Add a new kind of write barrier, one which has a bit per field; the
mutator that sets the bit will need to add the field's location (the
edge) to a remembered set. Here we just have the fast-path
implementation.
We use Treiber stacks to represent sets of blocks: blocks to sweep, full
blocks, and so on. This is fine as long as we are only adding to or
only removing from those sets, but as soon as we have concurrent add and
remove, we need to avoid the ABA problem.
Concurrent add and remove occurs for partly-full blocks, which are both
acquired and released by mutators; empty blocks, which can be added to
by heap growth at the same time as the mutator acquires them; and the
paged-out queue, which is also concurrent with heap growth/shrinkage.
The tests depend on libguile/guile-procedures.txt, for example via
documented? in bit-operations.test. Previously "make check -j..." in a
clean tree would fail because libguile/guile-procedures.txt is built by
./Makefile.am (rather than libguile/Makefile.am) so that it will have a
built module/ available, but when "." is not listed in SUBDIRS, it
builds last, and so the test-suite runs before guile-procedures.txt is
built.
To fix the problem add "." to SUBDIRS before the test-suite so that the
tests will be able depend on everything else, and move the existing
guile-procedures.txt target into libguile/ next to its
guile-procedures.texi dependency. That gives a better overview and
simplifies the recipe a bit. It also allows us to drop the explict
"all-local:" dependency, and to let the existing libguile/ code handle
the cleanup.
* Makefile.am (SUBDIRS): add . just before the test-suite.
(libguile/guile-procedures.txt): rely on libguile/Makefile.am.
(CLEANFILES): Drop libguile/procedures.txt.
* libguile/Makefile.am: (all-local): drop.
(libguile/guile-procedures.txt): move Makefile.am recipe here.
* module/language/cps/specialize-numbers.scm (sigbits-ref): New helper.
(invert-graph*): New helper.
(compute-significant-bits): When visiting a term changes computed
needed-bits for one of its definitions, we need to revisit the variables
that contributed to its result (the uses), because they might need more
bits as well. Previously we were doing this by enqueueing predecessors
to the term, which worked if the uses were defined in predecessors, or
if all defining terms were already in the worklist, which is the case
without loops. But with loops, when revisiting a term, you could see
that it causes sigbits to change, enqueue its predecessors, but then the
predecessors don't change anything and the fixpoint stops before
reaching the definitions of the variables we need. So instead we
compute the use-def graph and enqueue defs directly.
* module/language/cps/specialize-numbers.scm (next-power-of-two): Use
integer-length. No change.
(compute-significant-bits): Fix the fixpoint computation, which was
failing to complete in some cases with loops.
* module/language/cps/specialize-numbers.scm (specialize-operations):
Accept any operand to logand/immediate, provided the result is a u64 in
the right range.
* module/language/cps/types.scm
(ulogand, ulogand/immediate, ulogsub, ulogior, ulogxor): Where we have
u64 inputs, there's no need to `restrict!`; the range will come from the
definition.
* module/language/tree-il/peval.scm (peval)
(inlinable-kwargs-bug-fixup): Before 3.0.10, the inlinable exports pass
was incorrectly serializing functions with keyword arguments. This was
fixed in 2c645571b3, but that meant that
3.0.10 compiling against 3.0.9 binaries could raise an exception at
compile-time; whoops. Add a workaround so that 3.0.9 binaries still
work.
Fixes https://issues.guix.gnu.org/72936.
PCC with GC_PARALLEL=0 is exactly equivalent to SCC. Also now that PCC
will dynamically avoid atomic forwarding if parallelism is disabled at
run-time, there is no need to keep SCC around.
In GC, request mutators to stop before doing anything else; changes the
order of the event listener interface. Also, refactor mmc to look more
like pcc.
