The incremental build of Linux kernel is pretty slow when lots of
objects are compiled. The rebuild of allmodconfig may take a few
minutes even when none of the objects needs to be rebuilt.

The time-consuming part in the incremental build is the evaluation of
if_changed* macros since they are used in the recipes to compile C and
assembly source files into objects.

I notice the following code in if_changed* is expensive:

  $(filter-out $(PHONY) $(wildcard $^),$^)

In the incremental build, every object has its .*.cmd file, which
contains the auto-generated list of included headers. So, $^ are
expanded into the long list of the source file + included headers,
and $(wildcard $^) checks whether they exist.

It may not be clear why this check exists there.

Here is the record of my research.

[1] The first code addition into Kbuild

This code dates back to 2002. It is the pre-git era. So, I copy-pasted
it from the historical git tree.

| commit 4a6db079 (HEAD)
| Author: Kai Germaschewski <kai@tp1.ruhr-uni-bochum.de>
| Date:   Mon Jun 17 00:22:37 2002 -0500
|
|     kbuild: Handle removed headers
|
|     New and old way to handle dependencies would choke when a file
|     #include'd by other files was removed, since the dependency on it was
|     still recorded, but since it was gone, make has no idea what to do about
|     it (and would complain with "No rule to make <file> ...")
|
|     We now add targets for all the previously included files, so make will
|     just ignore them if they disappear.
|
| diff --git a/Rules.make b/Rules.make
| index 6ef827d3df39..7db5301ea7db 100644
| --- a/Rules.make
| +++ b/Rules.make
| @@ -446,7 +446,7 @@ if_changed = $(if $(strip $? \
|  # execute the command and also postprocess generated .d dependencies
|  # file
|
| -if_changed_dep = $(if $(strip $? \
| +if_changed_dep = $(if $(strip $? $(filter-out FORCE $(wildcard $^),$^)\
|                           $(filter-out $(cmd_$(1)),$(cmd_$@))\
|                           $(filter-out $(cmd_$@),$(cmd_$(1)))),\
|         @set -e; \
| diff --git a/scripts/fixdep.c b/scripts/fixdep.c
| index b5d7bee8efc7..db45bd1888c0 100644
| --- a/scripts/fixdep.c
| +++ b/scripts/fixdep.c
| @@ -292,7 +292,7 @@ void parse_dep_file(void *map, size_t len)
|                 exit(1);
|         }
|         memcpy(s, m, p-m); s[p-m] = 0;
| -       printf("%s: \\\n", target);
| +       printf("deps_%s := \\\n", target);
|         m = p+1;
|
|         clear_config();
| @@ -314,7 +314,8 @@ void parse_dep_file(void *map, size_t len)
|                 }
|                 m = p + 1;
|         }
| -       printf("\n");
| +       printf("\n%s: $(deps_%s)\n\n", target, target);
| +       printf("$(deps_%s):\n", target);
|  }
|
|  void print_deps(void)

The "No rule to make <file> ..." error can be solved by passing -MP to
the compiler, but I think the detection of header removal is a good
feature. When a header is removed, all source files that previously
included it should be re-compiled. This makes sure we has correctly
got rid of #include directives of it.

This is also related with the behavior of $?. The GNU Make manual says:

  $?
      The names of all the prerequisites that are newer than the target,
      with spaces between them.

This does not explain whether a non-existent prerequisite is considered
to be newer than the target.

At this point of time, GNU Make 3.7x was used, where the $? did not
include non-existent prerequisites. Therefore,

  $(filter-out FORCE $(wildcard $^),$^)

was useful to detect the header removal, and to rebuild the related
objects if it is the case.

[2] Change of $? behavior

Later, the behavior of $? was changed (fixed) to include prerequisites
that did not exist.

First, GNU Make commit 64e16d6c00a5 ("Various changes getting ready for
the release of 3.81.") changed it, but in the release test of 3.81, it
turned out to break the kernel build.

See these:

 - http://lists.gnu.org/archive/html/bug-make/2006-03/msg00003.html
 - https://savannah.gnu.org/bugs/?16002
 - https://savannah.gnu.org/bugs/?16051

Then, GNU Make commit 6d8d9b74d9c5 ("Numerous updates to tests for
issues found on Cygwin and Windows.") reverted it for the 3.81 release
to give Linux kernel time to adjust to the new behavior.

After the 3.81 release, GNU Make commit 7595f38f62af ("Fixed a number
of documentation bugs, plus some build/install issues:") re-added it.

[3] Adjustment to the new $? behavior on Kbuild side

Meanwhile, the kernel build was changed by commit 4f193362 ("kbuild:
change kbuild to not rely on incorrect GNU make behavior") to adjust to
the new $? behavior.

[4] GNU Make 3.82 released in 2010

GNU Make 3.82 was the first release that integrated the correct $?
behavior. At this point, Kbuild dealt with GNU Make versions with
different $? behaviors.

 3.81 or older:
    $? does not contain any non-existent prerequisite.
    $(filter-out $(PHONY) $(wildcard $^),$^) was useful to detect
    removed include headers.

 3.82 or newer:
    $? contains non-existent prerequisites. When a header is removed,
    it appears in $?. $(filter-out $(PHONY) $(wildcard $^),$^) became
    a redundant check.

With the correct $? behavior, we could have dropped the expensive
check for 3.82 or later, but we did not. (Maybe nobody noticed this
optimization.)

[5] The .SECONDARY special target trips up $?

Some time later, I noticed $? did not work as expected under some
circumstances. As above, $? should contain non-existent prerequisites,
but the ones specified as SECONDARY do not appear in $?.

I asked this in GNU Make ML, and it seems a bug:

  https://lists.gnu.org/archive/html/bug-make/2019-01/msg00001.html

Since commit 8e9b61b2 ("kbuild: move .SECONDARY special target to
Kbuild.include"), all files, including headers listed in .*.cmd files,
are treated as secondary.

So, we are back into the incorrect $? behavior.

If we Kbuild want to react to the header removal, we need to keep
$(filter-out $(PHONY) $(wildcard $^),$^) but this makes the rebuild
so slow.

[Summary]

 - I believe noticing the header removal and recompiling related objects
   is a nice feature for the build system.

 - If $? worked correctly, $(filter-out $(PHONY),$?) would be enough
   to detect the header removal.

 - Currently, $? does not work correctly when used with .SECONDARY,
   and Kbuild is hit by this bug.

 - I filed a bug report for this, but not fixed yet as of writing.

 - Currently, the header removal is detected by the following expensive
   code:

    $(filter-out $(PHONY) $(wildcard $^),$^)

 - I do not want to revert commit 8e9b61b2 ("kbuild: move
   .SECONDARY special target to Kbuild.include"). Specifying
   .SECONDARY globally is clean, and it matches to the Kbuild policy.

This commit proactively removes the expensive check since it makes the
incremental build faster. A downside is Kbuild will no longer be able
to notice the header removal.

You can confirm it by the full-build followed by a header removal, and
then re-build.

  $ make defconfig all
    [ full build ]
  $ rm include/linux/device.h
  $ make
    CALL    scripts/checksyscalls.sh
    CALL    scripts/atomic/check-atomics.sh
    DESCEND  objtool
    CHK     include/generated/compile.h
  Kernel: arch/x86/boot/bzImage is ready  (#11)
    Building modules, stage 2.
    MODPOST 12 modules

Previously, Kbuild noticed a missing header and emits a build error.
Now, Kbuild is fine with it. This is an unusual corner-case, not a big
deal. Once the $? bug is fixed in GNU Make, everything will work fine.
Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>