I think that WITHOUT_CLANG + WITH_CLANG_BOOTSTRAP (the latter is the default and
doesn't need to be explicit) is supposed to work.
I think that the combination should result in building clang as a cross-tool for
the rest of the build, but not building clang for the target.
However, it seems that the combination is broken at the moment in that the
cross-tool clang is not getting built and the host clang gets used for the build.
I have recently discovered this issue while trying to build releng/14.2 on a
main (aka 15-CURRENT) host with WITHOUT_CLANG in src.conf. Host and target
architectures are the same, amd64.
14.2 has clang 18 as a compiler, while main has clang 19.
The same problem exists for WITHOUT_TOOLCHAIN as well.
I think that this used to work. Or maybe I was just lucky and the compilers were
either the same or sufficiently compatible that the host compiler could compile
the branch code.
I think that
.if ${MK_CLANG} != "no"
SUBDIR+= clang
.endif
in usr.bin/clang/Makefile is the reason why the cross-tool clang is not built
when WITHOUT_CLANG is set.
A bit more of info is in the forwarded message and the thread to which it
belongs.
What's curious is that those lines are there since commit 8e1c989abbd1db4 "Don't
build and install {llvm,clang,lldb}-tblgen for the target", but I only noticed
the problem a few days ago. And apparenlty nobody else has seen it.
I'd imagine that the reported configuration is not too exotic.
So, not sure what and when get broken.
It could also be something with my build environment...
-------- Forwarded Message --------
Subject: Re: c++ error when trying to build releng/14.2 on 'main' host
Date: Thu, 10 Apr 2025 09:05:47 +0300
From: Andriy Gapon <a...@freebsd.org>
To: Dimitry Andric <d...@freebsd.org>
CC: toolch...@freebsd.org
On 09/04/2025 8:28 pm, Andriy Gapon wrote:
What's interesting is that I saw this during the build (make with -s option):
--------------------------------------------------------------
>>> stage 3: cross tools
--------------------------------------------------------------
===> lib/clang (obj,all,install)
===> lib/clang/libllvm (all)
===> lib/clang/libllvm (install)
===> usr.bin/clang (obj,all,install)
===> usr.bin/clang/lld (obj,all,install)
When I compared this to other builds, I noticed a missing bit:
===> usr.bin/clang/clang (all)
===> usr.bin/clang/clang (install)
usr.bin/clang/Makefile has this near the top:
.if ${MK_CLANG} != "no"
SUBDIR+= clang
.endif
If I read this right, it means that the actual clang is not built/installed if
WITHOUT_CLANG is configured.
Even in the cross-tools stage!
I am not sure how it worked before as I do not see any recent changes in that
direct area. Not sure when and what went wrong. Maybe it's something in one of
.mk include files, maybe something in my environment.
As hack I tried this change and it seems to have helped:
--- a/Makefile.inc1
+++ b/Makefile.inc1
@@ -787,6 +787,7 @@
# TOOLS_PREFIX set in BMAKE
XMAKE= ${BMAKE} \
TARGET=${TARGET} TARGET_ARCH=${TARGET_ARCH} \
+ MK_CLANG=${MK_CLANG_BOOTSTRAP} \
MK_LLDB=no \
MK_LLVM_BINUTILS=no \
MK_TESTS=no
I hope that more knowledgeable people can see what the problem could be,
wherever it is.
===> lib/libelftc (obj,all,install)
===> lib/libpe (obj,all,install)
===> usr.bin/elfctl (obj,all,install)
===> usr.bin/elfdump (obj,all,install)
===> usr.bin/objcopy (obj,all,install)
===> usr.bin/nm (obj,all,install)
===> usr.bin/size (obj,all,install)
===> usr.bin/strings (obj,all,install)
===> usr.bin/addr2line (obj,all,install)
===> cddl/lib/libctf (obj,all,install)
===> cddl/lib/libspl (obj,all,install)
===> cddl/usr.bin/ctfconvert (obj,all,install)
===> cddl/usr.bin/ctfmerge (obj,all,install)
===> stand/usb/tools (obj,all,install)
--
Andriy Gapon
