The safe-for-space pass could add clearing operations on "typed"
stack positions, which are known to contain a fixnum, flonum, or
extflonum. Non-clearing references, however, were not annotated to
indicate that clearing references were present, since clearing is
not expected on typed positions.
Along the lines of not expecting clearing, the bytecode validator's
encoding of the stack doesn't accomodate both "has a type" and "claims
never to be cleared", so it couldn't detect the bytecode compiler
bug. (Also, this problem didn't show up in the HOSC paper's model of
the validator, because the model pre-dates type tracking.)
Fix the bytecode compiler's space-safety pass so that it never inserts
clearing operations for typed stack positions. Then, the validator can
simply reject any attempt to clear a typed position.
Also, annotate applications generated by lifting as safe-for-space
tail calls.
Merge to v5.3.4
(cherry picked from commit 7ad1ddab64)
Conflicts:
src/racket/src/cstartup.inc
src/racket/src/schvers.h
Bug introduced by 7a8c2ff063: a tree can be deep enough that the
representation of the path to the current item can be too big to
fit into 32 bits. It will always fit in 64 bits, though.
Merge to 5.3.4
(cherry picked from commit f09d1d1ae8)
Conflicts:
src/racket/src/string.c
With this patch racket will never grow beyond of the defaults limits of
the OS and also it doesn't limit the stack size to a fixed value.
Merge to v5.3.4
(cherry picked from commit 5a566771e0)
The validator was not as smart as the compiler in determining
that a `let' expression could be relied on to produce a
constant-shaped function (without side effect or error) in the case
that a right-hand side expression is a call to a known structure
constructor or predicate.
Closes PR 13679
Merge to v5.3.4
(cherry picked from commit e49956e3ea)
Fix ephemeral-port support in `udp-bind!', and
change `udp-bind!' and `udp-connect!' to try address
resolutions in order to find one that works (which is
typically needed to auto-select an IPv4 or IPv6 variant
of an address).
(cherry picked from commit 842da32e48)
On OpenBSD, for example, the libffi link test doesn't work until
"-pthread" is in $LIBS.
Also, make `configure' complain if `--enable-libffi' is specified
and linking to libffi fails.
Allow a thread to be GCed when it is blocked on a place
channel for reading and the place channel's write end
is inaccessible.
GC is limited to threads that do not participate in cycles
of such threads, where the otherwise unerachable threads
are blocked on place channels that are reachable among the
set of threads. In other words, the GC finds the greatest
fix point (as measured by the threads to retain) instead of
least fix point --- which isn't what you want, but finding
the least fix point seems to require significant extra GC
machinery across places.
This improvement was intended to solve the same problem as
commit 7b0608c, but that case seems to run into the limitation
on cycles.
A tail call with certain kinds of primitives would fail to
clear local bindings in a detectable way. For example, a
tail call to `sync' that blocks could retain references
to unreachable data in the context of the `sync' call.
Primitives that can cause problems in the run-time system
are already identified as "imemdiate" primitives. The
safe-for-space pass now inserts clearing actions before a tail
call, unless the call it to a known immediate primitive or a
Racket-implemented function.
Clearing operations are now omitted before non-tail calls
to immediate operations like structure predicates.
The newly added clearing operations could affect performance,
but they probably won't, since the clear operations are still
avoided in tail-call cases that are otherwise fast. The newly
omitted clearing operations may improve performance.
The shortcut could be triggered in a bad case (first
argument as `#f' in non-timeout mode) and returned the
wrong result (void instead of the semaphore).
Enable extflonums in a MSVC build by relying on a MinGW-built DLL,
"longdouble.dll". The DLL is loaded on startup. If the DLL isn't
available, then `extflonum-available?' reports #f.
Instead of setting the floating-point mode globally to extended
precision, the mode is changed only just before (and restored right
after) extflonum arithmetic operations.
Using an impersonator property to represent an application mark was
a bad choice. The property gets propagated, so it is also on any later
chaperone layer, and then things go bad: the docs say that special
treatment is triggered by supplying an argument to `impersonate-property',
but it was actually triggered by the chaperone having the property.
Change the implementation to match the documentation. Using an
impersonator property to supply the mark should be regarded as a hack,
but now the implementaiton is at least consistent with the documentaiton.
"-rdynamic" doesn't work with GCC 4.7 on OpenBSD: "cc: error:
unrecognized command line option '-rdynamic'". The switch isn't
necessary because it is only a wrapper to "-Wl,--export-dynamic".
Look GCC bug 37454.
