I'm working with a makefile system using mcs as its build command. Briefly, mcs is invoked with:
source files: Passed directly;
resources: with -resource:;
a target: with -target: and can be exe, winexe, library or module.
And produces an *.exe, *.dll or *.netmodule. The -target: parameter is certain in the last two cases but an *.exe can be built with either exe or winexe.
Currently, each directory's makefile includes what boils down to this:
RESOLVED_SOURCES := $(addprefix $(srcdir)/,$(SOURCES))
RESOLVED_RESOURCES := $(addprefix $(srcdir)/,$(RESOURCES))
S_PARAM := $(RESOLVED_SOURCES)
R_PARAM := $(addprefix -resource:,$(RESOLVED_RESOURCES))
all: $(OUTPUT)
$(OUTPUT): $(RESOLVED_SOURCES) $(RESOLVED_RESOURCES)
$(MCS) $(MCSARGS) $(S_PARAM) $(R_PARAM) -target:$(TARGET) -out:$(OUTPUT)
And in each project directory the variables OUTPUT, TARGET, SOURCES and RESOURCES are defined before inclusion.
Even with everything I've left out it's fairly complicated and suffers from only being able to build a single concrete target where it's included, without duplicating the build rule.
To avoid the duplication and increase flexibility I'd like to define some pattern rules for this complex build command but it's difficult:
each compilation unit generally has many inputs;
there's no way to incrementally add each input;
some inputs need different handling.
e.g. For mcs these patterns seem possible:
%.exe: %.cs
$(MCS) $(MCSARGS) -target:$(target) -out:$# $<
%.dll: %.cs
$(MCS) $(MCSARGS) -target:library -out:$# $<
%.netmodule: %.cs
$(MCS) $(MCSARGS) -target:module -out:$# $<
As pattern rules go they aren't too bad but they aren't useful for most applications of mcs as there's only a single source file being passed.
If the user adds source file prerequisites for a target then the above $< could be swapped for $^ to pass them all but that won't work for resource files. The order only list could be used for passing resource file prerequisites but would prevent rebuilds when only resource files were updated.
Is there a general technique for turning this sort of complex command's invocation into a pattern rule or set of rules?
There are a number of problems:
the build command distinguishes more than one prerequisite type;
they must all be passed at once;
updates to any of them must cause a rebuild, they are all regular prerequisites.
Ideas:
delegating the inputs you want to separate to input files and then reading them back in allows you to have a simple target invoking the build;
using secondary expansion you can define those input files based on the contents of "pattern variables";
general search should work.
For example:
multi_A := foo
multi_B := bar
all: multi.foo
%.foo: .a.% .b.%
#echo $*: $(foreach _,$^,$(file < $_))
#touch $#
.SECONDEXPANSION:
.a.%: $$($$*_A)
$(file > $#,$^)
.b.%: $$($$*_B)
$(file > $#,$(addprefix -b:,$^))
To start with using secondary expansion seems dubious. Apart from that the immediate issue with this is that the user has to know that any regular prerequisites will be read and passed to the build command.
On the other hand, knowing that, the user can supplement a target with extra input parameters by depending on a target that writes them to another input file.
Related
Using GNU make, I am trying to solve a problem similar to make recipe execute twice — that is, to have a Makefile recipe run twice. In my case, however, the recipe is run under the .SECONDEXPANSION target, and the two different runs will be called with different parameters to generate different versions of the output file from the same input file. That is, with input file foo, this example Makefile should be callable via make foo.pdf or make foo.expanded.pdf to build one .pdf file, or make all to build both .pdf files:
.PHONY: all
all: foo.pdf foo.expanded.pdf
.SECONDEXPANSION:
%.expanded.pdf %.pdf: %
#echo building $(basename $#)
Of the two solutions given in that answer, the first is unsuitable because it always runs the rule twice; I want it run twice when the user asks for it.
The second solution posted there is conceptually what I am looking for and have implemented in the above example Makefile, with only the small problem that it doesn't work: although the all target lists both .pdf files as dependencies, only one is built when make all is run.
Is there a way to tell GNU make to build two different files using the same rule under a .SECONDEXPANSION?
EDIT: Clarified in problem description that the same input file is used to build both versions of the output file, and modified sample Makefile to include this dependency.
EDIT: I would like a solution as scalable as possible; that is, it should work if the input filename contains dots, specifying additional output file foo.reduced.pdf should require only adjusting the targets and recipe as appropriate, etc. This limits performing string surgery that relies on the filenames appearing exactly as given in this narrow example (e.g., changing the rule to %.pdf: $$(firstword $$(subst ., ,$$*)) fails if the input file could be either foo or foo.bar).
You are probably looking for Pattern-specific Variable Values. Let's assume your recipe depends on a make variable named BUILDFLAGS that takes value normal by default and special for the "expanded" targets. Then this Makefile:
BUILDER := builder
BUILDFLAGS := normal
.PHONY: all
all: foo.pdf foo.expanded.pdf
%.expanded.pdf: BUILDFLAGS := special
%.pdf:
$(BUILDER) $(BUILDFLAGS) $#
should do about what you want with the same rule for all targets, plus one pattern-specific variable value declaration. Replace builder, normal and special with what makes sense in your case. Demo:
$ make foo.pdf
builder normal foo.pdf
$ make foo.expanded.pdf
builder special foo.expanded.pdf
$ make
builder normal foo.pdf
builder special foo.expanded.pdf
Your problem has nothing to do with .SECONDEXPANSION. You can just drop that and the problem will be the same.
Your problem is that you are using a pattern rule with multiple target patterns, and expecting that it works similar to an explicit rule with multiple targets. But it does not (and in fact you cannot have a rule with both pattern and explicit targets).
For a pattern rule with multiple target patterns, Make matches the same pattern to all the %, including multiple times in the targets, and then assumes that it just has to execute the recipe with that pattern once, and it will make all the matched targets.
In your case the best way is to use multiple rules (I changed your recipe because using echo as a Make recipe is a bad idea):
.PHONY: all
all: foo.expanded.pdf foo.pdf
RECIPE = touch $#
%.expanded.pdf:
$(RECIPE)
%.pdf:
$(RECIPE)
I am aware of tools like CMake and GNU Autotools but I'm trying to write a universal build system myself, to use for my C and C++ projects. I'll briefly explain how it works and hopefully, someone can suggest either improvements or a better design altogether.
The build system proper lives in one of the project's subdirectories (I import it as a Git submodule). The project's root directory has a wrapper makefile that defines a couple of macros and includes the main makefile from said subdirectory. That does most of the work: it follows the directory organization scheme (i.e., it outputs libraries in lib, binaries in bin, etc.), it handles automatic dependencies for the source code and the DocBook documentation, and provides the de facto standard targets: all, test, clean, install, as well as others.
Here's what a wrapper makefile that builds two binaries, foo and bar, might look like:
# foo-specific macros
FOO_SRC_FILES = foo1.c foo2.c foo3.c
FOO_OBJ_FILES = $(FOO_SRC_FILES:.c=.o)
FOO_BIN_FILE = foo
# bar-specific macros
BAR_SRC_FILES = bar1.c bar2.c
BAR_OBJ_FILES = $(BAR_SRC_FILES:.c=.o)
BAR_BIN_FILE = bar
# Inform the build system about them
SRC_FILES = $(FOO_SRC_FILES) $(BAR_SRC_FILES)
OBJ_FILES = R(BAR_OBJ_FILES) $(BAR_OBJ_FILES)
BIN_FILES = $(FOO_BIN_FILE) $(BAR_BIN_FILE)
# Only install the binaries. If I were building a library, I would instead
# select the "lib" and perhaps "include" directories.
INSTALL = bin
INSTALL_DIR = /usr/share
# Use the build system
include build/build.mk
Now here's the problem. While build.mk can use pattern rules to create dependency and object files, there's only one OBJ_FILES and only one BIN_FILES. So if I put a pattern rule like the following in the build system that looks like this:
$(BIN_DIR)/$(BIN_FILES): $(OBJ_FILES:%=$(OBJ_DIR)/%) $(LIB_FILES:%=$(LIB_DIR)/%) | $(BIN_DIR)
$(CC) $(LDFLAGS) -o $# $(OBJ_FILES:%=$(OBJ_DIR)/%) -L $(LIB_DIR) $(LIB_FILES:lib%.a=-l %)
then foo would depend on and link with everything that bar does and vice versa. So what I end up doing is asking the user to put these rules in the wrapper makefile, even though they feel like they belong in build.mk:
$(BIN_DIR)/$(FOO_BIN_FILE): $(FOO_OBJ_FILES:%=$(OBJ_DIR)/%) $(FOO_LIB_FILES:%=$(LIB_DIR)/%) | $(BIN_DIR)
$(CC) $(LDFLAGS) -o $# $(FOO_OBJ_FILES:%=$(OBJ_DIR)/%) -L $(LIB_DIR) $(FOO_LIB_FILES:lib%.a=-l %)
$(BIN_DIR)/$(BAR_BIN_FILE): $(BAR_OBJ_FILES:%=$(OBJ_DIR)/%) $(BAR_LIB_FILES:%=$(LIB_DIR)/%) | $(BIN_DIR)
$(CC) $(LDFLAGS) -o $# $(BAR_OBJ_FILES:%=$(OBJ_DIR)/%) -L $(LIB_DIR) $(BAR_LIB_FILES:lib%.a=-l %)
The same issue applies to libraries as well, of course. The upside is that these rules can be copied and pasted almost verbatim; only the prefixes need to be changed (e.g., FOO or BAR).
Ideas to fix this include:
Asking the user to have separate wrapper makefiles for separate things (e.g., one for foo and another for bar) but that is just terrible.
Changing things up a little bit and then using m4 to do some preprocessing but I don't want to go through that unless a more elegant solution doesn't exist.
I would really appreciate some ideas.
PS: I know that the pattern matching expressions in the last two code samples can be replaced with text functions but those are GNU Make-specific. The style I used is more portable and is in fact on the list of additions for the next version of the POSIX standard.
I have begin to develop a similar system for my own C projects, but the logic I use does rely on some features which I believe are specific to GNU Make.
The main idea is to use a combinaison of $(eval) and $(call), by defining the logic of the build system, and then applying to the project tree.
To do so, I have in each of my directories and subdirectories a piece of Makefile of the following form, which I name Srcs.mk:
SRC := foo.c foo_bar.c bar.c
TARGET := foo_bar
SRC_DIR := src
OBJ_DIR := obj
I define a variable, which is in fact a macro, which is expanded with $(call) and then passed to $(eval). It's defined this way:
define get_local_variables
include Srcs.mk
$1SRC := $(SRC)
$1SRC_DIR := $(SRC_DIR)
$1OBJ_DIR := $(OBJ_DIR)
$1TARGET := $(TARGET)
TARGET :=
SRC :=
SRC_DIR :=
OBJ_DIR :=
$(call get_local_variables, $(DIR)) will expand to the above, with $1 replaced by the content of $(DIR). Then it will be treated as a Makefile fragment by $(eval)
This way, I fill per-directory variables for each of my directory.
I have then a handful or other rules which use this variables, using the same principles.
### Macros ###
obj = $(patsubst %.c,$($1OBJ_DIR)/%.o,$($1SRC))
define standard_rules
$($1TARGET): $(obj)
$$(LINK)
$(obj): $($1OBJ_DIR)/%.o:$($1SRC_DIR)/%.c | $($1OBJ_DIR)
$$(COMPILE)
endef
The variable are computed $(call), then expanded and read as makefile fragments by $(eval).
(I use static pattern rules but that it not intrinsic to the idea).
The whole idea is basically to define directories as a kind of namespace, with data attached to them, and then run function over them.
My actual system is a bit more complicated, but that the whole idea.
If you have a way to emulate $(eval) and $(call) (I think these are specific to GNU make, but not sure), you could try that approach.
You can also implement non recursive make this way, by adding a SUBDIRS variables in each directory and running recursively the same macro which is run on the current one. But it should been done carefully, not to mess it up with the order of expansion and evaluation in make.
So get_local_variables need to be evaluated before the rest of the macros are expanded.
(My project is visible on my Github account if you want to take a look, under make-build-system. But it is far from be complete enough^).
Be aware, though, that this is quite painful to debug when things go wrong. Make (at least, GNU) basically catch the error (when there is one) on the higher $(call) or $(eval) expansion.
I have developed my own non-recursive build system for GNU make, called prorab, where I solved the problem you described as follows.
The approach to solve your problem is somewhat similar to what #VannTen described in his answer, except that I use a macro to clean all state variables before defining build rules for the next binary.
For example, a makefile which builds two binaries could look like this:
include prorab.mk
this_name := AppName
this_ldlibs += -lsomelib1
this_cxxflags += -I../src -DDEBUG
this_srcs := main1.cpp MyClass1.cpp
$(eval $(prorab-build-app))
$(eval $(prorab-clear-this-vars))
this_name := AnotherppName
this_ldlibs += -lsomelib1
this_cxxflags += -I../src -DDEBUG
this_srcs := main2.cpp MyClass2.cpp
$(eval $(prorab-build-app))
So, in this example it will build two binaries: AppName and AnotherppName.
As you can see the build is configured by setting a number of this_-prefixed variables and the calling the $(eval $(prorab-build-app)) which expands to defining all the build, install, clean etc. rules.
Then a call to $(eval $(prorab-clear-this-vars)) clears all this_-prefixed variables, so that those can be defined again from scratch for the next binary, and so on.
Also, the very first line which includes the prorab.mk also cleans all this_-prefixed variables of course, so that the makefiles can be safely included into each other.
You can read more about that build system concepts here https://github.com/cppfw/prorab/blob/master/wiki/HomePage.adoc
I'm trying to write a Makefile where multiple source files (in my case they are markdown) create multiple target files (pdfs). However, the target files generated have extra characters in the file name that can't be predicted (it happens to be a version number encoded in the source), but ideally the Makefile would not have to read the source itself.
So, for example:
file1.md => file1-v1.pdf
file2.md => file2-v2.pdf
...
I can calculate source name given a target name (by excluding anything after the hyphen and adding .md), but cannot calculate target name given the source.
Is it possible to write a Makefile that builds only the targets where the source have been updated?
This will be ugly, but it will work.
As it often is with Make, our problem divides into these two problems:
1. construct a list of targets
2. build them
Suppose we have five md files which map to pdf files (whose names we don't know beforehand):
file1.md => file1-v1.pdf
file2.md => file2-v1.pdf
file3.md => file3-v1.pdf
file4.md => file4-v1.pdf
file5.md => file5-v1.pdf
We can't use the real output file names as targets, because we don't know them beforehand, but we see five input files and know that we must build one output file for each. For now, a fake target name will do:
file1-dummy.pdf: file1.md
zap file1.md
When Make executes this rule, it produces the file file1-v1.pdf. The fact that it doesn't produce a file named file1-dummy.pdf is disquieting, but not a serious problem. We can turn this into a pattern rule:
%-dummy.pdf: %.md
zap $<
Then all we have to do is turn the list of existing input files (file1.md, file2.md, ...) into a list of dummy targets (file1-dummy.pdf, file2-dummy.pdf, ...), and build them. So far, so good.
But suppose some of the output files already exist. If file2-v2.pdf already exists -- and is newer than file2.md -- then we would prefer that Make not rebuild it (by attempting to build file2-dummy.pdf). In that case we would prefer that file2-v2.pdf be in the target list, with a rule that worked like this:
file2-v2.pdf: file2.md
zap $<
This is not easy to turn into a pattern rule, because Make does not handle wildcards very well, and cannot cope with multiple wildcards in a single phrase, not without a lot of clumsiness. But there is a way to write one rule that will cover both cases. First note that we can obtain the part of a variable before the hyphen with this kludge:
$(basename $(subst -,.,$(VAR)))
Armed with this, and with secondary expansion, we can write a pattern rule that will work with both cases, and construct a target list that will exploit it:
# There are other ways to construct these two lists, but this will do.
MD := $(wildcard *.md)
PDF := $(wildcard *.pdf)
PDFROOTS := $(basename $(subst -,.,$(basename $(PDF))))
MDROOTS := $(filter-out $(PDFROOTS), $(basename $(MD)))
TARGETS:= $(addsuffix -foo.pdf, $(MDROOTS)) $(PDF)
.SECONDEXPANSION:
%.pdf: $$(basename $$(subst -,., $$*)).md
# perform actions on $<
Make's algorithm always starts with the final output product and works its way backwards to the source files, to see what needs to be updated.
Therefore, you HAVE to be able to enumerate the final output product as a target name and correlate that back to the inputs that generate that output, for make to work.
This is also why make is not a great tool for building Java, for example, since the output filenames don't map easily to the input file names.
So, you must have at least one target/prerequisite pair which is derivable (for implicit rules), or state-able (for explicit rules)--that is, known at the time you write the makefile. If you don't then a marker file is your only alternative. Note you CAN add extra generated, non-derivative prerequisites (for example, in compilers you can add header files as prerequisites that are not related to the source file name), in addition to the known prerequisite.
#Beta's answer is informative and helpful, but I needed a solution (using GNU Make 4.1) that worked when the destination filename bears no resemblance to the input filename, for example, if it is generated from its content. I came up with the following, which takes every file matching *.in, and creates a file by reading the contents of the source file, appending a .txt, and using it as a filename to create. (For example, if test.in exists and contains foo, the makefile will create a foo.txt file.)
SRCS := $(wildcard *.in)
.PHONY: all
all: all_s
define TXT_template =
$(2).txt: $(1)
touch $$#
ALL += $(2).txt
endef
$(foreach src,$(SRCS),$(eval $(call TXT_template, $(src), $(shell cat $(src)))))
.SECONDARY_EXPANSION:
all_s: $(ALL)
The explanation:
The define block defines the recipe needed to make the text file from the .in file. It's a function that takes two parameters; $(1) is the .in. file and $(2) is the contents of it, or the base of the output filename. Replace touch with whatever makes the output. We have to use $$# because eval will expand everything once, but we want $# to left after this expansion. Since we have to collect all the generated targets so we known what all the make, the ALL line accumulates the targets into one variable. The foreach line goes through each source file, calls the function with the source filename and the contents of the file (i.e. what we want to be the name of the target, here you'd normally use whatever script generates the desired filename), and then evaluates the resulting block, dynamically adding the recipe to make. Thanks to Beta for explaining .SECONDARY_EXPANSION; I needed it for reasons not entirely clear to me, but it works (putting all: $(ALL) at the top doesn't work). The all: at the top depends on the secondary expansion of all_s: at the bottom and somehow this magic makes it work. Comments welcome.
maybe try this ? or something along those lines
# makefile
SRCS=$(wildcard *.md)
PDFS=$(shell printf *.pdf)
$(PDFS): $(SRCS)
command ...
the printf *.pdf is meant to either expand to the first of the pdf files if they exist, else fail if they don't and that will signal to make that it should build. if this doesn't work i suggest maybe experimenting with find, ls or other listing tools (e.g. compgen, complete), maybe even in combination with xargs to get everything on one line.
I have another make-like tool that produces an XML as an artifact after parsing my makefile which I'll then further process with Python.
It'd simplify things for me - a lot - if I could have make consider every single prerequisite to be an actual target because then this other tool
will classify each and every file as a "job".
This is a fragment of my makefile:
.obj/eventlookupmodel.o: C:/Users/User1/Desktop/A/PROJ/src/AL2HMIBridge/LookupModels/eventlookupmodel.cpp C:\Users\User1\Desktop\A\PROJ\src\AL2HMIBridge\LookupModels\eventlookupmodel.h \
C:/Users/User1/Desktop/A/PROJ/qt5binaries/include/QtCore/qabstractitemmodel.h \
C:/Users/User1/Desktop/A/PROJ/qt5binaries/include/QtCore/qvariant.h \
...
I'd want for make to think I have a dummy rule for each prerequisite such as below:
C:/Users/User1/Desktop/A/PROJ/qt5binaries/include/QtCore/qvariant.h:
#echo target pre= $#
C:/Users/User1/Desktop/A/PROJ/qt5binaries/include/QtCore/qabstractitemmodel.h:
#echo target pre=$#
C:/Users/User1/Desktop/A/PROJ/src/AL2HMIBridge/LookupModels/eventlookupmodel.cpp :
#echo target pre=$#
C:\Users\User1\Desktop\A\PROJ\src\AL2HMIBridge\LookupModels\eventlookupmodel.h:
#echo target pre=$#
I don't care about the exact form of the rule just that each file is considered an actual target.
My method of passing in this rule would be by setting the MAKEFILES variable like so
make all MAKEFILES=Dummy.mk
with Dummy.mk containing this rule so that I do not modify the makefiles.
I've tried the following so far.
Dummy.mk:
%.h:
#echo header xyz = $#
%:
#echo other xyz= $#
This partially works.
I run make all --trace --print-data-base MAKEFILES=Dummy.mk and I can see that
make does "bind" the %.h: rule to the header files. In the --print-data-base section, I see that rule being assigned to the header files.
C:/Users/User1/Desktop/A/QNX_SDK/target/qnx6/usr/include/stddef.h:
# Implicit rule search has been done.
# Implicit/static pattern stem: 'C:/Users/User1/Desktop/A/QNX_SDK/target/qnx6/usr/include/stddef'
# Last modified 2016-05-27 12:39:16
# File has been updated.
# Successfully updated.
# recipe to execute (from '#$(QMAKE) top_builddir=C:/Users/User1/Desktop/A/HMI_FORGF/src/../lib/armle-v7/release/ top_srcdir=C:/Users/User1/Desktop/A/HMI_FORGF/ -Wall CONFIG+=release CONFIG+=qnx_build_release_with_symbols CONFIG+=rtc_build -o Makefile C:/Users/User1/Desktop/A/HMI_FORGF/src/HmiLogging/HmiLogging.pro
', line 2):
#echo header xyz = $#
However, I do NOT see the "echo header xyz $#"-rule being executed.
Regarding the %: rule, it is neither executed for the .cpp files nor "bound" to them in the --print-data-base section.
However, it is bound and executed for existing targets which have no suffix i.e.
all: library binary
binary: | library
ifs: | library
For the %: rule, the reason for this behavior is because of 10.5.5 Match-Anything Pattern Rules: If you do not mark the match-anything rule as terminal, then it is non-terminal. A non-terminal match-anything rule cannot apply to a file name that indicates a specific type of data. A file name indicates a specific type of data if some non-match-anything implicit rule target matches it.
If I make it non-terminal - no double colon - then the rule doesn't apply to built-in types like .cppunless I un-define the built-in rules that negate my intended %: rule.
If I make it terminal, "it does not apply unless its prerequisites actually exist". But a .h or .cpp doesn't technically have prerequisites; can I just create a dummy file and have that as its prerequisite?
NOTE: This has NOTHING to do with gcc -M generation. Yes the -M option would help in the specific case of header and source files but this question is for more generic targets and prerequisites that already exist in the makefile when make is launched.
This may take a few iterations. Try:
%.h: null
#echo header xyz = $#
%: null
#echo other xyz= $#
null:
#:
Try generating static pattern rules for the header files. See one of the answers to Make ignoring Prerequisite that doesn't exist.
Static pattern rules only apply to an explicit list of target files like this:
$(OBJECTS): %.o: %.c
*recipe here*
where the variable OBJECTS is defined earlier in the makefile to be a list of target files (separated by spaces), for example:
OBJECTS := src/fileA.c src/fileB.c src/fileC.c
Note that you can use the various make utility functions to build that list of target files. For example, $(wildcard pattern), $(addsuffix), etc.
You should also ensure that the recipe "touches" the header file to change the timestamp.
I've found that using static pattern rules instead of pattern rules fixes problems where make doesn’t build prerequisites that don’t exist, or deletes files that you want.
Here is an example of using wildcard to copy files from one directory to another.
# Copy images to build/images
img_files := $(wildcard src/images/*.png src/images/*.gif src/images/*.jpg \
src/images/*.mp3)
build_images := $(subst src/,$(BUILD_DIR)/,$(img_files))
$(build_images): $(BUILD_DIR)/images/% : src/images/%
mkdir -p $(dir $#)
cp -v -a $< $#
There are other make functions like addprefix that could be used to generate a more complex file specification.
Assume I have a make rule:
.PHONY:gen
gen: auto.template
generate-sources auto.template
that creates a bunch of files, for example auto1.src, auto2.src, auto3.src and so on.
If I now have rules to build targets from *.src files, like this:
$(patsubst %.src,%.target,$(wildcard *.src)): %.target: %.src
build $< > $#
How can I tell make to first execute the gen rule and then expand the preconditions for the second rule template? GNU extensions are welcome.
Note: I would like to keep it in one make invocation; A trivial solution to this would be to put the second rule in a secondary Makefile.secondrun and call $(MAKE) -f Makefile.secondrun after gen was processed. But I was wondering if there is a better option.
Building off Beta's answer, here's how you can do it using makefile remaking in GNU make, which is not the same thing as recursive make. Rather, it updates an included makefile using a rule in the main makefile, then restarts the original make instance. This is how *.d dependency files are typically generated and used.
# Get the list of auto-generated sources. If this file doesn't exist, or if it is older
# than auto.template, it will get built using the rule defined below, according to the
# standard behavior of GNU make. If autosrcs.mk is rebuilt, GNU make will automatically
# restart itself after autosrcs.mk is updated.
include autosrcs.mk
# Once we have the list of auto-generated sources, getting the list of targets to build
# from them is a simple pattern substitution.
TARGETS=$(patsubst %.src,%.target,$(AUTO_SRCS))
all: $(TARGETS)
# Rule describing how to build autosrcs.mk. This generates the sources, then computes
# the list of autogenerated sources and writes that to autosrcs.mk in the form of a
# make variable. Note that we use *shell* constructs to get the list of sources, not
# make constructs like $(wildcard), which could be expanded at the wrong time relative
# to when the source files are actually created.
autosrcs.mk: auto.template
./generate-sources auto.template
echo "AUTO_SRCS=`echo *.src`" > autosrcs.mk
# How to build *.target files from *.src files.
%.target: %.src
#echo 'build $< > $#'
Short answer: you can't. Make determines all of the rules it will have to execute before it executes any rule.
Longer answer: maybe you can. As you say, you can use recursive Make explicitly, or surreptitiously by, say, building a file which your makefile will include (I'm looking at you, Jack Kelly). Or if you could somehow obtain a list of the files which gen will build, you could write a rule around that. Or you could take a leap of faith like this:
%.target: %.src
build $< > $#
%.src: gen;