@c markers: BUG TODO @c Copyright (C) 1988-2014 Free Software Foundation, Inc. @c This is part of the GCC manual. @c For copying conditions, see the file gcc.texi. @node Passes @chapter Passes and Files of the Compiler @cindex passes and files of the compiler @cindex files and passes of the compiler @cindex compiler passes and files @cindex pass dumps This chapter is dedicated to giving an overview of the optimization and code generation passes of the compiler. In the process, it describes some of the language front end interface, though this description is no where near complete. @menu * Parsing pass:: The language front end turns text into bits. * Cilk Plus Transformation:: Transform Cilk Plus Code to equivalent C/C++. * Gimplification pass:: The bits are turned into something we can optimize. * Pass manager:: Sequencing the optimization passes. * Tree SSA passes:: Optimizations on a high-level representation. * RTL passes:: Optimizations on a low-level representation. * Optimization info:: Dumping optimization information from passes. @end menu @node Parsing pass @section Parsing pass @cindex GENERIC @findex lang_hooks.parse_file The language front end is invoked only once, via @code{lang_hooks.parse_file}, to parse the entire input. The language front end may use any intermediate language representation deemed appropriate. The C front end uses GENERIC trees (@pxref{GENERIC}), plus a double handful of language specific tree codes defined in @file{c-common.def}. The Fortran front end uses a completely different private representation. @cindex GIMPLE @cindex gimplification @cindex gimplifier @cindex language-independent intermediate representation @cindex intermediate representation lowering @cindex lowering, language-dependent intermediate representation At some point the front end must translate the representation used in the front end to a representation understood by the language-independent portions of the compiler. Current practice takes one of two forms. The C front end manually invokes the gimplifier (@pxref{GIMPLE}) on each function, and uses the gimplifier callbacks to convert the language-specific tree nodes directly to GIMPLE before passing the function off to be compiled. The Fortran front end converts from a private representation to GENERIC, which is later lowered to GIMPLE when the function is compiled. Which route to choose probably depends on how well GENERIC (plus extensions) can be made to match up with the source language and necessary parsing data structures. BUG: Gimplification must occur before nested function lowering, and nested function lowering must be done by the front end before passing the data off to cgraph. TODO: Cgraph should control nested function lowering. It would only be invoked when it is certain that the outer-most function is used. TODO: Cgraph needs a gimplify_function callback. It should be invoked when (1) it is certain that the function is used, (2) warning flags specified by the user require some amount of compilation in order to honor, (3) the language indicates that semantic analysis is not complete until gimplification occurs. Hum@dots{} this sounds overly complicated. Perhaps we should just have the front end gimplify always; in most cases it's only one function call. The front end needs to pass all function definitions and top level declarations off to the middle-end so that they can be compiled and emitted to the object file. For a simple procedural language, it is usually most convenient to do this as each top level declaration or definition is seen. There is also a distinction to be made between generating functional code and generating complete debug information. The only thing that is absolutely required for functional code is that function and data @emph{definitions} be passed to the middle-end. For complete debug information, function, data and type declarations should all be passed as well. @findex rest_of_decl_compilation @findex rest_of_type_compilation @findex cgraph_finalize_function In any case, the front end needs each complete top-level function or data declaration, and each data definition should be passed to @code{rest_of_decl_compilation}. Each complete type definition should be passed to @code{rest_of_type_compilation}. Each function definition should be passed to @code{cgraph_finalize_function}. TODO: I know rest_of_compilation currently has all sorts of RTL generation semantics. I plan to move all code generation bits (both Tree and RTL) to compile_function. Should we hide cgraph from the front ends and move back to rest_of_compilation as the official interface? Possibly we should rename all three interfaces such that the names match in some meaningful way and that is more descriptive than "rest_of". The middle-end will, at its option, emit the function and data definitions immediately or queue them for later processing. @node Cilk Plus Transformation @section Cilk Plus Transformation @cindex CILK_PLUS If Cilk Plus generation (flag @option{-fcilkplus}) is enabled, all the Cilk Plus code is transformed into equivalent C and C++ functions. Majority of this transformation occurs toward the end of the parsing and right before the gimplification pass. These are the major components to the Cilk Plus language extension: @itemize @bullet @item Array Notations: During parsing phase, all the array notation specific information is stored in @code{ARRAY_NOTATION_REF} tree using the function @code{c_parser_array_notation}. During the end of parsing, we check the entire function to see if there are any array notation specific code (using the function @code{contains_array_notation_expr}). If this function returns true, then we expand them using either @code{expand_array_notation_exprs} or @code{build_array_notation_expr}. For the cases where array notations are inside conditions, they are transformed using the function @code{fix_conditional_array_notations}. The C language-specific routines are located in @file{c/c-array-notation.c} and the equivalent C++ routines are in the file @file{cp/cp-array-notation.c}. Common routines such as functions to initialize built-in functions are stored in @file{array-notation-common.c}. @item Cilk keywords: @itemize @bullet @item @code{_Cilk_spawn}: The @code{_Cilk_spawn} keyword is parsed and the function it contains is marked as a spawning function. The spawning function is called the spawner. At the end of the parsing phase, appropriate built-in functions are added to the spawner that are defined in the Cilk runtime. The appropriate locations of these functions, and the internal structures are detailed in @code{cilk_init_builtins} in the file @file{cilk-common.c}. The pointers to Cilk functions and fields of internal structures are described in @file{cilk.h}. The built-in functions are described in @file{cilk-builtins.def}. During gimplification, a new "spawn-helper" function is created. The spawned function is replaced with a spawn helper function in the spawner. The spawned function-call is moved into the spawn helper. The main function that does these transformations is @code{gimplify_cilk_spawn} in @file{c-family/cilk.c}. In the spawn-helper, the gimplification function @code{gimplify_call_expr}, inserts a function call @code{__cilkrts_detach}. This function is expanded by @code{builtin_expand_cilk_detach} located in @file{c-family/cilk.c}. @item @code{_Cilk_sync}: @code{_Cilk_sync} is parsed like a keyword. During gimplification, the function @code{gimplify_cilk_sync} in @file{c-family/cilk.c}, will replace this keyword with a set of functions that are stored in the Cilk runtime. One of the internal functions inserted during gimplification, @code{__cilkrts_pop_frame} must be expanded by the compiler and is done by @code{builtin_expand_cilk_pop_frame} in @file{cilk-common.c}. @end itemize @end itemize Documentation about Cilk Plus and language specification is provided under the "Learn" section in @w{@uref{http://www.cilkplus.org/}}. It is worth mentioning that the current implementation follows ABI 1.1. @node Gimplification pass @section Gimplification pass @cindex gimplification @cindex GIMPLE @dfn{Gimplification} is a whimsical term for the process of converting the intermediate representation of a function into the GIMPLE language (@pxref{GIMPLE}). The term stuck, and so words like ``gimplification'', ``gimplify'', ``gimplifier'' and the like are sprinkled throughout this section of code. While a front end may certainly choose to generate GIMPLE directly if it chooses, this can be a moderately complex process unless the intermediate language used by the front end is already fairly simple. Usually it is easier to generate GENERIC trees plus extensions and let the language-independent gimplifier do most of the work. @findex gimplify_function_tree @findex gimplify_expr @findex lang_hooks.gimplify_expr The main entry point to this pass is @code{gimplify_function_tree} located in @file{gimplify.c}. From here we process the entire function gimplifying each statement in turn. The main workhorse for this pass is @code{gimplify_expr}. Approximately everything passes through here at least once, and it is from here that we invoke the @code{lang_hooks.gimplify_expr} callback. The callback should examine the expression in question and return @code{GS_UNHANDLED} if the expression is not a language specific construct that requires attention. Otherwise it should alter the expression in some way to such that forward progress is made toward producing valid GIMPLE@. If the callback is certain that the transformation is complete and the expression is valid GIMPLE, it should return @code{GS_ALL_DONE}. Otherwise it should return @code{GS_OK}, which will cause the expression to be processed again. If the callback encounters an error during the transformation (because the front end is relying on the gimplification process to finish semantic checks), it should return @code{GS_ERROR}. @node Pass manager @section Pass manager The pass manager is located in @file{passes.c}, @file{tree-optimize.c} and @file{tree-pass.h}. It processes passes as described in @file{passes.def}. Its job is to run all of the individual passes in the correct order, and take care of standard bookkeeping that applies to every pass. The theory of operation is that each pass defines a structure that represents everything we need to know about that pass---when it should be run, how it should be run, what intermediate language form or on-the-side data structures it needs. We register the pass to be run in some particular order, and the pass manager arranges for everything to happen in the correct order. The actuality doesn't completely live up to the theory at present. Command-line switches and @code{timevar_id_t} enumerations must still be defined elsewhere. The pass manager validates constraints but does not attempt to (re-)generate data structures or lower intermediate language form based on the requirements of the next pass. Nevertheless, what is present is useful, and a far sight better than nothing at all. Each pass should have a unique name. Each pass may have its own dump file (for GCC debugging purposes). Passes with a name starting with a star do not dump anything. Sometimes passes are supposed to share a dump file / option name. To still give these unique names, you can use a prefix that is delimited by a space from the part that is used for the dump file / option name. E.g. When the pass name is "ud dce", the name used for dump file/options is "dce". TODO: describe the global variables set up by the pass manager, and a brief description of how a new pass should use it. I need to look at what info RTL passes use first@enddots{} @node Tree SSA passes @section Tree SSA passes The following briefly describes the Tree optimization passes that are run after gimplification and what source files they are located in. @itemize @bullet @item Remove useless statements This pass is an extremely simple sweep across the gimple code in which we identify obviously dead code and remove it. Here we do things like simplify @code{if} statements with constant conditions, remove exception handling constructs surrounding code that obviously cannot throw, remove lexical bindings that contain no variables, and other assorted simplistic cleanups. The idea is to get rid of the obvious stuff quickly rather than wait until later when it's more work to get rid of it. This pass is located in @file{tree-cfg.c} and described by @code{pass_remove_useless_stmts}. @item OpenMP lowering If OpenMP generation (@option{-fopenmp}) is enabled, this pass lowers OpenMP constructs into GIMPLE. Lowering of OpenMP constructs involves creating replacement expressions for local variables that have been mapped using data sharing clauses, exposing the control flow of most synchronization directives and adding region markers to facilitate the creation of the control flow graph. The pass is located in @file{omp-low.c} and is described by @code{pass_lower_omp}. @item OpenMP expansion If OpenMP generation (@option{-fopenmp}) is enabled, this pass expands parallel regions into their own functions to be invoked by the thread library. The pass is located in @file{omp-low.c} and is described by @code{pass_expand_omp}. @item Lower control flow This pass flattens @code{if} statements (@code{COND_EXPR}) and moves lexical bindings (@code{BIND_EXPR}) out of line. After this pass, all @code{if} statements will have exactly two @code{goto} statements in its @code{then} and @code{else} arms. Lexical binding information for each statement will be found in @code{TREE_BLOCK} rather than being inferred from its position under a @code{BIND_EXPR}. This pass is found in @file{gimple-low.c} and is described by @code{pass_lower_cf}. @item Lower exception handling control flow This pass decomposes high-level exception handling constructs (@code{TRY_FINALLY_EXPR} and @code{TRY_CATCH_EXPR}) into a form that explicitly represents the control flow involved. After this pass, @code{lookup_stmt_eh_region} will return a non-negative number for any statement that may have EH control flow semantics; examine @code{tree_can_throw_internal} or @code{tree_can_throw_external} for exact semantics. Exact control flow may be extracted from @code{foreach_reachable_handler}. The EH region nesting tree is defined in @file{except.h} and built in @file{except.c}. The lowering pass itself is in @file{tree-eh.c} and is described by @code{pass_lower_eh}. @item Build the control flow graph This pass decomposes a function into basic blocks and creates all of the edges that connect them. It is located in @file{tree-cfg.c} and is described by @code{pass_build_cfg}. @item Find all referenced variables This pass walks the entire function and collects an array of all variables referenced in the function, @code{referenced_vars}. The index at which a variable is found in the array is used as a UID for the variable within this function. This data is needed by the SSA rewriting routines. The pass is located in @file{tree-dfa.c} and is described by @code{pass_referenced_vars}. @item Enter static single assignment form This pass rewrites the function such that it is in SSA form. After this pass, all @code{is_gimple_reg} variables will be referenced by @code{SSA_NAME}, and all occurrences of other variables will be annotated with @code{VDEFS} and @code{VUSES}; PHI nodes will have been inserted as necessary for each basic block. This pass is located in @file{tree-ssa.c} and is described by @code{pass_build_ssa}. @item Warn for uninitialized variables This pass scans the function for uses of @code{SSA_NAME}s that are fed by default definition. For non-parameter variables, such uses are uninitialized. The pass is run twice, before and after optimization (if turned on). In the first pass we only warn for uses that are positively uninitialized; in the second pass we warn for uses that are possibly uninitialized. The pass is located in @file{tree-ssa.c} and is defined by @code{pass_early_warn_uninitialized} and @code{pass_late_warn_uninitialized}. @item Dead code elimination This pass scans the function for statements without side effects whose result is unused. It does not do memory life analysis, so any value that is stored in memory is considered used. The pass is run multiple times throughout the optimization process. It is located in @file{tree-ssa-dce.c} and is described by @code{pass_dce}. @item Dominator optimizations This pass performs trivial dominator-based copy and constant propagation, expression simplification, and jump threading. It is run multiple times throughout the optimization process. It is located in @file{tree-ssa-dom.c} and is described by @code{pass_dominator}. @item Forward propagation of single-use variables This pass attempts to remove redundant computation by substituting variables that are used once into the expression that uses them and seeing if the result can be simplified. It is located in @file{tree-ssa-forwprop.c} and is described by @code{pass_forwprop}. @item Copy Renaming This pass attempts to change the name of compiler temporaries involved in copy operations such that SSA->normal can coalesce the copy away. When compiler temporaries are copies of user variables, it also renames the compiler temporary to the user variable resulting in better use of user symbols. It is located in @file{tree-ssa-copyrename.c} and is described by @code{pass_copyrename}. @item PHI node optimizations This pass recognizes forms of PHI inputs that can be represented as conditional expressions and rewrites them into straight line code. It is located in @file{tree-ssa-phiopt.c} and is described by @code{pass_phiopt}. @item May-alias optimization This pass performs a flow sensitive SSA-based points-to analysis. The resulting may-alias, must-alias, and escape analysis information is used to promote variables from in-memory addressable objects to non-aliased variables that can be renamed into SSA form. We also update the @code{VDEF}/@code{VUSE} memory tags for non-renameable aggregates so that we get fewer false kills. The pass is located in @file{tree-ssa-alias.c} and is described by @code{pass_may_alias}. Interprocedural points-to information is located in @file{tree-ssa-structalias.c} and described by @code{pass_ipa_pta}. @item Profiling This pass rewrites the function in order to collect runtime block and value profiling data. Such data may be fed back into the compiler on a subsequent run so as to allow optimization based on expected execution frequencies. The pass is located in @file{predict.c} and is described by @code{pass_profile}. @item Lower complex arithmetic This pass rewrites complex arithmetic operations into their component scalar arithmetic operations. The pass is located in @file{tree-complex.c} and is described by @code{pass_lower_complex}. @item Scalar replacement of aggregates This pass rewrites suitable non-aliased local aggregate variables into a set of scalar variables. The resulting scalar variables are rewritten into SSA form, which allows subsequent optimization passes to do a significantly better job with them. The pass is located in @file{tree-sra.c} and is described by @code{pass_sra}. @item Dead store elimination This pass eliminates stores to memory that are subsequently overwritten by another store, without any intervening loads. The pass is located in @file{tree-ssa-dse.c} and is described by @code{pass_dse}. @item Tail recursion elimination This pass transforms tail recursion into a loop. It is located in @file{tree-tailcall.c} and is described by @code{pass_tail_recursion}. @item Forward store motion This pass sinks stores and assignments down the flowgraph closer to their use point. The pass is located in @file{tree-ssa-sink.c} and is described by @code{pass_sink_code}. @item Partial redundancy elimination This pass eliminates partially redundant computations, as well as performing load motion. The pass is located in @file{tree-ssa-pre.c} and is described by @code{pass_pre}. Just before partial redundancy elimination, if @option{-funsafe-math-optimizations} is on, GCC tries to convert divisions to multiplications by the reciprocal. The pass is located in @file{tree-ssa-math-opts.c} and is described by @code{pass_cse_reciprocal}. @item Full redundancy elimination This is a simpler form of PRE that only eliminates redundancies that occur on all paths. It is located in @file{tree-ssa-pre.c} and described by @code{pass_fre}. @item Loop optimization The main driver of the pass is placed in @file{tree-ssa-loop.c} and described by @code{pass_loop}. The optimizations performed by this pass are: Loop invariant motion. This pass moves only invariants that would be hard to handle on RTL level (function calls, operations that expand to nontrivial sequences of insns). With @option{-funswitch-loops} it also moves operands of conditions that are invariant out of the loop, so that we can use just trivial invariantness analysis in loop unswitching. The pass also includes store motion. The pass is implemented in @file{tree-ssa-loop-im.c}. Canonical induction variable creation. This pass creates a simple counter for number of iterations of the loop and replaces the exit condition of the loop using it, in case when a complicated analysis is necessary to determine the number of iterations. Later optimizations then may determine the number easily. The pass is implemented in @file{tree-ssa-loop-ivcanon.c}. Induction variable optimizations. This pass performs standard induction variable optimizations, including strength reduction, induction variable merging and induction variable elimination. The pass is implemented in @file{tree-ssa-loop-ivopts.c}. Loop unswitching. This pass moves the conditional jumps that are invariant out of the loops. To achieve this, a duplicate of the loop is created for each possible outcome of conditional jump(s). The pass is implemented in @file{tree-ssa-loop-unswitch.c}. This pass should eventually replace the RTL level loop unswitching in @file{loop-unswitch.c}, but currently the RTL level pass is not completely redundant yet due to deficiencies in tree level alias analysis. The optimizations also use various utility functions contained in @file{tree-ssa-loop-manip.c}, @file{cfgloop.c}, @file{cfgloopanal.c} and @file{cfgloopmanip.c}. Vectorization. This pass transforms loops to operate on vector types instead of scalar types. Data parallelism across loop iterations is exploited to group data elements from consecutive iterations into a vector and operate on them in parallel. Depending on available target support the loop is conceptually unrolled by a factor @code{VF} (vectorization factor), which is the number of elements operated upon in parallel in each iteration, and the @code{VF} copies of each scalar operation are fused to form a vector operation. Additional loop transformations such as peeling and versioning may take place to align the number of iterations, and to align the memory accesses in the loop. The pass is implemented in @file{tree-vectorizer.c} (the main driver), @file{tree-vect-loop.c} and @file{tree-vect-loop-manip.c} (loop specific parts and general loop utilities), @file{tree-vect-slp} (loop-aware SLP functionality), @file{tree-vect-stmts.c} and @file{tree-vect-data-refs.c}. Analysis of data references is in @file{tree-data-ref.c}. SLP Vectorization. This pass performs vectorization of straight-line code. The pass is implemented in @file{tree-vectorizer.c} (the main driver), @file{tree-vect-slp.c}, @file{tree-vect-stmts.c} and @file{tree-vect-data-refs.c}. Autoparallelization. This pass splits the loop iteration space to run into several threads. The pass is implemented in @file{tree-parloops.c}. Graphite is a loop transformation framework based on the polyhedral model. Graphite stands for Gimple Represented as Polyhedra. The internals of this infrastructure are documented in @w{@uref{http://gcc.gnu.org/wiki/Graphite}}. The passes working on this representation are implemented in the various @file{graphite-*} files. @item Tree level if-conversion for vectorizer This pass applies if-conversion to simple loops to help vectorizer. We identify if convertible loops, if-convert statements and merge basic blocks in one big block. The idea is to present loop in such form so that vectorizer can have one to one mapping between statements and available vector operations. This pass is located in @file{tree-if-conv.c} and is described by @code{pass_if_conversion}. @item Conditional constant propagation This pass relaxes a lattice of values in order to identify those that must be constant even in the presence of conditional branches. The pass is located in @file{tree-ssa-ccp.c} and is described by @code{pass_ccp}. A related pass that works on memory loads and stores, and not just register values, is located in @file{tree-ssa-ccp.c} and described by @code{pass_store_ccp}. @item Conditional copy propagation This is similar to constant propagation but the lattice of values is the ``copy-of'' relation. It eliminates redundant copies from the code. The pass is located in @file{tree-ssa-copy.c} and described by @code{pass_copy_prop}. A related pass that works on memory copies, and not just register copies, is located in @file{tree-ssa-copy.c} and described by @code{pass_store_copy_prop}. @item Value range propagation This transformation is similar to constant propagation but instead of propagating single constant values, it propagates known value ranges. The implementation is based on Patterson's range propagation algorithm (Accurate Static Branch Prediction by Value Range Propagation, J. R. C. Patterson, PLDI '95). In contrast to Patterson's algorithm, this implementation does not propagate branch probabilities nor it uses more than a single range per SSA name. This means that the current implementation cannot be used for branch prediction (though adapting it would not be difficult). The pass is located in @file{tree-vrp.c} and is described by @code{pass_vrp}. @item Folding built-in functions This pass simplifies built-in functions, as applicable, with constant arguments or with inferable string lengths. It is located in @file{tree-ssa-ccp.c} and is described by @code{pass_fold_builtins}. @item Split critical edges This pass identifies critical edges and inserts empty basic blocks such that the edge is no longer critical. The pass is located in @file{tree-cfg.c} and is described by @code{pass_split_crit_edges}. @item Control dependence dead code elimination This pass is a stronger form of dead code elimination that can eliminate unnecessary control flow statements. It is located in @file{tree-ssa-dce.c} and is described by @code{pass_cd_dce}. @item Tail call elimination This pass identifies function calls that may be rewritten into jumps. No code transformation is actually applied here, but the data and control flow problem is solved. The code transformation requires target support, and so is delayed until RTL@. In the meantime @code{CALL_EXPR_TAILCALL} is set indicating the possibility. The pass is located in @file{tree-tailcall.c} and is described by @code{pass_tail_calls}. The RTL transformation is handled by @code{fixup_tail_calls} in @file{calls.c}. @item Warn for function return without value For non-void functions, this pass locates return statements that do not specify a value and issues a warning. Such a statement may have been injected by falling off the end of the function. This pass is run last so that we have as much time as possible to prove that the statement is not reachable. It is located in @file{tree-cfg.c} and is described by @code{pass_warn_function_return}. @item Leave static single assignment form This pass rewrites the function such that it is in normal form. At the same time, we eliminate as many single-use temporaries as possible, so the intermediate language is no longer GIMPLE, but GENERIC@. The pass is located in @file{tree-outof-ssa.c} and is described by @code{pass_del_ssa}. @item Merge PHI nodes that feed into one another This is part of the CFG cleanup passes. It attempts to join PHI nodes from a forwarder CFG block into another block with PHI nodes. The pass is located in @file{tree-cfgcleanup.c} and is described by @code{pass_merge_phi}. @item Return value optimization If a function always returns the same local variable, and that local variable is an aggregate type, then the variable is replaced with the return value for the function (i.e., the function's DECL_RESULT). This is equivalent to the C++ named return value optimization applied to GIMPLE@. The pass is located in @file{tree-nrv.c} and is described by @code{pass_nrv}. @item Return slot optimization If a function returns a memory object and is called as @code{var = foo()}, this pass tries to change the call so that the address of @code{var} is sent to the caller to avoid an extra memory copy. This pass is located in @code{tree-nrv.c} and is described by @code{pass_return_slot}. @item Optimize calls to @code{__builtin_object_size} This is a propagation pass similar to CCP that tries to remove calls to @code{__builtin_object_size} when the size of the object can be computed at compile-time. This pass is located in @file{tree-object-size.c} and is described by @code{pass_object_sizes}. @item Loop invariant motion This pass removes expensive loop-invariant computations out of loops. The pass is located in @file{tree-ssa-loop.c} and described by @code{pass_lim}. @item Loop nest optimizations This is a family of loop transformations that works on loop nests. It includes loop interchange, scaling, skewing and reversal and they are all geared to the optimization of data locality in array traversals and the removal of dependencies that hamper optimizations such as loop parallelization and vectorization. The pass is located in @file{tree-loop-linear.c} and described by @code{pass_linear_transform}. @item Removal of empty loops This pass removes loops with no code in them. The pass is located in @file{tree-ssa-loop-ivcanon.c} and described by @code{pass_empty_loop}. @item Unrolling of small loops This pass completely unrolls loops with few iterations. The pass is located in @file{tree-ssa-loop-ivcanon.c} and described by @code{pass_complete_unroll}. @item Predictive commoning This pass makes the code reuse the computations from the previous iterations of the loops, especially loads and stores to memory. It does so by storing the values of these computations to a bank of temporary variables that are rotated at the end of loop. To avoid the need for this rotation, the loop is then unrolled and the copies of the loop body are rewritten to use the appropriate version of the temporary variable. This pass is located in @file{tree-predcom.c} and described by @code{pass_predcom}. @item Array prefetching This pass issues prefetch instructions for array references inside loops. The pass is located in @file{tree-ssa-loop-prefetch.c} and described by @code{pass_loop_prefetch}. @item Reassociation This pass rewrites arithmetic expressions to enable optimizations that operate on them, like redundancy elimination and vectorization. The pass is located in @file{tree-ssa-reassoc.c} and described by @code{pass_reassoc}. @item Optimization of @code{stdarg} functions This pass tries to avoid the saving of register arguments into the stack on entry to @code{stdarg} functions. If the function doesn't use any @code{va_start} macros, no registers need to be saved. If @code{va_start} macros are used, the @code{va_list} variables don't escape the function, it is only necessary to save registers that will be used in @code{va_arg} macros. For instance, if @code{va_arg} is only used with integral types in the function, floating point registers don't need to be saved. This pass is located in @code{tree-stdarg.c} and described by @code{pass_stdarg}. @end itemize @node RTL passes @section RTL passes The following briefly describes the RTL generation and optimization passes that are run after the Tree optimization passes. @itemize @bullet @item RTL generation @c Avoiding overfull is tricky here. The source files for RTL generation include @file{stmt.c}, @file{calls.c}, @file{expr.c}, @file{explow.c}, @file{expmed.c}, @file{function.c}, @file{optabs.c} and @file{emit-rtl.c}. Also, the file @file{insn-emit.c}, generated from the machine description by the program @code{genemit}, is used in this pass. The header file @file{expr.h} is used for communication within this pass. @findex genflags @findex gencodes The header files @file{insn-flags.h} and @file{insn-codes.h}, generated from the machine description by the programs @code{genflags} and @code{gencodes}, tell this pass which standard names are available for use and which patterns correspond to them. @item Generation of exception landing pads This pass generates the glue that handles communication between the exception handling library routines and the exception handlers within the function. Entry points in the function that are invoked by the exception handling library are called @dfn{landing pads}. The code for this pass is located in @file{except.c}. @item Control flow graph cleanup This pass removes unreachable code, simplifies jumps to next, jumps to jump, jumps across jumps, etc. The pass is run multiple times. For historical reasons, it is occasionally referred to as the ``jump optimization pass''. The bulk of the code for this pass is in @file{cfgcleanup.c}, and there are support routines in @file{cfgrtl.c} and @file{jump.c}. @item Forward propagation of single-def values This pass attempts to remove redundant computation by substituting variables that come from a single definition, and seeing if the result can be simplified. It performs copy propagation and addressing mode selection. The pass is run twice, with values being propagated into loops only on the second run. The code is located in @file{fwprop.c}. @item Common subexpression elimination This pass removes redundant computation within basic blocks, and optimizes addressing modes based on cost. The pass is run twice. The code for this pass is located in @file{cse.c}. @item Global common subexpression elimination This pass performs two different types of GCSE depending on whether you are optimizing for size or not (LCM based GCSE tends to increase code size for a gain in speed, while Morel-Renvoise based GCSE does not). When optimizing for size, GCSE is done using Morel-Renvoise Partial Redundancy Elimination, with the exception that it does not try to move invariants out of loops---that is left to the loop optimization pass. If MR PRE GCSE is done, code hoisting (aka unification) is also done, as well as load motion. If you are optimizing for speed, LCM (lazy code motion) based GCSE is done. LCM is based on the work of Knoop, Ruthing, and Steffen. LCM based GCSE also does loop invariant code motion. We also perform load and store motion when optimizing for speed. Regardless of which type of GCSE is used, the GCSE pass also performs global constant and copy propagation. The source file for this pass is @file{gcse.c}, and the LCM routines are in @file{lcm.c}. @item Loop optimization This pass performs several loop related optimizations. The source files @file{cfgloopanal.c} and @file{cfgloopmanip.c} contain generic loop analysis and manipulation code. Initialization and finalization of loop structures is handled by @file{loop-init.c}. A loop invariant motion pass is implemented in @file{loop-invariant.c}. Basic block level optimizations---unrolling, peeling and unswitching loops--- are implemented in @file{loop-unswitch.c} and @file{loop-unroll.c}. Replacing of the exit condition of loops by special machine-dependent instructions is handled by @file{loop-doloop.c}. @item Jump bypassing This pass is an aggressive form of GCSE that transforms the control flow graph of a function by propagating constants into conditional branch instructions. The source file for this pass is @file{gcse.c}. @item If conversion This pass attempts to replace conditional branches and surrounding assignments with arithmetic, boolean value producing comparison instructions, and conditional move instructions. In the very last invocation after reload/LRA, it will generate predicated instructions when supported by the target. The code is located in @file{ifcvt.c}. @item Web construction This pass splits independent uses of each pseudo-register. This can improve effect of the other transformation, such as CSE or register allocation. The code for this pass is located in @file{web.c}. @item Instruction combination This pass attempts to combine groups of two or three instructions that are related by data flow into single instructions. It combines the RTL expressions for the instructions by substitution, simplifies the result using algebra, and then attempts to match the result against the machine description. The code is located in @file{combine.c}. @item Mode switching optimization This pass looks for instructions that require the processor to be in a specific ``mode'' and minimizes the number of mode changes required to satisfy all users. What these modes are, and what they apply to are completely target-specific. The code for this pass is located in @file{mode-switching.c}. @cindex modulo scheduling @cindex sms, swing, software pipelining @item Modulo scheduling This pass looks at innermost loops and reorders their instructions by overlapping different iterations. Modulo scheduling is performed immediately before instruction scheduling. The code for this pass is located in @file{modulo-sched.c}. @item Instruction scheduling This pass looks for instructions whose output will not be available by the time that it is used in subsequent instructions. Memory loads and floating point instructions often have this behavior on RISC machines. It re-orders instructions within a basic block to try to separate the definition and use of items that otherwise would cause pipeline stalls. This pass is performed twice, before and after register allocation. The code for this pass is located in @file{haifa-sched.c}, @file{sched-deps.c}, @file{sched-ebb.c}, @file{sched-rgn.c} and @file{sched-vis.c}. @item Register allocation These passes make sure that all occurrences of pseudo registers are eliminated, either by allocating them to a hard register, replacing them by an equivalent expression (e.g.@: a constant) or by placing them on the stack. This is done in several subpasses: @itemize @bullet @item The integrated register allocator (@acronym{IRA}). It is called integrated because coalescing, register live range splitting, and hard register preferencing are done on-the-fly during coloring. It also has better integration with the reload/LRA pass. Pseudo-registers spilled by the allocator or the reload/LRA have still a chance to get hard-registers if the reload/LRA evicts some pseudo-registers from hard-registers. The allocator helps to choose better pseudos for spilling based on their live ranges and to coalesce stack slots allocated for the spilled pseudo-registers. IRA is a regional register allocator which is transformed into Chaitin-Briggs allocator if there is one region. By default, IRA chooses regions using register pressure but the user can force it to use one region or regions corresponding to all loops. Source files of the allocator are @file{ira.c}, @file{ira-build.c}, @file{ira-costs.c}, @file{ira-conflicts.c}, @file{ira-color.c}, @file{ira-emit.c}, @file{ira-lives}, plus header files @file{ira.h} and @file{ira-int.h} used for the communication between the allocator and the rest of the compiler and between the IRA files. @cindex reloading @item Reloading. This pass renumbers pseudo registers with the hardware registers numbers they were allocated. Pseudo registers that did not get hard registers are replaced with stack slots. Then it finds instructions that are invalid because a value has failed to end up in a register, or has ended up in a register of the wrong kind. It fixes up these instructions by reloading the problematical values temporarily into registers. Additional instructions are generated to do the copying. The reload pass also optionally eliminates the frame pointer and inserts instructions to save and restore call-clobbered registers around calls. Source files are @file{reload.c} and @file{reload1.c}, plus the header @file{reload.h} used for communication between them. @cindex Local Register Allocator (LRA) @item This pass is a modern replacement of the reload pass. Source files are @file{lra.c}, @file{lra-assign.c}, @file{lra-coalesce.c}, @file{lra-constraints.c}, @file{lra-eliminations.c}, @file{lra-equivs.c}, @file{lra-lives.c}, @file{lra-saves.c}, @file{lra-spills.c}, the header @file{lra-int.h} used for communication between them, and the header @file{lra.h} used for communication between LRA and the rest of compiler. Unlike the reload pass, intermediate LRA decisions are reflected in RTL as much as possible. This reduces the number of target-dependent macros and hooks, leaving instruction constraints as the primary source of control. LRA is run on targets for which TARGET_LRA_P returns true. @end itemize @item Basic block reordering This pass implements profile guided code positioning. If profile information is not available, various types of static analysis are performed to make the predictions normally coming from the profile feedback (IE execution frequency, branch probability, etc). It is implemented in the file @file{bb-reorder.c}, and the various prediction routines are in @file{predict.c}. @item Variable tracking This pass computes where the variables are stored at each position in code and generates notes describing the variable locations to RTL code. The location lists are then generated according to these notes to debug information if the debugging information format supports location lists. The code is located in @file{var-tracking.c}. @item Delayed branch scheduling This optional pass attempts to find instructions that can go into the delay slots of other instructions, usually jumps and calls. The code for this pass is located in @file{reorg.c}. @item Branch shortening On many RISC machines, branch instructions have a limited range. Thus, longer sequences of instructions must be used for long branches. In this pass, the compiler figures out what how far each instruction will be from each other instruction, and therefore whether the usual instructions, or the longer sequences, must be used for each branch. The code for this pass is located in @file{final.c}. @item Register-to-stack conversion Conversion from usage of some hard registers to usage of a register stack may be done at this point. Currently, this is supported only for the floating-point registers of the Intel 80387 coprocessor. The code for this pass is located in @file{reg-stack.c}. @item Final This pass outputs the assembler code for the function. The source files are @file{final.c} plus @file{insn-output.c}; the latter is generated automatically from the machine description by the tool @file{genoutput}. The header file @file{conditions.h} is used for communication between these files. @item Debugging information output This is run after final because it must output the stack slot offsets for pseudo registers that did not get hard registers. Source files are @file{dbxout.c} for DBX symbol table format, @file{sdbout.c} for SDB symbol table format, @file{dwarfout.c} for DWARF symbol table format, files @file{dwarf2out.c} and @file{dwarf2asm.c} for DWARF2 symbol table format, and @file{vmsdbgout.c} for VMS debug symbol table format. @end itemize @node Optimization info @section Optimization info @include optinfo.texi