@c markers: BUG TODO

@c Copyright (C) 1988-2026 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.
* Gimplification pass::  The bits are turned into something we can optimize.
* Pass manager::         Sequencing the optimization passes.
* IPA passes::           Inter-procedural optimizations.
* 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 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.cc}.  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.cc}, @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 IPA passes
@section Inter-procedural optimization passes
@cindex IPA passes
@cindex inter-procedural optimization passes

The inter-procedural optimization (IPA) passes use call graph
information to perform transformations across function boundaries.
IPA is a critical part of link-time optimization (LTO) and
whole-program (WHOPR) optimization, and these passes are structured
with the needs of LTO and WHOPR in mind by dividing their operations
into stages.  For detailed discussion of the LTO/WHOPR IPA pass stages
and interfaces, see @ref{IPA}.

The following briefly describes the inter-procedural optimization (IPA)
passes, which are split into small IPA passes, regular IPA passes,
and late IPA passes, according to the LTO/WHOPR processing model.

@menu
* Small IPA passes::
* Regular IPA passes::
* Late IPA passes::
@end menu

@node Small IPA passes
@subsection Small IPA passes
@cindex small IPA passes
A small IPA pass is a pass derived from @code{simple_ipa_opt_pass}.
As described in @ref{IPA}, it does everything at once and 
defines only the @emph{Execute} stage.  During this
stage it accesses and modifies the function bodies.
No @code{generate_summary}, @code{read_summary}, or @code{write_summary}
hooks are defined.

@itemize @bullet
@item IPA free lang data

This pass frees resources that are used by the front end but are
not needed once it is done.  It is located in @file{tree.cc} and is described by
@code{pass_ipa_free_lang_data}.

@item IPA function and variable visibility

This is a local function pass handling visibilities of all symbols.  This
happens before LTO streaming, so @option{-fwhole-program} should be ignored
at this level.  It is located in @file{ipa-visibility.cc} and is described by
@code{pass_ipa_function_and_variable_visibility}.

@item IPA remove symbols

This pass performs reachability analysis and reclaims all unreachable nodes.
It is located in @file{passes.cc} and is described by
@code{pass_ipa_remove_symbols}.

@item IPA OpenACC

This is a pass group for OpenACC processing.  It is located in
@file{tree-ssa-loop.cc} and is described by @code{pass_ipa_oacc}.

@item IPA points-to analysis

This is a tree-based points-to analysis pass. The idea behind this analyzer
is to generate set constraints from the program, then solve the resulting
constraints in order to generate the points-to sets.  It is located in 
@file{tree-ssa-structalias.cc} and is described by @code{pass_ipa_pta}.

@item IPA OpenACC kernels

This is a pass group for processing OpenACC kernels regions.  It is a
subpass of the IPA OpenACC pass group that runs on offloaded functions
containing OpenACC kernels loops.  It is located in
@file{tree-ssa-loop.cc} and is described by
@code{pass_ipa_oacc_kernels}.

@item Target clone

This is a pass for parsing functions with multiple target attributes.
It is located in @file{multiple_target.cc} and is described by
@code{pass_target_clone}.

@item IPA auto profile

This pass uses AutoFDO profiling data to annotate the control flow graph.
It is located in @file{auto-profile.cc} and is described by
@code{pass_ipa_auto_profile}.

@item IPA tree profile

This pass does profiling for all functions in the call graph. 
It calculates branch
probabilities and basic block execution counts. It is located
in @file{tree-profile.cc} and is described by @code{pass_ipa_tree_profile}.

@item IPA free function summary

This pass is a small IPA pass when argument @code{small_p} is true.
It releases inline function summaries and call summaries.
It is located in @file{ipa-fnsummary.cc} and is described by
@code{pass_ipa_free_free_fn_summary}.

@item IPA increase alignment

This pass increases the alignment of global arrays to improve
vectorization. It is located in @file{tree-vectorizer.cc}
and is described by @code{pass_ipa_increase_alignment}.

@item IPA transactional memory

This pass is for transactional memory support.
It is located in @file{trans-mem.cc} and is described by
@code{pass_ipa_tm}.

@item IPA lower emulated TLS

This pass lowers thread-local storage (TLS) operations
to emulation functions provided by libgcc.
It is located in @file{tree-emutls.cc} and is described by
@code{pass_ipa_lower_emutls}.

@end itemize

@node Regular IPA passes
@subsection Regular IPA passes
@cindex regular IPA passes

A regular IPA pass is a pass derived from @code{ipa_opt_pass_d} that
is executed in WHOPR compilation. Regular IPA passes may have summary
hooks implemented in any of the LGEN, WPA or LTRANS stages (@pxref{IPA}).

@itemize @bullet
@item IPA whole program visibility

This pass performs various optimizations involving symbol visibility
with @option{-fwhole-program}, including symbol privatization,
discovering local functions, and dismantling comdat groups.  It is
located in @file{ipa-visibility.cc} and is described by
@code{pass_ipa_whole_program_visibility}.

@item IPA profile

The IPA profile pass propagates profiling frequencies across the call
graph.  It is located in @file{ipa-profile.cc} and is described by
@code{pass_ipa_profile}.

@item IPA identical code folding

This is the inter-procedural identical code folding pass.
The goal of this transformation is to discover functions
and read-only variables that have exactly the same semantics.  It is
located in @file{ipa-icf.cc} and is described by @code{pass_ipa_icf}.

@item IPA devirtualization

This pass performs speculative devirtualization based on the type
inheritance graph.  When a polymorphic call has only one likely target
in the unit, it is turned into a speculative call. It is located in
@file{ipa-devirt.cc} and is described by @code{pass_ipa_devirt}.

@item IPA constant propagation

The goal of this pass is to discover functions that are always invoked
with some arguments with the same known constant values and to modify
the functions accordingly.  It can also do partial specialization and
type-based devirtualization.  It is located in @file{ipa-cp.cc} and is
described by @code{pass_ipa_cp}.

@item IPA scalar replacement of aggregates

This pass can replace an aggregate parameter with a set of other parameters
representing part of the original, turning those passed by reference
into new ones which pass the value directly.  It also removes unused
function return values and unused function parameters.  This pass is
located in @file{ipa-sra.cc} and is described by @code{pass_ipa_sra}.

@item IPA constructor/destructor merge

This pass merges multiple constructors and destructors for static
objects into single functions.  It's only run at LTO time unless the
target doesn't support constructors and destructors natively.  The
pass is located in @file{ipa.cc} and is described by
@code{pass_ipa_cdtor_merge}.

@item IPA function summary

This pass provides function analysis for inter-procedural passes.
It collects estimates of function body size, execution time, and frame
size for each function.  It also estimates information about function
calls: call statement size, time and how often the parameters change
for each call.  It is located in @file{ipa-fnsummary.cc} and is
described by @code{pass_ipa_fn_summary}.

@item IPA inline

The IPA inline pass handles function inlining with whole-program
knowledge. Small functions that are candidates for inlining are
ordered in increasing badness, bounded by unit growth parameters.
Unreachable functions are removed from the call graph.  Functions called
once and not exported from the unit are inlined.  This pass is located in
@file{ipa-inline.cc} and is described by @code{pass_ipa_inline}.

@item IPA pure/const analysis

This pass marks functions as being either const (@code{TREE_READONLY}) or
pure (@code{DECL_PURE_P}).  The per-function information is produced
by @code{pure_const_generate_summary}, then the global information is computed
by performing a transitive closure over the call graph.   It is located in
@file{ipa-pure-const.cc} and is described by @code{pass_ipa_pure_const}.

@item IPA free function summary

This pass is a regular IPA pass when argument @code{small_p} is false.
It releases inline function summaries and call summaries.
It is located in @file{ipa-fnsummary.cc} and is described by
@code{pass_ipa_free_fn_summary}.

@item IPA reference

This pass gathers information about how variables whose scope is
confined to the compilation unit are used.  It is located in
@file{ipa-reference.cc} and is described by @code{pass_ipa_reference}.

@item IPA single use

This pass checks whether variables are used by a single function.
It is located in @file{ipa.cc} and is described by
@code{pass_ipa_single_use}.

@item IPA comdats

This pass looks for static symbols that are used exclusively
within one comdat group, and moves them into that comdat group. It is
located in @file{ipa-comdats.cc} and is described by
@code{pass_ipa_comdats}.

@end itemize

@node Late IPA passes
@subsection Late IPA passes
@cindex late IPA passes

Late IPA passes are simple IPA passes executed after
the regular passes.  In WHOPR mode the passes are executed after
partitioning and thus see just parts of the compiled unit.

@itemize @bullet
@item Materialize all clones

Once all functions from compilation unit are in memory, produce all clones
and update all calls.  It is located in @file{ipa.cc} and is described by
@code{pass_materialize_all_clones}.

@item IPA points-to analysis

Points-to analysis; this is the same as the points-to-analysis pass
run with the small IPA passes (@pxref{Small IPA passes}).

@item OpenMP simd clone

This is the OpenMP constructs' SIMD clone pass.  It creates the appropriate
SIMD clones for functions tagged as elemental SIMD functions.
It is located in @file{omp-simd-clone.cc} and is described by
@code{pass_omp_simd_clone}.

@end itemize

@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 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.cc} 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.cc} 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.cc} 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.cc}.  The lowering pass
itself is in @file{tree-eh.cc} 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.cc} and
is described by @code{pass_build_cfg}.

@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-into-ssa.cc} 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-uninit.cc}
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 lifetime 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.cc} 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.cc}
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.cc} and is described by @code{pass_forwprop}.

@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.cc} 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-structalias.cc} and is described by @code{pass_may_alias}.

Interprocedural points-to information is described by @code{pass_ipa_pta}.

@item Profiling

This pass instruments 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{tree-profile.cc} and
is described by @code{pass_ipa_tree_profile}.

@item Static profile estimation

This pass implements series of heuristics to guess propababilities
of branches.  The resulting predictions are turned into edge profile
by propagating branches across the control flow graphs.
The pass is located in @file{tree-profile.cc} 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.cc}
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.cc} 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.cc} 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.cc} 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.cc} 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.cc}
and is described by @code{pass_pre}.

@item CSE of reciprocals

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.cc} 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.cc} and
described by @code{pass_fre}.

@item Loop optimization

The main driver of the pass is placed in @file{tree-ssa-loop.cc}
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.cc}.

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.cc}.

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.cc}.

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.cc}.

Loop splitting.  If a loop contains a conditional statement that is
always true for one part of the iteration space and false for the other
this pass splits the loop into two, one dealing with one side the other
only with the other, thereby removing one inner-loop conditional.  The
pass is implemented in @file{tree-ssa-loop-split.cc}.

The optimizations also use various utility functions contained in
@file{tree-ssa-loop-manip.cc}, @file{cfgloop.cc}, @file{cfgloopanal.cc} and
@file{cfgloopmanip.cc}.

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.cc} (the main driver),
@file{tree-vect-loop.cc} and @file{tree-vect-loop-manip.cc} (loop specific parts
and general loop utilities), @file{tree-vect-slp} (loop-aware SLP
functionality), @file{tree-vect-stmts.cc}, @file{tree-vect-data-refs.cc} and
@file{tree-vect-slp-patterns.cc} containing the SLP pattern matcher.
Analysis of data references is in @file{tree-data-ref.cc}.

SLP Vectorization.  This pass performs vectorization of straight-line code. The
pass is implemented in @file{tree-vectorizer.cc} (the main driver),
@file{tree-vect-slp.cc}, @file{tree-vect-stmts.cc} and
@file{tree-vect-data-refs.cc}.

Autoparallelization.  This pass splits the loop iteration space to run
into several threads.  The pass is implemented in @file{tree-parloops.cc}.

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{https://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.cc} 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.cc} and is described
by @code{pass_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.cc} and described by
@code{pass_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.cc} and is
described by @code{pass_vrp}.

@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.cc} 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.cc} 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.cc} and is described by
@code{pass_tail_calls}.  The RTL transformation is handled by
@code{fixup_tail_calls} in @file{calls.cc}.

@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.cc} and
is described by @code{pass_warn_function_return}.

@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.cc} 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.cc} 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.cc} and is described by
@code{pass_return_slot}.

@item Optimize calls to @code{__builtin_object_size} or
@code{__builtin_dynamic_object_size}

This is a propagation pass similar to CCP that tries to remove calls to
@code{__builtin_object_size} when the upper or lower bound for the size
of the object can be computed at compile-time.  It also tries to replace
calls to @code{__builtin_dynamic_object_size} with an expression that
evaluates the upper or lower bound for the size of the object.  This
pass is located in @file{tree-object-size.cc} 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.cc} 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
the @file{graphile-*.cc} files and described by
@code{pass_graphite}.

@item Unrolling of small loops

This pass completely unrolls loops with few iterations.  The pass
is located in @file{tree-ssa-loop-ivcanon.cc} 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.cc}
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.cc} 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.cc} 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.cc} 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.cc},
@file{calls.cc},
@file{expr.cc},
@file{explow.cc},
@file{expmed.cc},
@file{function.cc},
@file{optabs.cc}
and @file{emit-rtl.cc}.
Also, the file
@file{insn-emit.cc}, 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.cc}.

@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.cc}, and there are support routines in @file{cfgrtl.cc}
and @file{jump.cc}.

@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.cc}.

@item Store forwarding avoidance

This pass attempts to reduce the overhead of store to load forwarding.
It detects when a load reads from one or more previous smaller stores and
then rearranges them so that the stores are done after the load.  The loaded
value is adjusted with a series of bit insert instructions so that it stays
the same.  The code is located in @file{avoid-store-forwarding.cc}.

@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.cc}.

@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.cc}, and the LCM routines
are in @file{lcm.cc}.

A third version of this pass is run on some targets to optimise assignments to
specific hard registers.  This can be used in cases where a register has a
single purpose, such as specifying a mode as an extra input for specific
instructions (@pxref{mode switching optimization} for another way of handling
instruction modes).

@item Loop optimization

This pass performs several loop related optimizations.
The source files @file{cfgloopanal.cc} and @file{cfgloopmanip.cc} contain
generic loop analysis and manipulation code.  Initialization and finalization
of loop structures is handled by @file{loop-init.cc}.
A loop invariant motion pass is implemented in @file{loop-invariant.cc}.
Basic block level optimizations---unrolling, and peeling loops---
are implemented in @file{loop-unroll.cc}.
Replacing of the exit condition of loops by special machine-dependent
instructions is handled by @file{loop-doloop.cc}.

@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.cc}.

@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.cc}.

@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.cc}.

@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.cc}.

@item Late instruction combination

This pass attempts to do further instruction combination, on top of
that performed by @file{combine.cc}.  Its current purpose is to
substitute definitions into all uses simultaneously, so that the
definition can be removed.  This differs from the forward propagation
pass, whose purpose is instead to simplify individual uses on the
assumption that the definition will remain.  It differs from
@file{combine.cc} in that there is no hard-coded limit on the number
of instructions that can be combined at once.  It also differs from
@file{combine.cc} in that it can move instructions, where necessary.

However, the pass is not in principle limited to this form of
combination.  It is intended to be a home for other, future
combination approaches as well.

The pass runs twice, once before register allocation and once after
register allocation.  The code is located in @file{late-combine.cc}.

@anchor{mode switching optimization}
@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.cc}.

@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.cc}.

@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.cc},
@file{sched-deps.cc}, @file{sched-ebb.cc}, @file{sched-rgn.cc} 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.cc}, @file{ira-build.cc},
@file{ira-costs.cc}, @file{ira-conflicts.cc}, @file{ira-color.cc},
@file{ira-emit.cc}, @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.cc} and @file{reload1.cc}, 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.cc}, @file{lra-assign.c}, @file{lra-coalesce.cc},
@file{lra-constraints.cc}, @file{lra-eliminations.cc},
@file{lra-lives.cc}, @file{lra-remat.cc}, @file{lra-spills.cc}, 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.cc}, and the various
prediction routines are in @file{predict.cc}.

@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.cc}.

@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.cc}.

@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.cc}.

@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.cc}.

@item Final

This pass outputs the assembler code for the function.  The source files
are @file{final.cc} plus @file{insn-output.cc}; 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{dwarfout.c} for
DWARF symbol table format, files @file{dwarf2out.cc} and @file{dwarf2asm.cc}
for DWARF2 symbol table format, and @file{vmsdbgout.cc} for VMS debug
symbol table format.

@end itemize

@node Optimization info
@section Optimization info
@include optinfo.texi