I gotta admit, I still find it a bit weird that curl.1 wouldn't be built as part of building the docs, regardless of if ENABLE_MANUAL is set or not.

However, there is also BUILD_CURL_EXE, and I can fully understand that if this is disabled, building the corresponding manual would be a bit weird. Could it be that this and ENABLE_MANUAL are a bit confounded?

curl.1 comes with the official source release tarball, so anyone using that
doesn't need ENABLE_MANUAL=ON to have curl.1. Same for tool_hugehelp.c.

ENABLE_MANUAL=ON has the effect of unconditionally rebuilding (I haven't
retested that after recent changes) curl.1, which breaks reproducibiliy in some
build envs. Not enabling it by default is a workaround for that (and not doing this
extra step for release builds, which is presumably most builds out there.).

We can rename BUILD_DOCS to BUILD_LIBCURL_DOCS to better reflect
its function. (possibly even ENABLE_MANUAL to ENABLE_CURL_MANUAL.)

As for BUILD_CURL_EXE, what about doing this?:

--- a/docs/CMakeLists.txt
+++ b/docs/CMakeLists.txt
@@ -25,6 +25,6 @@
 if(BUILD_DOCS)
   add_subdirectory(libcurl)
 endif()
-if(ENABLE_MANUAL)
+if(ENABLE_MANUAL AND BUILD_CURL_EXE)
   add_subdirectory(cmdline-opts)
 endif()

We can rename BUILD_DOCS to BUILD_LIBCURL_DOCS to better reflect its function. (possibly even ENABLE_MANUAL to ENABLE_CURL_MANUAL.)

That would be much less confusing, so yes please

As for BUILD_CURL_EXE, what about doing this?:

--- a/docs/CMakeLists.txt
+++ b/docs/CMakeLists.txt
@@ -25,6 +25,6 @@
 if(BUILD_DOCS)
   add_subdirectory(libcurl)
 endif()
-if(ENABLE_MANUAL)
+if(ENABLE_MANUAL AND BUILD_CURL_EXE)
   add_subdirectory(cmdline-opts)
 endif()

Looks good to me

Thanks, committed these updates.

curl.1 comes with the official source release tarball, so anyone using that doesn't need ENABLE_MANUAL=ON to have curl.1. Same for tool_hugehelp.c.

ENABLE_MANUAL=ON has the effect of unconditionally rebuilding (I haven't retested that after recent changes) curl.1, which breaks reproducibiliy in some build envs. Not enabling it by default is a workaround for that (and not doing this extra step for release builds, which is presumably most builds out there.).

Reviewing curl-for-win script again, this remains having another
downside: To avoid rebuilding existing curl.1 / tool_hugehelp.c
yet still embedding it into curl, the CPPFLAGS -DUSE_MANUAL
must be passed manually. This was so before this week's changes
and still true.

Doing the rebuild only when curl.1 (and/or tool_hugehelp.c)
are missing or outdated would be one solution; making the output
deterministic would be another (this would require dropping nroff,
which is on bagder's plans). Ideally both.
C USE_MANUAL also has the requirement that tool_hugehelp.c
does actually offer a hugehelp symbol in it, otherwise the build
breaks. Using src/tool_hugehelp.c.cvs satisfies this, but not the
empty stub generated by CMake (or autotools for that matter).

Perhaps an approach where we have a static tool_hugehelp.c in
the source tree, while expecting the manual content inside a
generated tool_hugehelp.h (depending on USE_MANUAL)
could help to avoid some failure modes. But touching any of that looks
like a deep rabbit-hole.

In any case this is left for a separate PR.

Re this discussion, I clearly don't have a deeper understanding of all the ramifications, so I defer to your knowledge.