aboutsummaryrefslogtreecommitdiffhomepage
path: root/dev
diff options
context:
space:
mode:
authorGravatar Pierre Letouzey <pierre.letouzey@inria.fr>2014-02-27 18:03:49 +0100
committerGravatar Pierre Letouzey <pierre.letouzey@inria.fr>2014-02-27 18:48:23 +0100
commite372b7244bcb8cc449196c29a7dabee6f7f84aa2 (patch)
treed249777e9bdc9ba03aaf1f4455f1a1aee3e8c274 /dev
parentca64cc032a3f03381d00d7c9b128688f3f920844 (diff)
Makefile: re-introduce 2 phases to avoid make strange -include's
Yet another revision of the build system. We avoid relying on the awkward include-which-fails-but-works-finally-after-a-retry feature of gnu make. This was working, but was quite hard to understand. Instead, we reuse the idea of two explicit phases (as in 2007 and its stage{1,2,3}), but in a lighter way. The main Makefile calls Makefile.build twice : - first for building grammar.cma (and q_constr.cmo), with a restricted set of .ml4 to consider (see variables BUILDGRAMMAR and ML4BASEFILES). - then on the true target(s) asked by the user (using the special variable MAKECMDGOALS). In pratice, this should change very little to the concrete developper's life, let me know otherwise. A few more messages of make due to the explicit first sub-call, but no more strange "not ready yet" messages... Btw: we should handle correctly now the parallel compilation of multiple targets (e.g. make -jX foo bar). As reported by Pierre B, this was triggering earlier two separate sub-calls to Makefile.build, one for foo, the other for bar, with possibly nasty interactions in case of parallelism. In addition, some cleanup of Makefile.build, removal of useless :: rules, etc etc.
Diffstat (limited to 'dev')
-rw-r--r--dev/doc/build-system.dev.txt130
-rw-r--r--dev/doc/build-system.txt248
2 files changed, 88 insertions, 290 deletions
diff --git a/dev/doc/build-system.dev.txt b/dev/doc/build-system.dev.txt
index 3d9cba143..af1120e97 100644
--- a/dev/doc/build-system.dev.txt
+++ b/dev/doc/build-system.dev.txt
@@ -1,21 +1,38 @@
-Since July 2007, Coq features a build system overhauled by Pierre
-Corbineau and Lionel Elie Mamane.
-
----------------------------------------------------------------------
-WARNING:
-In March 2010 this build system has been heavily adapted by Pierre
-Letouzey. In particular there no more explicit stage1,2. Stage3
-was removed some time ago when coqdep was splitted into coqdep_boot
-and full coqdep. Ideas are still similar to what is describe below,
-but:
-1) .ml4 are explicitely turned into .ml files, which stay after build
-2) we let "make" handle the inclusion of .d without trying to guess
- what could be done at what time. Some initial inclusions hence
- _fail_, but "make" tries again later and succeed.
-
-TODO: remove obsolete sections below and better describe the new approach
------------------------------------------------------------------------
+
+HISTORY:
+-------
+
+* July 2007 (Pierre Corbineau & Lionel Elie Mamane).
+ Inclusion of a build system with 3 explicit phases:
+ - Makefile.stage1: ocamldep, sed, camlp4 without Coq grammar extension
+ - Makefile.stage2: camlp4 with grammar.cma or q_constr.cmo
+ - Makefile.stage3: coqdep (.vo)
+
+* March 2010 (Pierre Letouzey).
+ Revised build system. In particular, no more stage1,2,3 :
+ - Stage3 was removed some time ago when coqdep was splitted into
+ coqdep_boot and full coqdep.
+ - Stage1,2 were replaced by brutal inclusion of all .d at the start
+ of Makefile.build, without trying to guess what could be done at
+ what time. Some initial inclusions hence _fail_, but "make" tries
+ again later and succeed.
+ - Btw, .ml4 are explicitely turned into .ml files, which stay after build.
+ By default, they are in binary ast format, see READABLE_ML4 option.
+
+* February 2014 (Pierre Letouzey).
+ Another revision of the build system. We avoid relying on the awkward
+ include-which-fails-but-works-finally-after-a-retry feature of gnu make.
+ This was working, but was quite hard to understand. Instead, we reuse
+ the idea of two explicit phases, but in a lighter way than in 2007.
+ The main Makefile calls Makefile.build twice :
+ - first for building grammar.cma (and q_constr.cmo), with a
+ restricted set of .ml4 (see variable BUILDGRAMMAR).
+ - then on the true target asked by the user.
+
+
+---------------------------------------------------------------------------
+
This file documents internals of the implementation of the build
system. For what a Coq developer needs to know about the build system,
@@ -24,68 +41,47 @@ see build-system.txt .
.ml4 files
----------
-.ml files corresponding to .ml4 files are created to keep ocamldep
-happy only. To ensure they are not used for compilation, they contain
-invalid OCaml.
-
+.ml4 are converted to .ml by camlp4. By default, they are produced
+in the binary ast format understood by ocamlc/ocamlopt/ocamldep.
+Pros:
+ - faster than parsing clear-text source file.
+ - no risk of editing them by mistake instead of the .ml4
+ - the location in the binary .ml are those of the initial .ml4,
+ hence errors are properly reported in the .ml4.
+Cons:
+ - This format may depend on your ocaml version, they should be
+ cleaned if you change your build environment.
+ - Unreadable in case you want to inspect this generated code.
+ For that, use make with the READABLE_ML4=1 option to switch to
+ clear-text generated .ml.
-multi-stage build
------------------
-Le processus de construction est séparé en trois étapes qui correspondent
-aux outils nécessaires pour calculer les dépendances de cette étape:
-
-stage1: ocamldep, sed , camlp4 sans fichiers de Coq
-stage2: camlp4 avec grammar.cma et/ou q_constr.cmo
-stage3: coqdep (.vo)
+Makefiles hierachy
+------------------
Le Makefile a été séparé en plusieurs fichiers :
-- Makefile: coquille vide qui délègue les cibles à la bonne étape sauf
- clean et les fichiers pour emacs (car ils sont en quelque sorte en
- "stage0": aucun calcul de dépendance nécessaire).
+- Makefile: coquille vide qui lançant Makefile.build sauf pour
+ clean et quelques petites choses ne nécessitant par de calculs
+ de dépendances.
- Makefile.common : définitions des variables (essentiellement des
listes de fichiers)
-- Makefile.build : les règles de compilation sans inclure de
- dépendances
-- Makefile.stage* : fichiers qui incluent les dépendances calculables
- à cette étape ainsi que Makefile.build.
-
-The build needs to be cut in stages because make will not take into
-account one include when making another include.
+- Makefile.build : contient les regles de compilation, ainsi que
+ le "include" des dépendances (restreintes ou non selon la variable
+ BUILDGRAMMAR).
+- Makefile.doc : regles specifiques à la compilation de la documentation.
Parallélisation
---------------
-Le découpage en étapes veut dire que le makefile est un petit peu
-moins parallélisable que strictement possible en théorie: par exemple,
-certaines choses faites en stage2 pourraient être faites en parallèle
-avec des choses de stage1. Nous essayons de minimiser cet effet, mais
-nous ne l'avons pas complètement éliminé parce que cela mènerait à un
-makefile très complexe. La minimisation est principalement que si on
-demande un objet spécifique (par exemple "make parsing/g_constr.cmx"),
-il est fait dans l'étape la plus basse possible (simplement), mais si
-un objet est fait comme dépendance de la cible demandée (par exemple
-dans un "make world"), il est fait le plus tard possible (par exemple,
-tout code OCaml non nécessaire pour coqdep ni grammar.cma ni
-q_constr.cmo est compilé en stage3 lors d'un "make world"; cela permet
-le parallélisme de compilation de code OCaml et de fichiers Coq (.v)).
-
-Le "(simplement)" ci-dessus veut dire que savoir si un fichier non
-nécessaire pour grammar.cma/q_constr.cmo peut en fait être fait en
-stage1 est compliqué avec make, alors nous retombons en général sur le
-stage2. La séparation entre le stage2 et stage3 est plus facile, donc
-l'optimisation ci-dessus s'y applique pleinement.
-
-En d'autres mots, nous avons au niveau conceptuel deux assignations
-d'étape pour chaque fichier:
-
- - l'étape la plus petite où nous savons qu'il peut être fait.
- - l'étape la plus grande où il peut être fait.
-
-Mais seule la première est gérée explicitement, la seconde est
-implicite.
+Il y a actuellement un double appel interne à "make -f Makefile.build",
+d'abord pour construire grammar.cma/q_constr.cmo, puis le reste.
+Cela signifie que ce makefile est un petit peu moins parallélisable
+que strictement possible en théorie: par exemple, certaines choses
+faites lors du second make pourraient être faites en parallèle avec
+le premier. En pratique, ce premier make va suffisemment vite pour
+que cette limitation soit peu gênante.
FIND_VCS_CLAUSE
diff --git a/dev/doc/build-system.txt b/dev/doc/build-system.txt
index d7cf396ff..31d9875ad 100644
--- a/dev/doc/build-system.txt
+++ b/dev/doc/build-system.txt
@@ -1,26 +1,8 @@
-Since July 2007, Coq features a build system overhauled by Pierre
-Corbineau and Lionel Elie Mamane.
-
----------------------------------------------------------------------
-WARNING:
-In March 2010 this build system has been heavily adapted by Pierre
-Letouzey. In particular there no more explicit stage1,2. Stage3
-was removed some time ago when coqdep was splitted into coqdep_boot
-and full coqdep. Ideas are still similar to what is describe below,
-but:
-1) .ml4 are explicitely turned into .ml files, which stay after build
-2) we let "make" handle the inclusion of .d without trying to guess
- what could be done at what time. Some initial inclusions hence
- _fail_, but "make" tries again later and succeed.
-
-TODO: remove obsolete sections below and better describe the new approach
------------------------------------------------------------------------
This file documents what a Coq developer needs to know about the build
system. If you want to enhance the build system itself (or are curious
-about its implementation details), see build-system.dev.txt .
-
-The build system is not at its optimal state, see TODO section.
+about its implementation details), see build-system.dev.txt, and in
+particular its initial HISTORY section.
FAQ: special features used in this Makefile
@@ -51,22 +33,10 @@ $(subst ...), $(patsubst ...), $(shell ...), $(foreach ...), $(if ...)
* Behavior of -include
If the file given to -include doesn't exist, make tries to build it,
-but doesn't care if this build fails. This can be quite surprising,
-see in particular the -include in Makefile.stage*
-
-
-Stages in build system
-----------------------
-
-The build system is separated into three stages, corresponding to the
-tool(s) necessary to compute the dependencies necessary at this stage:
-
-stage1: ocamldep, sed, camlp4 without Coq extensions
-stage2: camlp4 with grammar.cma and/or q_constr.cmo
-stage3: coqdep (.vo)
-
-The file "Makefile" itself serves as minimum stage for targets that
-should not need any dependency (such as *clean*).
+and even retries again if necessary, but doesn't care if this build
+finally fails. We used to rely on this "feature", but this should not
+be the case anymore. We kept "-include" instead of "include" for
+avoiding warnings about initially non-existant files.
Changes (for old-timers)
------------------------
@@ -97,37 +67,11 @@ save precious time:
- Always ask for what you want directly (e.g. bin/coqtop,
foo/bar.cmo, ...), don't do "make world" and interrupt
- it when it has done what you want. This will try to minimise the
- stage at which what you ask for is done (instead of maximising it
- in order to maximise parallelism of the build process).
-
+ it when it has done what you want.
For example, if you only want to test whether bin/coqtop still
builds (and eventually start it to test your bugfix or new
- feature), don't do "make world" and interrupt it when bin/coqtop is
- built. Use "make bin/coqtop" or "make coqbinaries" or something
- like that. This will avoid entering the stage 3, and cut build
- system overhead by 50% (1.2s instead of 2.4 on writer's machine).
-
- - If you want to avoid all .ml4 files being recompiled only because
- grammar.cma was rebuilt, do "make ml4depclean" once and then use
- NO_RECOMPILE_ML4=1.
-
- - The CM_STAGE1=1 option to make will build all .cm* files mentioned
- as targets on the command line in stage1. Whether this will work is
- your responsibility. It should work for .ml files that don't depend
- (nor directly nor indirectly through transitive closure of the
- dependencies) on any .ml4 file, or where those dependencies can be
- safely ignored in the current situation (e.g. all these .ml4 files
- don't need to be recompiled).
-
- This will avoid entering the stage2 (a reduction of 33% in
- overhead, 0.4s on the writer's machine).
-
- - To jump directly into a stage (e.g. because you know nothing is to
- be done in stage 1 or (1 and 2) or because you know that the target
- you give can be, in this situation, done in a lower stage than the
- build system dares to), use GOTO_STAGE=n. This will jump into stage
- n and try to do the targets you gave in that stage.
+ feature), use "make bin/coqtop" or "make coqbinaries" or something
+ like that.
- To disable all dependency recalculation, use the NO_RECALC_DEPS=1
option. It disables REcalculation of dependencies, not calculation
@@ -135,12 +79,6 @@ save precious time:
still created, but it is not updated every time the source file
(e.g. .ml) is changed.
-General speed improvements:
-
- - When building both the native and bytecode versions, the
- KEEP_ML4_PREPROCESSED=1 option may reduce global compilation time
- by running camlp4o only once on every .ml4 file, at the expense of
- readability of compilation error messages for .ml4 files.
Dependencies
------------
@@ -178,166 +116,30 @@ Targets for cleaning various parts:
.ml4 files
----------
-The camlp4-preprocessed version of FOO.ml4 is FOO.ml4-preprocessed and
-can be obtained with:
- make FOO.ml4-preprocessed
-
If a .ml4 file uses a grammar extension from Coq (such as grammar.cma
or q_constr.cmo), it must contain a line like:
(*i camlp4deps: "grammar.cma q_constr.cmo" i*)
-If it uses a standard grammar extension, it must contain a line like:
- (*i camlp4use: "pa_ifdef.cmo" i*)
-
-It can naturally contain both a camlp4deps and a camlp4use line. Both
-are used for preprocessing. It is thus _not_ necessary to add a
-specific rule for a .ml4 file in the Makefile.build just because it
-uses grammar extensions.
-
-By default, the build system is geared towards development that may
-use the Coq grammar extensions, but not development of Coq's grammar
-extensions themselves. This means that .ml4 files are compiled
-directly (using ocamlc/opt's -pp option), without use of an
-intermediary .ml (or .ml4-preprocessed) file. This is so that if a
-compilation error occurs, the location in the error message is a
-location in the .ml4 file. If you are modifying the grammar
-extensions, you may be more interested in the location of the error in
-the .ml4-preprocessed file, so that you can see what your new grammar
-extension made wrong. In that case, use the KEEP_ML4_PREPROCESSED=1
-option. This will make compilation of a .ml4 file a two-stage process:
-
-1) create the .ml4-preprocessed file with camlp4o
-2) compile it with straight ocamlc/opt without preprocessor
-
-and will instruct make not to delete .ml4-preprocessed files
-automatically just because they are intermediary files, so that you
-can inspect them.
-
-If you add a _new_ grammar extension to Coq:
-
- - if it can be built at stage1, that is the .ml4 file does not use a
- Coq grammar extension itself, then add it, and all .cmo files it
- needs to STAGE1_TARGETS and STAGE_ML4 in Makefile.common. See the
- handling of grammar.cma and q_constr.cmo for an example.
-
- - if it cannot be built at stage1, that is the .ml4 file itself needs
- to be preprocessed with a Coq camlp4 grammar extension, then,
- congratulations, you need to add a new stage between stage1 and
- stage2.
+The use of (*i camlp4use: ... i*) to mention uses of standard
+extension such as IFDEF has been discontinued, the Makefile now
+always calls camlp4 with pa_macros.cmo and a few others by default.
+
+For debugging a Coq grammar extension, it could be interesting
+to use the READABLE_ML4=1 option, otherwise the generated .ml are
+in an internal binary format (see build-system.dev.txt).
+
New files
---------
For a new file, in most cases, you just have to add it to the proper
-file list(s) in Makefile.common, such as ARITHVO or TACTICS.
+file list(s):
+ - For .ml, in the corresponding .mllib (e.g. kernel/kernel.mllib)
+ These files are also used by the experimental ocamlbuild plugin,
+ which is quite touchy about them : be careful with order,
+ duplicated entries, whitespace errors, and do not mention .mli there.
+ - For .v, in the corresponding vo.itarget (e.g theories/Init/vo.itarget)
+ - The definitions in Makefile.common might have to be adapted too.
+ - If your file needs a specific rule, add it to Makefile.build
The list of all ml4 files is not handled manually anymore.
-
-Exceptions are:
-
- - The file is necessary at stage1, that it is necessary to build the
- Coq camlp4 grammar extensions. In this case, make sure it ends up
- in STAGE1_CMO and (for .ml4 files) STAGE1_ML4. See the handling of
- grammar.cma and/or q_constr.cmo for an example.
-
- - if the file needs to be compiled with -rectypes, add it to
- RECTYPESML in Makefile.common. If it is a .ml4 file, implement
- RECTYPESML4 or '(*i ocamlflags i*)'; see TODO.
-
- - the file needs a specific Makefile entry; add it to Makefile.build
-
- - the files produced from the added file do not match an existing
- pattern or entry in "Makefile". (All the common cases of
- .ml{,i,l,y,4}, .v, .c, ... files that produces (respectively)
- .cm[iox], .vo, .glob, .o, ... files with the same basename are
- already covered.) In this case, see section "New targets".
-
-New targets
------------
-
-If you want to add:
-
- - a new PHONY target to the build system, that is a target that is
- not the name of the file it creates,
-
- - a normal target is not already mapped to a stage by "Makefile"
-
- then:
-
- - add the necessary rule to Makefile.build, if any
- - add the target to STAGEn_TARGETS, with n being the smallest stage
- it can be built at, that is:
- * 1 for OCaml code that doesn't use any Coq camlp4 grammar extension
- * 2 for OCaml code that uses (directly or indirectly) a Coq
- camlp4 grammar extension. Indirectly means a dependency of it
- does.
- * 3 for Coq (.v) code.
-
- *or*
-
- add a pattern matching the target to the pattern lists for the
- smallest stage it can be built at in "Makefile".
-
-TODO
-----
-
-delegate pa_extend.cmo to camlp4use statements and remove it from
-standard camlp4 options.
-
-maybe manage compilation flags (such as -rectypes or the CoqIDE ones)
-from a
- (*i ocamlflags: "-rectypes" i*)
-statement in the .ml(4) files themselves, like camlp4use. The CoqIDE
-files could have
- (*i ocamlflags: "${COQIDEFLAGS}" i*)
-and COQIDEFLAGS is still defined (and exported by) the Makefile.build.
-
-Clean up doc/Makefile
-
-config/Makefile looks like it contains a lot of unused variables,
-clean that up (are any maybe used by nightly scripts on
-pauillac?). Also, the COQTOP variable from config/Makefile (and used
-in contribs) has a very poorly chosen name, because "coqtop" is the
-name of a Coq executable! In the coq Makefiles, $(COQTOPEXE) is used
-to refer to that executable.
-
-Promote the granular .glob handling to official way of doing things
-for Coq developments, that is implement it in coq_makefile and the
-contribs. Here are a few hints:
-
->> Les fichiers de constantes produits par -dump-glob sont maintenant
->> produits par fichier et sont ensuite concaténés dans
->> glob.dump. Ilsont produits par défaut (avec les bonnes
->> dépendances).
-
-> C'est une chose que l'on voulait faire aussi.
-
-(J'ai testé et débogué ce concept sur CoRN dans les derniers mois.)
-
-> Est-ce que vous sauriez modifier coq_makefile pour qu'il procède de
-> la même façon
-
-Dans cette optique, il serait alors plus propre de changer coqdep pour
-qu'il produise directement l'output que nous mettons maintenant dans
-les .v.d (qui est celui de coqdoc post-processé avec sed).
-
-Si cette manière de gérer les glob devient le standard béni
-officiellement par "the Coq development team", ne voudrions nous pas
-changer coqc pour qu'il produise FOO.glob lors de la compilation de
-FOO.v par défaut (sans argument "-dump-glob")?
-
-> et que la production de a.html par coqdoc n'ait une dépendance qu'en
-> les a.v et a.glob correspondant ?
-
-Je crois que coqdoc exige un glob-dump unique, il convient donc de
-concaténer les .glob correspondants. Soit un glob-dump global par
-projet (par Makefile), soit un glob-dump global par .v(o), qui
-contient son .glob et ceux de tous les .v(o) atteignables par le
-graphe des dépendances. CoRN contient déjà un outil de calcul de
-partie atteignable du graphe des dépendances (il y est pour un autre
-usage, pour calculer les .v à mettre dans les différents tarballs sur
-http://corn.cs.ru.nl/download.html; les parties partielles sont
-définies par liste de fichiers .v + toutes leurs dépendances
-(in)directes), il serait alors adéquat de le mettre dans les tools de
-Coq.
-