Re: Build errors when building git on MacOS 11 (x86-64) and for M1 macs

From: Jeff King <peff@peff.net>
Date: Jun 10, 2021, 14:36 UTC
Message-ID: <YMIjaytMYslKjK6z@coredump.intra.peff.net>
In-Reply-To: <87sg1p24zc.fsf@evledraar.gmail.com>

On Thu, Jun 10, 2021 at 03:52:09PM +0200, Ævar Arnfjörð Bjarmason wrote:

Show 8 quoted lines

> I have no access to such a system, but I think think I see the problem
> from what you've supplied here.
> 
> You supplied a CFLAGS=-I/opt/chef-workstation/embedded/include which has
> an archive.h file that's unrelated to the archive.h file git expects.
> 
> Thus we include that and find things unrelated to git, and error when we
> encounter a function we expected to have declared in our own archive.h.

Yeah, I agree that is likely the problem.

> The solution is something like defining a config.mak file where you add
> flags with BASIC_CFLAGS +=, not =. See config.mak.uname for an
> example. You'll then add new directories after our own -I.

That would work, though I think in general we should consider BASIC_CFLAGS to be an internal thing, and encourage people to use the standard CFLAGS, etc.

> This is arguably a bug in git's Makefile in that we should have that
> "-I." in an ESSENTIAL_CFLAGS variable or something, I can't think of a
> scenario where git would compile without it. That or things in builtin/
> should include e.g. ../archive.h, perhaps such a thing isn't portable.

Yeah, we do say "-I." explicitly and unconditionally. I think we'd just want to make sure it's before any user-provided flags. We could do something like this:

diff --git a/Makefile b/Makefile
index 7d719ece46..b5794560e9 100644
--- a/Makefile
+++ b/Makefile
@@ -1241,7 +1241,7 @@ ALL_COMMANDS_TO_INSTALL += git-upload-archive$(X)
 ALL_COMMANDS_TO_INSTALL += git-upload-pack$(X)
 endif
 
-ALL_CFLAGS = $(DEVELOPER_CFLAGS) $(CPPFLAGS) $(CFLAGS)
+ALL_CFLAGS = $(DEVELOPER_CFLAGS) $(CPPFLAGS) $(BASIC_CFLAGS) $(CFLAGS)
 ALL_LDFLAGS = $(LDFLAGS)
 
 comma := ,
@@ -2106,7 +2106,6 @@ PAGER_ENV_CQ = "$(subst ",\",$(subst \,\\,$(PAGER_ENV)))"
 PAGER_ENV_CQ_SQ = $(subst ','\'',$(PAGER_ENV_CQ))
 BASIC_CFLAGS += -DPAGER_ENV='$(PAGER_ENV_CQ_SQ)'
 
-ALL_CFLAGS += $(BASIC_CFLAGS)
 ALL_LDFLAGS += $(BASIC_LDFLAGS)
 
 export DIFF TAR INSTALL DESTDIR SHELL_PATH

but I imagine that it is helpful for other BASIC_CFLAGS to be
overrideable by the user's CFLAGS (especially other system-level -I
directives we may include!). So yes, having an $(ESSENTIAL_CFLAGS) would
work. Or perhaps even just:

diff --git a/Makefile b/Makefile
index 7d719ece46..72517db31e 100644
--- a/Makefile
+++ b/Makefile
@@ -1202,7 +1202,7 @@ endif
 CFLAGS = -g -O2 -Wall
 LDFLAGS =
 CC_LD_DYNPATH = -Wl,-rpath,
-BASIC_CFLAGS = -I.
+BASIC_CFLAGS =
 BASIC_LDFLAGS =
 
 # library flags
@@ -1241,7 +1241,7 @@ ALL_COMMANDS_TO_INSTALL += git-upload-archive$(X)
 ALL_COMMANDS_TO_INSTALL += git-upload-pack$(X)
 endif
 
-ALL_CFLAGS = $(DEVELOPER_CFLAGS) $(CPPFLAGS) $(CFLAGS)
+ALL_CFLAGS = -I. $(DEVELOPER_CFLAGS) $(CPPFLAGS) $(CFLAGS)
 ALL_LDFLAGS = $(LDFLAGS)
 
 comma := ,

since this "-I." really is special.

> I think (but have not confirmed) that you probably got this far because
> your compiler will stick -I. at the /end/ of the flags implicitly (or is
> that standardized C behavior? I can't remember).

The standard just says (for quoted #include lines):

  The named source file is searched for in an implementation-defined
  manner.

Most compilers put the directory of the source file at the start of the
search path. E.g., gcc says under -I:

  1. For the quote form of the include directive, the directory of the
     current file is searched first.

Which is why we need "-I." at all; we are finding "archive.h" from
"builtin/*.c". I'm not sure what "current file" there means, or how
portable it is. Is it the source file being filed, or the file
containing the #include directive?

I'd hope it's consistently the latter. Otherwise "foo.h" which includes
"bar.h" cannot be included as "../foo.h" (from builtin/foo.c) and as
"foo.h" (from top-level foo.c).

If so, then yeah, using "../archive.h" (and dropping -I. entirely) would
be an option. Which is nice, because it makes things less magical and
more predictable (think what confusion we'd see if we introduced
"archive.h" into builtin/ ourselves).

We do use "../foo.h" in lots of places already. But the fact that we
_also_ have "-I." means I don't think we have any data on the "what does
current file mean" question above. My gut feeling is that it probably
does consistently work as we'd want it to, though.

-Peff