22 May
2020
22 May
'20
10:10 a.m.
Richard Salz <rich.salz(a)gmail.com> wrote:
And what about John Gilmore making all bsd user it?
And the multiple usenix
tutorials?
I think Rich is referring to the time in 1987-8 when I spent some time
compiling the entire BSD distribution sources with the Vax version of
gcc. This was a volunteer effort on my part so that Berkeley could
adopt GCC to replace PCC. They got an ANSI C compiler, and avoided AT&T
copyright restrictions on Yet Another critical piece of Berkeley Unix.
GNU got an extensive test of GCC which moved it out of "beta" status.
I ended up taking extensive notes, and wrote a 1988 paper about the
experience, which I submitted to USENIX. But it was rejected, on the
theory that porting code (even ancient crufty Unix code) through new
compilers wasn't research. Indeed, I recall Kirk McKusick remarking to
me around that time that even Unix kernel ports to new architectures
were so routine as to not be research in his opinion.
Oddly, I was easily able to find that paper (thanks to Kryder's Law), so
I have appended it verbatim below (in troff with -ms macros). In short,
I found about a dozen bugs in GCC, which RMS fixed; and many hundreds of
bugs in the 4.3BSD Unix sources, which I fixed and Keith merged upstream.
Note the quaint footnoted homage to distributed collaboration, which was
still novel back then in the pre-Covid, pre-public-Internet, 2400 baud
modem era.
John
.TL
Porting Berkeley
.UX
through the GNU C Compiler
.AU
John Gilmore
.AI
Grasshopper Group
San Francisco, CA, USA 94117
gnu(a)toad.com
.AB
We have ported UC Berkeley's latest
.UX
sources through the GNU C Compiler,
a free draft-ANSI compatible compiler written by Richard Stallman and available from the
Free
Software Foundation. In the process, we made Berkeley
.UX
more compatible
with the draft ANSI C standard, and tested the GNU C Compiler
for its full production release.
We describe the impact of various ANSI C changes on the Berkeley
.UX
sources, the kinds of non-portable code that the conversion uncovered,
and how we fixed them. We also briefly explore some limitations in the tools
used to build a
.UX
system.
.AE
.SH
Introduction
.PP
The GNU C Compiler (GCC) is a complete C compiler, compatible with the draft
ANSI standard, and
available in source from the Free Software Foundation (FSF). It was written by
Richard Stallman
in 1986 and 1987, and is (at this writing) in its
18th release. It is a major component of the GNU (``GNU's Not
.UX '')
project, whose aim
is to build a complete
.UX -like
software system,
available in source to anyone who wants it.
The compiler produces good code \(em better than most commercial
compilers \(em and has been ported to the Vax, MC680X0,
and NS32XXX.
.PP
Berkeley
.UX ,
from the Computer Systems Research Group (CSRG) at the University
of California at Berkeley,
had its start in the 1970's with a prerelease
.UX
Version 7, and
has been improving ever since. The current sources derive from the
1978 AT&T ``32V'' release, a V7 variant for the Vax. CSRG has produced
four major releases for the Vax
\(em 3, 4.1, 4.2, and 4.3BSD. These releases have set the
standard for high powered
.UX
systems for many years, and continue
to offer an improved alternative to the flat-tasting AT&T
.UX
releases.
.PP
However, Berkeley's C compiler is based on an old version of PCC,
the Portable C Compiler from AT&T. There was little chance that anyone
would provide ANSI C language extensions in this compiler, or do significant
work on optimizing the generated code. By merging the GNU C compiler
into the Berkeley release, we provided these new features to Berkeley
Unix users at a low cost,
while offering the GNU project an important test case for GNU C.
.SH
Goals
.PP
The major goal for the project is to move GCC out of ``beta test'' and
into ``production'' status,
by demonstrating that a successful
.UX
port can be based on it.
.PP
We are also providing a better maintained
compiler for Berkeley
.UX .
GCC already produces better
object code then the previous compiler,
has a more modern internal structure, and supports useful features
such as function prototype declarations.
It is also maintained by a large collection of people around the world,
who contribute their fixes and enhancements to the master sources.
Regular releases by the
Free Software Foundation encourage distribution of the improvements.
In contrast, PCC
is proprietary to AT&T, and few fixes are widely distributed, except as
part of infrequent and expensive AT&T releases.
.PP
We are producing a
.UX
source tree which can be compiled
by
.I both
the old and the new compilers. This is partly for convenience during the port,
partly in case the project suffers long delays,
and partly because Berkeley
.UX
also runs on the Tahoe, a fast Vax-like machine
built by Computer Consoles, which
GCC does not yet support.
We are avoiding the introduction of new
.B #ifdef 's,
instead rewriting the code so that it does not depend
on the features of either compiler.
.PP
We have to constantly remind ourselves to minimize the changes required.
It's too easy to get lost in a maze of twisty
.UX
code, all desperately
needing improvement.
.PP
Whenever we have to make a change, we have moved in the direction of
ANSI
C and POSIX compatability.
.SH
People
.PP
The project was conceived by John Gilmore, and endorsed
by Keith Bostic and Mike Karels of CSRG, and Richard Stallman of FSF.
John did the major grunt work and provided fixes to the
.UX
code.
Keith and Mike provided machine
resources, collaborated
on major decisions, and arbitrated the style and content of the changes
to
.UX .
Richard provided quick turnaround on compiler bug fixes and problem
solving.
This setup worked extremely well.
.PP
We started work on 17 December 1987, and are not yet done at the
time of writing (19 February 1988). About 9 days of my time, 2 of Keith's,
half a day of Mike's, and XXX days of Richard's have gone into the
project so far.
.SH
Working Style
.PP
Most of the work was done over networks, in a loosely coordinated
style which was hard to concieve of only a few years ago.\(dg
.FS \(dg
Much of the free software work that is happening these days occurs in this
manner, and I would like to publicly thank the original DARPA pioneers who gave
birth to this vision of wide area, computer mediated collaborative work.
.FE
John worked in San Francisco,
Keith in Berkeley, and Richard in Cambridge. Keith set up an account and
a copy of the source tree on
.I vangogh ,
a Vax 8600 at Berkeley.
John spent a few
days in front of a Sun at Berkeley getting things straight, but did
most of the work by dialing in at 2400 baud from his office in San Francisco.
When we modified
.UX
source files, Keith
checked the changes and merged them back into the master
.UX
sources on another machine at Berkeley. When we found an apparent
bug in GCC, we isolated a small
excerpt or test program to demonstrate the bug, and forwarded it to Richard by Internet
electronic
mail.
Bug fixes came back as new GCC releases, which were FTP'd over the Internet
from MIT. Ongoing status reports, discussions, and scheduling were done
by \fIuucp\fP and Internet electronic mail.
.PP
At this writing, we have used four GCC releases (1.15 through 1.18).
For each
GCC release, we did a ``pass'' over the
.UX
source tree;
one such pass included an updated source tree as well.
Each GCC
release was built, tested, and installed on
.I vangogh
without trouble.
Then we ran
.I "make clean; make"
on the source tree, and examined 500K to 800K of resulting
output. Keith Bostic's Makefiles did an excellent job of
automating this process, though we ran into some problems with the
.UX
compilation model in general, and limitations in
.I make
in particular.
.SH
ANSI Language Changes
.PP
The problems encountered during the port fell into two general categories.
Some of the code was not written portably and failed in the new environment.
Other code was written portably for its time, but failed because ANSI C
has redefined parts of the language. In some cases it was hard to tell
the difference; the consensus on what is ``portable code'' changes over
time, and on some points there is no agreement.
.PP
The major ANSI C problem was the generation of
.B "character constants in cpp" .
The traditional
.UX
C preprocessor (\fIcpp\fP), written by John F. Reiser, would
substitute a macro's parameters into like-named substrings even inside
single or double quotes in the macro definition. For example:
.DS
#define CTRL(c) ('c'&037)
#define CEOF CTRL(d)
.DE
In an attempt to make things easier for tokenizing preprocessors,
ANSI C has changed the
rules here, and there is in fact
.I no
way to generate a character constant containing a macro argument.
(There is a way to generate a character
.I string ,
e.g. double-quoted string, but not a single-quoted character.
We consider this a bug in ANSI C.)
Fixing this required altering both the macro definition and each reference
to the macro:
.DS
#define CTRL(c) (c&037)
#define CEOF CTRL('d')
.DE
This required changes in about 10 system include files and in about 45
source modules. Many user programs turned out to depend on the undocumented
.B CTRL
macro, defined in
.B <sys/ttychars.h> ,
and since all its callers had to change, all those programs did too.
.PP
Another \fIcpp\fP problem involved
.B "token concatenation" .
No formal facilities were provided for this in the old \fIcpp\fP, but many
users discovered that with code like this, from the /etc/passwd scanning code:
.DS
#define EXPAND(e) passwd.pw_/**/e = tp; while (*tp++ = *cp++);
EXPAND(name);
EXPAND(passwd);
.DE
they could cause a macro argument to be concatenated with another argument,
or with preexisting text, to make a single name. In one case
(\fIphantasia\fP),
the Makefile provided half of a quoted string as a command line
.B #define ,
and the source text provided the other half!
ANSI C
does not allow a preprocessor to concatenate tokens in these ways, instead
providing a newly invented
.B ##
operator, and new rules requiring the compiler to concatenate adjacent
character strings. Again,
it was impossible to write
a macro that works with both old and new compilers, and we didn't want
to uglify our code with
.B "#ifdef __STDC__" ;
our solution was to
rewrite both the macros and all their callers, to avoid ever having to
concatenate tokens:
.DS
#define EXPAND(e) passwd.e = tp; while (*tp++ = *cp++);
EXPAND(pw_name);
EXPAND(pw_passwd);
.DE
Mostly the token concatenation was used as a typing convenience, so this
was not a problem. It involved changes to five modules.
We found no clean solution for
.I phantasia ;
a fix will probably involve rewriting it to do explicit
string concatenations at runtime.
.PP
Changes to the
.B "scope of externals"
provided another set of widely scattered changes. If an external
identifier is declared from inside a function, PCC causes that declaration
to be visible to the entire remaining text of the source file.
This also applies to functions which are implicitly declared
when they first appear in an expression.
This
behaviour was not explicitly sanctioned by K&R,
but it was condoned (pg. 206, 2nd paragraph), and many programs depended on it.
ANSI C changed the scope rules to be more consistent; if you declare an
external identifier in a local block, the declaration has no effect outside
the block. We moved extern declarations to global scope, or added global
function declarations, in 38 files to handle this.
.PP
A number of programs used
.B "new keywords"
such as \fIsigned\fP or \fIconst\fP as identifiers. We renamed the identifiers
in 9 modules.
.PP
The Fortran libraries used a \fBtypedef name as a formal parameter\fP
to a set of functions. ANSI C has disallowed this, since it complicates
the parsing of the new prototype-style function declarations. We renamed
the parameter in 8 modules.
.PP
Three modules used a \fBtypedef with modifiers\fP, e.g.:
.DS
typedef int CONSZ;
x = (unsigned CONSZ) y;
.DE
This has been repudiated by ANSI C. We fixed it by making the original
typedef \fBunsigned\fP where possible, or by
creating a second typedef for ``U_CONSZ''.
.SH
Non-Portable Constructs
.PP
The worst non-portable construct we found in the
.UX
sources was the use of
.B "pointers to non-members" .
There was plenty of code as bad as:
.DS
int *foo;
foo->memb = 5
if (foo->humbug >= -1) bah();
.DE
and, in many cases, \fImemb\fP and \fIhumbug\fP are not even members of
the same struct!
Such code seems to have been written with a ``BCPL'' mentality, assuming
that all pointers are really the same thing and it doesn't matter what their
type is. Early C implementations lacked the
.B union
declarator,
and did not distinguish between the members of different structures.
Exploiting this has been considered
bad practice for years, and lint checks for it,
though many
.UX
compilers do not. We found a lot of it in old code, though newer
code did not lack for examples either.
Fixing this problem caused the most work,
because we had to figure out what each untyped or mistyped pointer was
.I really
being used for, then fix its type, and whatever references to it were
inconsistent with that type. We changed 5 modules due to this.
One program, \fIefl\fP, would have required so much work
that we abandoned it, since we could
not find anyone using it.
.PP
Another problem was caused by existing uses of
.B "cpp on non-C sources" .
Various assembler language modules were being preprocessed by \fIcpp\fP,
probably
because there is no standard macro assembler for
.UX .
These modules are
carefully arranged to avoid confusing the old \fIcpp\fP; for example,
assembler language comments are introduced by
.B # ,
but indented so that \fIcpp\fP will not treat them as control lines.
ANSI \fIcpp\fP's handle white space on both sides of the ``#'', so
indentation no longer hides these comments. Also, the ANSI rules
to require the preprocessor to keep track of which
material is inside single and double quotes and which is outside;
the old \fIcpp\fP terminated a character string or constant at the next
unescaped newline. Vax assembler language uses unmatched quotes
when specifying single ASCII characters, such as in immediate operands.
This causes an ANSI \fIcpp\fP to stop processing # directives at that point,
until it finds another
unmatched quote. We chose to alter the assembler modules to avoid
stumbling over these features in ANSI C preprocessors, without fixing the
larger problem of using a C-specific preprocessor on non-C text.
.PP
In addition to embedded C preprocessor statements in assembler
sources, we had to deal with
.B "asm() constructs"
in C source. Some system-dependent routines were written in C
with intermixed assembler code, producing a mess when compiled with
anything but the original compiler. Other routines, such as
.I compress ,
drop in an
.B asm()
here or there as an optimization. Still more modules, including the kernel,
run a
.I sed
script over the assembler code generated by the C compiler, before
assembling and linking it. There is no general solution to these
problems. GCC has added an asm() facility that is independent of
the compiler's register allocation strategy, but programs using this are
incompatible with the old C compiler.
We are investigating
a possible fix involving
changing all these places to use e.g.
.B "#include <machine/inline.h>"
which, in GCC, would define inline code containing asm()s, while
in PCC, declarations of (slower) external functions would be generated.
.PP
.I Troff
used
.B "multi-character constants"
in its font tables; we fixed it with a macro for building an int out of two
characters. A Fortran library module used the character constant
.B 'EOF' ,
presumably a typo for
.B EOF ;
and \fIrogue\fP defined the character '\300' as a possible command letter.
While ANSI C permits multiple character constants, they are implementation
defined, and GCC wisely defines them to be invalid (as the standard should
have done).
.PP
Some programs tried to declare functions or variables,
.B "omitting both type and storage class" .
This usage is not even valid in K&R, though PCC accepts it. We fixed this in
about 15
modules, by adding ``int'' to the declarations. There were two other modules
where this check uncovered inadvertent use of ``;'' in a declaration list
where ``,'' was intended.
.PP
GCC provides better error checking in a few ways, and caught a number
of bugs caused by misunderstood
.B "sign extension" .
It warns ``comparison is always 0 due to limited range of data type''
for constructs like:
.DS
char c;
if (c == 0x80) foo();
.DE
If a signed character contains the bit pattern 0x80, using it in an
expression causes it to be
sign-extended to 0xFFFFFF80, which does not equal 0x00000080.
Bugs of this sort were fixed, typically by casting the 0x80 to (char),
in 5 modules.
.PP
Changes to the rules for \fBparsing declarations\fP made us fix two modules
where the last declaration in a struct was immediately followed by a
closing brace, without a semicolon. Three more modules needed changes
because the rules for where braces are required in struct or array
initializers have changed. Four programs defined a \fBstruct foo\fP
and then referenced it as a \fBunion foo\fP, or vice verse. Two programs
declared \fBregister struct foo bar;\fP and then took bar's address, which
is not allowed for register variables!
.PP
Thirteen programs had miscellaneous \fBpointer usage bugs\fP
fixed. Two more were
comparing pointers to \fB-1\fP; these were changed to use zero as a
flag value instead.
.PP
In ANSI C, local variables in use at a
.B setjmp()
are no longer guaranteed to be preserved when a
.B longjmp()
occurs, unless they are declared \fBvolatile\fP. This
is not a problem for the Vax port, since the Vax longjmp()
will continue to restore the registers, but gcc warns about this
situation, since code that assumes restoration is not portable.
We have not yet worked on fixes for this.
.PP
Five or ten other miscellaneous bugs were caught and fixed.
.SH
Least portable
.UX
code
.PP
The process of porting software inevitably uncovers
a few files that cause a disproportionate share of problems.
For our port,
the clear winner is
.I efl ,
the Extended Fortran Language, by Stu Feldman.
It defines ``\fBtypedef int * ptr;\fP'' in a header file,
and then uses a ``ptr'' to point to anything.
GCC produced
1600 lines of errors messages on this program alone, and three modules
of it caused compiler core dumps. We ended
up deciding to abandon support for it rather than attempt to clean
it up.
.PP
A runner-up is
.I pcc ,
the Portable C Compiler itself, by Steven C. Johnson.
It caused GCC to coredump twice, tickled another GCC parsing bug,
and contained the modified typedef and sign extension problems mentioned above.
.PP
Third place goes to
.I monop ,
the Monopoly\(dg
.FS \(dg
Trademark of Parker Brothers
.FE
game, by Ken Arnold. This
program used a variety of typed pointers, but the main pointer to
a set of structs was declared as a \fBchar *\fP. Another part of
the code initialized an array of struct pointers with integer values,
then a small loop at the beginning of the game would read out these
integers and replace them with corresponding ``real'' struct pointers.
It took about two days to face up to the job and about a day to clean
it up.
.PP
Honorable mention for silly mistakes goes to the
.I indent
program, by someone at the University of Illinois.
It contain the only instance of
.B "a + = b"
(with a space between + and =), and was the only module
to terminate its
.B #include
directives with a semicolon.
It also contained a comparison between a character and the value 0200,
a value that a signed 8-bit char can never hold.
.SH
Results
.PP
We are pleased with the results so far. Most of the
.UX
code compiled
without problems, and the parts which we have executed are free from
code generation bugs.
The worst of the ANSI C changes only required roughly fifty modules
to be changed, and there were only two problems of this magnitude.
A total of
twenty bugs in gcc were located so far, and most of them are now fixed.
We expected several times this many bugs; the compiler is in better
shape than any of us expected.
.PP
Many minor type problems and ``nit'' incompatabilities with ANSI C have
been removed from the
.UX
sources.
.SH
Future Results
.PP
\fI(This section will move to \fBResults\fP for the final paper.)\fP
.PP
We expect that the size of the
.UX
binaries will be significantly less than
with the previous compiler, but at the current stage of the project
we can't easily confirm the expectation.
.PP
When the system compiled with GCC is in everyday use at Berkeley, GCC
will be relabeled as a full production-quality compiler, which will
encourage its wider use.
.SH
Non-Results
.PP
We have not attempted to make Berkeley
.UX
fully ANSI C compliant.
In particular, we have retained preprocessor comments (#endif FOO)
as well as machine-specific \fB#define\fP's (#ifdef vax). GCC supports
these features without trouble, even though ANSI C does not.
.PP
The
.UX
kernel has not yet been ported to gcc. Other people are working on
this, compiling one module at a time and running it for a while before
moving on to the next. We will merge their work with
ours once we have the rest of the system in a stable state.
.PP
Pieces of the Portable C Compiler are still being used inside
.I "lint, f77" ,
and
.I pc .
Eventually someone will write Fortran and Pascal front-ends for gcc;
this has already been done for C++. So far nobody has created a GNU
\fIlint\fP, but it is an obvious project.
.PP
CSRG has ported Berkeley
.UX
to the Tahoe, a fast Vax-like machine
built by Computer Consoles and resold by Harris and others. We are looking
for someone to do a Tahoe port of gcc, to replace the PCC supplied by CCI.
.SH
Problems in Building
.UX
.PP
.UX
compilers traditionally look in certain global places in the
file system for their libraries, include files, etc. This is a problem
when cross-compiling, or when building a new
.UX
release (which almost
amounts to the same thing). While it is possible to provide a new
default directory for
.B #include
files, if a source program
.B #include s
a file that is not in the cross-compilation include files,
the C compiler will erroneously use the one from /usr/include.
There should be a switch that turns off \fIall\fP the built-in include
file and library pathnames, and only uses those specified on the
compiler's command line.
.PP
However, there is still the problem of getting those switches to the
compiler's command line.
.I Make
is a great tool for dealing with one directory's worth of files,
but as
.UX
has evolved, \fImake\fP has not kept up. Indeed, it has fallen behind;
Makefiles that worked perfectly well five years ago will no longer
work because each manufacturer (AT&T especially) has hacked up their
.I make
to include harmful, gratuitous, and mutually incompatible changes.
The result is that a Makefile that works on your system is unlikely
to work on your neighbor's system, unless they are from the same manufacturer,
and you happen to use the same login shell.
.PP
.I Make
works poorly on nested directory structures, too.
As an example, we could find no way to change ``cc'' to ``gcc'' in all
the
Makefiles used to build Berkeley
.UX
(short of text-editing them all).
In a single directory, you can say
.I "make CC=gcc" ,
but this change is not propagated to subdirectories. You can manually
propagate that change one level by saying
.I "make CC=gcc MFLAGS='CC=gcc'"
but that only goes one level (at least in Berkeley's version of
.I make ).
We ended up putting a copy of gcc in a private
.I bin
directory, named
.I cc ,
and putting that directory on the front of the search path.
(When we later wanted to override CFLAGS as well, \fI~/bin/cc\fP
became a shell script that invokes
.I "gcc -W" ).
.PP
Another problem with
.I make
is that even if it was instructed to ignore errors (with -i or -k), it exits
if it can't locate a file that something else depends upon. This has the
effect of ``pruning'' a potentially large section
of the source hierarchy, and the
only warning is an unobtrusive
message buried among 500K of other output.
.PP
Of course, if someone was to fix these bugs in \fImake\fP, they would
be creating yet another incompatible version.
I have been watching the papers on the ``new makes'' and so far there
doesn't seem to be one that handles deeply nested
source trees in a clean and consistent fashion, or is otherwise
so much better than \fImake\fP that it's worth the effort to switch.
I think it is time to look for a completely new paradigm for
software compilation control. I don't have any major insights on where
to go from here, but it is clear to me that
.I make
and its derivatives have reached their useful limits.
.SH
Availability
.PP
These changes will be available to recipients of Berkeley's next software
distribution, whenever that is. We will also make diffs available
to others involved in porting
.UX
to ANSI C. We suspect that most of the
problems we solved have already been handled in one or another
.UX
port, but the work had to be duplicated because either it was not
sent back to Berkeley or AT&T, or the changes were not accepted. (AT&T
has a history of pretending that
.UX
bugs do not exist, and
Berkeley has limited manpower).
.SH
Future Work
.PP
Future projects include building a complete set of ANSI C and POSIX
compatible include files and libraries (including function prototypes),
and converting the existing sources to use them. An eventual goal
is to produce a fully standard-conforming
.UX
system \(em not only in
the interface provided to users, but with sources which will compile
and run on any standard-conforming compiler and libraries.
.PP
The success of this collaboration between GNU and CSRG has encouraged further
cooperation. Both parties feel that AT&T licensing
is a problem; most recipients of CSRG releases have old
.UX
licenses,
and are unwilling to upgrade to more expensive and more onerous AT&T
licenses. However, new AT&T releases include some features which would
be useful in Berkeley
.UX .
The GNU project is working to provide
early reimplementations of these features, such as improved shells and
``make'' commands. In return, CSRG is working to release software to
the public which has previously been held to be ``
.UX
licensed'' even though
it was not derived from AT&T code, such as the implementation
of TCP/IP, and many of the Berkeley utility programs.
.SH
References
.LP
\fIDraft Proposed American National Standard \(em Programming Language C\fP,
ANSI X3.J11, draft of October 1, 1986 (update for new draft when out).
CBEMA, 311 First Street NW #1500, Washington DC 20001.
.LP
\fI4.3BSD Manual Set\fP,
Computer Systems Research Group, University of California
at Berkeley.
.LP
Fowler, Glenn S., ``The Fourth Generation Make'', Usenix conference
proceedings, Summer 1985, page 159. (More references on ``make''
are provided in this paper.)
.LP
Hume, Andrew, ``Mk: a successor to make'', Usenix conference
proceedings, Summer 1987, page 445.
.LP
Kernighan, Brian W. and Ritchie, Dennis M., ``\fIThe
C Programming Language\fP'', Prentice-Hall, 1978.