\input texinfo
@setfilename cppinternals.info
@settitle The GNU C Preprocessor Internals

@ifinfo
@dircategory Programming
@direntry
* Cpplib: (cppinternals).      Cpplib internals.
@end direntry
@end ifinfo

@c @smallbook
@c @cropmarks
@c @finalout
@setchapternewpage odd
@ifinfo
This file documents the internals of the GNU C Preprocessor.

Copyright 2000, 2001 Free Software Foundation, Inc.

Permission is granted to make and distribute verbatim copies of
this manual provided the copyright notice and this permission notice
are preserved on all copies.

@ignore
Permission is granted to process this file through Tex and print the
results, provided the printed document carries copying permission
notice identical to this one except for the removal of this paragraph
(this paragraph not being relevant to the printed manual).

@end ignore
Permission is granted to copy and distribute modified versions of this
manual under the conditions for verbatim copying, provided also that
the entire resulting derived work is distributed under the terms of a
permission notice identical to this one.

Permission is granted to copy and distribute translations of this manual
into another language, under the above conditions for modified versions.
@end ifinfo

@titlepage
@c @finalout
@title Cpplib Internals
@subtitle Last revised September 2001
@subtitle for GCC version 3.1
@author Neil Booth
@page
@vskip 0pt plus 1filll
@c man begin COPYRIGHT
Copyright @copyright{} 2000, 2001
Free Software Foundation, Inc.

Permission is granted to make and distribute verbatim copies of
this manual provided the copyright notice and this permission notice
are preserved on all copies.

Permission is granted to copy and distribute modified versions of this
manual under the conditions for verbatim copying, provided also that
the entire resulting derived work is distributed under the terms of a
permission notice identical to this one.

Permission is granted to copy and distribute translations of this manual
into another language, under the above conditions for modified versions.
@c man end
@end titlepage
@contents
@page

@node Top
@top
@chapter Cpplib---the core of the GNU C Preprocessor

The GNU C preprocessor in GCC 3.x has been completely rewritten.  It is
now implemented as a library, cpplib, so it can be easily shared between
a stand-alone preprocessor, and a preprocessor integrated with the C,
C++ and Objective-C front ends.  It is also available for use by other
programs, though this is not recommended as its exposed interface has
not yet reached a point of reasonable stability.

The library has been written to be re-entrant, so that it can be used
to preprocess many files simultaneously if necessary.  It has also been
written with the preprocessing token as the fundamental unit; the
preprocessor in previous versions of GCC would operate on text strings
as the fundamental unit.

This brief manual documents some of the internals of cpplib, and a few
tricky issues encountered.  It also describes certain behaviour we would
like to preserve, such as the format and spacing of its output.

@menu
* Conventions::         Conventions used in the code.
* Lexer::               The combined C, C++ and Objective-C Lexer.
* Hash Nodes::          All identifiers are entered into a hash table.
* Macro Expansion::     Macro expansion algorithm.
* Token Spacing::       Spacing and paste avoidance issues.
* Line Numbering::      Tracking location within files.
* Guard Macros::        Optimizing header files with guard macros.
* Files::               File handling.
* Index::               Index.
@end menu

@node Conventions
@unnumbered Conventions
@cindex interface
@cindex header files

cpplib has two interfaces---one is exposed internally only, and the
other is for both internal and external use.

The convention is that functions and types that are exposed to multiple
files internally are prefixed with @samp{_cpp_}, and are to be found in
the file @file{cpphash.h}.  Functions and types exposed to external
clients are in @file{cpplib.h}, and prefixed with @samp{cpp_}.  For
historical reasons this is no longer quite true, but we should strive to
stick to it.

We are striving to reduce the information exposed in @file{cpplib.h} to the
bare minimum necessary, and then to keep it there.  This makes clear
exactly what external clients are entitled to assume, and allows us to
change internals in the future without worrying whether library clients
are perhaps relying on some kind of undocumented implementation-specific
behaviour.

@node Lexer
@unnumbered The Lexer
@cindex lexer
@cindex newlines
@cindex escaped newlines

@section Overview
The lexer is contained in the file @file{cpplex.c}.  It is a hand-coded
lexer, and not implemented as a state machine.  It can understand C, C++
and Objective-C source code, and has been extended to allow reasonably
successful preprocessing of assembly language.  The lexer does not make
an initial pass to strip out trigraphs and escaped newlines, but handles
them as they are encountered in a single pass of the input file.  It
returns preprocessing tokens individually, not a line at a time.

It is mostly transparent to users of the library, since the library's
interface for obtaining the next token, @code{cpp_get_token}, takes care
of lexing new tokens, handling directives, and expanding macros as
necessary.  However, the lexer does expose some functionality so that
clients of the library can easily spell a given token, such as
@code{cpp_spell_token} and @code{cpp_token_len}.  These functions are
useful when generating diagnostics, and for emitting the preprocessed
output.

@section Lexing a token
Lexing of an individual token is handled by @code{_cpp_lex_direct} and
its subroutines.  In its current form the code is quite complicated,
with read ahead characters and such-like, since it strives to not step
back in the character stream in preparation for handling non-ASCII file
encodings.  The current plan is to convert any such files to UTF-8
before processing them.  This complexity is therefore unnecessary and
will be removed, so I'll not discuss it further here.

The job of @code{_cpp_lex_direct} is simply to lex a token.  It is not
responsible for issues like directive handling, returning lookahead
tokens directly, multiple-include optimisation, or conditional block
skipping.  It necessarily has a minor r@^ole to play in memory
management of lexed lines.  I discuss these issues in a separate section
(@pxref{Lexing a line}).

The lexer places the token it lexes into storage pointed to by the
variable @var{cur_token}, and then increments it.  This variable is
important for correct diagnostic positioning.  Unless a specific line
and column are passed to the diagnostic routines, they will examine the
@var{line} and @var{col} values of the token just before the location
that @var{cur_token} points to, and use that location to report the
diagnostic.

The lexer does not consider whitespace to be a token in its own right.
If whitespace (other than a new line) precedes a token, it sets the
@code{PREV_WHITE} bit in the token's flags.  Each token has its
@var{line} and @var{col} variables set to the line and column of the
first character of the token.  This line number is the line number in
the translation unit, and can be converted to a source (file, line) pair
using the line map code.

The first token on a logical, i.e.@: unescaped, line has the flag
@code{BOL} set for beginning-of-line.  This flag is intended for
internal use, both to distinguish a @samp{#} that begins a directive
from one that doesn't, and to generate a call-back to clients that want
to be notified about the start of every non-directive line with tokens
on it.  Clients cannot reliably determine this for themselves: the first
token might be a macro, and the tokens of a macro expansion do not have
the @code{BOL} flag set.  The macro expansion may even be empty, and the
next token on the line certainly won't have the @code{BOL} flag set.

New lines are treated specially; exactly how the lexer handles them is
context-dependent.  The C standard mandates that directives are
terminated by the first unescaped newline character, even if it appears
in the middle of a macro expansion.  Therefore, if the state variable
@var{in_directive} is set, the lexer returns a @code{CPP_EOF} token,
which is normally used to indicate end-of-file, to indicate
end-of-directive.  In a directive a @code{CPP_EOF} token never means
end-of-file.  Conveniently, if the caller was @code{collect_args}, it
already handles @code{CPP_EOF} as if it were end-of-file, and reports an
error about an unterminated macro argument list.

The C standard also specifies that a new line in the middle of the
arguments to a macro is treated as whitespace.  This white space is
important in case the macro argument is stringified.  The state variable
@code{parsing_args} is non-zero when the preprocessor is collecting the
arguments to a macro call.  It is set to 1 when looking for the opening
parenthesis to a function-like macro, and 2 when collecting the actual
arguments up to the closing parenthesis, since these two cases need to
be distinguished sometimes.  One such time is here: the lexer sets the
@code{PREV_WHITE} flag of a token if it meets a new line when
@code{parsing_args} is set to 2.  It doesn't set it if it meets a new
line when @code{parsing_args} is 1, since then code like

@smallexample
#define foo() bar
foo
baz
@end smallexample

@noindent would be output with an erroneous space before @samp{baz}:

@smallexample
foo
 baz
@end smallexample

This is a good example of the subtlety of getting token spacing correct
in the preprocessor; there are plenty of tests in the test-suite for
corner cases like this.

The lexer is written to treat each of @samp{\r}, @samp{\n}, @samp{\r\n}
and @samp{\n\r} as a single new line indicator.  This allows it to
transparently preprocess MS-DOS, Macintosh and Unix files without their
needing to pass through a special filter beforehand.

We also decided to treat a backslash, either @samp{\} or the trigraph
@samp{??/}, separated from one of the above newline indicators by
non-comment whitespace only, as intending to escape the newline.  It
tends to be a typing mistake, and cannot reasonably be mistaken for
anything else in any of the C-family grammars.  Since handling it this
way is not strictly conforming to the ISO standard, the library issues a
warning wherever it encounters it.

Handling newlines like this is made simpler by doing it in one place
only.  The function @code{handle_newline} takes care of all newline
characters, and @code{skip_escaped_newlines} takes care of arbitrarily
long sequences of escaped newlines, deferring to @code{handle_newline}
to handle the newlines themselves.

The most painful aspect of lexing ISO-standard C and C++ is handling
trigraphs and backlash-escaped newlines.  Trigraphs are processed before
any interpretation of the meaning of a character is made, and unfortunately
there is a trigraph representation for a backslash, so it is possible for
the trigraph @samp{??/} to introduce an escaped newline.

Escaped newlines are tedious because theoretically they can occur
anywhere---between the @samp{+} and @samp{=} of the @samp{+=} token,
within the characters of an identifier, and even between the @samp{*}
and @samp{/} that terminates a comment.  Moreover, you cannot be sure
there is just one---there might be an arbitrarily long sequence of them.

So, for example, the routine that lexes a number, @code{parse_number},
cannot assume that it can scan forwards until the first non-number
character and be done with it, because this could be the @samp{\}
introducing an escaped newline, or the @samp{?} introducing the trigraph
sequence that represents the @samp{\} of an escaped newline.  If it
encounters a @samp{?} or @samp{\}, it calls @code{skip_escaped_newlines}
to skip over any potential escaped newlines before checking whether the
number has been finished.

Similarly code in the main body of @code{_cpp_lex_direct} cannot simply
check for a @samp{=} after a @samp{+} character to determine whether it
has a @samp{+=} token; it needs to be prepared for an escaped newline of
some sort.  Such cases use the function @code{get_effective_char}, which
returns the first character after any intervening escaped newlines.

The lexer needs to keep track of the correct column position, including
counting tabs as specified by the @option{-ftabstop=} option.  This
should be done even within C-style comments; they can appear in the
middle of a line, and we want to report diagnostics in the correct
position for text appearing after the end of the comment.

@anchor{Invalid identifiers}
Some identifiers, such as @code{__VA_ARGS__} and poisoned identifiers,
may be invalid and require a diagnostic.  However, if they appear in a
macro expansion we don't want to complain with each use of the macro.
It is therefore best to catch them during the lexing stage, in
@code{parse_identifier}.  In both cases, whether a diagnostic is needed
or not is dependent upon the lexer's state.  For example, we don't want
to issue a diagnostic for re-poisoning a poisoned identifier, or for
using @code{__VA_ARGS__} in the expansion of a variable-argument macro.
Therefore @code{parse_identifier} makes use of state flags to determine
whether a diagnostic is appropriate.  Since we change state on a
per-token basis, and don't lex whole lines at a time, this is not a
problem.

Another place where state flags are used to change behaviour is whilst
lexing header names.  Normally, a @samp{<} would be lexed as a single
token.  After a @code{#include} directive, though, it should be lexed as
a single token as far as the nearest @samp{>} character.  Note that we
don't allow the terminators of header names to be escaped; the first
@samp{"} or @samp{>} terminates the header name.

Interpretation of some character sequences depends upon whether we are
lexing C, C++ or Objective-C, and on the revision of the standard in
force.  For example, @samp{::} is a single token in C++, but in C it is
two separate @samp{:} tokens and almost certainly a syntax error.  Such
cases are handled by @code{_cpp_lex_direct} based upon command-line
flags stored in the @code{cpp_options} structure.

Once a token has been lexed, it leads an independent existence.  The
spelling of numbers, identifiers and strings is copied to permanent
storage from the original input buffer, so a token remains valid and
correct even if its source buffer is freed with @code{_cpp_pop_buffer}.
The storage holding the spellings of such tokens remains until the
client program calls cpp_destroy, probably at the end of the translation
unit.

@anchor{Lexing a line}
@section Lexing a line
@cindex token run

When the preprocessor was changed to return pointers to tokens, one
feature I wanted was some sort of guarantee regarding how long a
returned pointer remains valid.  This is important to the stand-alone
preprocessor, the future direction of the C family front ends, and even
to cpplib itself internally.

Occasionally the preprocessor wants to be able to peek ahead in the
token stream.  For example, after the name of a function-like macro, it
wants to check the next token to see if it is an opening parenthesis.
Another example is that, after reading the first few tokens of a
@code{#pragma} directive and not recognising it as a registered pragma,
it wants to backtrack and allow the user-defined handler for unknown
pragmas to access the full @code{#pragma} token stream.  The stand-alone
preprocessor wants to be able to test the current token with the
previous one to see if a space needs to be inserted to preserve their
separate tokenization upon re-lexing (paste avoidance), so it needs to
be sure the pointer to the previous token is still valid.  The
recursive-descent C++ parser wants to be able to perform tentative
parsing arbitrarily far ahead in the token stream, and then to be able
to jump back to a prior position in that stream if necessary.

The rule I chose, which is fairly natural, is to arrange that the
preprocessor lex all tokens on a line consecutively into a token buffer,
which I call a @dfn{token run}, and when meeting an unescaped new line
(newlines within comments do not count either), to start lexing back at
the beginning of the run.  Note that we do @emph{not} lex a line of
tokens at once; if we did that @code{parse_identifier} would not have
state flags available to warn about invalid identifiers (@pxref{Invalid
identifiers}).

In other words, accessing tokens that appeared earlier in the current
line is valid, but since each logical line overwrites the tokens of the
previous line, tokens from prior lines are unavailable.  In particular,
since a directive only occupies a single logical line, this means that
the directive handlers like the @code{#pragma} handler can jump around
in the directive's tokens if necessary.

Two issues remain: what about tokens that arise from macro expansions,
and what happens when we have a long line that overflows the token run?

Since we promise clients that we preserve the validity of pointers that
we have already returned for tokens that appeared earlier in the line,
we cannot reallocate the run.  Instead, on overflow it is expanded by
chaining a new token run on to the end of the existing one.

The tokens forming a macro's replacement list are collected by the
@code{#define} handler, and placed in storage that is only freed by
@code{cpp_destroy}.  So if a macro is expanded in our line of tokens,
the pointers to the tokens of its expansion that we return will always
remain valid.  However, macros are a little trickier than that, since
they give rise to three sources of fresh tokens.  They are the built-in
macros like @code{__LINE__}, and the @samp{#} and @samp{##} operators
for stringifcation and token pasting.  I handled this by allocating
space for these tokens from the lexer's token run chain.  This means
they automatically receive the same lifetime guarantees as lexed tokens,
and we don't need to concern ourselves with freeing them.

Lexing into a line of tokens solves some of the token memory management
issues, but not all.  The opening parenthesis after a function-like
macro name might lie on a different line, and the front ends definitely
want the ability to look ahead past the end of the current line.  So
cpplib only moves back to the start of the token run at the end of a
line if the variable @code{keep_tokens} is zero.  Line-buffering is
quite natural for the preprocessor, and as a result the only time cpplib
needs to increment this variable is whilst looking for the opening
parenthesis to, and reading the arguments of, a function-like macro.  In
the near future cpplib will export an interface to increment and
decrement this variable, so that clients can share full control over the
lifetime of token pointers too.

The routine @code{_cpp_lex_token} handles moving to new token runs,
calling @code{_cpp_lex_direct} to lex new tokens, or returning
previously-lexed tokens if we stepped back in the token stream.  It also
checks each token for the @code{BOL} flag, which might indicate a
directive that needs to be handled, or require a start-of-line call-back
to be made.  @code{_cpp_lex_token} also handles skipping over tokens in
failed conditional blocks, and invalidates the control macro of the
multiple-include optimization if a token was successfully lexed outside
a directive.  In other words, its callers do not need to concern
themselves with such issues.

@node Hash Nodes
@unnumbered Hash Nodes
@cindex hash table
@cindex identifiers
@cindex macros
@cindex assertions
@cindex named operators

When cpplib encounters an ``identifier'', it generates a hash code for
it and stores it in the hash table.  By ``identifier'' we mean tokens
with type @code{CPP_NAME}; this includes identifiers in the usual C
sense, as well as keywords, directive names, macro names and so on.  For
example, all of @code{pragma}, @code{int}, @code{foo} and
@code{__GNUC__} are identifiers and hashed when lexed.

Each node in the hash table contain various information about the
identifier it represents.  For example, its length and type.  At any one
time, each identifier falls into exactly one of three categories:

@itemize @bullet
@item Macros

These have been declared to be macros, either on the command line or
with @code{#define}.  A few, such as @code{__TIME__} are built-ins
entered in the hash table during initialisation.  The hash node for a
normal macro points to a structure with more information about the
macro, such as whether it is function-like, how many arguments it takes,
and its expansion.  Built-in macros are flagged as special, and instead
contain an enum indicating which of the various built-in macros it is.

@item Assertions

Assertions are in a separate namespace to macros.  To enforce this, cpp
actually prepends a @code{#} character before hashing and entering it in
the hash table.  An assertion's node points to a chain of answers to
that assertion.

@item Void

Everything else falls into this category---an identifier that is not
currently a macro, or a macro that has since been undefined with
@code{#undef}.

When preprocessing C++, this category also includes the named operators,
such as @code{xor}.  In expressions these behave like the operators they
represent, but in contexts where the spelling of a token matters they
are spelt differently.  This spelling distinction is relevant when they
are operands of the stringizing and pasting macro operators @code{#} and
@code{##}.  Named operator hash nodes are flagged, both to catch the
spelling distinction and to prevent them from being defined as macros.
@end itemize

The same identifiers share the same hash node.  Since each identifier
token, after lexing, contains a pointer to its hash node, this is used
to provide rapid lookup of various information.  For example, when
parsing a @code{#define} statement, CPP flags each argument's identifier
hash node with the index of that argument.  This makes duplicated
argument checking an O(1) operation for each argument.  Similarly, for
each identifier in the macro's expansion, lookup to see if it is an
argument, and which argument it is, is also an O(1) operation.  Further,
each directive name, such as @code{endif}, has an associated directive
enum stored in its hash node, so that directive lookup is also O(1).

@node Macro Expansion
@unnumbered Macro Expansion Algorithm

@c TODO

@node Token Spacing
@unnumbered Token Spacing
@cindex paste avoidance
@cindex spacing
@cindex token spacing

First, let's look at an issue that only concerns the stand-alone
preprocessor: we want to guarantee that re-reading its preprocessed
output results in an identical token stream.  Without taking special
measures, this might not be the case because of macro substitution.  For
example:

@smallexample
#define PLUS +
#define EMPTY
#define f(x) =x=
+PLUS -EMPTY- PLUS+ f(=)
        @expansion{} + + - - + + = = =
@emph{not}
        @expansion{} ++ -- ++ ===
@end smallexample

One solution would be to simply insert a space between all adjacent
tokens.  However, we would like to keep space insertion to a minimum,
both for aesthetic reasons and because it causes problems for people who
still try to abuse the preprocessor for things like Fortran source and
Makefiles.

For now, just notice that the only places we need to be careful about
@dfn{paste avoidance} are when tokens are added (or removed) from the
original token stream.  This only occurs because of macro expansion, but
care is needed in many places: before @strong{and} after each macro
replacement, each argument replacement, and additionally each token
created by the @samp{#} and @samp{##} operators.

Let's look at how the preprocessor gets whitespace output correct
normally.  The @code{cpp_token} structure contains a flags byte, and one
of those flags is @code{PREV_WHITE}.  This is flagged by the lexer, and
indicates that the token was preceded by whitespace of some form other
than a new line.  The stand-alone preprocessor can use this flag to
decide whether to insert a space between tokens in the output.

Now consider the following:

@smallexample
#define add(x, y, z) x + y +z;
sum = add (1,2, 3);
        @expansion{} sum = 1 + 2 +3;
@end smallexample

The interesting thing here is that the tokens @samp{1} and @samp{2} are
output with a preceding space, and @samp{3} is output without a
preceding space, but when lexed none of these tokens had that property.
Careful consideration reveals that @samp{1} gets its preceding
whitespace from the space preceding @samp{add} in the macro
@emph{invocation}, @samp{2} gets its whitespace from the space preceding
the parameter @samp{y} in the macro @emph{replacement list}, and
@samp{3} has no preceding space because parameter @samp{z} has none in
the replacement list.

Once lexed, tokens are effectively fixed and cannot be altered, since
pointers to them might be held in many places, in particular by
in-progress macro expansions.  So instead of modifying the two tokens
above, the preprocessor inserts a special token, which I call a
@dfn{padding token}, into the token stream in front of every macro
expansion and expanded macro argument, to indicate that the subsequent
token should assume its @code{PREV_WHITE} flag from a different
@dfn{source token}.  In the above example, the source tokens are
@samp{add} in the macro invocation, and @samp{y} and @samp{z} in the
macro replacement list, respectively.

It is quite easy to get multiple padding tokens in a row, for example if
a macro's first replacement token expands straight into another macro.

@smallexample
#define foo bar
#define bar baz
[foo]
        @expansion{} [baz]
@end smallexample

Here, two padding tokens with sources @samp{foo} between the brackets,
and @samp{bar} from foo's replacement list, are generated.  Clearly the
first padding token is the one that matters.  But what if we happen to
leave a macro expansion?  Adjusting the above example slightly:

@smallexample
#define foo bar
#define bar EMPTY baz
#define EMPTY
[foo] EMPTY;
        @expansion{} [ baz] ;
@end smallexample

As shown, now there should be a space before baz and the semicolon.  Our
initial algorithm fails for the former, because we would see three
padding tokens, one per macro invocation, followed by @samp{baz}, which
would have inherit its spacing from the original source, @samp{foo},
which has no leading space.  Note that it is vital that cpplib get
spacing correct in these examples, since any of these macro expansions
could be stringified, where spacing matters.

So, I have demonstrated that not just entering macro and argument
expansions, but leaving them requires special handling too.  So cpplib
inserts a padding token with a @code{NULL} source token when leaving
macro expansions and after each replaced argument in a macro's
replacement list.  It also inserts appropriate padding tokens on either
side of tokens created by the @samp{#} and @samp{##} operators.

Now we can see the relationship with paste avoidance: we have to be
careful about paste avoidance in exactly the same locations we take care
to get white space correct.  This makes implementation of paste
avoidance easy: wherever the stand-alone preprocessor is fixing up
spacing because of padding tokens, and it turns out that no space is
needed, it has to take the extra step to check that a space is not
needed after all to avoid an accidental paste.  The function
@code{cpp_avoid_paste} advises whether a space is required between two
consecutive tokens.  To avoid excessive spacing, it tries hard to only
require a space if one is likely to be necessary, but for reasons of
efficiency it is slightly conservative and might recommend a space where
one is not strictly needed.

@node Line Numbering
@unnumbered Line numbering
@cindex line numbers

The preprocessor takes great care to ensure it keeps track of both the
position of a token in the source file, for diagnostic purposes, and
where it should appear in the output file, because using CPP for other
languages like assembler requires this.  The two positions may differ
for the following reasons:

@itemize @bullet
@item
Escaped newlines are deleted, so lines spliced in this way are joined to
form a single logical line.

@item
A macro expansion replaces the tokens that form its invocation, but any
newlines appearing in the macro's arguments are interpreted as a single
space, with the result that the macro's replacement appears in full on
the same line that the macro name appeared in the source file.  This is
particularly important for stringification of arguments---newlines
embedded in the arguments must appear in the string as spaces.
@end itemize

The source file location is maintained in the @code{lineno} member of the
@code{cpp_buffer} structure, and the column number inferred from the
current position in the buffer relative to the @code{line_base} buffer
variable, which is updated with every newline whether escaped or not.

@c FINISH THIS

@node Guard Macros
@unnumbered The Multiple-Include Optimization

@c TODO

@node Files
@unnumbered File Handling
@cindex files

Fairly obviously, the file handling code of cpplib resides in the file
@file{cppfiles.c}.  It takes care of the details of file searching,
opening, reading and caching, for both the main source file and all the
headers it recursively includes.

The basic strategy is to minimize the number of system calls.  On many
systems, the basic @code{open ()} and @code{fstat ()} system calls can
be quite expensive.  For every @code{#include}-d file, we need to try
all the directories in the search path until we find a match.  Some
projects, such as glibc, pass twenty or thirty include paths on the
command line, so this can rapidly become time consuming.

For a header file we have not encountered before we have little choice
but to do this.  However, it is often the case that the same headers are
repeatedly included, and in these cases we try to avoid repeating the
filesystem queries whilst searching for the correct file.

For each file we try to open, we store the constructed path in a splay
tree.  This path first undergoes simplification by the function
@code{_cpp_simplify_pathname}.  For example,
@file{/usr/include/bits/../foo.h} is simplified to
@file{/usr/include/foo.h} before we enter it in the splay tree and try
to @code{open ()} the file.  CPP will then find subsequent uses of
@file{foo.h}, even as @file{/usr/include/foo.h}, in the splay tree and
save system calls.

Further, it is likely the file contents have also been cached, saving a
@code{read ()} system call.  We don't bother caching the contents of
header files that are re-inclusion protected, and whose re-inclusion
macro is defined when we leave the header file for the first time.  If
the host supports it, we try to map suitably large files into memory,
rather than reading them in directly.

The include paths are internally stored on a null-terminated
singly-linked list, starting with the @code{"header.h"} directory search
chain, which then links into the @code{<header.h>} directory chain.

Files included with the @code{<foo.h>} syntax start the lookup directly
in the second half of this chain.  However, files included with the
@code{"foo.h"} syntax start at the beginning of the chain, but with one
extra directory prepended.  This is the directory of the current file;
the one containing the @code{#include} directive.  Prepending this
directory on a per-file basis is handled by the function
@code{search_from}.

Note that a header included with a directory component, such as
@code{#include "mydir/foo.h"} and opened as
@file{/usr/local/include/mydir/foo.h}, will have the complete path minus
the basename @samp{foo.h} as the current directory.

Enough information is stored in the splay tree that CPP can immediately
tell whether it can skip the header file because of the multiple include
optimisation, whether the file didn't exist or couldn't be opened for
some reason, or whether the header was flagged not to be re-used, as it
is with the obsolete @code{#import} directive.

For the benefit of MS-DOS filesystems with an 8.3 filename limitation,
CPP offers the ability to treat various include file names as aliases
for the real header files with shorter names.  The map from one to the
other is found in a special file called @samp{header.gcc}, stored in the
command line (or system) include directories to which the mapping
applies.  This may be higher up the directory tree than the full path to
the file minus the base name.

@node Index
@unnumbered Index
@printindex cp

@bye