When performing inlining LLVM unfortunately "leaks" types (the
types survive and are usable, but a new round of inlining will
recreate new structurally equivalent types). This accumulation
will over time amount to a memory leak which for some queries
can be large enough to trigger the OOM process killer.
To avoid accumulation of types, all IR related data is stored
in an LLVMContextRef which is dropped and recreated in order
to release all types. Dropping and recreating incurs overhead,
so it will be done only after 100 queries. This is a heuristic
which might be revisited, but until we can get the size of the
context from LLVM we are flying a bit blind.
This issue has been reported several times, there may be more
references to it in the archives on top of the threads linked
below.
This is a backpatch of 9dce22033d5 to all supported branches.
Reported-By: Justin Pryzby <pryzby@telsasoft.com>
Reported-By: Kurt Roeckx <kurt@roeckx.be>
Reported-By: Jaime Casanova <jcasanov@systemguards.com.ec>
Reported-By: Lauri Laanmets <pcspets@gmail.com>
Author: Andres Freund and Daniel Gustafsson
Discussion: https://postgr.es/m/7acc8678-df5f-4923-9cf6-e843131ae89d@www.fastmail.com
Discussion: https://postgr.es/m/20201218235607.GC30237@telsasoft.com
Discussion: https://postgr.es/m/CAPH-tTxLf44s3CvUUtQpkDr1D8Hxqc2NGDzGXS1ODsfiJ6WSqA@mail.gmail.com
Backpatch-through: v12
Remove use of LLVMGetElementType() and provide the type of all pointers
to LLVMBuildXXX() functions when emitting IR, as required by modern LLVM
versions[1].
* For LLVM <= 14, we'll still use the old LLVMBuildXXX() functions.
* For LLVM == 15, we'll continue to do the same, explicitly opting
out of opaque pointer mode.
* For LLVM >= 16, we'll use the new LLVMBuildXXX2() functions that take
the extra type argument.
The difference is hidden behind some new IR emitting wrapper functions
l_load(), l_gep(), l_call() etc. The change is mostly mechanical,
except that at each site the correct type had to be provided.
In some places we needed to do some extra work to get functions types,
including some new wrappers for C++ APIs that are not yet exposed by in
LLVM's C API, and some new "example" functions in llvmjit_types.c
because it's no longer possible to start from the function pointer type
and ask for the function type.
Back-patch to 12, because it's a little tricker in 11 and we agreed not
to put the latest LLVM support into the upcoming final release of 11.
[1] https://llvm.org/docs/OpaquePointers.html
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Ronan Dunklau <ronan.dunklau@aiven.io>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKGKNX_%3Df%2B1C4r06WETKTq0G4Z_7q4L4Fxn5WWpMycDj9Fw%40mail.gmail.com
llvm_release_context() called llvm_enter_fatal_on_oom(), but was missing
the corresponding llvm_leave_fatal_on_oom() call. As a result, if JIT was
used at all, we were almost always in the "fatal-on-oom" state.
It only makes a difference if you use an extension written in C++, and
run out of memory in a C++ 'new' call. In that case, you would get a
PostgreSQL FATAL error, instead of the default behavior of throwing a
C++ exception.
Back-patch to all supported versions.
Reviewed-by: Daniel Gustafsson
Discussion: https://www.postgresql.org/message-id/54b78cca-bc84-dad8-4a7e-5b56f764fab5@iki.fi
While on it, newlines are removed from the end of two elog() strings.
The others are simple grammar mistakes. One comment in pg_upgrade
referred incorrectly to sequences since a7e5457.
Author: Justin Pryzby
Discussion: https://postgr.es/m/20221230231257.GI1153@telsasoft.com
Backpatch-through: 11
In the same spirit as 6301c3ada, fix some more places where we were
using list_delete_first() in a loop and thereby risking O(N^2)
behavior. It's not clear that the lists manipulated in these spots
can get long enough to be really problematic ... but it's not clear
that they can't, either, and the fixes are simple enough.
As before, back-patch to v13.
Discussion: https://postgr.es/m/CD2F0E7F-9822-45EC-A411-AE56F14DEA9F@amazon.com
If an allocation failed within LLVM it is not safe to call back into LLVM as
LLVM is not generally safe against exceptions / stack-unwinding. Thus errors
while in LLVM code are promoted to FATAL. However llvm_shutdown() did call
back into LLVM even in such cases, while llvm_release_context() was careful
not to do so.
We cannot generally skip shutting down LLVM, as that can break profiling. But
it's OK to do so if there was an error from within LLVM.
Reported-By: Jelte Fennema <Jelte.Fennema@microsoft.com>
Author: Andres Freund <andres@anarazel.de>
Author: Justin Pryzby <pryzby@telsasoft.com>
Discussion: https://postgr.es/m/AM5PR83MB0178C52CCA0A8DEA0207DC14F7FF9@AM5PR83MB0178.EURPRD83.prod.outlook.com
Backpatch: 11-, where jit was introduced
LLVM 13 (due out in September) has changed the semantics of
LLVMOrcAbsoluteSymbols(), so we need to bump some reference counts to
avoid a double-free that causes crashes and bad query results.
A proactive change seems necessary to avoid having a window of time
where our respective latest releases would interact badly. It's
possible that the situation could change before then, though.
Thanks to Fabien Coelho for monitoring bleeding edge LLVM and Andres
Freund for tracking down the change.
Back-patch to 11, where the JIT code arrived.
Discussion: https://postgr.es/m/CA%2BhUKGLEy8mgtN7BNp0ooFAjUedDTJj5dME7NxLU-m91b85siA%40mail.gmail.com
Also "make reformat-dat-files".
The only change worthy of note is that pgindent messed up the formatting
of launcher.c's struct LogicalRepWorkerId, which led me to notice that
that struct wasn't used at all anymore, so I just took it out.
It is error prone (see 5da871bfa1b) and verbose to manually create function
types. Add a helper that can reference a function pointer type via
llvmjit_types.c and and convert existing instances of manual creation.
Author: Andres Freund <andres@anarazel.de>
Reviewed-By: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/20201207212142.wz5tnbk2jsaqzogb@alap3.anarazel.de
LLVM 12, to be released in a few months, made some breaking changes to
the Orc JIT interface. OrcV2 eventually will make it easier to support
features like concurrent JIT compilation, but this commit only allows
to compile against LLVM 12.
This commit is a bit bigger than desirable. That partially is because
the V2 interface is more granular than V1 interface, but also because
I chose to make some minor changes to < LLVM 12 code to keep the code
somewhat readable.
The LLVM 12 support will need to be backpatched. I plan to do so after
the patch stewed on the buildfarm for a few days.
Author: Andres Freund
Discussion: https://postgr.es/m/20201016011244.pmyvr3ee2gbzplq4@alap3.anarazel.de
Unfortunately in LLVM 3.9 LLVMGetAttributeCountAtIndex(func, index)
crashes when called with an index that has 0 attributes. Since there's
no way to work around this in the C API, add a small C++ wrapper doing
so.
The only reason this didn't fail before 72559438f92 is that there
always are function attributes...
Author: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20201016001254.w2nfj7gd74jmb5in@alap3.anarazel.de
Backpatch: 11-, like 72559438f92
Previously we only copied the function attributes. That caused problems at
least on s390x: Because we didn't copy the 'zeroext' attribute for
ExecAggTransReparent()'s *IsNull parameters, expressions invoking it didn't
ensure that the upper bytes of the registers were zeroed. In the - relatively
rare - cases where not, ExecAggTransReparent() wrongly ended up in the
newValueIsNull branch due to the register not being zero. Subsequently causing
a crash.
It's quite possible that this would cause problems on other platforms, and in
other places than just ExecAggTransReparent() on s390x.
Thanks to Christoph (and the Debian project) for providing me with access to a
s390x machine, allowing me to debug this.
Reported-By: Christoph Berg
Author: Andres Freund
Discussion: https://postgr.es/m/20201015083246.kie5726xerdt3ael@alap3.anarazel.de
Backpatch: 11-, where JIT was added
This seems more correct, because other before_shmem_exit calls may
expect the infrastructure that is needed to run queries and access the
database to be working, and also because this cleanup has nothing to
do with shared memory.
There are no known user-visible consequences to this, though, apart
from what was previous fixed by commit
303640199d0436c5e7acdf50b837a027b5726594 and back-patched as commit
bcbc27251d35336a6442761f59638138a772b839 and commit
f7013683d9bb663a6a917421b1374306a32f165b, so for now, no back-patch.
Bharath Rupireddy
Discussion: http://postgr.es/m/CALj2ACWk7j4F2v2fxxYfrroOF=AdFNPr1WsV+AGtHAFQOqm_pw@mail.gmail.com
The main benefit of doing so is that this allows llvm to ensure that
types match - previously that'd only be detected by a crash within the
called function. There were a number of cases where we passed a
superfluous parameter...
To avoid needing to add all the functions to llvmjit.{c,h}, instead
get them from the llvm module for llvmjit_types.c. Also use that for
the functions from llvmjit_types already in llvmjit.h.
Author: Soumyadeep Chakraborty and Andres Freund
Discussion: https://postgr.es/m/CADwEdooww3wZv-sXSfatzFRwMuwa186LyTwkBfwEW6NjtooBPA@mail.gmail.com
Similar to commits 7e735035f2 and dddf4cdc33, this commit makes the order
of header file inclusion consistent for backend modules.
In the passing, removed a couple of duplicate inclusions.
Author: Vignesh C
Reviewed-by: Kuntal Ghosh and Amit Kapila
Discussion: https://postgr.es/m/CALDaNm2Sznv8RR6Ex-iJO6xAdsxgWhCoETkaYX=+9DW3q0QCfA@mail.gmail.com
Before this change FunctionCallInfoData, the struct arguments etc for
V1 function calls are stored in, always had space for
FUNC_MAX_ARGS/100 arguments, storing datums and their nullness in two
arrays. For nearly every function call 100 arguments is far more than
needed, therefore wasting memory. Arg and argnull being two separate
arrays also guarantees that to access a single argument, two
cachelines have to be touched.
Change the layout so there's a single variable-length array with pairs
of value / isnull. That drastically reduces memory consumption for
most function calls (on x86-64 a two argument function now uses
64bytes, previously 936 bytes), and makes it very likely that argument
value and its nullness are on the same cacheline.
Arguments are stored in a new NullableDatum struct, which, due to
padding, needs more memory per argument than before. But as usually
far fewer arguments are stored, and individual arguments are cheaper
to access, that's still a clear win. It's likely that there's other
places where conversion to NullableDatum arrays would make sense,
e.g. TupleTableSlots, but that's for another commit.
Because the function call information is now variable-length
allocations have to take the number of arguments into account. For
heap allocations that can be done with SizeForFunctionCallInfoData(),
for on-stack allocations there's a new LOCAL_FCINFO(name, nargs) macro
that helps to allocate an appropriately sized and aligned variable.
Some places with stack allocation function call information don't know
the number of arguments at compile time, and currently variably sized
stack allocations aren't allowed in postgres. Therefore allow for
FUNC_MAX_ARGS space in these cases. They're not that common, so for
now that seems acceptable.
Because of the need to allocate FunctionCallInfo of the appropriate
size, older extensions may need to update their code. To avoid subtle
breakages, the FunctionCallInfoData struct has been renamed to
FunctionCallInfoBaseData. Most code only references FunctionCallInfo,
so that shouldn't cause much collateral damage.
This change is also a prerequisite for more efficient expression JIT
compilation (by allocating the function call information on the stack,
allowing LLVM to optimize it away); previously the size of the call
information caused problems inside LLVM's optimizer.
Author: Andres Freund
Reviewed-By: Tom Lane
Discussion: https://postgr.es/m/20180605172952.x34m5uz6ju6enaem@alap3.anarazel.de
We usually don't change the name of structs between the struct name
itself and the name of the typedef. Additionally, structs that are
usually used via a typedef that hides being a pointer, are commonly
suffixed Data. Change tupdesc code to follow those convention.
This is triggered by a future patch that intends to forward declare
TupleDescData in another header - keeping with the naming scheme makes
that easier to understand.
Author: Andres Freund
Discussion: https://postgr.es/m/20190114000701.y4ttcb74jpskkcfb@alap3.anarazel.de
This commit completes the work prepared in 1a0586de36, splitting the
old TupleTableSlot implementation (which could store buffer, heap,
minimal and virtual slots) into four different slot types. As
described in the aforementioned commit, this is done with the goal of
making tuple table slots extensible, to allow for pluggable table
access methods.
To achieve runtime extensibility for TupleTableSlots, operations on
slots that can differ between types of slots are performed using the
TupleTableSlotOps struct provided at slot creation time. That
includes information from the size of TupleTableSlot struct to be
allocated, initialization, deforming etc. See the struct's definition
for more detailed information about callbacks TupleTableSlotOps.
I decided to rename TTSOpsBufferTuple to TTSOpsBufferHeapTuple and
ExecCopySlotTuple to ExecCopySlotHeapTuple, as that seems more
consistent with other naming introduced in recent patches.
There's plenty optimization potential in the slot implementation, but
according to benchmarking the state after this commit has similar
performance characteristics to before this set of changes, which seems
sufficient.
There's a few changes in execReplication.c that currently need to poke
through the slot abstraction, that'll be repaired once the pluggable
storage patchset provides the necessary infrastructure.
Author: Andres Freund and Ashutosh Bapat, with changes by Amit Khandekar
Discussion: https://postgr.es/m/20181105210039.hh4vvi4vwoq5ba2q@alap3.anarazel.de
This mainly de-duplicates code. As evaluating a system variable isn't
the hottest path and the current inline implementation ends up calling
out to an external function anyway, this is OK from a performance POV.
The main motivation for de-duplicating is the upcoming slot
abstraction work, after which there's not guaranteed to be a HeapTuple
backing the slot.
Author: Andres Freund, Amit Khandekar
Discussion: https://postgr.es/m/20181105210039.hh4vvi4vwoq5ba2q@alap3.anarazel.de
Previously, when using parallel query, EXPLAIN (ANALYZE)'s JIT
compilation timings did not include the overhead from doing so on the
workers. Fix that.
We do so by simply aggregating the cost of doing JIT compilation on
workers and the leader together. Arguably that's not quite accurate,
because the total time spend doing so is spent in parallel - but it's
hard to do much better. For additional detail, when VERBOSE is
specified, the stats for workers are displayed separately.
Author: Amit Khandekar and Andres Freund
Discussion: https://postgr.es/m/CAJ3gD9eLrz51RK_gTkod+71iDcjpB_N8eC6vU2AW-VicsAERpQ@mail.gmail.com
Backpatch: 11-
During the work of upstreaming my previous patches for gdb and perf
support the API changed. Adapt. Normally this wouldn't necessarily be
something to backpatch, but the previous API wasn't upstream, and at
least the gdb support is quite useful for debugging.
Author: Andres Freund
Backpatch: 11, where LLVM based JIT support was added.
This provides infrastructure to allow JITed code to inline code
implemented in C. This e.g. can be postgres internal functions or
extension code.
This already speeds up long running queries, by allowing the LLVM
optimizer to optimize across function boundaries. The optimization
potential currently doesn't reach its full potential because LLVM
cannot optimize the FunctionCallInfoData argument fully away, because
it's allocated on the heap rather than the stack. Fixing that is
beyond what's realistic for v11.
To be able to do that, use CLANG to convert C code to LLVM bitcode,
and have LLVM build a summary for it. That bitcode can then be used to
to inline functions at runtime. For that the bitcode needs to be
installed. Postgres bitcode goes into $pkglibdir/bitcode/postgres,
extensions go into equivalent directories. PGXS has been modified so
that happens automatically if postgres has been compiled with LLVM
support.
Currently this isn't the fastest inline implementation, modules are
reloaded from disk during inlining. That's to work around an apparent
LLVM bug, triggering an apparently spurious error in LLVM assertion
enabled builds. Once that is resolved we can remove the superfluous
read from disk.
Docs will follow in a later commit containing docs for the whole JIT
feature.
Author: Andres Freund
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
Instead using memset to set tts_isnull, call the new
slot_getmissingattrs().
Also fix a bug (= instead of >=) in the code generation. Normally = is
correct, but when repeatedly deforming fields not in a
tuple (e.g. deform up to natts + 1 and then natts + 2) >= is needed.
Discussion: https://postgr.es/m/20180328010053.i2qvsuuusst4lgmc@alap3.anarazel.de
This is mostly done to be able to validate features and fixes
submitted to LLVM. Given the size of these changes that seems
acceptable.
Author: Andres Freund
Due to the differing APIs between versions, I forgot to deallocate the
generated module in older LLVM versions, leading to a memory leak.
Author: Andres Freund
Performing JIT compilation for deforming gains performance benefits
over unJITed deforming from compile-time knowledge of the tuple
descriptor. Fixed column widths, NOT NULLness, etc can be taken
advantage of.
Right now the JITed deforming is only used when deforming tuples as
part of expression evaluation (and obviously only if the descriptor is
known). It's likely to be beneficial in other cases, too.
By default tuple deforming is JITed whenever an expression is JIT
compiled. There's a separate boolean GUC controlling it, but that's
expected to be primarily useful for development and benchmarking.
Docs will follow in a later commit containing docs for the whole JIT
feature.
Author: Andres Freund
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
The LLVM JIT provider uses clang to synchronize types between normal C
code and runtime generated code. Clang represents stdbool.h style
booleans in return values & parameters differently from booleans
stored in variables.
Thus the expression compilation code from 2a0faed9d needs to be
adapted to 9a95a77d9. Instead of hardcoding i8 as the type for
booleans (which already was wrong on some edge case platforms!), use
postgres' notion of a boolean as used for storage and for parameters.
Per buildfarm animal xenodermus.
Author: Andres Freund
In addition to the interpretation of expressions (which back
evaluation of WHERE clauses, target list projection, aggregates
transition values etc) support compiling expressions to native code,
using the infrastructure added in earlier commits.
To avoid duplicating a lot of code, only support emitting code for
cases that are likely to be performance critical. For expression steps
that aren't deemed that, use the existing interpreter.
The generated code isn't great - some architectural changes are
required to address that. But this already yields a significant
speedup for some analytics queries, particularly with WHERE clauses
filtering a lot, or computing multiple aggregates.
Author: Andres Freund
Tested-By: Thomas Munro
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
Disable JITing for VALUES() nodes.
VALUES() nodes are only ever executed once. This is primarily helpful
for debugging, when forcing JITing even for cheap queries.
Author: Andres Freund
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
This commit introduces the ability to actually generate code using
LLVM. In particular, this adds:
- Ability to emit code both in heavily optimized and largely
unoptimized fashion
- Batching facility to allow functions to be defined in small
increments, but optimized and emitted in executable form in larger
batches (for performance and memory efficiency)
- Type and function declaration synchronization between runtime
generated code and normal postgres code. This is critical to be able
to access struct fields etc.
- Developer oriented jit_dump_bitcode GUC, for inspecting / debugging
the generated code.
- per JitContext statistics of number of functions, time spent
generating code, optimizing, and emitting it. This will later be
employed for EXPLAIN support.
This commit doesn't yet contain any code actually generating
functions. That'll follow in later commits.
Documentation for GUCs added, and for JIT in general, will be added in
later commits.
Author: Andres Freund, with contributions by Pierre Ducroquet
Testing-By: Thomas Munro, Peter Eisentraut
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
This commit introduces:
1) JIT provider abstraction, which allows JIT functionality to be
implemented in separate shared libraries. That's desirable because
it allows to install JIT support as a separate package, and because
it allows experimentation with different forms of JITing.
2) JITContexts which can be, using functions introduced in follow up
commits, used to emit JITed functions, and have them be cleaned up
on error.
3) The outline of a LLVM JIT provider, which will be fleshed out in
subsequent commits.
Documentation for GUCs added, and for JIT in general, will be added in
later commits.
Author: Andres Freund, with architectural input from Jeff Davis
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
