makepp_speedup -- How to make makepp faster
So you think makepp is slow? It has gotten noticeably faster, but granted, it's still slow, especially if you come from GNU make. This is because it conscientiously checks all those things where gmake gives you a headache, ignoring lots of dependencies (the "I think I need to gmake clean to get rid of a mysterious bug" syndrome). If you suspect some Perl code you added to your makefiles might be at fault, take a look at perl_performance.
But there are a few things you can do to squeeze out more speed. Some of the things are labeled unsafe, in the sence that you're asking makepp not to check or do certain things, which you think are not needed. If these things would have been necessary, the build may not be correct. Luckily this problem will be temporary, however. It will get corrected as soon as you let makepp do all checks.
You can combine several of these tips to increase the time gain even more.
The stand-alone utility makeppreplay, mppr repeats things that makepp has already done, without any overhead.
Within version 5.8, all are roughly the same, only 5.8.7 is a bit faster. Tuning your Perl can also help, like not compiling it for 64 bits, which makepp doesn't need. For example ActiveState's build (http://www.activestate.com/activeperl) of 5.8.7 for Linux is faster than the Perl 5.8.7 that comes with SuSE Linux 10.0.
Each additional file you include is doubly penalizing. On the one hand, the compiler must look for and at all those files. You don't notice this so much, because it's just a little extra per compiler call. On the other hand makepp must look too, to find dependencies and figure out whether they incur a rebuild. Then it can seem to stall, while it is digesting a lot of dependencies at once.
An absolutely deadly variant is the project master include file, which in turn conveniently includes anything you might need. The result is that any header file change leads to a full build. Even without a change, makepp must think about all those headers again, for every source you compile. Just a tiny effort, since this is cached, but thousands of files can make this staggering.
It may be cumbersome to figure out the minimal set of includes, and to cleanup those no longer needed, but it really pays off. If anybody knows a tool that can identify which files get included unnecessarily, I'd be glad to mention it here!
If you have a default target which makes several programs, then makepp will have to check all their dependencies, right down to the smallest header file. But maybe you want to test your change with only one of those programs.
Then you would call makepp with an explicit target. The less modules or headers all those programs have in common, the greater the benefit of not letting makepp check them all.
Say your top level Makeppfile has this rule:
$(phony all): proggie1 proggie2 $(only_phony_targets */**/all)
Then you would call things like
$ makepp proggie2 $ makepp proggie1 dir/subdir/proggie27
Makepp looks for makefiles (unless you specify them explicitly on the command line or with
load-makefile) in the order RootMakeppfile, RootMakeppfile.mk, Makeppfile and Makeppfile.mk, followed by the classical makefile names. (The .mk variants are for purely suffix-based systems.)
So, if you use RootMakeppfile at the root of your build tree, and Makeppfile everywhere else, the files will be found slightly faster. Makepp will also have a slightly smaller memory consumption (caching the fact that the other names don't exist), which also means speed through less memory management.
Likewise if you have a statement
there will first be an attempt to find standard.makepp, so you might as well use that name.
Makepp keeps track not only of existent files, but also of any it learns to create. (That's why it offers reliable wildcards like *.o.) The price for this power is a lot of management. So, if you tell it how to create a .o from a .c, that's fine, because it will happen for most if not all candidates.
But if you tell it how to link any suffixless executable from a like named .o, that's expensive, because it will probably only happen for a small part of them (those that contain a main function), but the basis will get laid for all. You have to weigh the comfort of a linker pattern rule, against the efficiency of individual linker rules.
If you don't use any of them, you should also turn off the builtin rules with:
makepp_no_builtin = 1
If you do use them, but, for the reasons explained above, not the builtin linker rules, you should turn those off with:
makepp_no_builtin_linker = 1
Makepp offers very convenient possibilities of being extended through Perl. But if you write some functions, commands or statements in a file and include that from dozens of makefiles, you will get dozens of copies of them all in memory. And they will be read dozens of times by the makepp parser, which is a bit slower than Perl's.
In this situation it is better to put your own functions into a module.
If you have several developers working on the same machine or if you change to and fro between sets of build options, this is for you. Repositories allow you to offer a central reference where you only need to build what is locally different. A build cache simply collects all produced files, and reuses them as appropriate, with less planning needed. The latter page also describes the differences.
If your build is so big that makepp is having a hard time digesting all the information and if you can find a way of splitting it up into smaller independent parts, sandboxes might give you better parallelism than the
Makepp's logging feature is very powerful for tracking down bugs in the build system, or for analyzing your dependencies. Whenever you don't do these things, you can save quite a bit of formatting and I/O with
--stop) allows makepp to start its work while you are still editing. It will repeatedly suspend itself when it gets to the point analyzing the dependencies. You decide when you're ready to let it go on. On our huge project this saves half a minute, and that's only when we have a CPU to ourselves.
This method has two potential drawbacks:
proggie: *.o $(LD) $(inputs) -o $(output)
If you add another source file, or a file from which makepp knows how to generate a source, then
*.o should match the object that produces. But, if this file was added after starting makepp, it will not, because the wildcard was expanded too early.
In both of these cases you should kill the prestarted makepp and start it anew.
--gullible tells makepp to believe that a rule changes what it says it will, neither less nor more. Not performing these checks can save a few percent of makepp's CPU time. And the Disk I/O savings is especially welcome on network file systems. If you do nightly full builds in an empty directory with the
--repository option, but without the
--gullible option, you can be fairly sure that your rule set is consistent. Then this option shouldn't hurt in your daytime work.
These methods are unsafe if you give makepp the wrong hints. But everything will again be fine, however, as soon as you let makepp do all the checks, by not passing it any limiting options. For this reason I suggest using these hints to get quick intermediate builds, and use lunchtime and nights to let makepp do its job thoroughly.
This is the same tip of using explicit targets discussed under "Build as Little as You Need" above. But it becomes more dangerous, if you do it because you are sure that your change will not affect any of the other programs. Then they will not be built, even though it might have been necessary.
--dont-build is very powerful for speeding makepp up a lot. If you know one or more directories, which you are sure are unaffected by any change you made since the last time, you can issue
--dont-build options for them. This can save makepp a lot of dependency analysis. But it will not build anything in those directories, even if it should have.
This is the same as "Know where not to build", but instead of an exclusion list, you supply an inclusion list. The trick is that a
--do-build option, with a
--dont-build=/ option or under a
RootMakeppfile(.mk) directory without a
--dont-build option on a higher level directory means: build nothing except what I tell you to. This is what users of traditional makes are looking for when they want to build just one directory:
$ makepp --do-build=dir/subdir
or, if you don't have a
$ makepp --dont-build=/ --do-build=dir/subdir
The difference is that any default target in the top level Makeppfile, i.e. link commands are also executed this way. If you don't want that, you must give an explicit target, which is automatically also marked for
$ makepp --do-build=dir1/subdir dir2/proggie
An extreme variant is asking makepp not to build anything but what you tell it to. This is not so dangerous if you changed no include files, only modules, and you know which programs they go into.
Say you have only changed
src/b.cpp and these are linked directly into one program. Dot is the current directory including all subdirectories.
$ makepp --dont-build=. src/a.o src/b.o proggie1
Or equivalently, because a
--do-build option, without a
--dont-build option on a higher level directory implies
--dont-build for the root of the build tree:
$ makepp --do-build=src/a.o src/b.o proggie1
You can do something like the following in your Shell's $ENV file or .profile to save typing (csh users replace '=' with ' '):
alias mppb='makepp --do-build' alias mppsb='makepp --stop --do-build'
Then the last example becomes:
$ mppb src/a.o src/b.o proggie1
Modern computers, especially servers, typically have a high mean time between failure. If this is the case for you, and you have lots of RAM to spare, you can save the time you wait for I/O. You should edit on a real disk, or replicate your edits there quickly. But the build results are reproducible, so they can reside in RAM. If you don't want to risk rebuilding, you can always replicate to disk after each build or at night. You should not do this during the build, as you might catch partially written files, just as if the machine had crashed.
If you have a system and/or storage unit with good caching and RAID, the gain might not be so big.
Daniel Pfeiffer <email@example.com>