Building packages targeting psABIs

Starting with version 2.33, the GNU C library (glibc) grew the capability to search for shared libraries using additional paths, based on the hardware capabilities of the machine running the code. This was a great boon for x86_64, which was first released in 2003, and has seen many changes in the capabilities of the hardware since then. While it is extremely common for Linux distributions to compile for a baseline which encompasses all of an architecture, there is performance being left on the table by targeting such an old specification and not one of the newer revisions.

One option used internally in glibc and in some other performance-critical libraries is indirect functions, or IFUNCs (see also here) The loader, ld.so uses them to pick function implementations optimized for the available CPU at load time. GCC's functional multi-versioning (FMV) generates several optimized versions of functions, using the IFUNC mechanism so the approprate one is selected at load time. These are strategies which most performance-sensitive libraries do, but not all of them.

With the --tune using package transformation option, Guix implements so-called package multi-versioning which creates package variants using compiler flags set to use optimizations targeted for a specific CPU.

Finally - and we're getting to the central topic of this post! - glibc since version 2.33 supports another approach: ld.so would search not just the /lib folder, but also the glibc-hwcaps folders, which for x86_64 included /lib/glibc-hwcaps/x86-64-v2, /lib/glibc-hwcaps/x86-64-v3 and /lib/glibc-hwcaps/x86-64-v4, corresponding to the psABI micro-architectures of the x86_64 architecture (psABI stands for processor supplement of the application binary interface and refers to the document that specifies among other things those x86-64-v* levels). This means that if a library was compiled against the baseline of the architecture then it should be installed in /lib, but if it were compiled a second time, this time using (depending on the build instructions) -march=x86-64-v2, then the libraries could be installed in /lib/glibc-hwcaps/x86-64-v2 and then glibc, using ld.so, would choose the correct library at runtime.

These micro-architectures aren't a perfect match for the different hardware available, it is often the case that a particular CPU would satisfy the requirements of one tier and part of the next but would therefore only be able to use the optimizations provided by the first tier and not by the added features that the CPU also supports.

This of course shouldn't be a problem in Guix; it's possible, and even encouraged, to adjust packages to be more useful for one's needs. The problem comes from the search paths: ld.so will only search for the glibc-hwcaps directory if it has already found the base library in the preceding /lib directory. This isn't a problem for distributions following the File System Hierarchy (FHS), but for Guix we will need to ensure that all the different versions of the library will be in the same output.

With a little bit of planning this turns out to not be as hard as it sounds. Lets take for example, the GNU Scientific Library, gsl, a math library which helps with all sorts of numerical analysis. First we create a procedure to generate our 3 additional packages, corresponding to the psABIs that are searched for in the glibc-hwcaps directory.

(define (gsl-hwabi psabi)
  (package/inherit gsl
    (name (string-append "gsl-" psabi))
    (arguments
     (substitute-keyword-arguments (package-arguments gsl)
       ((#:make-flags flags #~'())
        #~(append (list (string-append "CFLAGS=-march=" #$psabi)
                        (string-append "CXXFLAGS=-march=" #$psabi))
                  #$flags))
       ((#:configure-flags flags #~'())
        #~(append (list (string-append "--libdir=" #$output
                                       "/lib/glibc-hwcaps/" #$psabi))
                  #$flags))
       ;; The building machine can't necessarily run the code produced.
       ((#:tests? _ #t) #f)
       ((#:phases phases #~%standard-phases)
        #~(modify-phases #$phases
            (add-after 'install 'remove-extra-files
              (lambda _
                (for-each (lambda (dir)
                            (delete-file-recursively (string-append #$output dir)))
                          (list (string-append "/lib/glibc-hwcaps/" #$psabi "/pkgconfig")
                                "/bin" "/include" "/share"))))))))
    (supported-systems '("x86_64-linux" "powerpc64le-linux"))
    (properties `((hidden? . #t)
                  (tunable? . #f)))))

We remove some directories and any binaries since we only want the libraries produced from the package; we want to use the headers and any other bits from the main package. We then combine all of the pieces together to produce a package which can take advantage of the hardware on which it is run:

(define-public gsl-hwcaps
  (package/inherit gsl
    (name "gsl-hwcaps")
    (arguments
     (substitute-keyword-arguments (package-arguments gsl)
       ((#:phases phases #~%standard-phases)
        #~(modify-phases #$phases
            (add-after 'install 'install-optimized-libraries
              (lambda* (#:key inputs outputs #:allow-other-keys)
                (let ((hwcaps "/lib/glibc-hwcaps/"))
                  (for-each
                    (lambda (psabi)
                      (copy-recursively
                        (string-append (assoc-ref inputs (string-append "gsl-" psabi))
                                       hwcaps psabi)
                        (string-append #$output hwcaps psabi))
                  '("x86-64-v2" "x86-64-v3" "x86-64-v4"))))))))
    (native-inputs
     (modify-inputs (package-native-inputs gsl)
                    (append (gsl-hwabi "x86-64-v2")
                            (gsl-hwabi "x86-64-v3")
                            (gsl-hwabi "x86-64-v4"))))
    (supported-systems '("x86_64-linux"))
    (properties `((tunable? . #f)))))

In this case the size of the final package is increased by about 13 MiB, from 5.5 MiB to 18 MiB. It is up to you if the speed-up from providing an optimized library is worth the size trade-off.

To use this package as a replacement build input in a package package-input-rewriting/spec is a handy tool:

(define use-glibc-hwcaps
 (package-input-rewriting/spec
   ;; Replace some packages with ones built targeting custom packages build
   ;; with glibc-hwcaps support.
   `(("gsl" . ,(const gsl-hwcaps)))))

(define-public inkscape-with-hwcaps
  (package
    (inherit (use-glibc-hwcaps inkscape))
    (name "inkscape-with-hwcaps")))

Of the Guix supported architectures, x86_64-linux and powerpc64le-linux can both benefit from this new capability.

Through the magic of newer versions of GCC and LLVM it is safe to use these libraries in place of the standard libraries while compiling packages; these compilers know about the glibc-hwcap directories and will purposefully link against the base library during build time, with glibc's ld.so choosing the optimized library at runtime.

One possible use case for these libraries is creating guix packs of packages to run on other systems. By substituting these libraries it becomes possible to crate a guix pack which will have better performance than a standard package used in a guix pack. This works even when the included libraries don't make use of the IFUNCs from glibc or functional multi-versioning from GCC. Providing optimized yet portable pre-compiled binaries is a great way to take advantage of this feature.

About GNU Guix

GNU Guix is a transactional package manager and an advanced distribution of the GNU system that respects user freedom. Guix can be used on top of any system running the Hurd or the Linux kernel, or it can be used as a standalone operating system distribution for i686, x86_64, ARMv7, AArch64 and POWER9 machines.

In addition to standard package management features, Guix supports transactional upgrades and roll-backs, unprivileged package management, per-user profiles, and garbage collection. When used as a standalone GNU/Linux distribution, Guix offers a declarative, stateless approach to operating system configuration management. Guix is highly customizable and hackable through Guile programming interfaces and extensions to the Scheme language.

Související témata:

Packaging

Pokud není uvedeno jinak, příspěvky na blogu na těchto stránkách jsou chráněny autorskými právy příslušných autorů a zveřejněny za podmínek licence CC-BY-SA 4.0 a podmínek licence GNU Free Documentation License (verze 1.3 nebo novější, bez invariantních částí, bez textů na přední straně obálky a bez textů na zadní straně obálky).