bug#39315: Too aggressive in lining

bug#39315: Too aggressive in lining

[Stefan Israelsson Tampe]

[-- Attachment #1: Type: text/plain, Size: 19438 bytes --]

---------- Forwarded message ---------
From: Stefan Israelsson Tampe <stefan.itampe@gmail.com>
Date: Mon, Jan 27, 2020 at 8:38 PM
Subject: Re: GNU Guile 3.0.0 released
To: guile-devel <guile-devel@gnu.org>, Andy Wingo <wingo@pobox.com>


This does not look right (for guile 3.0.0)

---------------------------------
;b.scm:
(define-module (b)
  #:export (%f g))

(define %f 1)

(define (g)
  (pk %f))

---------------------------------
;c:scm:
(define-module (c)
  #:use-module (b)
  #:export (f))

(define (f x)
  (set! %f x)
  (g))
--------------------------------
Interaction:
scheme@(guile-user)> (use-modules (c))
scheme@(guile-user)> (f 2)

;;; (1)
$1 = 1







On Thu, Jan 16, 2020 at 12:04 PM Andy Wingo <wingo@pobox.com> wrote:

> We are delighted to announce GNU Guile release 3.0.0, the first in the
> new 3.0 stable release series.
>
> Compared to the previous stable series (2.2.x), Guile 3.0 adds support
> for just-in-time native code generation, speeding up all Guile programs.
> See the NEWS extract at the end of the mail for full details.
>
>
> The Guile web page is located at http://gnu.org/software/guile/, and
> among other things, it contains a copy of the Guile manual and pointers
> to more resources.
>
> Guile is an implementation of the Scheme programming language, packaged
> for use in a wide variety of environments.  In addition to implementing
> the R5RS, R6RS, and R7RS Scheme standards, Guile includes full access to
> POSIX system calls, networking support, multiple threads, dynamic
> linking, a foreign function call interface, powerful string processing,
> and HTTP client and server implementations.
>
> Guile can run interactively, as a script interpreter, and as a Scheme
> compiler to VM bytecode.  It is also packaged as a library so that
> applications can easily incorporate a complete Scheme interpreter/VM.
> An application can use Guile as an extension language, a clean and
> powerful configuration language, or as multi-purpose "glue" to connect
> primitives provided by the application.  It is easy to call Scheme code
> from C code and vice versa.  Applications can add new functions, data
> types, control structures, and even syntax to Guile, to create a
> domain-specific language tailored to the task at hand.
>
> Guile 3.0.0 can be installed in parallel with Guile 2.2.x; see
>
> http://www.gnu.org/software/guile/manual/html_node/Parallel-Installations.html
> .
>
> A more detailed NEWS summary follows these details on how to get the
> Guile sources.
>
> Here are the compressed sources:
>   http://ftp.gnu.org/gnu/guile/guile-3.0.0.tar.lz   (10MB)
>   http://ftp.gnu.org/gnu/guile/guile-3.0.0.tar.xz   (12MB)
>   http://ftp.gnu.org/gnu/guile/guile-3.0.0.tar.gz   (21MB)
>
> Here are the GPG detached signatures[*]:
>   http://ftp.gnu.org/gnu/guile/guile-3.0.0.tar.lz.sig
>   http://ftp.gnu.org/gnu/guile/guile-3.0.0.tar.xz.sig
>   http://ftp.gnu.org/gnu/guile/guile-3.0.0.tar.gz.sig
>
> Use a mirror for higher download bandwidth:
>   http://www.gnu.org/order/ftp.html
>
> Here are the SHA256 checksums:
>
>   e28c450d11f7335769f607214f9b79547400881ddbbc9805ccf3ce2121aa97e0
> guile-3.0.0.tar.lz
>   c9138d6595a9f69bf9733d0bc2d3b9f3d8b79f35f289006912b3361cb0510c75
> guile-3.0.0.tar.xz
>   049b286849fa9764fac781071c4ec9daef707da51e5050ffb498d7bf6422da2f
> guile-3.0.0.tar.gz
>
> [*] Use a .sig file to verify that the corresponding file (without the
> .sig suffix) is intact.  First, be sure to download both the .sig file
> and the corresponding tarball.  Then, run a command like this:
>
>   gpg --verify guile-3.0.0.tar.gz.sig
>
> If that command fails because you don't have the required public key,
> then run this command to import it:
>
>   gpg --keyserver keys.gnupg.net --recv-keys
> 4FD4D288D445934E0A14F9A5A8803732E4436885
>
> and rerun the 'gpg --verify' command.
>
> This release was bootstrapped with the following tools:
>   Autoconf 2.69
>   Automake 1.16.1
>   Libtool 2.4.6
>   Gnulib v0.1-1157-gb03f418
>   Makeinfo 6.7
>
> An extract from NEWS follows.
>
>
> Changes in 3.0.0 (since the stable 2.2 series):
>
> * Notable changes
>
> ** Just-in-time code generation
>
> Guile programs now run up to 4 times faster, relative to Guile 2.2,
> thanks to just-in-time (JIT) native code generation.  Notably, this
> brings the performance of "eval" as written in Scheme back to the level
> of "eval" written in C, as in the days of Guile 1.8.
>
> See "Just-In-Time Native Code" in the manual, for more information.  JIT
> compilation will be enabled automatically and transparently.  To disable
> JIT compilation, configure Guile with `--enable-jit=no' or
> `--disable-jit'.  The default is `--enable-jit=auto', which enables the
> JIT if it is available.  See `./configure --help' for more.
>
> JIT compilation is enabled by default on x86-64, i686, ARMv7, and
> AArch64 targets.
>
> ** Lower-level bytecode
>
> Relative to the virtual machine in Guile 2.2, Guile's VM instruction set
> is now more low-level.  This allows it to express more advanced
> optimizations, for example type check elision or integer
> devirtualization, and makes the task of JIT code generation easier.
>
> Note that this change can mean that for a given function, the
> corresponding number of instructions in Guile 3.0 may be higher than
> Guile 2.2, which can lead to slowdowns when the function is interpreted.
> We hope that JIT compilation more than makes up for this slight
> slowdown.
>
> ** Interleaved internal definitions and expressions allowed
>
> It used to be that internal definitions had to precede all expressions
> in their bodies.  This restriction has been relaxed.  If an expression
> precedes an internal definition, it is treated as if it were a
> definition of an unreferenced variable.  For example, the expression
> `(foo)' transforms to the equivalent of `(define _ (begin (foo) #f))',
> if it precedes other definitions.
>
> This change improves the readability of Guile programs, as it used to be
> that program indentation tended to increase needlessly to allow nested
> `let' and `letrec' to re-establish definition contexts after initial
> expressions, for example for type-checks on procedure arguments.
>
> ** Record unification
>
> Guile used to have a number of implementations of structured data types
> in the form of "records": a core facility, SRFI-9 (records), SRFI-35
> (condition types -- a form of records) and R6RS records.  These
> facilities were not compatible, as they all were built in different
> ways.  This had the unfortunate corollary that SRFI-35 conditions were
> not compatible with R6RS conditions.  To fix this problem, we have now
> added the union of functionality from all of these record types into
> core records: single-inheritance subtyping, mutable and immutable
> fields, and so on.  See "Records" in the manual, for full details.
>
> R6RS records, SRFI-9 records, and the SRFI-35 and R6RS exception types
> have been accordingly "rebased" on top of core records.
>
> ** Reimplementation of exceptions
>
> Since Guile's origins 25 years ago, `throw' and `catch' have been the
> primary exception-handling primitives.  However these primitives have
> two problems.  One is that it's hard to handle exceptions in a
> structured way using `catch'.  Few people remember what the
> corresponding `key' and `args' are that an exception handler would see
> in response to a call to `error', for example.  In practice, this
> results in more generic catch-all exception handling than one might
> like.
>
> The other problem is that `throw', `catch', and especially
> `with-throw-handler' are quite unlike what the rest of the Scheme world
> uses.  R6RS and R7RS, for example, have mostly converged on
> SRFI-34-style `with-exception-handler' and `raise' primitives, and
> encourage the use of SRFI-35-style structured exception objects to
> describe the error.  Guile's R6RS layer incorporates an adapter between
> `throw'/`catch' and structured exception handling, but it didn't apply
> to SRFI-34/SRFI-35, and we would have to duplicate it for R7RS.
>
> In light of these considerations, Guile has now changed to make
> `with-exception-handler' and `raise-exception' its primitives for
> exception handling and defined a hierarchy of R6RS-style exception types
> in its core.  SRFI-34/35, R6RS, and the exception-handling components of
> SRFI-18 (threads) have been re-implemented in terms of this core
> functionality.  There is also a a compatibility layer that makes it so
> that exceptions originating in `throw' can be handled by
> `with-exception-hander', and vice-versa for `raise-exception' and
> `catch'.
>
> Generally speaking, users will see no difference.  The one significant
> difference is that users of SRFI-34 will see more exceptions flowing
> through their `with-exception-handler'/`guard' forms, because whereas
> before they would only see exceptions thrown by SRFI-34, now they will
> see exceptions thrown by R6RS, R7RS, or indeed `throw'.
>
> Guile's situation is transitional.  Most exceptions are still signalled
> via `throw'.  These will probably migrate over time to
> `raise-exception', while preserving compatibility of course.
>
> See "Exceptions" in the manual, for full details on the new API.
>
> ** `guard' no longer unwinds the stack for clause tests
>
> SRFI-34, and then R6RS and R7RS, defines a `guard' form that is a
> shorthand for `with-exception-handler'.  The cond-like clauses for the
> exception handling are specified to run with the continuation of the
> `guard', while any re-propagation of the exception happens with the
> continuation of the original `raise'.
>
> In practice, this means that one needs full `call-with-continuation' to
> implement the specified semantics, to be able to unwind the stack to the
> cond clauses, then rewind if none match.  This is not only quite
> expensive, it is also error-prone as one usually doesn't want to rewind
> dynamic-wind guards in an exceptional situation.  Additionally, as
> continuations bind tightly to the current thread, it makes it impossible
> to migrate a subcomputation with a different thread if a `guard' is live
> on the stack, as is done in Fibers.
>
> Guile now works around these issues by running the test portion of the
> guard expressions within the original `raise' continuation, and only
> unwinding once a test matches.  This is an incompatible semantic change
> but we think the situation is globally much better, and we expect that
> very few people will be affected by the change.
>
> ** Optimization of top-level bindings within a compilation unit
>
> At optimization level 2 and above, Guile's compiler is now allowed to
> inline top-level definitions within a compilation unit.  See
> "Declarative Modules" in the manual, for full details.  This change can
> improve the performance of programs with many small top-level
> definitions by quite a bit!
>
> At optimization level 3 and above, Guile will assume that any top-level
> binding in a declarative compilation unit that isn't exported from a
> module can be completely inlined into its uses.  (Prior to this change,
> -O3 was the same as -O2.)  Note that with this new
> `seal-private-bindings' pass, private declarative bindings are no longer
> available for access from the first-class module reflection API.  The
> optimizations afforded by this pass can be useful when you need a speed
> boost, but having them enabled at optimization level 3 means they are
> not on by default, as they change Guile's behavior in ways that users
> might not expect.
>
> ** By default, GOOPS classes are not redefinable
>
> It used to be that all GOOPS classes were redefinable, at least in
> theory.  This facility was supported by an indirection in all "struct"
> instances, even though only a subset of structs would need redefinition.
> We wanted to remove this indirection, in order to speed up Guile
> records, allow immutable Guile records to eventually be described by
> classes, and allow for some optimizations in core GOOPS classes that
> shouldn't be redefined anyway.
>
> Thus in GOOPS now there are classes that are redefinable and classes
> that aren't.  By default, classes created with GOOPS are not
> redefinable.  To make a class redefinable, it should be an instance of
> `<redefinable-class>'.  See "Redefining a Class" in the manual for more
> information.
>
> ** Define top-level bindings for aux syntax: `else', `=>', `...', `_'
>
> These auxiliary syntax definitions are specified to be defined in the
> R6RS and the R7RS.  They were previously unbound, even in the R6RS
> modules.  This change is not anticipated to cause any incompatibility
> with existing Guile code, and improves things for R6RS and R7RS users.
>
> ** Conventional gettext alias is now `G_'
>
> Related to the last point, since the "Fix literal matching for
> module-bound literals" change in the 2.2 series, it was no longer
> possible to use the conventional `_' binding as an alias for `gettext',
> because a local `_' definition would prevent `_' from being recognized
> as auxiliary syntax for `match', `syntax-rules', and similar.  The new
> recommended conventional alias for `gettext' is `G_'.
>
> ** Add --r6rs command-line option
>
> The new `install-r6rs!' procedure adapts Guile's defaults to be more
> R6RS-compatible.  This procedure is called if the user passes `--r6rs'
> as a command-line argument.  See "R6RS Incompatibilities" in the manual,
> for full details.
>
> ** Add support for R7RS
>
> Thanks to Göran Weinholt and OKUMURA Yuki, Guile now implements the R7RS
> modules.  As the R7RS library syntax is a subset of R6RS, to use R7RS
> you just `(import (scheme base))' and off you go.  As with R6RS also,
> there are some small lexical incompatibilities regarding hex escapes;
> see "R6RS Support" in the manual, for full details.
>
> Also as with R6RS, there is an `install-r7rs!' procedure and a `--r7rs'
> command-line option.
>
> ** Add #:re-export-and-replace argument to `define-module'
>
> This new keyword specifies a set of bindings to re-export, but also
> marks them as intended to replace core bindings.  See "Creating Guile
> Modules" in the manual, for full details.
>
> Note to make this change, we had to change the way replacement flags are
> stored, to being associated with modules instead of individual variable
> objects.  This means that users who #:re-export an imported binding that
> was already marked as #:replace by another module will now see warnings,
> as they need to use #:re-export-and-replace instead.
>
> ** `define-module' #:autoload no longer pulls in the whole module
>
> One of the ways that a module can use another is "autoloads".  For
> example:
>
>   (define-module (a) #:autoload (b) (make-b))
>
> In this example, module `(b)' will only be imported when the `make-b'
> identifier is referenced.  However besides the imprecision about when a
> given binding is actually referenced, this mechanism used to cause the
> whole imported module to become available, not just the specified
> bindings.  This has now been changed to only import the specified bindings.
>
> This is a backward-incompatible change.  The fix is to mention all
> bindings of interest in the autoload clause.  Feedback is welcome.
>
> ** Improve SRFI-43 vector-fill!
>
> SRFI-43 vector-fill! now has the same performance whether an optional
> range is provided or not, and is also provided in core.  As a side
> effect, vector-fill! and vector_fill_x no longer work on non-vector
> rank-1 arrays.  Such cases were handled incorrectly before; for example,
> prior to this change:
>
>   (define a (make-vector 10 'x))
>   (define b (make-shared-array a (lambda (i) (list (* 2 i))) 5))
>   (vector-fill! b 'y)
>
>   => #1(y y y x x)
>
> This is now an error.  Instead, use array-fill!.
>
> ** `iota' in core and SRFI-1 `iota' are the same
>
> Previously, `iota' in core would not accept start and step arguments and
> would return an empty list for negative count. Now there is only one
> `iota' function with the extended semantics of SRFI-1.  Note that as an
> incompatible change, core `iota' no longer accepts a negative count.
>
> ** Improved Transport Layer Security (TLS) support in (web client)
>
> `http-request', `http-get', and related procedures from (web client) are
> able to access content over TLS ("HTTPS") since Guile 2.2.  However,
> that support lacked important facilities, which are now available.
>
> First, these procedures now have a #:verify-certificates?  parameter to
> enable or disable the verification of X.509 server certificates.  The
> new `x509-certificate-directory' SRFI-39 parameter specifies X.509
> certificates are searched for.  Second, HTTPS proxies are now supported
> (in addition to HTTP proxies) and the new `current-https-proxy'
> parameter controls that.  See "Web Client" in the manual for details.
>
> * New deprecations
>
> ** scm_t_uint8, etc deprecated in favor of C99 stdint.h
>
> It used to be that Guile defined its own `scm_t_uint8' because C99
> `uint8_t' wasn't widely enough available.  Now Guile finally made the
> change to use C99 types, both internally and in Guile's public headers.
>
> Note that this also applies to SCM_T_UINT8_MAX, SCM_T_INT8_MIN, for intN
> and uintN for N in 8, 16, 32, and 64.  Guile also now uses ptrdiff_t
> instead of scm_t_ptrdiff, and similarly for intmax_t, uintmax_t,
> intptr_t, and uintptr_t.
>
> ** The two-argument form of `record-constructor'
>
> Calling `record-constructor' with two arguments (the record type and a
> list of field names) is deprecated.  Instead, call with just one
> argument, and provide a wrapper around that constructor if needed.
>
> * Incompatible changes
>
> ** All deprecated code removed
>
> All code deprecated in Guile 2.2 has been removed.  See older NEWS, and
> check that your programs can compile without linker warnings and run
> without runtime warnings.  See "Deprecation" in the manual.
>
> In particular, the function `scm_generalized_vector_get_handle' which
> was deprecated in 2.0.9 but remained in 2.2, has now finally been
> removed. As a replacement, use `scm_array_get_handle' to get a handle
> and `scm_array_handle_rank' to check the rank.
>
> ** Remove "self" field from vtables and "redefined" field from classes
>
> These fields were used as part of the machinery for class redefinition
> and is no longer needed.
>
> ** VM hook manipulation simplified
>
> The low-level mechanism to instrument a running virtual machine for
> debugging and tracing has been simplified.  See "VM Hooks" in the
> manual, for more.
>
> * Changes to the distribution
>
> ** New effective version
>
> The "effective version" of Guile is now 3.0, which allows parallel
> installation with other effective versions (for example, the older Guile
> 2.2).  See "Parallel Installations" in the manual for full details.
> Notably, the `pkg-config' file is now `guile-3.0', and there are new
> `guile-3' and `guile-3.0' features for `cond-expand'.
>
>

[-- Attachment #2: Type: text/html, Size: 22951 bytes --]