[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[Pierre Neidhardt]

* gnu/packages/llvm.scm (llvm)[arguments]: Set configure-flags to build
a dynamic library bundle in the "lib" output.
Add phases to move the /bin and /include directories to the "out" output.

The goal of this change is to reduce the closure size of LLVM dependents.

- The dynamic library bundles saves a few dozen MiB over the separate dynamic
  libraries.

- Removing the /bin and the /include directories from the dependent input
  saves about 35 MiB for LLVM 10.
---
 gnu/packages/llvm.scm | 65 +++++++++++++++++++++++++++++++++++--------
 1 file changed, 53 insertions(+), 12 deletions(-)

diff --git a/gnu/packages/llvm.scm b/gnu/packages/llvm.scm
index b7bc21ea6e..3e9d428b9f 100644
--- a/gnu/packages/llvm.scm
+++ b/gnu/packages/llvm.scm
@@ -99,7 +99,7 @@ as \"x86_64-linux\"."
        (base32
         "1pwgm6cr0xr5a0hrbqs1zvsvvjvy0yq1y47c96804wcs795s90yz"))))
     (build-system cmake-build-system)
-    (outputs '("out" "opt-viewer"))
+    (outputs '("out" "opt-viewer" "lib"))
     (native-inputs
      `(("python" ,python-2) ;bytes->str conversion in clang>=3.7 needs python-2
        ("perl"   ,perl)))
@@ -108,12 +108,18 @@ as \"x86_64-linux\"."
     (propagated-inputs
      `(("zlib" ,zlib)))                 ;to use output from llvm-config
     (arguments
-     `(#:configure-flags '("-DCMAKE_SKIP_BUILD_RPATH=FALSE"
-                           "-DCMAKE_BUILD_WITH_INSTALL_RPATH=FALSE"
-                           "-DBUILD_SHARED_LIBS:BOOL=TRUE"
-                           "-DLLVM_ENABLE_FFI:BOOL=TRUE"
-                           "-DLLVM_REQUIRES_RTTI=1" ; For some third-party utilities
-                           "-DLLVM_INSTALL_UTILS=ON") ; Needed for rustc.
+     `(#:configure-flags (list "-DCMAKE_SKIP_BUILD_RPATH=FALSE"
+                               "-DCMAKE_BUILD_WITH_INSTALL_RPATH=FALSE"
+                               ;; LLVM cannot enable BUILD_SHARED_LIBS with LLVM_LINK_LLVM_DYLIB.
+                               ;; "-DBUILD_SHARED_LIBS:BOOL=TRUE"
+                               "-DLLVM_BUILD_LLVM_DYLIB=ON"
+                               "-DLLVM_LINK_LLVM_DYLIB=ON"
+                               (string-append "-DCMAKE_INSTALL_PREFIX=" (assoc-ref %outputs "lib"))
+                               (string-append "-DCMAKE_INSTALL_RPATH=" (assoc-ref %outputs "lib")
+                                              "/lib")
+                               "-DLLVM_ENABLE_FFI:BOOL=TRUE"
+                               "-DLLVM_REQUIRES_RTTI=1" ; For some third-party utilities
+                               "-DLLVM_INSTALL_UTILS=ON") ; Needed for rustc.
 
        ;; Don't use '-g' during the build, to save space.
        #:build-type "Release"
@@ -128,14 +134,49 @@ as \"x86_64-linux\"."
              (setenv "LD_LIBRARY_PATH"
                      (string-append (getcwd) "/lib"))
              #t))
-         (add-after 'install 'install-opt-viewer
+         (add-after 'install 'install-bin
            (lambda* (#:key outputs #:allow-other-keys)
              (let* ((out (assoc-ref outputs "out"))
+                    (out-lib (string-append out "/lib"))
+                    (lib-output (assoc-ref outputs "lib"))
+                    (lib-bin (string-append lib-output "/bin")))
+               (mkdir-p out)
+               (rename-file (string-append lib-output "/bin")
+                            (string-append out "/bin"))
+               ;; llvm-config is required by most lib dependents.  It's only a
+               ;; few KiB, so it does not warrant a separate output.
+               (mkdir-p lib-bin)
+               (rename-file (string-append out "/bin/llvm-config")
+                            (string-append lib-bin "/llvm-config"))
+               (rename-file (string-append lib-output "/include")
+                            (string-append out "/include"))
+               (mkdir-p out-lib)
+               (if (file-exists? (string-append lib-output "/lib/cmake"))
+                   (rename-file (string-append lib-output "/lib/cmake")
+                                (string-append out-lib "/cmake"))
+                   ;; The cmake files change location in llvm 3.9.
+                   (begin
+                     (mkdir-p (string-append out "/share/llvm"))
+                     (rename-file (string-append lib-output "/share/llvm/cmake")
+                                  (string-append out "/share/llvm/cmake"))))
+               (for-each
+                (lambda (file)
+                  (rename-file file
+                               (string-append out-lib "/" (basename file))))
+                (find-files (string-append lib-output "/lib") "\\.a$"))
+               (for-each
+                (lambda (file)
+                  (symlink file
+                           (string-append out-lib "/" (basename file))))
+                (find-files (string-append lib-output "/lib") "\\.so")))
+             #t))
+         (add-after 'install 'install-opt-viewer
+           (lambda* (#:key outputs #:allow-other-keys)
+             (let* ((lib-output (assoc-ref outputs "lib"))
                     (opt-viewer-out (assoc-ref outputs "opt-viewer"))
-                    (opt-viewer-share-dir (string-append opt-viewer-out "/share"))
-                    (opt-viewer-dir (string-append opt-viewer-share-dir "/opt-viewer")))
-               (mkdir-p opt-viewer-share-dir)
-               (rename-file (string-append out "/share/opt-viewer")
+                    (opt-viewer-dir (string-append opt-viewer-out "/share/opt-viewer")))
+               (mkdir-p (dirname opt-viewer-dir))
+               (rename-file (string-append lib-output "/share/opt-viewer")
                             opt-viewer-dir))
              #t)))))
     (home-page "https://www.llvm.org")

base-commit: 0e1428ac5dc3a7f1aa68988dd88885009e9706a6
-- 
2.27.0

[bug#42576] [PATCH 2/2] gnu: llvm-3.9.1: Move libraries to "lib" output.

[Pierre Neidhardt]

* gnu/packages/llvm.scm (llvm-3.9.1)[outputs]: Add "lib".
---
 gnu/packages/llvm.scm | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/gnu/packages/llvm.scm b/gnu/packages/llvm.scm
index 3e9d428b9f..1c099bbdd3 100644
--- a/gnu/packages/llvm.scm
+++ b/gnu/packages/llvm.scm
@@ -690,7 +690,7 @@ components which highly leverage existing libraries in the larger LLVM Project."
       (sha256
        (base32
         "1vi9sf7rx1q04wj479rsvxayb6z740iaz3qniwp266fgp5a07n8z"))))
-    (outputs '("out"))
+    (outputs '("out" "lib"))
     (arguments
      (substitute-keyword-arguments (package-arguments llvm)
        ((#:phases phases)
-- 
2.27.0

[bug#42576] Acknowledgement ([PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.)

[Pierre Neidhardt]

[-- Attachment #1: Type: text/plain, Size: 895 bytes --]

This patch is meant for core-updates since it rebuilds every LLVm
dependent, so more than 1000+ packages.  I've only tested against master
though.

It's not ready yet.

Since LLVM@10 takes a long time to compile, it's cumbersome to iterate
against it.
Instead, I found out that LLVM@3.5 builds much faster and the `pure'
package is a great candidate for testing.

I've added ("llvm" llvm-3.5 "lib") to the dependencies of pure.
But ("llvm" llvm-3.5) must be included as a native input because that's
where the C headers (include) files are.

The end result still depends on llvm "out" because the .so files has
references to the headers.

We could move the headers to a separate output, but LLVM@10 headers are
17MiB big already, so I'd rather not include them in the closure of
every package.

Any idea how to remove them?

-- 
Pierre Neidhardt
https://ambrevar.xyz/

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 487 bytes --]

[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[Danny Milosavljevic]

[-- Attachment #1: Type: text/plain, Size: 450 bytes --]

Hmm, glib for example moves the binaries to a "bin" output while retaining the
libraries in "out".

Any reason this patch moves the libraries and not the binaries?  Especially
since there are no end-user facing programs in llvm anyway (programmer-facing
maybe).  I mean we can...

Also, I would have thought that those llvm programs reference the llvm
libraries anyway, and thus in the end no space could be saved.  Is this
not the case?


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[Pierre Neidhardt]

[-- Attachment #1: Type: text/plain, Size: 675 bytes --]

Hi Danny!

Danny Milosavljevic <dannym@scratchpost.org> writes:

> Hmm, glib for example moves the binaries to a "bin" output while retaining the
> libraries in "out".
>
> Any reason this patch moves the libraries and not the binaries?

Ludo suggested this way I think out of consistency with the rest.

> Also, I would have thought that those llvm programs reference the llvm
> libraries anyway, and thus in the end no space could be saved.  Is this
> not the case?

This patch not about saving space for LLVM programs, but packages that depend
on LLVM libraries and which don't need the programs.

Cheers!

-- 
Pierre Neidhardt
https://ambrevar.xyz/

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 487 bytes --]

[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[Jakub Kądziołka]

[-- Attachment #1: Type: text/plain, Size: 913 bytes --]

On Tue, Jul 28, 2020 at 11:58:22AM +0200, Pierre Neidhardt wrote:
> * gnu/packages/llvm.scm (llvm)[arguments]: Set configure-flags to build
> a dynamic library bundle in the "lib" output.
> Add phases to move the /bin and /include directories to the "out" output.
> 
> The goal of this change is to reduce the closure size of LLVM dependents.
> 
> - The dynamic library bundles saves a few dozen MiB over the separate dynamic
>   libraries.
> 
> - Removing the /bin and the /include directories from the dependent input
>   saves about 35 MiB for LLVM 10.

Pierre,

please note that cmake seems to store a list of files installed by the
package, which broke cmake-using dependents of clang when we attempted
to change the set of files installed: http://issues.guix.gnu.org/41872

Did you try building something that depends on LLVM and uses
cmake-build-system?

Regards,
Jakub Kądziołka

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[Pierre Neidhardt]

[-- Attachment #1: Type: text/plain, Size: 1039 bytes --]

Hi Jakub,

Jakub Kądziołka <kuba@kadziolka.net> writes:

> Pierre,
>
> please note that cmake seems to store a list of files installed by the
> package, which broke cmake-using dependents of clang when we attempted
> to change the set of files installed: http://issues.guix.gnu.org/41872
>
> Did you try building something that depends on LLVM and uses
> cmake-build-system?

It is broken indeed, this is one of the things that need to be fixed
before we can merge this patch.

In issue 41872 the problem is with missing .a files.
A similar issue occurs here since we move files around, but the CMake
files are not aware of the move.

One possible fix would be to patch the CMake files with the new locations.
This is rather inelegant though.

A better fix would be to configure CMake to produce the various files
directly to the right location, e.g. the binary files, the headers and
the libraries to their own respective outputs.

Any clue if we can do that?

-- 
Pierre Neidhardt
https://ambrevar.xyz/

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 487 bytes --]

[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[Sarah Morgensen]

Hello Pierre,

Thanks for your pioneering work on this!

Pierre Neidhardt <mail@ambrevar.xyz> writes:

> Hi Jakub,
>
> Jakub Kądziołka <kuba@kadziolka.net> writes:
>
>> Pierre,
>>
>> please note that cmake seems to store a list of files installed by the
>> package, which broke cmake-using dependents of clang when we attempted
>> to change the set of files installed: http://issues.guix.gnu.org/41872
>>
>> Did you try building something that depends on LLVM and uses
>> cmake-build-system?
>
> It is broken indeed, this is one of the things that need to be fixed
> before we can merge this patch.
>
> In issue 41872 the problem is with missing .a files.
> A similar issue occurs here since we move files around, but the CMake
> files are not aware of the move.
>
> One possible fix would be to patch the CMake files with the new locations.
> This is rather inelegant though.
>
> A better fix would be to configure CMake to produce the various files
> directly to the right location, e.g. the binary files, the headers and
> the libraries to their own respective outputs.
>
> Any clue if we can do that?

There has been some recent work on this in the LLVM project [0] and in
Nix [1][2], based on the `GnuInstallDirs' CMake module.  It looks like
this would be doable for us, especially if we move in the direction of
'dev' outputs [3].

[0] https://reviews.llvm.org/D99484
[1] https://github.com/NixOS/nikpkgs/pull/111487
[2] https://github.com/NixOS/nixpkgs/tree/master/pkgs/development/compilers/llvm/12
[3] https://lists.gnu.org/archive/html/guix-devel/2021-09/msg00107.html

Perhaps it's time to revive this effort (particularly for clang, which
is a behemoth)?

--
Sarah

[bug#42576] [PATCH 1/2] gnu: llvm: Move dynamic libraries to a separate "lib" output.

[Pierre Neidhardt]

[-- Attachment #1: Type: text/plain, Size: 208 bytes --]

Hello Sarah, thanks for the update.

Unfortunately I won't be able to commit much time to the Guix project at
the moment, but if you want to take over this initiative, you're more
than welcome :)

Good luck!

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 511 bytes --]