1 @c Copyright (c) 2004 Free Software Foundation, Inc.
2 @c Free Software Foundation, Inc.
3 @c This is part of the GCC manual.
4 @c For copying conditions, see the file gcc.texi.
6 @c ---------------------------------------------------------------------
8 @c ---------------------------------------------------------------------
11 @chapter Analysis and Optimization of GIMPLE Trees
13 @cindex Optimization infrastructure for GIMPLE
15 GCC uses three main intermediate languages to represent the program
16 during compilation: GENERIC, GIMPLE and RTL@. GENERIC is a
17 language-independent representation generated by each front end. It
18 is used to serve as an interface between the parser and optimizer.
19 GENERIC is a common representation that is able to represent programs
20 written in all the languages supported by GCC@.
22 GIMPLE and RTL are used to optimize the program. GIMPLE is used for
23 target and language independent optimizations (e.g., inlining,
24 constant propagation, tail call elimination, redundancy elimination,
25 etc). Much like GENERIC, GIMPLE is a language independent, tree based
26 representation. However, it differs from GENERIC in that the GIMPLE
27 grammar is more restrictive: expressions contain no more than 3
28 operands (except function calls), it has no control flow structures
29 and expressions with side-effects are only allowed on the right hand
30 side of assignments. See the chapter describing GENERIC and GIMPLE
33 This chapter describes the data structures and functions used in the
34 GIMPLE optimizers (also known as ``tree optimizers'' or ``middle
35 end''). In particular, it focuses on all the macros, data structures,
36 functions and programming constructs needed to implement optimization
40 * GENERIC:: A high-level language-independent representation.
41 * GIMPLE:: A lower-level factored tree representation.
42 * Annotations:: Attributes for statements and variables.
43 * Statement Operands:: Variables referenced by GIMPLE statements.
44 * SSA:: Static Single Assignment representation.
45 * Alias analysis:: Representing aliased loads and stores.
52 The purpose of GENERIC is simply to provide a language-independent way of
53 representing an entire function in trees. To this end, it was necessary to
54 add a few new tree codes to the back end, but most everything was already
55 there. If you can express it with the codes in @code{gcc/tree.def}, it's
58 Early on, there was a great deal of debate about how to think about
59 statements in a tree IL. In GENERIC, a statement is defined as any
60 expression whose value, if any, is ignored. A statement will always
61 have @code{TREE_SIDE_EFFECTS} set (or it will be discarded), but a
62 non-statement expression may also have side effects. A
63 @code{CALL_EXPR}, for instance.
65 It would be possible for some local optimizations to work on the
66 GENERIC form of a function; indeed, the adapted tree inliner works
67 fine on GENERIC, but the current compiler performs inlining after
68 lowering to GIMPLE (a restricted form described in the next section).
69 Indeed, currently the frontends perform this lowering before handing
70 off to @code{tree_rest_of_compilation}, but this seems inelegant.
72 If necessary, a front end can use some language-dependent tree codes
73 in its GENERIC representation, so long as it provides a hook for
74 converting them to GIMPLE and doesn't expect them to work with any
75 (hypothetical) optimizers that run before the conversion to GIMPLE.
76 The intermediate representation used while parsing C and C++ looks
77 very little like GENERIC, but the C and C++ gimplifier hooks are
78 perfectly happy to take it as input and spit out GIMPLE.
84 GIMPLE is a simplified subset of GENERIC for use in optimization. The
85 particular subset chosen (and the name) was heavily influenced by the
86 SIMPLE IL used by the McCAT compiler project at McGill University,
87 though we have made some different choices. For one thing, SIMPLE
88 doesn't support @code{goto}; a production compiler can't afford that
91 GIMPLE retains much of the structure of the parse trees: lexical
92 scopes are represented as containers, rather than markers. However,
93 expressions are broken down into a 3-address form, using temporary
94 variables to hold intermediate values. Also, control structures are
97 In GIMPLE no container node is ever used for its value; if a
98 @code{COND_EXPR} or @code{BIND_EXPR} has a value, it is stored into a
99 temporary within the controlled blocks, and that temporary is used in
100 place of the container.
102 The compiler pass which lowers GENERIC to GIMPLE is referred to as the
103 @samp{gimplifier}. The gimplifier works recursively, replacing complex
104 statements with sequences of simple statements.
106 @c Currently, the only way to
107 @c tell whether or not an expression is in GIMPLE form is by recursively
108 @c examining it; in the future there will probably be a flag to help avoid
109 @c redundant work. FIXME FIXME
114 * GIMPLE Expressions::
117 * Rough GIMPLE Grammar::
121 @subsection Interfaces
122 @cindex gimplification
124 The tree representation of a function is stored in
125 @code{DECL_SAVED_TREE}. It is lowered to GIMPLE by a call to
126 @code{gimplify_function_tree}.
128 If a front end wants to include language-specific tree codes in the tree
129 representation which it provides to the back end, it must provide a
130 definition of @code{LANG_HOOKS_GIMPLIFY_EXPR} which knows how to
131 convert the front end trees to GIMPLE. Usually such a hook will involve
132 much of the same code for expanding front end trees to RTL. This function
133 can return fully lowered GIMPLE, or it can return GENERIC trees and let the
134 main gimplifier lower them the rest of the way; this is often simpler.
136 The C and C++ front ends currently convert directly from front end
137 trees to GIMPLE, and hand that off to the back end rather than first
138 converting to GENERIC. Their gimplifier hooks know about all the
139 @code{_STMT} nodes and how to convert them to GENERIC forms. There
140 was some work done on a genericization pass which would run first, but
141 the existence of @code{STMT_EXPR} meant that in order to convert all
142 of the C statements into GENERIC equivalents would involve walking the
143 entire tree anyway, so it was simpler to lower all the way. This
144 might change in the future if someone writes an optimization pass
145 which would work better with higher-level trees, but currently the
146 optimizers all expect GIMPLE.
148 A front end which wants to use the tree optimizers (and already has
149 some sort of whole-function tree representation) only needs to provide
150 a definition of @code{LANG_HOOKS_GIMPLIFY_EXPR}, call
151 @code{gimplify_function_tree} to lower to GIMPLE, and then hand off to
152 @code{tree_rest_of_compilation} to compile and output the function.
154 You can tell the compiler to dump a C-like representation of the GIMPLE
155 form with the flag @code{-fdump-tree-gimple}.
158 @subsection Temporaries
161 When gimplification encounters a subexpression which is too complex, it
162 creates a new temporary variable to hold the value of the subexpression,
163 and adds a new statement to initialize it before the current statement.
164 These special temporaries are known as @samp{expression temporaries}, and are
165 allocated using @code{get_formal_tmp_var}. The compiler tries to
166 always evaluate identical expressions into the same temporary, to simplify
167 elimination of redundant calculations.
169 We can only use expression temporaries when we know that it will not be
170 reevaluated before its value is used, and that it will not be otherwise
171 modified@footnote{These restrictions are derived from those in Morgan 4.8.}.
172 Other temporaries can be allocated using
173 @code{get_initialized_tmp_var} or @code{create_tmp_var}.
175 Currently, an expression like @code{a = b + 5} is not reduced any
176 further. We tried converting it to something like
181 but this bloated the representation for minimal benefit. However, a
182 variable which must live in memory cannot appear in an expression; its
183 value is explicitly loaded into a temporary first. Similarly, storing
184 the value of an expression to a memory variable goes through a
187 @node GIMPLE Expressions
188 @subsection Expressions
189 @cindex GIMPLE Expressions
191 In general, expressions in GIMPLE consist of an operation and the
192 appropriate number of simple operands; these operands must either be a
193 GIMPLE rvalue (@code{is_gimple_val}), i.e. a constant or a register
194 variable. More complex operands are factored out into temporaries, so
205 The same rule holds for arguments to a @code{CALL_EXPR}.
207 The target of an assignment is usually a variable, but can also be an
208 @code{INDIRECT_REF} or a compound lvalue as described below.
211 * Compound Expressions::
213 * Conditional Expressions::
214 * Logical Operators::
217 @node Compound Expressions
218 @subsubsection Compound Expressions
219 @cindex Compound Expressions
221 The left-hand side of a C comma expression is simply moved into a separate
224 @node Compound Lvalues
225 @subsubsection Compound Lvalues
226 @cindex Compound Lvalues
228 Currently compound lvalues involving array and structure field references
229 are not broken down; an expression like @code{a.b[2] = 42} is not reduced
230 any further (though complex array subscripts are). This restriction is a
231 workaround for limitations in later optimizers; if we were to convert this
239 alias analysis would not remember that the reference to @code{T1[2]} came
240 by way of @code{a.b}, so it would think that the assignment could alias
241 another member of @code{a}; this broke @code{struct-alias-1.c}. Future
242 optimizer improvements may make this limitation unnecessary.
244 @node Conditional Expressions
245 @subsubsection Conditional Expressions
246 @cindex Conditional Expressions
248 A C @code{?:} expression is converted into an @code{if} statement with
249 each branch assigning to the same temporary. So,
263 Note that in GIMPLE, @code{if} statements are also represented using
264 @code{COND_EXPR}, as described below.
266 @node Logical Operators
267 @subsubsection Logical Operators
268 @cindex Logical Operators
270 Except when they appear in the condition operand of a @code{COND_EXPR},
271 logical `and' and `or' operators are simplified as follows:
272 @code{a = b && c} becomes
281 Note that @code{T1} in this example cannot be an expression temporary,
282 because it has two different assignments.
285 @subsection Statements
288 Most statements will be assignment statements, represented by
289 @code{MODIFY_EXPR}. A @code{CALL_EXPR} whose value is ignored can
290 also be a statement. No other C expressions can appear at statement level;
291 a reference to a volatile object is converted into a @code{MODIFY_EXPR}.
292 In GIMPLE form, type of @code{MODIFY_EXPR} is not meaningful. Instead, use type
295 There are also several varieties of complex statements.
299 * Statement Sequences::
302 * Selection Statements::
305 * GIMPLE Exception Handling::
309 @subsubsection Blocks
312 Block scopes and the variables they declare in GENERIC and GIMPLE are
313 expressed using the @code{BIND_EXPR} code, which in previous versions of
314 GCC was primarily used for the C statement-expression extension.
316 Variables in a block are collected into @code{BIND_EXPR_VARS} in
317 declaration order. Any runtime initialization is moved out of
318 @code{DECL_INITIAL} and into a statement in the controlled block. When
319 gimplifying from C or C++, this initialization replaces the
322 Variable-length arrays (VLAs) complicate this process, as their size often
323 refers to variables initialized earlier in the block. To handle this, we
324 currently split the block at that point, and move the VLA into a new, inner
325 @code{BIND_EXPR}. This strategy may change in the future.
327 @code{DECL_SAVED_TREE} for a GIMPLE function will always be a
328 @code{BIND_EXPR} which contains declarations for the temporary variables
329 used in the function.
331 A C++ program will usually contain more @code{BIND_EXPR}s than there are
332 syntactic blocks in the source code, since several C++ constructs have
333 implicit scopes associated with them. On the other hand, although the C++
334 front end uses pseudo-scopes to handle cleanups for objects with
335 destructors, these don't translate into the GIMPLE form; multiple
336 declarations at the same level use the same BIND_EXPR.
338 @node Statement Sequences
339 @subsubsection Statement Sequences
340 @cindex Statement Sequences
342 Multiple statements at the same nesting level are collected into a
343 @code{STATEMENT_LIST}. Statement lists are modified and traversed
344 using the interface in @samp{tree-iterator.h}.
346 @node Empty Statements
347 @subsubsection Empty Statements
348 @cindex Empty Statements
350 Whenever possible, statements with no effect are discarded. But if they
351 are nested within another construct which cannot be discarded for some
352 reason, they are instead replaced with an empty statement, generated by
353 @code{build_empty_stmt}. Initially, all empty statements were shared,
354 after the pattern of the Java front end, but this caused a lot of trouble in
357 An empty statement is represented as @code{(void)0}.
363 At one time loops were expressed in GIMPLE using @code{LOOP_EXPR}, but
364 now they are lowered to explicit gotos.
366 @node Selection Statements
367 @subsubsection Selection Statements
368 @cindex Selection Statements
370 A simple selection statement, such as the C @code{if} statement, is
371 expressed in GIMPLE using a void @code{COND_EXPR}. If only one branch is
372 used, the other is filled with an empty statement.
374 Normally, the condition expression is reduced to a simple comparison. If
375 it is a shortcut (@code{&&} or @code{||}) expression, however, we try to
376 break up the @code{if} into multiple @code{if}s so that the implied shortcut
377 is taken directly, much like the transformation done by @code{do_jump} in
380 A @code{SWITCH_EXPR} in GIMPLE contains the condition and a
381 @code{TREE_VEC} of @code{CASE_LABEL_EXPR}s describing the case values
382 and corresponding @code{LABEL_DECL}s to jump to. The body of the
383 @code{switch} is moved after the @code{SWITCH_EXPR}.
389 Other jumps are expressed by either @code{GOTO_EXPR} or @code{RETURN_EXPR}.
391 The operand of a @code{GOTO_EXPR} must be either a label or a variable
392 containing the address to jump to.
394 The operand of a @code{RETURN_EXPR} is either @code{NULL_TREE} or a
395 @code{MODIFY_EXPR} which sets the return value. It would be nice to
396 move the @code{MODIFY_EXPR} into a separate statement, but the special
397 return semantics in @code{expand_return} make that difficult. It may
398 still happen in the future, perhaps by moving most of that logic into
399 @code{expand_assignment}.
402 @subsubsection Cleanups
405 Destructors for local C++ objects and similar dynamic cleanups are
406 represented in GIMPLE by a @code{TRY_FINALLY_EXPR}. When the controlled
407 block exits, the cleanup is run.
409 @code{TRY_FINALLY_EXPR} complicates the flow graph, since the cleanup
410 needs to appear on every edge out of the controlled block; this
411 reduces the freedom to move code across these edges. Therefore, the
412 EH lowering pass which runs before most of the optimization passes
413 eliminates these expressions by explicitly adding the cleanup to each
416 @node GIMPLE Exception Handling
417 @subsubsection Exception Handling
418 @cindex GIMPLE Exception Handling
420 Other exception handling constructs are represented using
421 @code{TRY_CATCH_EXPR}. The handler operand of a @code{TRY_CATCH_EXPR}
422 can be a normal statement to be executed if the controlled block throws an
423 exception, or it can have one of two special forms:
426 @item A @code{CATCH_EXPR} executes its handler if the thrown exception
427 matches one of the allowed types. Multiple handlers can be
428 expressed by a sequence of @code{CATCH_EXPR} statements.
429 @item An @code{EH_FILTER_EXPR} executes its handler if the thrown
430 exception does not match one of the allowed types.
433 Currently throwing an exception is not directly represented in GIMPLE,
434 since it is implemented by calling a function. At some point in the future
435 we will want to add some way to express that the call will throw an
436 exception of a known type.
438 Just before running the optimizers, the compiler lowers the high-level
439 EH constructs above into a set of @samp{goto}s, magic labels, and EH
440 regions. Continuing to unwind at the end of a cleanup is represented
441 with a @code{RESX_EXPR}.
444 @subsection GIMPLE Example
445 @cindex GIMPLE Example
448 struct A @{ A(); ~A(); @};
455 int j = (--i, i ? 0 : 1);
457 for (int x = 42; x > 0; --x)
524 @node Rough GIMPLE Grammar
525 @subsection Rough GIMPLE Grammar
526 @cindex Rough GIMPLE Grammar
531 DECL_SAVED_TREE -> block
534 BIND_EXPR_VARS -> DECL chain
535 BIND_EXPR_BLOCK -> BLOCK
540 op0 -> non-compound-stmt
562 op2 -> TREE_VEC of CASE_LABEL_EXPRs
565 op0 -> LABEL_DECL | '*' ID
593 op0 -> _DECL | '&' _DECL
601 varname : compref | _DECL
602 lhs: varname | '*' _DECL
603 pseudo-lval: _DECL | '*' _DECL
606 op0 -> compref | pseudo-lval
608 op0 -> compref | pseudo-lval
611 condition : val | val relop val
622 unop: '+' | '-' | '!' | '~'
624 binop: relop | '-' | '+' | '/' | '*'
625 | '%' | '&' | '|' | '<<' | '>>' | '^'
627 relop: All tree codes of class '<'
634 The optimizers need to associate attributes with statements and
635 variables during the optimization process. For instance, we need to
636 know what basic block does a statement belong to or whether a variable
637 has aliases. All these attributes are stored in data structures
638 called annotations which are then linked to the field @code{ann} in
639 @code{struct tree_common}.
641 Presently, we define annotations for statements (@code{stmt_ann_t}),
642 variables (@code{var_ann_t}) and SSA names (@code{ssa_name_ann_t}).
643 Annotations are defined and documented in @file{tree-flow.h}.
646 @node Statement Operands
647 @section Statement Operands
649 @cindex virtual operands
650 @cindex real operands
651 @findex get_stmt_operands
654 Almost every GIMPLE statement will contain a reference to a variable
655 or memory location. Since statements come in different shapes and
656 sizes, their operands are going to be located at various spots inside
657 the statement's tree. To facilitate access to the statement's
658 operands, they are organized into arrays associated inside each
659 statement's annotation. Each element in an operand array is a pointer
660 to a @code{VAR_DECL}, @code{PARM_DECL} or @code{SSA_NAME} tree node.
661 This provides a very convenient way of examining and replacing
664 Data flow analysis and optimization is done on all tree nodes
665 representing variables. Any node for which @code{SSA_VAR_P} returns
666 nonzero is considered when scanning statement operands. However, not
667 all @code{SSA_VAR_P} variables are processed in the same way. For the
668 purposes of optimization, we need to distinguish between references to
669 local scalar variables and references to globals, statics, structures,
670 arrays, aliased variables, etc. The reason is simple, the compiler
671 can gather complete data flow information for a local scalar. On the
672 other hand, a global variable may be modified by a function call, it
673 may not be possible to keep track of all the elements of an array or
674 the fields of a structure, etc.
676 The operand scanner gathers two kinds of operands: @dfn{real} and
677 @dfn{virtual}. An operand for which @code{is_gimple_reg} returns true
678 is considered real, otherwise it is a virtual operand. We also
679 distinguish between uses and definitions. An operand is used if its
680 value is loaded by the statement (e.g., the operand at the RHS of an
681 assignment). If the statement assigns a new value to the operand, the
682 operand is considered a definition (e.g., the operand at the LHS of
685 Virtual and real operands also have very different data flow
686 properties. Real operands are unambiguous references to the
687 full object that they represent. For instance, given
696 Since @code{a} and @code{b} are non-aliased locals, the statement
697 @code{a = b} will have one real definition and one real use because
698 variable @code{b} is completely modified with the contents of
699 variable @code{a}. Real definition are also known as @dfn{killing
700 definitions}. Similarly, the use of @code{a} reads all its bits.
702 In contrast, virtual operands are used with variables that can have
703 a partial or ambiguous reference. This includes structures, arrays,
704 globals, and aliased variables. In these cases, we have two types of
705 definitions. For globals, structures, and arrays, we can determine from
706 a statement whether a variable of these types has a killing definition.
707 If the variable does, then the statement is marked as having a
708 @dfn{must definition} of that variable. However, if a statement is only
709 defining a part of the variable (ie. a field in a structure), or if we
710 know that a statement might define the variable but we cannot say for sure,
711 then we mark that statement as having a @dfn{may definition}. For
727 The assignment @code{*p = 5} may be a definition of @code{a} or
728 @code{b}. If we cannot determine statically where @code{p} is
729 pointing to at the time of the store operation, we create virtual
730 definitions to mark that statement as a potential definition site for
731 @code{a} and @code{b}. Memory loads are similarly marked with virtual
732 use operands. Virtual operands are shown in tree dumps right before
733 the statement that contains them. To request a tree dump with virtual
734 operands, use the @option{-vops} option to @option{-fdump-tree}:
754 Notice that @code{V_MAY_DEF} operands have two copies of the referenced
755 variable. This indicates that this is not a killing definition of
756 that variable. In this case we refer to it as a @dfn{may definition}
757 or @dfn{aliased store}. The presence of the second copy of the
758 variable in the @code{V_MAY_DEF} operand will become important when the
759 function is converted into SSA form. This will be used to link all
760 the non-killing definitions to prevent optimizations from making
761 incorrect assumptions about them.
763 Operands are collected by @file{tree-ssa-operands.c}. They are stored
764 inside each statement's annotation and can be accessed with
765 @code{DEF_OPS}, @code{USE_OPS}, @code{V_MAY_DEF_OPS},
766 @code{V_MUST_DEF_OPS} and @code{VUSE_OPS}. The following are all the
767 accessor macros available to access USE operands. To access all the
768 other operand arrays, just change the name accordingly:
770 @defmac USE_OPS (@var{ann})
771 Returns the array of operands used by the statement with annotation
775 @defmac STMT_USE_OPS (@var{stmt})
776 Alternate version of USE_OPS that takes the statement @var{stmt} as
780 @defmac NUM_USES (@var{ops})
781 Return the number of USE operands in array @var{ops}.
784 @defmac USE_OP_PTR (@var{ops}, @var{i})
785 Return a pointer to the @var{i}th operand in array @var{ops}.
788 @defmac USE_OP (@var{ops}, @var{i})
789 Return the @var{i}th operand in array @var{ops}.
792 The following function shows how to print all the operands of a given
797 print_ops (tree stmt)
800 v_may_def_optype v_may_defs;
801 v_must_def_optype v_must_defs;
807 get_stmt_operands (stmt);
808 ann = stmt_ann (stmt);
810 defs = DEF_OPS (ann);
811 for (i = 0; i < NUM_DEFS (defs); i++)
812 print_generic_expr (stderr, DEF_OP (defs, i), 0);
814 uses = USE_OPS (ann);
815 for (i = 0; i < NUM_USES (uses); i++)
816 print_generic_expr (stderr, USE_OP (uses, i), 0);
818 v_may_defs = V_MAY_DEF_OPS (ann);
819 for (i = 0; i < NUM_V_MAY_DEFS (v_may_defs); i++)
820 print_generic_expr (stderr, V_MAY_DEF_OP (v_may_defs, i), 0);
822 v_must_defs = V_MUST_DEF_OPS (ann);
823 for (i = 0; i < NUM_V_MUST_DEFS (v_must_defs); i++)
824 print_generic_expr (stderr, V_MUST_DEF_OP (v_must_defs, i), 0);
826 vuses = VUSE_OPS (ann);
827 for (i = 0; i < NUM_VUSES (vuses); i++)
828 print_generic_expr (stderr, VUSE_OP (vuses, i), 0);
832 To collect the operands, you first need to call
833 @code{get_stmt_operands}. Since that is a potentially expensive
834 operation, statements are only scanned if they have been marked
835 modified by a call to @code{modify_stmt}. So, if your pass replaces
836 operands in a statement, make sure to call @code{modify_stmt}.
840 @section Static Single Assignment
842 @cindex static single assignment
844 Most of the tree optimizers rely on the data flow information provided
845 by the Static Single Assignment (SSA) form. We implement the SSA form
846 as described in @cite{R. Cytron, J. Ferrante, B. Rosen, M. Wegman, and
847 K. Zadeck. Efficiently Computing Static Single Assignment Form and the
848 Control Dependence Graph. ACM Transactions on Programming Languages
849 and Systems, 13(4):451-490, October 1991}.
851 The SSA form is based on the premise that program variables are
852 assigned in exactly one location in the program. Multiple assignments
853 to the same variable create new versions of that variable. Naturally,
854 actual programs are seldom in SSA form initially because variables
855 tend to be assigned multiple times. The compiler modifies the program
856 representation so that every time a variable is assigned in the code,
857 a new version of the variable is created. Different versions of the
858 same variable are distinguished by subscripting the variable name with
859 its version number. Variables used in the right-hand side of
860 expressions are renamed so that their version number matches that of
861 the most recent assignment.
863 We represent variable versions using @code{SSA_NAME} nodes. The
864 renaming process in @file{tree-ssa.c} wraps every real and
865 virtual operand with an @code{SSA_NAME} node which contains
866 the version number and the statement that created the
867 @code{SSA_NAME}. Only definitions and virtual definitions may
868 create new @code{SSA_NAME} nodes.
870 Sometimes, flow of control makes it impossible to determine what is the
871 most recent version of a variable. In these cases, the compiler
872 inserts an artificial definition for that variable called
873 @dfn{PHI function} or @dfn{PHI node}. This new definition merges
874 all the incoming versions of the variable to create a new name
875 for it. For instance,
885 # a_4 = PHI <a_1, a_2, a_3>
889 Since it is not possible to determine which of the three branches
890 will be taken at runtime, we don't know which of @code{a_1},
891 @code{a_2} or @code{a_3} to use at the return statement. So, the
892 SSA renamer creates a new version @code{a_4} which is assigned
893 the result of ``merging'' @code{a_1}, @code{a_2} and @code{a_3}.
894 Hence, PHI nodes mean ``one of these operands. I don't know
897 The following macros can be used to examine PHI nodes
899 @defmac PHI_RESULT (@var{phi})
900 Returns the @code{SSA_NAME} created by PHI node @var{phi} (i.e.,
904 @defmac PHI_NUM_ARGS (@var{phi})
905 Returns the number of arguments in @var{phi}. This number is exactly
906 the number of incoming edges to the basic block holding @var{phi}@.
909 @defmac PHI_ARG_ELT (@var{phi}, @var{i})
910 Returns a tuple representing the @var{i}th argument of @var{phi}@.
911 Each element of this tuple contains an @code{SSA_NAME} @var{var} and
912 the incoming edge through which @var{var} flows.
915 @defmac PHI_ARG_EDGE (@var{phi}, @var{i})
916 Returns the incoming edge for the @var{i}th argument of @var{phi}.
919 @defmac PHI_ARG_DEF (@var{phi}, @var{i})
920 Returns the @code{SSA_NAME} for the @var{i}th argument of @var{phi}.
924 @subsection Preserving the SSA form
925 @findex vars_to_rename
926 @cindex preserving SSA form
927 Some optimization passes make changes to the function that
928 invalidate the SSA property. This can happen when a pass has
929 added new variables or changed the program so that variables that
930 were previously aliased aren't anymore.
932 Whenever something like this happens, the affected variables must
933 be renamed into SSA form again. To do this, you should mark the
934 new variables in the global bitmap @code{vars_to_rename}. Once
935 your pass has finished, the pass manager will invoke the SSA
936 renamer to put the program into SSA once more.
938 @subsection Examining @code{SSA_NAME} nodes
939 @cindex examining SSA_NAMEs
941 The following macros can be used to examine @code{SSA_NAME} nodes
943 @defmac SSA_NAME_DEF_STMT (@var{var})
944 Returns the statement @var{s} that creates the @code{SSA_NAME}
945 @var{var}. If @var{s} is an empty statement (i.e., @code{IS_EMPTY_STMT
946 (@var{s})} returns @code{true}), it means that the first reference to
947 this variable is a USE or a VUSE@.
950 @defmac SSA_NAME_VERSION (@var{var})
951 Returns the version number of the @code{SSA_NAME} object @var{var}.
955 @subsection Walking use-def chains
957 @deftypefn {Tree SSA function} void walk_use_def_chains (@var{var}, @var{fn}, @var{data})
959 Walks use-def chains starting at the @code{SSA_NAME} node @var{var}.
960 Calls function @var{fn} at each reaching definition found. Function
961 @var{FN} takes three arguments: @var{var}, its defining statement
962 (@var{def_stmt}) and a generic pointer to whatever state information
963 that @var{fn} may want to maintain (@var{data}). Function @var{fn} is
964 able to stop the walk by returning @code{true}, otherwise in order to
965 continue the walk, @var{fn} should return @code{false}.
967 Note, that if @var{def_stmt} is a @code{PHI} node, the semantics are
968 slightly different. For each argument @var{arg} of the PHI node, this
972 @item Walk the use-def chains for @var{arg}.
973 @item Call @code{FN (@var{arg}, @var{phi}, @var{data})}.
976 Note how the first argument to @var{fn} is no longer the original
977 variable @var{var}, but the PHI argument currently being examined.
978 If @var{fn} wants to get at @var{var}, it should call
979 @code{PHI_RESULT} (@var{phi}).
982 @subsection Walking the dominator tree
984 @deftypefn {Tree SSA function} void walk_dominator_tree (@var{walk_data}, @var{bb})
986 This function walks the dominator tree for the current CFG calling a
987 set of callback functions defined in @var{struct dom_walk_data} in
988 @file{domwalk.h}. The call back functions you need to define give you
989 hooks to execute custom code at various points during traversal:
992 @item Once to initialize any local data needed while processing
993 @var{bb} and its children. This local data is pushed into an
994 internal stack which is automatically pushed and popped as the
995 walker traverses the dominator tree.
997 @item Once before traversing all the statements in the @var{bb}.
999 @item Once for every statement inside @var{bb}.
1001 @item Once after traversing all the statements and before recursing
1002 into @var{bb}'s dominator children.
1004 @item It then recurses into all the dominator children of @var{bb}.
1006 @item After recursing into all the dominator children of @var{bb} it
1007 can, optionally, traverse every statement in @var{bb} again
1008 (i.e., repeating steps 2 and 3).
1010 @item Once after walking the statements in @var{bb} and @var{bb}'s
1011 dominator children. At this stage, the block local data stack
1016 @node Alias analysis
1017 @section Alias analysis
1019 @cindex flow-sensitive alias analysis
1020 @cindex flow-insensitive alias analysis
1022 Alias analysis proceeds in 3 main phases:
1025 @item Points-to and escape analysis.
1027 This phase walks the use-def chains in the SSA web looking for
1031 @item Assignments of the form @code{P_i = &VAR}
1032 @item Assignments of the form P_i = malloc()
1033 @item Pointers and ADDR_EXPR that escape the current function.
1036 The concept of `escaping' is the same one used in the Java world.
1037 When a pointer or an ADDR_EXPR escapes, it means that it has been
1038 exposed outside of the current function. So, assignment to
1039 global variables, function arguments and returning a pointer are
1042 This is where we are currently limited. Since not everything is
1043 renamed into SSA, we lose track of escape properties when a
1044 pointer is stashed inside a field in a structure, for instance.
1045 In those cases, we are assuming that the pointer does escape.
1047 We use escape analysis to determine whether a variable is
1048 call-clobbered. Simply put, if an ADDR_EXPR escapes, then the
1049 variable is call-clobbered. If a pointer P_i escapes, then all
1050 the variables pointed-to by P_i (and its memory tag) also escape.
1052 @item Compute flow-sensitive aliases
1054 We have two classes of memory tags. Memory tags associated with
1055 the pointed-to data type of the pointers in the program. These
1056 tags are called ``type memory tag'' (TMT). The other class are
1057 those associated with SSA_NAMEs, called ``name memory tag'' (NMT).
1058 The basic idea is that when adding operands for an INDIRECT_REF
1059 *P_i, we will first check whether P_i has a name tag, if it does
1060 we use it, because that will have more precise aliasing
1061 information. Otherwise, we use the standard type tag.
1063 In this phase, we go through all the pointers we found in
1064 points-to analysis and create alias sets for the name memory tags
1065 associated with each pointer P_i. If P_i escapes, we mark
1066 call-clobbered the variables it points to and its tag.
1069 @item Compute flow-insensitive aliases
1071 This pass will compare the alias set of every type memory tag and
1072 every addressable variable found in the program. Given a type
1073 memory tag TMT and an addressable variable V@. If the alias sets
1074 of TMT and V conflict (as computed by may_alias_p), then V is
1075 marked as an alias tag and added to the alias set of TMT@.
1078 For instance, consider the following function:
1097 After aliasing analysis has finished, the type memory tag for
1098 pointer @code{p} will have two aliases, namely variables @code{a} and
1100 Every time pointer @code{p} is dereferenced, we want to mark the
1101 operation as a potential reference to @code{a} and @code{b}.
1112 # p_1 = PHI <p_4(1), p_6(2)>;
1114 # a_7 = V_MAY_DEF <a_3>;
1115 # b_8 = V_MAY_DEF <b_5>;
1118 # a_9 = V_MAY_DEF <a_7>
1128 In certain cases, the list of may aliases for a pointer may grow
1129 too large. This may cause an explosion in the number of virtual
1130 operands inserted in the code. Resulting in increased memory
1131 consumption and compilation time.
1133 When the number of virtual operands needed to represent aliased
1134 loads and stores grows too large (configurable with @option{--param
1135 max-aliased-vops}), alias sets are grouped to avoid severe
1136 compile-time slow downs and memory consumption. The alias
1137 grouping heuristic proceeds as follows:
1140 @item Sort the list of pointers in decreasing number of contributed
1143 @item Take the first pointer from the list and reverse the role
1144 of the memory tag and its aliases. Usually, whenever an
1145 aliased variable Vi is found to alias with a memory tag
1146 T, we add Vi to the may-aliases set for T@. Meaning that
1147 after alias analysis, we will have:
1150 may-aliases(T) = @{ V1, V2, V3, ..., Vn @}
1153 This means that every statement that references T, will get
1154 @code{n} virtual operands for each of the Vi tags. But, when
1155 alias grouping is enabled, we make T an alias tag and add it
1156 to the alias set of all the Vi variables:
1159 may-aliases(V1) = @{ T @}
1160 may-aliases(V2) = @{ T @}
1162 may-aliases(Vn) = @{ T @}
1165 This has two effects: (a) statements referencing T will only get
1166 a single virtual operand, and, (b) all the variables Vi will now
1167 appear to alias each other. So, we lose alias precision to
1168 improve compile time. But, in theory, a program with such a high
1169 level of aliasing should not be very optimizable in the first
1172 @item Since variables may be in the alias set of more than one
1173 memory tag, the grouping done in step (2) needs to be extended
1174 to all the memory tags that have a non-empty intersection with
1175 the may-aliases set of tag T@. For instance, if we originally
1176 had these may-aliases sets:
1179 may-aliases(T) = @{ V1, V2, V3 @}
1180 may-aliases(R) = @{ V2, V4 @}
1183 In step (2) we would have reverted the aliases for T as:
1186 may-aliases(V1) = @{ T @}
1187 may-aliases(V2) = @{ T @}
1188 may-aliases(V3) = @{ T @}
1191 But note that now V2 is no longer aliased with R@. We could
1192 add R to may-aliases(V2), but we are in the process of
1193 grouping aliases to reduce virtual operands so what we do is
1194 add V4 to the grouping to obtain:
1197 may-aliases(V1) = @{ T @}
1198 may-aliases(V2) = @{ T @}
1199 may-aliases(V3) = @{ T @}
1200 may-aliases(V4) = @{ T @}
1203 @item If the total number of virtual operands due to aliasing is
1204 still above the threshold set by max-alias-vops, go back to (2).