OCaml 4.10 released

Dear OCaml users,

We have the pleasure of celebrating the birthday of Francis Ronalds
by announcing the release of OCaml version 4.10.0.

Some of the highlights in this release are:

  • A new best-fit allocator for the major heap which reduces both GC cost and
    memory usage.
  • Some preliminary runtime work for OCaml multicore
  • Immutable strings are now enforced at configuration time
  • User-defined indexing operators for multidimensional arrays
  • Coming soon: statmemprof, a new statistical memory profiler.
    The external API will be release next version.
  • Various improvements to the manual
  • More precise exhaustiveness check for GADTs
  • Many bug fixes

Merlin, the OCaml editor service, is not yet available for this release.
We will publish a follow-up announcement when Merlin is ready.

This release is (or soon will be) available as a set of OPAM switches,
and as a source download here:

https://caml.inria.fr/pub/distrib/ocaml-4.10/

Happy hacking,

— Florian Angeletti for the OCaml team.

OCaml 4.10.0 (21 February 2020)

(Changes that can break existing programs are marked with a “*”)

Language features

  • #7757, #1726: multi-indices for extended indexing operators:
    a.%{0;1;2} desugars to ( .%{ ;.. } ) a [|0;1;2|]
    (Florian Angeletti, review by Gabriel Radanne)
  • [breaking change] #1859, #9117: enforce safe (immutable) strings by removing
    the -unsafe-string option by default. This can be overridden by
    a configure-time option (available since 4.04 in 2016):
    –disable-force-safe-string since 4.08, -no-force-safe-since
    between 4.07 and 4.04.
    In the force-safe-string mode (now the default), the return type of the
    String_val macro in C stubs is const char* instead of
    char*. This change may break C FFI code.
    (Kate Deplaix)
  • #6662, #8908: allow writing “module _ = E” to ignore module expressions
    (Thomas Refis, review by Gabriel Radanne)

Runtime system:

  • #8809, #9292: Add a best-fit allocator for the major heap; still
    experimental, it should be much better than current allocation
    policies (first-fit and next-fit) for programs with large heaps,
    reducing both GC cost and memory usage.
    This new best-fit is not (yet) the default; set it explicitly with
    OCAMLRUNPARAM=“a=2” (or Gc.set from the program). You may also want
    to increase the space_overhead parameter of the GC (a percentage,
    80 by default), for example OCAMLRUNPARAM=“o=85”, for optimal
    speed.
    (Damien Doligez, review by Stephen Dolan, Jacques-Henri Jourdan,
    Xavier Leroy, Leo White)
  • [breaking change] #8713, #8940, #9115, #9143, #9202, #9251:
    Introduce a state table in the runtime to contain the global variables.
    (The Multicore runtime will have one such state for each domain.)

    This changes the status of some internal variables of the OCaml runtime;
    in many cases the header file originally defining the internal variable
    provides a compatibility macro with the old name, but programs
    re-defining those variables by hand need to be fixed.

    (KC Sivaramakrishnan and Stephen Dolan,
    compatibility hacking by David Allsopp, Florian Angeletti, Kate Deplaix,
    Jacques Garrigue, Guillaume Munch-Maccagnoni and Nicolás Ojeda Bär,
    review by David Allsopp, Alain Frisch, Nicolas Ojeda Bar,
    Gabriel Scherer, Damien Doligez, and Guillaume Munch-Maccagnoni)

  • #8993: New C functions caml_process_pending_actions{,_exn} in
    caml/signals.h, intended for executing all pending actions inside
    long-running C functions (requested minor and major collections,
    signal handlers, finalisers, and memprof callbacks). The function
    caml_process_pending_actions_exn returns any exception arising
    during their execution, allowing resources to be cleaned-up before
    re-raising.
    (Guillaume Munch-Maccagnoni, review by Jacques-Henri Jourdan,
    Stephen Dolan, and Gabriel Scherer)
  • [breaking change] #8691, #8897, #9027: Allocation functions are now guaranteed not to
    trigger any OCaml callback when called from C. In long-running C
    functions, this can be replaced with calls to
    caml_process_pending_actions at safe points.
    Side effect of this change: in bytecode mode, polling for
    asynchronous callbacks is performed at every minor heap allocation,
    in addition to function calls and loops as in previous OCaml
    releases.
    (Jacques-Henri Jourdan, review by Stephen Dolan, Gabriel Scherer and
    Guillaume Munch-Maccagnoni)

  • [breaking change] #9037: caml_check_urgent_gc is now guaranteed not to trigger any
    finaliser. In long-running C functions, this can be replaced
    with calls to caml_process_pending_actions at safe points.
    (Guillaume Munch-Maccagnoni, review by Jacques-Henri Jourdan and
    Stephen Dolan)

  • #8619: Ensure Gc.minor_words remains accurate after a GC.
    (Stephen Dolan, Xavier Leroy and David Allsopp,
    review by Xavier Leroy and Gabriel Scherer)

  • #8667: Limit GC credit to 1.0
    (Leo White, review by Damien Doligez)

  • #8670: Fix stack overflow detection with systhreads
    (Stephen Dolan, review by Xavier Leroy, Anil Madhavapeddy, Gabriel Scherer,
    Frédéric Bour and Guillaume Munch-Maccagnoni)

  • [breaking change] #8711: The major GC hooks are no longer allowed to interact with the
    OCaml heap.
    (Jacques-Henri Jourdan, review by Damien Doligez)
  • #8630: Use abort() instead of exit(2) in caml_fatal_error, and add
    the new hook caml_fatal_error_hook.
    (Jacques-Henri Jourdan, review by Xavier Leroy)

  • #8641: Better call stacks when a C call is involved in byte code mode
    (Jacques-Henri Jourdan, review by Xavier Leroy)

  • #8634, #8668, #8684, #9103 (originally #847): Statistical memory profiling.
    In OCaml 4.10, support for allocations in the minor heap in native
    mode is not available, and callbacks for promotions and
    deallocations are not available.
    Hence, there is not any public API for this feature yet.
    (Jacques-Henri Jourdan, review by Stephen Dolan, Gabriel Scherer
    and Damien Doligez)

  • #9268, #9271: Fix bytecode backtrace generation with large integers present.
    (Stephen Dolan and Mark Shinwell, review by Gabriel Scherer and
    Jacques-Henri Jourdan)

Standard library:

  • #8760: List.concat_map : ('a -> 'b list) -> 'a list -> 'b list
    (Gabriel Scherer, review by Daniel Bünzli and Thomas Refis)

  • #8832: List.find_map : ('a -> 'b option) -> 'a list -> 'b option
    (Gabriel Scherer, review by Jeremy Yallop, Nicolás Ojeda Bär
    and Daniel Bünzli)

  • #7672, #1492: Add Filename.quote_command to produce properly-quoted
    commands for execution by Sys.command.
    (Xavier Leroy, review by David Allsopp and Damien Doligez)

  • #8971: Add Filename.null, the conventional name of the “null” device.
    (Nicolás Ojeda Bär, review by Xavier Leroy and Alain Frisch)

  • #8651: add ‘%#F’ modifier in printf to output OCaml float constants
    in hexadecimal
    (Pierre Roux, review by Gabriel Scherer and Xavier Leroy)

  • #8657: Optimization in [Array.make] when initializing with unboxed
    or young values.
    (Jacques-Henri Jourdan, review by Gabriel Scherer and Stephen Dolan)

  • #8716: Optimize [Array.fill] and [Hashtbl.clear] with a new runtime primitive
    (Alain Frisch, review by David Allsopp, Stephen Dolan and Damien Doligez)

  • #8530: List.sort: avoid duplicate work by chop
    (Guillaume Munch-Maccagnoni, review by David Allsopp, Damien Doligez and
    Gabriel Scherer)

Other libraries:

  • #1939, #2023: Implement Unix.truncate and Unix.ftruncate on Windows.
    (Florent Monnier and Nicolás Ojeda Bär, review by David Allsopp)

Code generation and optimizations:

  • #8806: Add an [@@immediate64] attribute for types that are known to
    be immediate only on 64 bit platforms
    (Jérémie Dimino, review by Vladimir Keleshev)

  • #9028, #9032: Fix miscompilation by no longer assuming that
    untag_int (tag_int x) = x in Cmmgen; the compilation of (n lsl 1) + 1,
    for example, would be incorrect if evaluated with a large value for n.
    (Stephen Dolan, review by Vincent Laviron and Xavier Leroy)

  • #8672: Optimise Switch code generation on booleans.
    (Stephen Dolan, review by Pierre Chambart)

  • #8990: amd64: Emit 32bit registers for Iconst_int when we can
    (Xavier Clerc, Tom Kelly and Mark Shinwell, review by Xavier Leroy)

  • #2322: Add pseudo-instruction Ladjust_trap_depth to replace
    dummy Lpushtrap generated in linearize
    (Greta Yorsh and Vincent Laviron, review by Xavier Leroy)

  • #8707: Simplif: more regular treatment of Tupled and Curried functions
    (Gabriel Scherer, review by Leo White and Alain Frisch)

  • #8526: Add compile-time option -function-sections in ocamlopt to emit
    each function in a separate named text section on supported targets.
    (Greta Yorsh, review by Pierre Chambart)

  • #2321: Eliminate dead ICatch handlers
    (Greta Yorsh, review by Pierre Chambart and Vincent Laviron)

  • #8919: lift mutable lets along with immutable ones
    (Leo White, review by Pierre Chambart)

  • #8909: Graph coloring register allocator: the weights put on
    preference edges should not be divided by 2 in branches of
    conditional constructs, because it is not good for performance
    and because it leads to ignoring preference edges with 0 weight.
    (Eric Stavarache, review by Xavier Leroy)

  • #9006: int32 code generation improvements
    (Stephen Dolan, designed with Greta Yorsh, review by Xavier Clerc,
    Xavier Leroy and Alain Frisch)

  • #9041: amd64: Avoid stall in sqrtsd by clearing destination.
    (Stephen Dolan, with thanks to Andrew Hunter, Will Hasenplaugh,
    Spiros Eliopoulos and Brian Nigito. Review by Xavier Leroy)

  • #2165: better unboxing heuristics for let-bound identifiers
    (Alain Frisch, review by Vincent Laviron and Gabriel Scherer)

  • #8735: unbox across static handlers
    (Alain Frisch, review by Vincent Laviron and Gabriel Scherer)

Manual and documentation:

  • #8718, #9089: syntactic highlighting for code examples in the manual
    (Florian Angeletti, report by Anton Kochkov, review by Gabriel Scherer)

  • #9101: add links to section anchor before the section title,
    make the name of those anchor explicits.
    (Florian Angeletti, review by Daniel Bünzli, Sébastien Hinderer,
    and Gabriel Scherer)

  • #9257, cautionary guidelines for using the internal runtime API
    without too much updating pain.
    (Florian Angeletti, review by Daniel Bünzli, Guillaume Munch-Maccagnoni
    and KC Sivaramakrishnan)

  • #8950: move local opens in pattern out of the extension chapter
    (Florian Angeletti, review and suggestion by Gabriel Scherer)

  • #9088, #9097: fix operator character classes
    (Florian Angelettion, review by Gabriel Scherer,
    report by Clément Busschaert)

  • #9169: better documentation for the best-fit allocation policy
    (Gabriel Scherer, review by Guillaume Munch-Maccagnoni
    and Florian Angeletti)

Compiler user-interface and warnings:

  • #8833: Hint for (type) redefinitions in toplevel session
    (Florian Angeletti, review by Gabriel Scherer)

  • #2127, #9185: Refactor lookup functions
    Included observable changes:

    • makes the location of usage warnings and alerts for constructors more
      precise
    • don’t warn about a constructor never being used to build values when it
      has been defined as private
      (Leo White, Hugo Heuzard review by Thomas Refis, Florian Angeletti)
  • #8702, #8777: improved error messages for fixed row polymorphic variants
    (Florian Angeletti, report by Leo White, review by Thomas Refis)

  • #8844: Printing faulty constructors, inline records fields and their types
    during type mismatches. Also slightly changed other type mismatches error
    output.
    (Mekhrubon Turaev, review by Florian Angeletti, Leo White)

  • #8885: Warn about unused local modules
    (Thomas Refis, review by Alain Frisch)

  • #8872: Add ocamlc option “-output-complete-exe” to build a self-contained
    binary for bytecode programs, containing the runtime and C stubs.
    (Stéphane Glondu, Nicolás Ojeda Bär, review by Jérémie Dimino and Daniel
    Bünzli)

  • #8874: add tests for typechecking error messages and pack them into
    pretty-printing boxes.
    (Oxana Kostikova, review by Gabriel Scherer)

  • #8891: Warn about unused functor parameters
    (Thomas Refis, review by Gabriel Radanne)

  • #8903: Improve errors for first-class modules
    (Leo White, review by Jacques Garrigue)

  • #8914: clarify the warning on unboxable types used in external primitives (61)
    (Gabriel Scherer, review by Florian Angeletti, report on the Discourse forum)

  • #9046: disable warning 30 by default
    This outdated warning complained on label/constructor name conflicts
    within a mutually-recursive type declarations; there is now no need
    to complain thanks to type-based disambiguation.
    (Gabriel Scherer)

Tools:

  • [breaking change] #6792, #8654 ocamldebug now supports programs using Dynlink. This
    changes ocamldebug messages, which may break compatibility
    with older emacs modes.
    (Whitequark and Jacques-Henri Jourdan, review by Gabriel Scherer
    and Xavier Clerc)
  • #8621: Make ocamlyacc a Windows Unicode application
    (David Allsopp, review by Nicolás Ojeda Bär)
  • [breaking change] #8834, ocaml: adhere to the XDG base directory specification to
    locate an .ocamlinit file. Reads an $XDG_CONFIG_HOME/ocaml/init.ml
    file before trying to lookup ~/.ocamlinit. On Windows the behaviour
    is unchanged.
    (Daniel C. Bünzli, review by David Allsopp, Armaël Guéneau and
    Nicolás Ojeda Bär)
  • #9113: ocamldoc: fix the rendering of multi-line code blocks
    in the ‘man’ backend.
    (Gabriel Scherer, review by Florian Angeletti)

  • #9127, #9130: ocamldoc: fix the formatting of closing brace in record types.
    (David Allsopp, report by San Vu Ngoc)

  • #9181: make objinfo work on Cygwin and look for the caml_plugin_header
    symbol in both the static and the dynamic symbol tables.
    (Sébastien Hinderer, review by Gabriel Scherer and David Allsopp)

Build system:

  • #8840: use ocaml{c,opt}.opt when available to build internal tools
    On my machine this reduces parallel-build times from 3m30s to 2m50s.
    (Gabriel Scherer, review by Xavier Leroy and Sébastien Hinderer)

  • #8650: ensure that “make” variables are defined before use;
    revise generation of config/util.ml to better quote special characters
    (Xavier Leroy, review by David Allsopp)

  • #8690, #8696: avoid rebuilding the world when files containing primitives
    change.
    (Stephen Dolan, review by Gabriel Scherer, Sébastien Hinderer and
    Thomas Refis)

  • #8835: new configure option --disable-stdlib-manpages to disable building
    and installation of the library manpages.
    (David Allsopp, review by Florian Angeletti and Gabriel Scherer)

  • #8837: build manpages using ocamldoc.opt when available
    cuts the manpages build time from 14s to 4s
    (Gabriel Scherer, review by David Allsopp and Sébastien Hinderer,
    report by David Allsopp)

  • #8843, #8841: fix use of off_t on 32-bit systems.
    (Stephen Dolan, report by Richard Jones, review by Xavier Leroy)

  • #8947, #9134, #9302, #9311: fix/improve support for the BFD library
    (Sébastien Hinderer, review by Damien Doligez and David Allsopp)

  • #8951: let make’s default target build the compiler
    (Sébastien Hinderer, review by David Allsopp)

  • #8995: allow developers to specify frequently-used configure options in
    Git (ocaml.configure option) and a directory for host-specific, shareable
    config.cache files (ocaml.configure-cache option). See HACKING.adoc for
    further details.
    (David Allsopp, review by Gabriel Scherer)

  • #9136: Don’t propagate Cygwin-style prefix from configure to
    Makefile.config on Windows ports.
    (David Allsopp, review by Sébastien Hinderer)

Internal/compiler-libs changes:

  • #8828: Added abstractions for variants, records, constructors, fields and
    extension constructor types mismatch.
    (Mekhrubon Turaev, review by Florian Angeletti, Leo White and Gabriel Scherer)

  • #7927, #8527: Replace long tuples into records in typeclass.ml
    (Ulugbek Abdullaev, review by David Allsopp and Gabriel Scherer)

  • #1963: split cmmgen into generic Cmm helpers and clambda transformations
    (Vincent Laviron, review by Mark Shinwell)

  • #1901: Fix lexing of character literals in comments
    (Pieter Goetschalckx, review by Damien Doligez)

  • #1932: Allow octal escape sequences and identifiers containing apostrophes
    in ocamlyacc actions and comments.
    (Pieter Goetschalckx, review by Damien Doligez)

  • #2288: Move middle end code from [Asmgen] to [Clambda_middle_end] and
    [Flambda_middle_end]. Run [Un_anf] from the middle end, not [Cmmgen].
    (Mark Shinwell, review by Pierre Chambart)

  • #8692: Remove Misc.may_map and similar
    (Leo White, review by Gabriel Scherer and Thomas Refis)

  • #8677: Use unsigned comparisons in amd64 and i386 emitter of Lcondbranch3.
    (Greta Yorsh, review by Xavier Leroy)

  • #8766: Parmatch: introduce a type for simplified pattern heads
    (Gabriel Scherer and Thomas Refis, review by Stephen Dolan and
    Florian Angeletti)

  • #8774: New implementation of Env.make_copy_of_types
    (Alain Frisch, review by Thomas Refis, Leo White and Jacques Garrigue)

  • #7924: Use a variant instead of an int in Bad_variance exception
    (Rian Douglas, review by Gabriel Scherer)

  • #8890: in -dtimings output, show time spent in C linker clearly
    (Valentin Gatien-Baron)

  • #8910, #8911: minor improvements to the printing of module types
    (Gabriel Scherer, review by Florian Angeletti)

  • #8913: ocamltest: improve ‘promote’ implementation to take
    skipped lines/bytes into account
    (Gabriel Scherer, review by Sébastien Hinderer)

  • #8908: Use an option instead of a string for module names ("_" becomes None),
    and a dedicated type for functor parameters: “()” maps to “Unit” (instead of
    “*”).
    (Thomas Refis, review by Gabriel Radanne)

  • #8928: Move contains_calls and num_stack_slots from Proc to Mach.fundecl
    (Greta Yorsh, review by Florian Angeletti and Vincent Laviron)

  • #8959, #8960, #8968, #9023: minor refactorings in the typing of patterns:

    • refactor the {let,pat}_bound_idents* functions
    • minor bugfix in type_pat
    • refactor the generic pattern-traversal functions
      in Typecore and Typedtree
    • restrict the use of Need_backtrack
      (Gabriel Scherer and Florian Angeletti,
      review by Thomas Refis and Gabriel Scherer)
  • #9030: clarify and document the parameter space of type_pat
    (Gabriel Scherer and Florian Angeletti and Jacques Garrigue,
    review by Florian Angeletti and Thomas Refis)

  • #8975: “ocamltests” files are no longer required or used by
    “ocamltest”. Instead, any text file in the testsuite directory containing a
    valid “TEST” block will be automatically included in the testsuite.
    (Nicolás Ojeda Bär, review by Gabriel Scherer and Sébastien Hinderer)

  • #8992: share argument implementations between executables
    (Florian Angeletti, review by Gabriel Scherer)

  • #9015: fix fatal error in pprint_ast (#8789)
    (Damien Doligez, review by …)

Bug fixes:

  • #5673, #7636: unused type variable causes generalization error
    (Jacques Garrigue and Leo White, review by Leo White,
    reports by Jean-Louis Giavitto and Christophe Raffalli)

  • #6922, #8955: Fix regression with -principal type inference for inherited
    methods, allowing to compile ocamldoc with -principal
    (Jacques Garrigue, review by Leo White)

  • #7925, #8611: fix error highlighting for exceptionally
    long toplevel phrases
    (Kyle Miller, reported by Armaël Guéneau, review by Armaël Guéneau
    and Nicolás Ojeda Bär)

  • #8622: Don’t generate #! headers over 127 characters.
    (David Allsopp, review by Xavier Leroy and Stephen Dolan)

  • #8715: minor bugfixes in CamlinternalFormat; removes the unused
    and misleading function CamlinternalFormat.string_of_formatting_gen
    (Gabriel Scherer and Florian Angeletti,
    review by Florian Angeletti and Gabriel Radanne)

  • #8792, #9018: Possible (latent) bug in Ctype.normalize_type
    removed incrimined Btype.log_type, replaced by Btype.set_type
    (Jacques Garrigue, report by Alain Frisch, review by Thomas Refis)

  • #8856, #8860: avoid stackoverflow when printing cyclic type expressions
    in some error submessages.
    (Florian Angeletti, report by Mekhrubon Turaev, review by Leo White)

  • #8875: fix missing newlines in the output from MSVC invocation.
    (Nicolás Ojeda Bär, review by Gabriel Scherer)

  • #8921, #8924: Fix stack overflow with Flambda
    (Vincent Laviron, review by Pierre Chambart and Leo White,
    report by Aleksandr Kuzmenko)

  • #8892, #8895: fix the definition of Is_young when CAML_INTERNALS is not
    defined.
    (David Allsopp, review by Xavier Leroy)

  • #8896: deprecate addr typedef in misc.h
    (David Allsopp, suggestion by Xavier Leroy)

  • #8981: Fix check for incompatible -c and -o options.
    (Greta Yorsh, review by Damien Doligez)

  • #9019, #9154: Unsound exhaustivity of GADTs from incomplete unification
    Also fixes bug found by Thomas Refis in #9012
    (Jacques Garrigue, report and review by Leo White, Thomas Refis)

  • #9031: Unregister Windows stack overflow handler while shutting
    the runtime down.
    (Dmitry Bely, review by David Allsopp)

  • #9051: fix unregistered local root in win32unix/select.c (could result in
    select returning file_descr-like values which weren’t in the original sets)
    and correct initialisation of some blocks allocated with caml_alloc_small.
    (David Allsopp, review by Xavier Leroy)

  • #9073, #9120: fix incorrect GC ratio multiplier when allocating custom blocks
    with caml_alloc_custom_mem in runtime/custom.c
    (Markus Mottl, review by Gabriel Scherer and Damien Doligez)

  • #9209, #9212: fix a development-version regression caused by #2288
    (Kate Deplaix and David Allsopp, review by Sébastien Hinderer
    and Gabriel Scherer )

  • #9218, #9269: avoid a rare wrong module name error with “-annot” and
    inline records.
    (Florian Angeletti, review by Gabriel Scherer, report by Kate Deplaix)

  • #9261: Fix a soundness bug in Rec_check, new in 4.10 (from #8908)
    (Vincent Laviron, review by Jeremy Yallop and Gabriel Scherer)

29 Likes

Thanks to everyone who worked hard on this release. In particular, @octachron did all the release management, and @kit-ty-kate did tremendous work on the monitoring of OPAM packages and the health of the OPAM ecosystem on 4.10.

Speaking as a compiler maintainer, we never get as much testing of release-candidates as we would like, but this is one of the better-prepared releases we have had in a couple years, in large part thanks to the work of Florian and Kate.

And it was not particularly easy this time: we are accelerating the pace of changes ot the runtime, with preliminary work for Multicore and for statmemprof; many low-level C stuff in the ecosystem needed careful compatibility studies (no-thanks for silently relying on private runtime internals!), and we can easily predict that this is going to keep happening in the next few releases.

14 Likes

Indeed, many thanks to everyone who leapt in to make 4.10 ready in opam in such record time! Just a note that the CI Docker images are now also rebuilt for x86_64, arm32/64 and ppc64le to reflect the 4.10 release, so feel free to start using them. https://hub.docker.com/r/ocaml/opam2/tags

8 Likes

The Merlin team has just released a preview version of Merlin which is compatible with 4.10.0
(Merlin is an editor service that provides modern IDE features for OCaml) .

This is a preview version:

  • the support for short-path is disabled
  • only OCaml 4.10.0 is supported in this preview

It can be installed via opam with the usual

 $ opam install merlin
12 Likes

The link to the release page is broken: https://ocaml.org/releases/4.10.0.html

That URL is linked from https://ocaml.org/docs/install.html and https://ocaml.org/releases/.

There is unmerged PR fixing this:https://github.com/ocaml/ocaml.org/pull/1106

There are also a couple minor PRs that should be merged too:

Seems also the certificate of https://ocaml.org was expired at March, 6

The link https://ocaml.org/releases/4.10.0.html is fixed now.

3 Likes

I’m happy to report that we are using this new allocator in (some of) our systems with good results.

With the space overhead set so that the overall runtime is steady with the next-fit allocator of 4.09, we see around 25% reduction in peak memory consumption. In the other direction, setting overhead to hold peak memory consumption steady, we see around 5% overall CPU time speedups. (This is using 4.11.1, and for full info this system does a significant amount of Marshaling and when using the next-fit allocator is slightly slower with 4.11 than 4.09.)

These are wonderful results, thank you very very much for this work @damiendoligez and team!

In case it is interesting, runtime holds steady with an overhead value around 120, that is, higher than the manual’s suggested 100. And more interestingly, peak memory usage holds steady with a significantly higher overhead value of around 250. So, as always, it can be helpful for everyone to benchmark their systems with a spectrum of settings.

5 Likes

And see @stedolan’s much more thorough blog post.