| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| ... | |
| |
|
|
|
|
| |
Sets `MiscFlags.disableDelayedOsMemoryReturn`.
See the added `Note [MADV_FREE and MADV_DONTNEED]` for details.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a part of GHC Proposal #25: "Offer more array resizing primitives".
Resources related to the proposal:
- Discussion: https://github.com/ghc-proposals/ghc-proposals/pull/121
- Proposal: https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0025-resize-boxed.rst
Only shrinkSmallMutableArray# is implemented as a primop since a
library-space implementation of resizeSmallMutableArray# (in GHC.Exts)
is no less efficient than a primop would be. This may be replaced by
a primop in the future if someone devises a strategy for growing
arrays in-place. The library-space implementation always copies the
array when growing it.
This commit also tweaks the documentation of the deprecated
sizeofMutableByteArray#, removing the mention of concurrency. That
primop is unsound even in single-threaded applications. Additionally,
the non-negativity assertion on the existing shrinkMutableByteArray#
primop has been removed since this predicate is trivially always true.
|
| |\
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
This introduces a concurrent mark & sweep garbage collector to manage the old
generation. The concurrent nature of this collector typically results in
significantly reduced maximum and mean pause times in applications with large
working sets.
Due to the large and intricate nature of the change I have opted to
preserve the fully-buildable history, including merge commits, which is
described in the "Branch overview" section below.
Collector design
================
The full design of the collector implemented here is described in detail
in a technical note
> B. Gamari. "A Concurrent Garbage Collector For the Glasgow Haskell
> Compiler" (2018)
This document can be requested from @bgamari.
The basic heap structure used in this design is heavily inspired by
> K. Ueno & A. Ohori. "A fully concurrent garbage collector for
> functional programs on multicore processors." /ACM SIGPLAN Notices/
> Vol. 51. No. 9 (presented at ICFP 2016)
This design is intended to allow both marking and sweeping
concurrent to execution of a multi-core mutator. Unlike the Ueno design,
which requires no global synchronization pauses, the collector
introduced here requires a stop-the-world pause at the beginning and end
of the mark phase.
To avoid heap fragmentation, the allocator consists of a number of
fixed-size /sub-allocators/. Each of these sub-allocators allocators into
its own set of /segments/, themselves allocated from the block
allocator. Each segment is broken into a set of fixed-size allocation
blocks (which back allocations) in addition to a bitmap (used to track
the liveness of blocks) and some additional metadata (used also used
to track liveness).
This heap structure enables collection via mark-and-sweep, which can be
performed concurrently via a snapshot-at-the-beginning scheme (although
concurrent collection is not implemented in this patch).
Implementation structure
========================
The majority of the collector is implemented in a handful of files:
* `rts/Nonmoving.c` is the heart of the beast. It implements the entry-point
to the nonmoving collector (`nonmoving_collect`), as well as the allocator
(`nonmoving_allocate`) and a number of utilities for manipulating the heap.
* `rts/NonmovingMark.c` implements the mark queue functionality, update
remembered set, and mark loop.
* `rts/NonmovingSweep.c` implements the sweep loop.
* `rts/NonmovingScav.c` implements the logic necessary to scavenge the
nonmoving heap.
Branch overview
===============
```
* wip/gc/opt-pause:
| A variety of small optimisations to further reduce pause times.
|
* wip/gc/compact-nfdata:
| Introduce support for compact regions into the non-moving
|\ collector
| \
| \
| | * wip/gc/segment-header-to-bdescr:
| | | Another optimization that we are considering, pushing
| | | some segment metadata into the segment descriptor for
| | | the sake of locality during mark
| | |
| * | wip/gc/shortcutting:
| | | Support for indirection shortcutting and the selector optimization
| | | in the non-moving heap.
| | |
* | | wip/gc/docs:
| |/ Work on implementation documentation.
| /
|/
* wip/gc/everything:
| A roll-up of everything below.
|\
| \
| |\
| | \
| | * wip/gc/optimize:
| | | A variety of optimizations, primarily to the mark loop.
| | | Some of these are microoptimizations but a few are quite
| | | significant. In particular, the prefetch patches have
| | | produced a nontrivial improvement in mark performance.
| | |
| | * wip/gc/aging:
| | | Enable support for aging in major collections.
| | |
| * | wip/gc/test:
| | | Fix up the testsuite to more or less pass.
| | |
* | | wip/gc/instrumentation:
| | | A variety of runtime instrumentation including statistics
| | / support, the nonmoving census, and eventlog support.
| |/
| /
|/
* wip/gc/nonmoving-concurrent:
| The concurrent write barriers.
|
* wip/gc/nonmoving-nonconcurrent:
| The nonmoving collector without the write barriers necessary
| for concurrent collection.
|
* wip/gc/preparation:
| A merge of the various preparatory patches that aren't directly
| implementing the GC.
|
|
* GHC HEAD
.
.
.
```
|
| | |\ \
| | | |
| | | |
| | | | |
wip/gc/everything2
|
| | | | | |
|
| | | | | |
|
| | | |/ |
|
| | |/ |
|
| | |\ \ |
|
| | | |/ |
|
| | | | |
|
| | |/
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
This introduces a few events to mark key points in the nonmoving
garbage collection cycle. These include:
* `EVENT_CONC_MARK_BEGIN`, denoting the beginning of a round of
marking. This may happen more than once in a single major collection
since we the major collector iterates until it hits a fixed point.
* `EVENT_CONC_MARK_END`, denoting the end of a round of marking.
* `EVENT_CONC_SYNC_BEGIN`, denoting the beginning of the post-mark
synchronization phase
* `EVENT_CONC_UPD_REM_SET_FLUSH`, indicating that a capability has
flushed its update remembered set.
* `EVENT_CONC_SYNC_END`, denoting that all mutators have flushed their
update remembered sets.
* `EVENT_CONC_SWEEP_BEGIN`, denoting the beginning of the sweep portion
of the major collection.
* `EVENT_CONC_SWEEP_END`, denoting the end of the sweep portion of the
major collection.
|
| | |
| |
| |
| |
| |
| | |
This required some fiddling around with the location of forward
declarations since the C sources generated by GHC's C backend only
includes Stg.h.
|
| | | |
|
| | | |
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
This extends the non-moving collector to allow concurrent collection.
The full design of the collector implemented here is described in detail
in a technical note
B. Gamari. "A Concurrent Garbage Collector For the Glasgow Haskell
Compiler" (2018)
This extension involves the introduction of a capability-local
remembered set, known as the /update remembered set/, which tracks
objects which may no longer be visible to the collector due to mutation.
To maintain this remembered set we introduce a write barrier on
mutations which is enabled while a concurrent mark is underway.
The update remembered set representation is similar to that of the
nonmoving mark queue, being a chunked array of `MarkEntry`s. Each
`Capability` maintains a single accumulator chunk, which it flushed
when it (a) is filled, or (b) when the nonmoving collector enters its
post-mark synchronization phase.
While the write barrier touches a significant amount of code it is
conceptually straightforward: the mutator must ensure that the referee
of any pointer it overwrites is added to the update remembered set.
However, there are a few details:
* In the case of objects with a dirty flag (e.g. `MVar`s) we can
exploit the fact that only the *first* mutation requires a write
barrier.
* Weak references, as usual, complicate things. In particular, we must
ensure that the referee of a weak object is marked if dereferenced by
the mutator. For this we (unfortunately) must introduce a read
barrier, as described in Note [Concurrent read barrier on deRefWeak#]
(in `NonMovingMark.c`).
* Stable names are also a bit tricky as described in Note [Sweeping
stable names in the concurrent collector] (`NonMovingSweep.c`).
We take quite some pains to ensure that the high thread count often seen
in parallel Haskell applications doesn't affect pause times. To this end
we allow thread stacks to be marked either by the thread itself (when it
is executed or stack-underflows) or the concurrent mark thread (if the
thread owning the stack is never scheduled). There is a non-trivial
handshake to ensure that this happens without racing which is described
in Note [StgStack dirtiness flags and concurrent marking].
Co-Authored-by: Ömer Sinan Ağacan <omer@well-typed.com>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
This implements the core heap structure and a serial mark/sweep
collector which can be used to manage the oldest-generation heap.
This is the first step towards a concurrent mark-and-sweep collector
aimed at low-latency applications.
The full design of the collector implemented here is described in detail
in a technical note
B. Gamari. "A Concurrent Garbage Collector For the Glasgow Haskell
Compiler" (2018)
The basic heap structure used in this design is heavily inspired by
K. Ueno & A. Ohori. "A fully concurrent garbage collector for
functional programs on multicore processors." /ACM SIGPLAN Notices/
Vol. 51. No. 9 (presented by ICFP 2016)
This design is intended to allow both marking and sweeping
concurrent to execution of a multi-core mutator. Unlike the Ueno design,
which requires no global synchronization pauses, the collector
introduced here requires a stop-the-world pause at the beginning and end
of the mark phase.
To avoid heap fragmentation, the allocator consists of a number of
fixed-size /sub-allocators/. Each of these sub-allocators allocators into
its own set of /segments/, themselves allocated from the block
allocator. Each segment is broken into a set of fixed-size allocation
blocks (which back allocations) in addition to a bitmap (used to track
the liveness of blocks) and some additional metadata (used also used
to track liveness).
This heap structure enables collection via mark-and-sweep, which can be
performed concurrently via a snapshot-at-the-beginning scheme (although
concurrent collection is not implemented in this patch).
The mark queue is a fairly straightforward chunked-array structure.
The representation is a bit more verbose than a typical mark queue to
accomodate a combination of two features:
* a mark FIFO, which improves the locality of marking, reducing one of
the major overheads seen in mark/sweep allocators (see [1] for
details)
* the selector optimization and indirection shortcutting, which
requires that we track where we found each reference to an object
in case we need to update the reference at a later point (e.g. when
we find that it is an indirection). See Note [Origin references in
the nonmoving collector] (in `NonMovingMark.h`) for details.
Beyond this the mark/sweep is fairly run-of-the-mill.
[1] R. Garner, S.M. Blackburn, D. Frampton. "Effective Prefetch for
Mark-Sweep Garbage Collection." ISMM 2007.
Co-Authored-By: Ben Gamari <ben@well-typed.com>
|
| | |
| |
| |
| | |
This flag will enable the use of a non-moving oldest generation.
|
| | |\ \
| | | |
| | | |
| | | | |
'wip/gc/aligned-block-allocation' into wip/gc/preparation
|
| | | |/
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
This implements support for block group allocations which are aligned to
an integral number of blocks.
This will be used by the nonmoving garbage collector, which uses the
block allocator to allocate the segments which back its heap. These
segments are a fixed number of blocks in size, with each segment being
aligned to the segment size boundary. This allows us to easily find the
segment metadata stored at the beginning of the segment.
|
| | | | |
|
| | | |
| | |
| | |
| | |
| | | |
This were previously quite unclear and will change a bit under the
non-moving collector so let's clear this up now.
|
| | |/
| |
| |
| |
| |
| |
| | |
Namely ensure that block descriptors are initialized with valid
generation numbers.
Co-Authored-By: Ben Gamari <ben@well-typed.com>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
With this change it is possible to reconstruct the timing portion of a
`.prof` file after the fact. By logging the stacks at each time point
a more precise executation trace of the program can be observed rather
than all identical cost centres being identified in the report.
There are two new events:
1. `EVENT_PROF_BEGIN` - emitted at the start of profiling to communicate
the tick interval
2. `EVENT_PROF_SAMPLE_COST_CENTRE` - emitted on each tick to communicate the
current call stack.
Fixes #17322
|
| | |
| |
| |
| |
| |
| | |
This patch adds support for the s390x architecture for the LLVM code
generator. The patch includes a register mapping of STG registers onto
s390x machine registers which enables a registerised build.
|
| |/ |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- No need to distinguish between gcc-llvm and clang. First of all,
gcc-llvm is quite old and surely unmaintained by now. Second of all,
none of the code actually care about that distinction!
Now, it does make sense to consider C multiple frontends for LLVMs in
the form of clang vs clang-cl (same clang, yes, but tweaked
interface). But this is better handled in terms of "gccish vs
mvscish" and "is LLVM", yielding 4 combinations. Therefore, I don't
think it is useful saving the existing code for that.
- Get the remaining CC_LLVM_BACKEND, and also TABLES_NEXT_TO_CODE in
mk/config.h the normal way, rather than hacking it post-hoc. No point
keeping these special cases around for now reason.
- Get rid of hand-rolled `die` function and just use `AC_MSG_ERROR`.
- Abstract check + flag override for unregisterised and tables next to
code.
Oh, and as part of the above I also renamed/combined some variables
where it felt appropriate.
- GccIsClang -> CcLlvmBackend. This is for `AC_SUBST`, like the other
Camal case ones. It was never about gcc-llvm, or Apple's renamed clang,
to be clear.
- llvm_CC_FLAVOR -> CC_LLVM_BACKEND. This is for `AC_DEFINE`, like the
other all-caps snake case ones. llvm_CC_FLAVOR was just silly
indirection *and* an odd name to boot.
|
| |
|
|
|
|
|
|
|
| |
This allows the stage1 compiler (which needs to run on the build
platform and produce code for the host) to depend upon properties of the
target. This is wrong. However, it's no more wrong than it was
previously and @Erichson2314 is working on fixing this so I'm going to
remove the guard so we can finally bootstrap HEAD with ghc-8.8 (see
issue #17146).
|
| | |
|
| | |
|
| |
|
|
| |
To avoid polluting the macro namespace
|
| |
|
|
| |
They are only used in a file we construct directly, so just skip CPP.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The generated headers are now generated per stage, which means we can
skip hacks like `ghc_boot_platform.h` and just have that be the stage 0
header as proper. In general, stages are to be embraced: freely generate
everything in each stage but then just build what you depend on, and
everything is symmetrical and efficient. Trying to avoid stages because
bootstrapping is a mind bender just creates tons of bespoke
mini-mind-benders that add up to something far crazier.
Hadrian was pretty close to this "stage-major" approach already, and so
was fairly easy to fix. Make needed more work, however: it did know
about stages so at least there was a scaffold, but few packages except
for the compiler cared, and the compiler used its own counting system.
That said, make and Hadrian now work more similarly, which is good for
the transition to Hadrian. The merits of embracing stage aside, the
change may be worthy for easing that transition alone.
|
| |
|
|
| |
Zeros heap memory after gc freed it.
|
| |
|
|
|
| |
It doesn't need it, and it shouldn't need it or else multi-target will
break.
|
| | |
|
| |
|
|
|
|
|
| |
This commit starts renaming some flip bit related functions for the
generalised heap traversal code and adds provitions for sharing the
per-closure profiling header field currently used exclusively for retainer
profiling with other heap traversal profiling modes.
|
| |
|
|
|
|
|
| |
The `defined(DEBUG_RETAINER) == true` branch doesn't even compile anymore
because 1) retainerSet was renamed to RetainerSet and 2) even if I fix that
the context in Rts.h seems to have changed such that it's not in scope. If
3) I fix that 'flip' is still not in scope :) At that point I just gave up.
|
| |
|
|
|
|
|
| |
This updates the documentation of the MIN_PAYLOAD_SIZE constant and adds
a new Note [Mark bits in mark-compact collector] explaning why the
mark-compact collector uses two bits per objet and why we need
MIN_PAYLOAD_SIZE.
|
| |
|
|
|
|
|
| |
Until 0472f0f6a92395d478e9644c0dbd12948518099f there was a meaningful
host vs target distinction (though it wasn't used right, in genapply).
After that, they did not differ in meaningful ways, so it's best to just
only keep one.
|
| |
|
|
|
|
|
|
|
|
| |
This patch adds a new eventlog event which indicates the start of
a biographical profiler sample. These are different to normal events as
they also include the timestamp of when the census took place. This is
because the LDV profiler only emits samples at the end of the run.
Now all the different profiling modes emit consumable events to the
eventlog.
|
| |
|
|
|
|
| |
Add StgToCmm module hierarchy. Platform modules that are used in several
other places (NCG, LLVM codegen, Cmm transformations) are put into
GHC.Platform.
|
| |
|
|
|
|
|
|
|
| |
Some where using `True` / `False`, a legacy of when they were in
`Config.hs`. See #16914 / d238d3062a9858 for a similar problem.
Also clean up the configure variables names for consistency and clarity
while we're at it. "Target" makes clear we are talking about outputted
code, not where GHC itself runs.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
`TablesNextToCode` is now a substituted by configure, where it has the
correct defaults and error handling. Nowhere else needs to duplicate
that, though we may want the compiler to to guard against bogus settings
files.
I renamed it from `GhcEnableTablesNextToCode` to `TablesNextToCode` to:
- Help me guard against any unfixed usages
- Remove any lingering connotation that this flag needs to be combined
with `GhcUnreigsterised`.
Original reviewers:
Original subscribers: TerrorJack, rwbarton, carter
Original Differential Revision: https://phabricator.haskell.org/D5082
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Effects as I measured them:
RTS Size: +0.1%
Compile times: -0.5%
Runtine nofib: -1.1%
Nofib runtime result seems to mostly come from the `CS` benchmark
which is very sensible to alignment changes so this is likely over
represented.
However the compile time changes are realistic.
This is related to #16961.
|
| |
|
|
|
| |
Now that the target macros are not being used, we remove them. This
prevents target hardcoding regressions.
|
| |
|
|
|
|
|
| |
Unfortunately this will require more work; register allocation is
quite broken.
This reverts commit acd795583625401c5554f8e04ec7efca18814011.
|
| |
|
|
|
| |
Instead following @angerman's suggestion put them in the config file.
Maybe we could re-key llvm-targets someday, but this is good for now.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
These prevent multi-target builds. They were gotten rid of in 3 ways:
1. In the compiler itself, replacing `#if` with runtime `if`. In these
cases, we care about the target platform still, but the target platform
is dynamic so we must delay the elimination to run time.
2. In the compiler itself, replacing `TARGET` with `HOST`. There was
just one bit of this, in some code splitting strings representing lists
of paths. These paths are used by GHC itself, and not by the compiled
binary. (They are compiler lookup paths, rather than RPATHS or something
that does matter to the compiled binary, and thus would legitamentally
be target-sensative.) As such, the path-splitting method only depends on
where GHC runs and not where code it produces runs. This should have
been `HOST` all along.
3. Changing the RTS. The RTS doesn't care about the target platform,
full stop.
4. `includes/stg/HaskellMachRegs.h` This file is also included in the
genapply executable. This is tricky because the RTS's host platform
really is that utility's target platform. so that utility really really
isn't multi-target either. But at least it isn't an installed part of
GHC, but just a one-off tool when building the RTS. Lying with the
`HOST` to a one-off program (genapply) that isn't installed doesn't seem so bad.
It's certainly better than the other way around of lying to the RTS
though not to genapply. The RTS is more important, and it is installed,
*and* this header is installed as part of the RTS.
|
| |
|
|
|
|
|
| |
This adds support for constructing vector types from Float#, Double# etc
and performing arithmetic operations on them
Cleaned-Up-By: Ben Gamari <ben@well-typed.com>
|
