Debugging dumps.
Delete our main hashtable.
Free asserted bitmaps and stacks.
Free the value-handle array.
|
static |
Hash table with expressions made available during the renaming process. When an assignment of the form X_i = EXPR is found, the statement is stored in this table. If the same expression EXPR is later found on the RHS of another statement, it is replaced with X_i (thus performing global redundancy elimination). Similarly as we pass through conditionals we record the conditional itself as having either a true or false value in this table.
|
static |
|
static |
Track whether or not we have changed the control flow graph.
Stack of dest,src pairs that need to be restored during finalization.
A NULL entry is used to mark the end of pairs which need to be restored during finalization of this block.
|
static |
Bitmap of blocks that have had EH statements cleaned. We should remove their dead edges eventually.
|
static |
Referenced by free_all_edge_infos().