diff --git a/README.md b/README.md index 9f17003..f85c5d2 100644 --- a/README.md +++ b/README.md @@ -4,11 +4,11 @@ program. There are many such competitors, because many people over the years have been dissatisfied with make's limitations. However, of all the replacements I've seen, only redo captures the essential simplicity and -flexibility of make, while eliminating its flaws. To my great surprise, it +flexibility of make, while avoiding its flaws. To my great surprise, it manages to do this while being simultaneously simpler than make, more flexible than make, and more powerful than make. -Although I wrote redo and I would love to take credit for it, this magical +Although I wrote redo and I would love to take credit for it, the magical simplicity and flexibility comes because I copied verbatim a design by Daniel J. Bernstein (creator of qmail and djbdns, among many other useful things). He posted some very terse notes on his web site at one point @@ -24,25 +24,27 @@ way](http://thedjbway.b0llix.net/future.html), by Wayne Marshall. After I found out about djb redo, I searched the Internet for any sign that other people had discovered what I had: a hidden, unimplemented gem of -brilliant code design. I found only one interesting link: Alan Grosskurth, whose -[Master's thesis at the University of -Waterloo](http://grosskurth.ca/papers/mmath-thesis.pdf) was about top-down -software rebuilding, that is, djb redo. He wrote his own (admittedly slow) -implementation in about 250 lines of shell script. +brilliant code design. I found only one interesting link: Alan Grosskurth, +whose [Master's thesis at the University of Waterloo](http://grosskurth.ca/papers/mmath-thesis.pdf) +was about top-down software rebuilding, that is, djb redo. He wrote his +own (admittedly slow) implementation in about 250 lines of shell script. If you've ever thought about rewriting GNU make from scratch, the idea of doing it in 250 lines of shell script probably didn't occur to you. redo is so simple that it's actually possible. For testing, I actually wrote an -even more minimal version (which always rebuilds everything instead of -checking dependencies) in only 70 lines of shell. +even more minimal version,which always rebuilds everything instead of +checking dependencies, in 99 lines of shell (less than 2 kbytes). The design is simply that good. -This implementation of redo is called `redo` for the same reason that there -are 75 different versions of `make` that are all called `make`. It's just +My implementation of redo is called `redo` for the same reason that there +are 75 different versions of `make` that are all called `make`. It's somehow easier that way. Hopefully it will turn out to be compatible with the other implementations, should there be any. +My extremely minimal implementation, called `do`, is in the `minimal/` +directory of this repository. + # License @@ -60,7 +62,11 @@ get all the speed of non-recursive `make` (only check dependencies once per run) combined with all the cleanliness of recursive `make` (you don't have code from one module stomping on code from another module). -But the easiest way to show it is with an example. +(Disclaimer: my current implementation is not as fast as `make` for some +things, because it's written in python. Eventually I'll rewrite it an C and +it'll be very, very fast.) + +The easiest way to show it is with an example. Create a file called default.o.do: redo-ifchange $1.c @@ -110,9 +116,8 @@ the autodependencies? The filename `default.o.do` means "run this script to generate a .o file unless there's a more specific whatever.o.do script that applies." -The amazing innovation in redo - and really, this is the key innovation that -makes everything else work - is that declaring a dependency is just another -shell command. The `redo-ifchange` command means, "build each of my +The key thing to understand about redo is that declaring a dependency is just +another shell command. The `redo-ifchange` command means, "build each of my arguments. If any of them or their dependencies ever change, then I need to run the *current script* over again." @@ -126,13 +131,14 @@ just once right at the start of your project build. But best of all, as you can see in `default.o.do`, you can declare a dependency *after* building the program. In C, you get your best dependency information by trying to actually build, since that's how you find out which -headers you need. redo offers the following simple insight: you don't actually -care what the dependencies are *before* you build the file; if the file +headers you need. redo is based on the following simple insight: +you don't actually +care what the dependencies are *before* you build the target; if the target doesn't exist, you obviously need to build it. Then, the build script itself can provide the dependency information however it wants; unlike in `make`, you don't need a special dependency syntax at all. You can even -declare your dependencies after building, which makes everything much -simpler. +declare some of your dependencies after building, which makes C-style +autodependencies much simpler. (GNU make supports putting some of your dependencies in include files, and auto-reloading those include files if they change. But this is very @@ -142,16 +148,18 @@ changes. With redo, you can just read the script from top to bottom. A `redo-ifchange` call is like calling a function, which you can also read from top to bottom.) + # One script per file? Can't I just put it all in one big Redofile like make does? One of my favourite features of redo is that it doesn't add any new syntax; -the syntax of redo is *just* the syntax of sh. +the syntax of redo is *exactly* the syntax of sh... because sh is the program +interpreting your .do file. Also, it's surprisingly useful to have each build script in its own file; that way, you can declare a dependency on just that one build script instead of the entire Makefile, and you won't have to rebuild everything just -because of a one-line Makefile change. (Some build tools avoid that -by tracking exactly which variables and commands were used to do the build. +because of a one-line Makefile change. (Some build tools avoid that same +problem by tracking which variables and commands were used to do the build. But that's more complex, more error prone, and slower.) Still, it would be rather easy to make a "Redofile" parser that just has a @@ -167,9 +175,8 @@ into their own files. You could even write a .do file to do it. It's not obvious that this would be a real improvement however. -See djb's [Target files depend on build -scripts](http://cr.yp.to/redo/honest-script.html) article for more -information. +See djb's [Target files depend on build scripts](http://cr.yp.to/redo/honest-script.html) +article for more information. # What are the three parameters ($1, $2, $3) to a .do file? @@ -201,20 +208,23 @@ You should use $1 and $2 only in constructing input filenames and dependencies; never modify the file named by $1 in your script. Only ever write to the file named by $3. That way redo can guarantee proper dependency -management and atomicity. +management and atomicity. (For convenience, you can write +to stdout instead of $3 if you want.) For example, you could compile a .c file into a .o file like this, from a script named `default.o.do`: - + + redo-ifchange $1.c gcc -o $3 -c $1.c -Note that $2, the .o extension, is rarely useful. +Note that $2, the output file's .o extension, is rarely useful +since you always know what it is. FIXME: djb's design documentation doesn't clearly describe $1 and $2, although it's clear that $3 is the output filename. We may have guessed $1 and $2, particularly $2, incorrectly, so we might have to change their meanings -later. +later in order to be compatible with djb's implementation. # What happens to the stdin/stdout/stderr in a redo file? @@ -244,13 +254,13 @@ Not currently. There's nothing fundamentally preventing us from allowing it. However, it seems easier to reason about your build process if you *aren't* auto-generating your build scripts on the fly. -This might change. +This might change someday. # Do end users have to have redo installed in order to build my project? -No. We include a very short (70 lines, as of this writing) shell script -called `do` in the `minimal` subdirectory of the redo project. `do` is like +No. We include a very short (99 lines, as of this writing) shell script +called `do` in the `minimal/` subdirectory of the redo project. `do` is like `redo` (and it works with the same `*.do` scripts), except it doesn't understand dependencies; it just always rebuilds everything from the top. @@ -264,11 +274,14 @@ matters, you could just include it with your project. # How does redo store dependencies? -FIXME: -Currently, in a directory called `.redo` that's full of text files. This -isn't really optimal, so it will change eventually. Please consider the -storage format undocumented (but feel free to poke around and look; it's -simple enough). +At the toplevel of your project, redo creates a directory +named `.redo`. That directory contains a sqlite3 database +with dependency information. + +The format of the `.redo` directory is undocumented because +it may change at any time. If you really need to make a +took that pokes around in there, please ask on the mailing +list if we can standardize something for you. # If a target didn't change, how to I prevent dependents from being rebuilt? @@ -280,28 +293,35 @@ identical. With `make`, which makes build decisions based on timestamps, you would simply have the ./configure script write to config.h.new, then only -overwrite config.h with that if the two files are different. +overwrite config.h with that if the two files are different. +However, that's a bit tedious. -FIXME: -This is not possible in the current version of `redo`. redo knows whenever it -rebuilds a target and doesn't bother re-checking dependencies after that; -even if the file didn't technically change, it is considered "rebuilt," -which means all its dependants now need to be rebuilt. +With `redo`, there's an easier way. You can have a +config.do script that looks like this: -The advantage of this method is you can't accidentally prevent the -rebuilding of things just by marking the target files as "newer" or marking -the source files as "older" (as sometimes happens when you extract an old -tarball or backup on top of your source code files). The disadvantage is -unnecessary rebuilding of some stuff sometimes. + redo-ifchange autogen.sh *.ac + ./autogen.sh + ./configure + cat config.h configure Makefile | redo-stamp + +Now any of your other .do files can depend on a target called +`config`. `config` gets rebuilt automatically if any of +your autoconf input files are changed (or if someone does +`redo config` to force it). But because of the call to +redo-stamp, `config` is only considered to have changed if +the contents of config.h, configure, or Makefile are +different than they were before. -We will have to find a solution to this before redo 1.0. - -See also the next question. +(Note that you might actually want to break this .do up into a +few phases: for example, one that runs aclocal, one that +runs autoconf, and one that runs ./configure. That way +your build can always do the minimum amount of work +necessary.) -# What about checksum-based dependencies instead of timestamp-based ones? +# Why not always use checksum-based dependencies instead of timestamps? -Some build systems keep a checksum of target files and rebuild dependants +Some build systems keep a checksum of target files and rebuild dependents only when the target changes. This is appealing in some cases; for example, with ./configure generating config.h, it could just go ahead and generate config.h; the build system would be smart enough to rebuild or not rebuild @@ -309,24 +329,23 @@ dependencies automatically. This keeps build scripts simple and gets rid of the need for people to re-implement file comparison over and over in every project or for multiple files in the same project. -I think this would be a good addition to `redo` - and not a very difficult -one. +There are disadvantages to using checksums for everything, +however: -Probably we should add a new command similar to `redo-ifchange`; let's call -it `redo-ifsum` or `redo-ifdiff`. That command would verify checksums -instead of timestamps. +- calculating checksums for every output file adds time to + the build; + +- it makes it hard to *force* things to rebuild when you + know you absolutely want that; + +- targets that are just used for aggregation (ie. they + don't produce any output of their own) would always have + the same checksum - the checksum of a zero-byte file - + which causes confusing results. -Sometimes you don't want to use checksums for verification; for example, in -some complicated build systems, you want to create empty `something.stamp` -files to indicate that some big complex operation has completed -successfully. But empty files all have the same checksum, so perhaps you'd -rather just use a timestamp comparison in that case. (Alternatively, you -could fill the file with data - maybe a series of checksums - indicating the -state of the data that was produced. If that data changed, the stamp would -then be out of date.) - -FIXME: This requires a bit more thought before we commit to any particular -option. +Thus, we made the decision to only use checksums for +targets that explicitly call `redo-stamp` (see previous +question). # Can my .do files be written in a language other than sh? @@ -337,29 +356,41 @@ shell to run your .do script. But that opens new problems, like figuring out what is the equivalent of the `redo-ifchange` command in python. Do you just run it in a subprocess? That might be unnecessarily slow. And so on. -Right now, `redo` explicitly runs `sh -c filename.do`. The main reasons for +Right now, `redo` explicitly runs `sh filename.do`. The main reasons for this are to make the #!/ line optional, and so you don't have to remember to `chmod +x` your .do files. # Can a single .do script generate multiple outputs? -FIXME: Not presently. This seems like a useful feature though. +FIXME: Yes, but this is a bit imperfect. -For example, you could have a file called `default.do.do` that would -generate .do files from a `Redofile`. Then you wouldn't have to argue with -the `redo` maintainers about whether putting stuff into a single `Redofile` -is better than the current behaviour. +For example, compiling a .java file produces a bunch of .class +files, but exactly which files? It depends on the content +of the .java file. Ideally, we would like to allow our .do +file to compile the .java file, note which .class files +were generated, and tell redo about it for dependency +checking. -Right now you could do that, except you would want to parse the `Redofile` -only once and produce a bunch of `.do` files from that single action. But -you would still want `redo` to know which `.do` files were produced, so it -could rerun the splitter script, if `Redofile` ever changed, before using -one of the generated `.do` files. +However, this ends up being confusing; if myprog depends +on foo.class, we know that foo.class was generated from +bar.java only *after* bar.java has been compiled. But how +do you know, the first time someone asks to build myprog, +where foo.class is supposed to come from? -It would also be useful, again, with ./configure: typically running the -configure script produces several output files, and it would be nice to -declare dependencies on all of them. +So we haven't thought about this enough yet. + +Note that it's *okay* for a .do file to produce targets +other than the advertised one; you just have to be careful. +You could have a default.javac.do that runs 'javac +$1.java', and then have your program depend on a bunch of .javac +files. Just be careful not to depend on the .class files +themselves, since redo won't know how to regenerate them. + +This feature would also be useful, again, with ./configure: +typically running the configure script produces several +output files, and it would be nice to declare dependencies +on all of them. # Recursive make is considered harmful. Isn't redo even *more* recursive? @@ -370,33 +401,41 @@ by Peter Miller. Yes, redo is recursive, in the sense that every target is built by its own `.do` file, and every `.do` file is a shell script being run recursively from other shell scripts, which might call back into `redo`. In fact, it's -even more recursive than recursive make. +even more recursive than recursive make. There is no +non-recursive way to use redo. -However the reason recursive make is considered harmful is that each +However, the reason recursive make is considered harmful is that each instance of make has no access to the dependency information seen by the other instances. Each one starts from its own Makefile, which only has a -partial picture of what's going on; moreover, each one has to stat a lot of -the same files over again, leading to slowness. That's the thesis of the -"considered harmful" paper. +partial picture of what's going on; moreover, each one has to +stat() a lot of the same files over again, leading to slowness. That's +the thesis of the "considered harmful" paper. -On the other hand, nobody has written a paper about it, but non-recursive -make should also be considered harmful. The problem is Makefiles aren't +Nobody has written a paper about it, but *non-recursive* +make should also be considered harmful! The problem is Makefiles aren't very "hygienic" or "modular"; if you're not running make recursively, then your one copy of make has to know *everything* about *everything* in your entire project. Every variable in make is global, so every variable defined in *any* of your Makefiles is visible in *all* of your Makefiles. Every little private function or macro is visible everywhere. In a huge project made up of multiple projects from multiple vendors, that's just not okay. +Plus, if all your Makefiles are tangled together, make has +to read and parse the entire mess even to build the +smallest, simplest target file, making it slow. -`redo` deftly manages to dodge both sets of problems. First of all, +`redo` deftly dodges both the problems of recursive make +and the problems of non-recursive make. First of all, dependency information is shared through a global persistent `.redo` -directory, which is accessed by all your `redo` instances at once. +database, which is accessed by all your `redo` instances at once. Dependencies created or checked by one instance can be immediately used by another instance. And there's locking to prevent two instances from building the same target at the same time. So you get all the "global -dependency" knowledge of non-recursive make. +dependency" knowledge of non-recursive make. And it's a +binary file, so you can just grab the dependency +information you need right now, rather than going through +everything linearly. -But also, every `.do` script is entirely hygienic and traceable; `redo` +Also, every `.do` script is entirely hygienic and traceable; `redo` discourages the use of global environment variables, suggesting that you put settings into files (which can have timestamps and dependencies) instead. So you also get all the hygiene and modularity advantages of recursive make. @@ -408,8 +447,9 @@ non-recursive Makefile setup with a bunch of included files, you end up with lots and lots of rules that can all be executed in a random order; tracing becomes impossible. Recursive make tries to compensate for this by breaking the rules into subsections, but that ends up with all the "considered harmful" -paper's complaints. `redo` just runs from top to bottom in a nice tree, so -it's traceable no matter how many layers you have. +paper's complaints. `redo` runs your scripts from top to bottom in a +nice tree, so it's traceable no matter how many layers you have. + # How do I set environment variables that affect the entire build? @@ -429,14 +469,9 @@ create a file called `compile.do`: redo-ifchange config.sh . ./config.sh - echo "gcc -c -o \$3 $1.c $CFLAGS" + echo "gcc -c -o \$3 $1.c $CFLAGS" >$3 chmod a+x $3 -(Note that if a .do script produces data to stdout, like we have here, then -`redo` will use that to atomically update the target file. So you don't -have to redirect it yourself, or worry about using a temp file and then -renaming it.) - Then, your `default.o.do` can simply look like this: redo-ifchange compile $1.c @@ -457,13 +492,15 @@ command line is being used. As a bonus, all the variable expansions only need to be done once: when generating the ./compile program. With make, it would be recalculating -expansions every time it compiles a file. +expansions every time it compiles a file. Because of the +way make does expansions as macros instead of as normal +variables, this can be slow. + # How do I write a default.o.do that works for both C and C++ source? -See the previous answer for why you will probably use a compile.do instead. -Then your default.o.do looks like it does in the previous answer. We can -then upgrade compile.do to look something like this: +We can upgrade the compile.do from the previous answer to +look something like this: redo-ifchange config.sh . ./config.sh @@ -485,14 +522,18 @@ in make: Then it has to do all the same checks. Except make has even *more* implicit rules than that, so it ends up trying and discarding lots of possibilities -before it actually builds your program. +before it actually builds your program. Is there a %.s? A +%.cpp? A %.pas? It needs to look for *all* of them, and +it gets slow. The more implicit rules you have, the slower +make gets. -So in this case, it's not implicit at all; you're specifying exactly how to +In redo, it's not implicit at all; you're specifying exactly how to decide whether it's a C program or a C++ program, and what to do in each -case. You can also share the two gcc command lines between the two rules, +case. Plus you can share the two gcc command lines between the two rules, which is hard in make. (In GNU make you can use macro functions, but the syntax for those is ugly.) + # Can I just rebuild a part of the project? Absolutely! Although `redo` runs "top down" in the sense of one .do file @@ -504,7 +545,7 @@ you start, `redo` will be able to build all the dependencies in the right order. Unlike non-recursive make, you don't have to jump through any strange hoops -(like adding a fake Makefile in each directory that does `make -C ${TOPDIR}` +(like adding, in each directory, a fake Makefile that does `make -C ${TOPDIR}` back up to the main non-recursive Makefile). redo just uses `filename.do` to build `filename`, or uses `default*.do` if the specific `filename.do` doesn't exist. @@ -518,6 +559,31 @@ And it will work exactly like this: cd ../utils redo foo.o + +In make, if you run + + make ../utils/foo.o + +it means to look in ./Makefile for a rule called +../utils/foo.o... and it probably doesn't have such a +rule. On the other hand, if you run + + cd ../utils + make foo.o + +it means to look in ../utils/Makefile and look for a rule +called foo.o. And that might do something totally +different! redo combines these two forms and does +the right thing in both cases. + +Note: redo will always change to the directory containing +the target before trying to build it. So if you do + + redo ../utils/foo.o + +the .do file will be run with its current directory set to +../utils. Thus, the .do file's runtime environment is +always reliable. # Can my filenames have spaces in them? @@ -538,8 +604,8 @@ No. # What if my .c file depends on a generated .h file? This problem arises as follows. foo.c includes config.h, and config.h is -created by running ./configure. The second part is easy; just write a config.h.do -that depends on the existence of configure (which is created by +created by running ./configure. The second part is easy; just write a +config.h.do that depends on the existence of configure (which is created by configure.do, which probably runs autoconf). The first part, however, is not so easy. Normally, the headers that a C @@ -549,46 +615,44 @@ you do redo foo.o -There's no way for redo to know that compiling foo.c into foo.o depends on -first generating config.h. +There's no way for redo to *automatically* know that compiling foo.c +into foo.o depends on first generating config.h. -FIXME: -There seem to be a few workarounds for this, but none are very good. You -can explicitly make all .o files depend on config.h (which will cause things -to get recompiled unnecessarily). You can just tell people to run -./configure before running redo (which is what make users typically do). +Since most .h files are *not* auto-generated, the easiest +thing to do is probably to just add a line like this to +your default.o.do: -You can try to guess dependencies using `gcc -MM -MG` (-MG means don't die -if headers aren't present), but that's not actually very helpful, because of -the compiler's header include paths: if I #include "foo.h", which directory -is redo supposed to look at to find foo.h? All of them? Normally a C -file's autodependencies have full paths attached, so this would be special. + redo-ifchange config.h + +Sometimes a specific solution is much easier than a general +one. +If you really want to solve the general case, [djb has a solution for his own projects](http://cr.yp.to/redo/honest-nonfile.html), which is a simple script that looks through C files to pull out #include lines. He assumes that `#include ` is a system header (thus not subject to being built) and `#include "file.h"` is in the current directory (thus easy to -find). Unfortunately this isn't really a complete solution in the general -case. - -Maybe just telling redo to try to generate the file in all possible -locations is the best way to go. But we need to think about this more. - -In the meantime, you'll have to explicitly specify such dependencies in your -*.o.do files. +find). Unfortunately this isn't really a complete +solution, but at least it would be able to redo-ifchange a +required header before compiling a program that requires +that header. -# Why don't you by default print the commands as they are run? +# Why doesn't redo by default print the commands as they are run? make prints the commands it runs as it runs them. redo doesn't, although -you can get this behaviour with `redo -v`. +you can get this behaviour with `redo -v` or `redo -x`. +(The difference between -v and -x is the same as it is in +sh... because we simply forward those options onward to sh +as it runs your .do script.) -The main reason we made this decision is that the commands get pretty long +The main reason we don't do this by default is that the commands get +pretty long winded (a compiler command line might be multiple lines of repeated gibberish) and, on large projects, it's hard to actually see the progress of the overall build. Thus, make users often work hard to have make hide the -command output in order to make the log "more useful." +command output in order to make the log "more readable." The reduced output is a pain with make, however, because if there's ever a problem, you're left wondering exactly what commands were run at what time, @@ -628,15 +692,19 @@ you can cut-and-paste a line from the build script and rerun it directly. So if you ever want to debug what happened at a particular step, you can choose to run only that step in verbose mode: - $ ./redo t/c.c.c.b.b -v + $ redo t/c.c.c.b.b -x redo t/c.c.c.b.b - redo-ifchange "$1$2.a" - echo a-to-b - cat "$1$2.a" + * sh -ex default.b.do c.c.c.b .b c.c.c.b.b.redo2.tmp + + redo-ifchange c.c.c.b.b.a + + echo a-to-b + + cat c.c.c.b.b.a + + ./sleep 1.1 + redo t/c.c.c.b.b (done) + If you're using an autobuilder or something that logs build results for future examination, you should probably set it to always run redo with -the -v option. +the -x option. # Is redo compatible with autoconf? @@ -654,14 +722,15 @@ Hells no. You can thank me later. But see next question. # Is redo compatible with make? Yes. If you have an existing Makefile (for example, in one of your -subprojects), you can just call make to build that subproject. +subprojects), you can just call make from a .do script to build that +subproject. In a file called myproject.stamp.do: redo-ifchange $(find myproject -name '*.[ch]') make -C myproject all -So, to amend our answer to the previous question, you can include +So, to amend our answer to the previous question, you *can* use automake-generated Makefiles as part of your redo-based project. @@ -670,7 +739,27 @@ automake-generated Makefiles as part of your redo-based project. Yes! redo implements the same jobserver protocol as GNU make, which means that redo running under make -j, or make running under redo -j, will do the right thing. Thus, it's safe to mix-and-match redo and make in a recursive -build system. Just make sure you declare your dependencies correctly. +build system. + +Just make sure you declare your dependencies correctly; +redo won't know all the specific dependencies included in +your Makefile, and make won't know your redo dependencies, +of course. + +One way of cheating is to just have your make.do script +depend on *all* the source files of a subproject, like +this: + + make -C subproject all + find subproject -name '*.[ch]' | xargs redo-ifchange + +Now if any of the .c or .h files in subproject are changed, +your make.do will run, which calls into the subproject to +rebuild anything that might be needed. Worst case, if the +dependencies are too generous, we end up calling 'make all' +more often than necessary. But 'make all' probably runs +pretty fast when there's nothing to do, so that's not so +bad. # What about distributed builds? @@ -696,12 +785,24 @@ same preparsed headers over and over again). At the time, he promised to open source this tool eventually. It would be pretty fun to play with it. -FIXME: However, it won't work as easily with redo as it did with make. With +The problem: + +This idea won't work as easily with redo as it did with make. With make, a separate copy of $SHELL is launched for each step of the build (and gets migrated to the remote machine), but make runs only on your local machine, so it can control parallelism and avoid building the same target -from multiple machines, and so on. With redo, since the entire script runs -inside the shell, we'd have to do the parallelism another way. +from multiple machines, and so on. The key to the above +distribution mechanism is it can send files to the remote +machine at the beginning of the $SHELL, and send them back +when the $SHELL exits, and know that nobody cares about +them in the meantime. With redo, since the entire script runs +inside a shell (and the shell might not exit until the very +end of the build), we'd have to do the parallelism some +other way. + +I'm sure it's doable, however. One nice thing about redo +is that the source code is so small compared to make: you +can just rewrite it. # How fast is redo compared to make? @@ -709,64 +810,65 @@ inside the shell, we'd have to do the parallelism another way. FIXME: The current version of redo is written in python and has not been optimized. So right now, it's usually a bit slower. Not too embarrassingly slower, -though, and the slowness really only strikes the first time you build a -project. +though, and the slowness mostly only strikes when you're +building a project from scratch. For incrementally building only the changed parts of the project, redo can be much faster than make, because it can check all the dependencies up front and doesn't need to repeatedly parse and re-parse the Makefile (as recursive make needs to do). -More bad news, however: the current version of redo also pointlessly -re-evaluates the same dependencies over and over. Unlike with make, there -is no good reason for this, so it'll be fixed eventually. +redo's sqlite3-based dependency database is very fast (and +it would be even faster if we rewrite redo in C instead of +python). Better still, it would be possible to write an +inotify daemon that can update the dependency database in +real time; if you're running the daemon, you can run 'redo' +from the toplevel and if your build is clean, it could return +instantly, no matter how many dependencies you have. -Even better, in 'redo -j' mode, it sometimes rebuilds the -same dependency more than once. Not at the same time, so -your build won't break, but it'll just be unnecessarily -slow. Obviously this should be fixed. +On my machine, redo can currently check about 10,000 +dependencies per second. As an example, a program that +depends on every single .c or .h file in the Linux kernel +2.6.36 repo (about 36000 files) can be checked in about 4 +seconds. -Furthermore, redo's current dependency database (in the `.redo` directory) -is not very efficient, and thrashes a lot of filesystem metadata, which is -especially slow on some filesystems, notably ext3. We'll want to improve that -by using a "real" data structure eventually. With a "real" data structure, -a daemon, and inotify, you could actually get redo's dependency evaluation -to run instantly. On very large projects with tens of thousands of files to -stat() before we can even start building, that could make a pretty big -difference. With make, an inotify implementation would be pretty hard to -do (since just parsing the Makefiles of such a project is complicated, and -there's no guarantee the dependencies are the same as last time). But with -the way redo works, this kind of optimization would be pretty easy. +Rewritten in C, dependency checking would probably go about +10 times faster still. -We'll probably have to rewrite redo in C eventually to speed it up further. -This won't be very hard; the design is so simple that it should be easy to -write in any language. It's just *even easier* in python, -which was good for writing a prototype. +This probably isn't too hard; the design of redo is so simple that +it should be easy to write in any language. It's just +*even easier* in python, which was good for writing the +prototype and debugging the parallelism and locking rules. -Most of the so-called slowness at the moment (it's really not that bad) -is because redo-ifchange (and also sh itself) need to be fork'd and -exec'd over and over during the build process. The `minimal/do` script -shows a way around this; redo-ifchange could be a shell function -instead of a standalone program. Eliminating the need for extra -fork/exec in the common case could actually get us much faster than -make even when doing an initial build; make executes $SHELL for every -command, but with redo, there is a shell running at all times anyway, -so if we're very careful we could optimize that out. +Most of the slowness at the moment is because redo-ifchange +(and also sh itself) need to be fork'd and exec'd over and +over during the build process. As a point of reference, on my computer, I can fork-exec -redo-ifchange.py about 87 times per second; an empty python program, -about 100 times per second; an empty C program, about 1000 times per -second; an empty make, about 300 times per second. So if I could -compile 87 files per second, which I can't, then python overhead -would be 50%. Also, if you're using redo -j on a multicore machine, all -the python forking happens in parallel with everything else, so that's -87 per second per core. Nevertheless, that's still slower than make and -should be fixed. +redo-ifchange.py about 87 times per second; an empty python +program, about 100 times per second; an empty C program, +about 1000 times per second; an empty make, about 300 times +per second. So if I could compile 87 files per second with +gcc, which I can't because gcc is slower than that, then +python overhead would be 50%. Since gcc is slower than +that, python overhead is generally much less - more like +10%. + +Also, if you're using redo -j on a multicore machine, all +the python forking happens in parallel with everything +else, so that's 87 per second per core. Nevertheless, +that's still slower than make and should be fixed. + +(On the other hand, all this measurement is confounded +because redo's more fine-grained dependencies mean you can +have more parallelism. So if you have a lot of CPU cores, redo +might build *faster* than make just because it makes better +use of them.) # What's missing? How can I help? -redo is thoroughly incomplete and probably has numerous bugs. Just what you +redo is incomplete and probably has numerous bugs. Just what you always wanted in a build system, I know. What's missing? Search for the word FIXME in this document; anything with a