LLVM 22 has the fix that we copied into our tree in commit 9044fc1d and
a new function to reach it[1][2], so we only need to use our copy for
Aarch64 + LLVM < 22. The only change to the final version that our copy
didn't get is a new LLVM_ABI macro, but that isn't appropriate for us.
Our copy is hopefully now frozen and would only need maintenance if bugs
are found in the upstream code.
Non-Aarch64 systems now also use the new API with LLVM 22. It allocates
all sections with one contiguous mmap() instead of one per
section. We could have done that earlier, but commit 9044fc1d wanted to
limit the blast radius to the affected systems. We might as well
benefit from that small improvement everywhere now that it is available
out of the box.
We can't delete our copy until LLVM 22 is our minimum supported version,
or we switch to the newer JITLink API for at least Aarch64.
[1] https://github.com/llvm/llvm-project/pull/71968
[2] https://github.com/llvm/llvm-project/pull/174307
Backpatch-through: 14
Discussion: https://postgr.es/m/CA%2BhUKGJTumad75o8Zao-LFseEbt%3DenbUFCM7LZVV%3Dc8yg2i7dg%40mail.gmail.com
While JIT can speed up large analytical queries, it can also cause
serious performance issues on otherwise very fast queries. Compiling
and optimizing the expressions may be so expensive, it completely
outweighs the JIT benefits for shorter queries.
Ideally, we'd address this in the cost model, but the part deciding
whether to enable JIT for a query is rather simple, partially because we
don't have any reliable estimates of how expensive the LLVM compilation
and optimization is.
Sometimes seemingly unrelated changes (for example a couple additional
INSERTs into a table) increase the cost just enough to enable JIT,
resulting in a performance cliff.
Because of these risks, most large-scale deployments already disable JIT
by default. Notably, this includes all hyperscalers.
This commit changes our default to align with that established practice.
If we improve the JIT (be it better costing or cheaper execution), we
can consider enabling it by default again.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Discussion: https://postgr.es/m/DG1VZJEX1AQH.2EH4OKGRUDB71@jeltef.nl
Remove a bunch of #include lines from execnodes.h. Most of these
requier suitable typedefs to be added, so that it still compiles
standalone. In one case, the fix is to move a struct definition to the
one .c file where it is needed.
Also some light clean up in plannodes.h and genam.h, though not as
extensive as in execnodes.h.
Author: Álvaro Herrera <alvherre@kurilemu.de>
Author: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/202603131240.ihwqdxnj7w2o@alvherre.pgsql
This commit includes various optimizations to improve the performance of
tuple deformation.
We now precalculate CompactAttribute's attcacheoff, which allows us to
remove the code from the deform routines which was setting the
attcacheoff. Setting the attcacheoff is now handled by
TupleDescFinalize(), which must be called before the TupleDesc is used for
anything. Having TupleDescFinalize() means we can store the first
attribute in the TupleDesc which does not have an offset cached. That
allows us to add a dedicated deforming loop to deform all attributes up
to the final one with an attcacheoff set, or up to the first NULL
attribute, whichever comes first.
Here we also improve tuple deformation performance of tuples with NULLs.
Previously, if the HEAP_HASNULL bit was set in the tuple's t_infomask,
deforming would, one-by-one, check each and every bit in the NULL bitmap
to see if it was zero. Now, we process the NULL bitmap 1 byte at a time
rather than 1 bit at a time to find the attnum with the first NULL. We
can now deform the tuple without checking for NULLs up to just before that
attribute.
We also record the maximum attribute number which is guaranteed to exist
in the tuple, that is, has a NOT NULL constraint and isn't an
atthasmissing attribute. When deforming only attributes prior to the
guaranteed attnum, we've no need to access the tuple's natt count. As an
additional optimization, we only count fixed-width columns when
calculating the maximum guaranteed column, as this eliminates the need to
emit code to fetch byref types in the deformation loop for guaranteed
attributes.
Some locations in the code deform tuples that have yet to go through NOT
NULL constraint validation. We're unable to perform the guaranteed
attribute optimization when that's the case. This optimization is opt-in
via the TupleTableSlot using the TTS_FLAG_OBEYS_NOT_NULL_CONSTRAINTS
flag.
This commit also adds a more efficient way of populating the isnull
array by using a bit-wise SWAR trick which performs multiplication on the
inverse of the tuple's bitmap byte and masking out all but the lower bit
of each of the boolean's byte. This results in much more optimal code
when compared to determining the NULLness via att_isnull(). 8 isnull
elements are processed at once using this method, which means we need to
round the tts_isnull array size up to the next 8 bytes. The palloc code
does this anyway, but the round-up needed to be formalized so as not to
overwrite the sentinel byte in MEMORY_CONTEXT_CHECKING builds. Doing
this also allows the NULL-checking deforming loop to more efficiently
check the isnull array, rather than doing the bit-wise processing for each
attribute that att_isnull() does.
The level of performance improvement from these changes seems to vary
depending on the CPU architecture. Apple's M chips seem particularly
fond of the changes, with some of the tested deform-heavy queries going
over twice as fast as before. With x86-64, the speedups aren't quite as
large. With tables containing only a small number of columns, the
speedups will be less.
Author: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: John Naylor <johncnaylorls@gmail.com>
Reviewed-by: Amit Langote <amitlangote09@gmail.com>
Reviewed-by: Zsolt Parragi <zsolt.parragi@percona.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Junwang Zhao <zhjwpku@gmail.com>
Discussion: https://postgr.es/m/CAApHDvpoFjaj3%2Bw_jD5uPnGazaw41A71tVJokLDJg2zfcigpMQ%40mail.gmail.com
With LLVM >= 17, transform passes are provided as a string to
LLVMRunPasses. Only two strings were used: "default<O3>" and
"default<O0>,mem2reg".
With previous LLVM versions, an additional inline pass was added when
JIT inlining was enabled without optimization. With LLVM >= 17, the code
would go through llvm_inline, prepare the functions for inlining, but
the generated bitcode would be the same due to the missing inline pass.
This patch restores the previous behavior by adding an inline pass when
inlining is enabled but no optimization is done.
This fixes an oversight introduced by 76200e5e when support for LLVM 17
was added.
Backpatch-through: 14
Author: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Pierre Ducroquet <p.psql@pinaraf.info>
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Discussion: https://postgr.es/m/CAO6_XqrNjJnbn15ctPv7o4yEAT9fWa-dK15RSyun6QNw9YDtKg%40mail.gmail.com
jit_profiling_support=true captures profile data for Linux perf. On
other platforms, LLVMCreatePerfJITEventListener() returns NULL and the
attempt to register the listener would crash.
Fix by ignoring the setting in that case. The documentation already
says that it only has an effect if perf support is present, and we
already did the same for older LLVM versions that lacked support.
No field reports, unsurprisingly for an obscure developer-oriented
setting. Noticed in passing while working on commit 1a28b4b4.
Backpatch-through: 14
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKGJgB6gvrdDohgwLfCwzVQm%3DVMtb9m0vzQn%3DCwWn-kwG9w%40mail.gmail.com
We currently require LLVM 14, so these probes for LLVM 9 functions
always succeeded. Even when the features aren't enabled in an LLVM
build, dummy functions are defined (a problem for a later commit).
The whole PGAC_CHECK_LLVM_FUNCTIONS macro and Meson equivalent are
removed, because we switched to testing LLVM_VERSION_MAJOR at compile
time in subsequent work and these were the last holdouts. That suits
the nature of LLVM API evolution better, and also allows for strictly
mechanical pruning in future commits like 820b5af7 and 972c2cd2. They
advanced the minimum LLVM version but failed to spot these.
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKGJgB6gvrdDohgwLfCwzVQm%3DVMtb9m0vzQn%3DCwWn-kwG9w%40mail.gmail.com
REL_18_STABLE and master have commit ee485912, so they always use the
newer LLVM opaque pointer functions. Drop -Wno-deprecated-declarations
(commit a56e7b660) for code under jit/llvm in those branches, to catch
any new deprecation warnings that arrive in future version of LLVM.
Older branches continued to use functions marked deprecated in LLVM 14
and 15 (ie switched to the newer functions only for LLVM 16+), as a
precaution against unforeseen compatibility problems with bitcode
already shipped. In those branches, the comment about warning
suppression is updated to explain that situation better. In theory we
could suppress warnings only for LLVM 14 and 15 specifically, but that
isn't done here.
Backpatch-through: 14
Reported-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/1407185.1766682319%40sss.pgh.pa.us
An array of LLVMBasicBlockRef is allocated with the size used for an
element being "LLVMBasicBlockRef *" rather than "LLVMBasicBlockRef".
LLVMBasicBlockRef is a type that refers to a pointer, so this did not
directly cause a problem because both should have the same size, still
it is incorrect.
This issue has been spotted while reviewing a different patch, and
exists since 2a0faed9d7, so backpatch all the way down.
Discussion: https://postgr.es/m/CA+hUKGLngd9cKHtTUuUdEo2eWEgUcZ_EQRbP55MigV2t_zTReg@mail.gmail.com
Backpatch-through: 14
The idea is to encourage more the use of these new routines across the
tree, as these offer stronger type safety guarantees than palloc().
This batch of changes includes most of the trivial changes suggested by
the author for src/backend/.
A total of 334 files are updated here. Among these files, 48 of them
have their build change slightly; these are caused by line number
changes as the new allocation formulas are simpler, shaving around 100
lines of code in total.
Similar work has been done in 0c3c5c3b06 and 31d3847a37.
Author: David Geier <geidav.pg@gmail.com>
Discussion: https://postgr.es/m/ad0748d4-3080-436e-b0bc-ac8f86a3466a@gmail.com
In 2a0faed9d7, which added JIT compilation support for expressions, I
accidentally used sizeof(LLVMBasicBlockRef *) instead of
sizeof(LLVMBasicBlockRef) as part of computing the size of an allocation. That
turns out to have no real negative consequences due to LLVMBasicBlockRef being
a pointer itself (and thus having the same size). It still is wrong and
confusing, so fix it.
Reported by coverity.
Backpatch-through: 13
When JIT deformed tuples (controlled via the jit_tuple_deforming GUC),
types narrower than sizeof(Datum) would be zero-extended up to Datum
width. This wasn't the same as what fetch_att() does in the standard
tuple deforming code. Logically the values are the same when fetching
via the DatumGet*() marcos, but negative numbers are not the same in
binary form.
In the report, the problem was manifesting itself with:
ERROR: could not find memoization table entry
in a query which had a "Cache Mode: binary" Memoize node. However, it's
currently unclear what else is affected. Anything that uses
datum_image_eq() or datum_image_hash() on a Datum from a tuple deformed by
JIT could be affected, but it may not be limited to that.
The fix for this is simple: use signed extension instead of zero
extension.
Many thanks to Emmanuel Touzery for reporting this issue and providing
steps and backup which allowed the problem to easily be recreated.
Reported-by: Emmanuel Touzery <emmanuel.touzery@plandela.si>
Author: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/DB8P194MB08532256D5BAF894F241C06393F3A@DB8P194MB0853.EURP194.PROD.OUTLOOK.COM
Backpatch-through: 13
Commit 2a600a93 made Datum 8 bytes wide everywhere. It was no longer
appropriate to use TypeSizeT on 32 bit systems, and JIT compilation
would fail with various type check errors. Introduce a separate
LLVMTypeRef with the name TypeDatum. TypeSizeT is still used in some
places for actual size_t values.
Reported-by: Dmitry Mityugov <d.mityugov@postgrespro.ru>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Tested-by: Dmitry Mityugov <d.mityugov@postgrespro.ru>
Discussion: https://postgr.es/m/0a9f0be59171c2e8f1b3bc10f4fcf267%40postgrespro.ru
For many optional libraries, we extract the -L and -l switches needed
to link the library from a helper program such as llvm-config. In
some cases we put the resulting -L switches into LDFLAGS ahead of
-L switches specified via --with-libraries. That risks breaking
the user's intention for --with-libraries.
It's not such a problem if the library's -L switch points to a
directory containing only that library, but on some platforms a
library helper may "helpfully" offer a switch such as -L/usr/lib
that points to a directory holding all standard libraries. If the
user specified --with-libraries in hopes of overriding the standard
build of some library, the -L/usr/lib switch prevents that from
happening since it will come before the user-specified directory.
To fix, avoid inserting these switches directly into LDFLAGS during
configure, instead adding them to LIBDIRS or SHLIB_LINK. They will
still eventually get added to LDFLAGS, but only after the switches
coming from --with-libraries.
The same problem exists for -I switches: those coming from
--with-includes should appear before any coming from helper programs
such as llvm-config. We have not heard field complaints about this
case, but it seems certain that a user attempting to override a
standard library could have issues.
The changes for this go well beyond configure itself, however,
because many Makefiles have occasion to manipulate CPPFLAGS to
insert locally-desirable -I switches, and some of them got it wrong.
The correct ordering is any -I switches pointing at within-the-
source-tree-or-build-tree directories, then those from the tree-wide
CPPFLAGS, then those from helper programs. There were several places
that risked pulling in a system-supplied copy of libpq headers, for
example, instead of the in-tree files. (Commit cb36f8ec2 fixed one
instance of that a few months ago, but this exercise found more.)
The Meson build scripts may or may not have any comparable problems,
but I'll leave it to someone else to investigate that.
Reported-by: Charles Samborski <demurgos@demurgos.net>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/70f2155f-27ca-4534-b33d-7750e20633d7@demurgos.net
Backpatch-through: 13
The previous minimum was to maintain support for Python 3.5, but we
now require Python 3.6 anyway (commit 45363fca63), so that reason is
obsolete. A small raise to Meson 0.57 allows getting rid of a fair
amount of version conditionals and silences some future-deprecated
warnings.
With the version bump, the following deprecation warnings appeared and
are fixed:
WARNING: Project targets '>=0.57' but uses feature deprecated since '0.55.0': ExternalProgram.path. use ExternalProgram.full_path() instead
WARNING: Project targets '>=0.57' but uses feature deprecated since '0.56.0': meson.build_root. use meson.project_build_root() or meson.global_build_root() instead.
It turns out that meson 0.57.0 and 0.57.1 are buggy for our use, so
the minimum is actually set to 0.57.2. This is specific to this
version series; in the future we won't necessarily need to be this
precise.
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://www.postgresql.org/message-id/flat/42e13eb0-862a-441e-8d84-4f0fd5f6def0%40eisentraut.org
This cleans up some loose ends left by commit e8ca9ed1d. I hadn't
looked closely enough at these places before, but now I have.
The use of double-quoted #includes for Perl headers in plperl_system.h
seems to be simply a mistake introduced in 6c944bf3c and faithfully
copied forward since then. (I had thought possibly it was required
by some weird Windows build setup, but there's no evidence of that in
our history.)
The occurrences in SectionMemoryManager.h and SectionMemoryManager.cpp
evidently stem from those files' origin as LLVM code. It's
understandable that LLVM would treat their own files as needing
double-quoted #includes; but they're still system headers to us.
I also applied the same check to *.c files, and found a few other
random incorrect usages in both directions.
Our ECPG headers and test files routinely use angle brackets to refer
to ECPG headers. I left those usages alone, since it seems reasonable
for an ECPG user to regard those headers as system headers.
This allows them to be added without scanning the table, and validating
them afterwards without holding access exclusive lock on the table after
any violating rows have been deleted or fixed.
Doing ALTER TABLE ... SET NOT NULL for a column that has an invalid
not-null constraint validates that constraint. ALTER TABLE .. VALIDATE
CONSTRAINT is also supported. There are various checks on whether an
invalid constraint is allowed in a child table when the parent table has
a valid constraint; this should match what we do for enforced/not
enforced constraints.
pg_attribute.attnotnull is now only an indicator for whether a not-null
constraint exists for the column; whether it's valid or invalid must be
queried in pg_constraint. Applications can continue to query
pg_attribute.attnotnull as before, but now it's possible that NULL rows
are present in the column even when that's set to true.
For backend internal purposes, we cache the nullability status in
CompactAttribute->attnullability that each tuple descriptor carries
(replacing CompactAttribute.attnotnull, which was a mirror of
Form_pg_attribute.attnotnull). During the initial tuple descriptor
creation, based on the pg_attribute scan, we set this to UNRESTRICTED if
pg_attribute.attnotnull is false, or to UNKNOWN if it's true; then we
update the latter to VALID or INVALID depending on the pg_constraint
scan. This flag is also copied when tupledescs are copied.
Comparing tuple descs for equality must also compare the
CompactAttribute.attnullability flag and return false in case of a
mismatch.
pg_dump deals with these constraints by storing the OIDs of invalid
not-null constraints in a separate array, and running a query to obtain
their properties. The regular table creation SQL omits them entirely.
They are then dealt with in the same way as "separate" CHECK
constraints, and dumped after the data has been loaded. Because no
additional pg_dump infrastructure was required, we don't bump its
version number.
I decided not to bump catversion either, because the old catalog state
works perfectly in the new world. (Trying to run with new catalog state
and the old server version would likely run into issues, however.)
System catalogs do not support invalid not-null constraints (because
commit 14e87ffa5c didn't allow them to have pg_constraint rows
anyway.)
Author: Rushabh Lathia <rushabh.lathia@gmail.com>
Author: Jian He <jian.universality@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Tested-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Discussion: https://postgr.es/m/CAGPqQf0KitkNack4F5CFkFi-9Dqvp29Ro=EpcWt=4_hs-Rt+bQ@mail.gmail.com
It seems potentially useful to label our shared libraries with version
information, now that a facility exists for retrieving that. This
patch labels them with the PG_VERSION string. There was some
discussion about using semantic versioning conventions, but that
doesn't seem terribly helpful for modules with no SQL-level presence;
and for those that do have SQL objects, we typically expect them
to support multiple revisions of the SQL definitions, so it'd still
not be very helpful.
I did not label any of src/test/modules/. It seems unnecessary since
we don't install those, and besides there ought to be someplace that
still provides test coverage for the original PG_MODULE_MAGIC macro.
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/dd4d1b59-d0fe-49d5-b28f-1e463b68fa32@gmail.com
Many STRICT function calls will have one or two arguments, in which
case we can speed up checking for NULL input by avoiding setting up
a loop over the arguments. This adds EEOP_FUNCEXPR_STRICT_1 and the
corresponding EEOP_FUNCEXPR_STRICT_2 for functions with one and two
arguments respectively.
Author: Andres Freund <andres@anarazel.de>
Co-authored-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://postgr.es/m/415721CE-7D2E-4B74-B5D9-1950083BA03E@yesql.se
Discussion: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
Knowing when the side-effects of an expression is the intended result
of the execution, rather than the returnvalue, is important for being
able generate more efficient JITed code. This replaces EEOP_DONE with
two new steps: EEOP_DONE_RETURN and EEOP_DONE_NO_RETURN. Expressions
which return a value should use the former step; expressions used for
their side-effects which don't return value should use the latter.
Author: Andres Freund <andres@anarazel.de>
Co-authored-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://postgr.es/m/415721CE-7D2E-4B74-B5D9-1950083BA03E@yesql.se
Discussion: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
Instead of deciding at runtime whether to read from casetest.value
or caseValue_datum, split EEOP_CASE_TESTVAL into two opcodes and
make the decision during expression compilation. Similarly for
EEOP_DOMAIN_TESTVAL. This actually results in net less code,
mainly because llvmjit_expr.c's code for handling these opcodes
gets shorter. The performance gain is doubtless negligible, but
this seems worth changing anyway on grounds of simplicity and
understandability.
Author: Andreas Karlsson <andreas@proxel.se>
Co-authored-by: Xing Guo <higuoxing@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CACpMh+AiBYAWn+D1aU7Rsy-V1tox06Cbc0H3qA7rwL5zdJ=anQ@mail.gmail.com
This allows the RETURNING list of INSERT/UPDATE/DELETE/MERGE queries
to explicitly return old and new values by using the special aliases
"old" and "new", which are automatically added to the query (if not
already defined) while parsing its RETURNING list, allowing things
like:
RETURNING old.colname, new.colname, ...
RETURNING old.*, new.*
Additionally, a new syntax is supported, allowing the names "old" and
"new" to be changed to user-supplied alias names, e.g.:
RETURNING WITH (OLD AS o, NEW AS n) o.colname, n.colname, ...
This is useful when the names "old" and "new" are already defined,
such as inside trigger functions, allowing backwards compatibility to
be maintained -- the interpretation of any existing queries that
happen to already refer to relations called "old" or "new", or use
those as aliases for other relations, is not changed.
For an INSERT, old values will generally be NULL, and for a DELETE,
new values will generally be NULL, but that may change for an INSERT
with an ON CONFLICT ... DO UPDATE clause, or if a query rewrite rule
changes the command type. Therefore, we put no restrictions on the use
of old and new in any DML queries.
Dean Rasheed, reviewed by Jian He and Jeff Davis.
Discussion: https://postgr.es/m/CAEZATCWx0J0-v=Qjc6gXzR=KtsdvAE7Ow=D=mu50AgOe+pvisQ@mail.gmail.com
RowCompareType served as a way to describe the fundamental meaning of
an operator, notionally independent of an operator class (although so
far this was only really supported for btrees). Its original purpose
was for use inside RowCompareExpr, and it has also found some small
use outside, such as for get_op_btree_interpretation().
We want to expand this now, as a more general way to describe operator
semantics for other index access methods, including gist (to improve
GistTranslateStratnum()) and others not written yet. To avoid future
confusion, we rename the type to CompareType and the symbols from
ROWCOMPARE_XXX to COMPARE_XXX to reflect their more general purpose.
Reviewed-by: Mark Dilger <mark.dilger@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
Here we convert CompactAttribute.attalign from a char, which is directly
derived from pg_attribute.attalign into a uint8, which stores the number
of bytes to align the column's value by in the tuple.
This allows tuple deformation and tuple size calculations to move away
from using the inefficient att_align_nominal() macro, which manually
checks each TYPALIGN_* char to translate that into the alignment bytes
for the given type. Effectively, this commit changes those to TYPEALIGN
calls, which are branchless and only perform some simple arithmetic with
some bit-twiddling.
The removed branches were often mispredicted by CPUs, especially so in
real-world tables which often contain a mishmash of different types
with different alignment requirements.
Author: David Rowley
Reviewed-by: Andres Freund, Victor Yegorov
Discussion: https://postgr.es/m/CAApHDvrBztXP3yx=NKNmo3xwFAFhEdyPnvrDg3=M0RhDs+4vYw@mail.gmail.com
The new compact_attrs array stores a few select fields from
FormData_pg_attribute in a more compact way, using only 16 bytes per
column instead of the 104 bytes that FormData_pg_attribute uses. Using
CompactAttribute allows performance-critical operations such as tuple
deformation to be performed without looking at the FormData_pg_attribute
element in TupleDesc which means fewer cacheline accesses.
For some workloads, tuple deformation can be the most CPU intensive part
of processing the query. Some testing with 16 columns on a table
where the first column is variable length showed around a 10% increase in
transactions per second for an OLAP type query performing aggregation on
the 16th column. However, in certain cases, the increases were much
higher, up to ~25% on one AMD Zen4 machine.
This also makes pg_attribute.attcacheoff redundant. A follow-on commit
will remove it, thus shrinking the FormData_pg_attribute struct by 4
bytes.
Author: David Rowley
Reviewed-by: Andres Freund, Victor Yegorov
Discussion: https://postgr.es/m/CAApHDvrBztXP3yx=NKNmo3xwFAFhEdyPnvrDg3=M0RhDs+4vYw@mail.gmail.com
The C standard says that sizeof(bool) is implementation-defined, but we
know of no current systems where it is not 1. The last known systems
seem to have been Apple macOS/PowerPC 10.5 and Microsoft Visual C++ 4,
both long defunct.
PostgreSQL has always required sizeof(bool) == 1 for the definition of
bool that it used, but previously it would define its own type if the
system-provided bool had a different size. That was liable to cause
memory layout problems when interacting with system and third-party
libraries on (by now hypothetical) computers with wider _Bool, and now
C23 has introduced a new problem by making bool a built-in datatype
(like C++), so the fallback code doesn't even compile. We could
probably work around that, but then we'd be writing new untested code
for a computer that doesn't exist.
Instead, delete the unreachable and C23-uncompilable fallback code, and
let existing static assertions fail if the system-provided bool is too
wide. If we ever get a problem report from a real system, then it will
be time to figure out what to do about it in a way that also works on
modern compilers.
Note on C++: Previously we avoided including <stdbool.h> or trying to
define a new bool type in headers that might be included by C++ code.
These days we might as well just include <stdbool.h> unconditionally:
it should be visible to C++11 but do nothing, just as in C23. We
already include <stdint.h> without C++ guards in c.h, and that falls
under the same C99-compatibility section of the C++11 standard as
<stdbool.h>, so let's remove the guards here too.
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/3198438.1731895163%40sss.pgh.pa.us
If passed a read-write expanded object pointer, the EEOP_NULLIF
code would hand that same pointer to the equality function
and then (unless equality was reported) also return the same
pointer as its value. This is no good, because a function that
receives a read-write expanded object pointer is fully entitled
to scribble on or even delete the object, thus corrupting the
NULLIF output. (This problem is likely unobservable with the
equality functions provided in core Postgres, but it's easy to
demonstrate with one coded in plpgsql.)
To fix, make sure the pointer passed to the equality function
is read-only. We can still return the original read-write
pointer as the NULLIF result, allowing optimization of later
operations.
Per bug #18722 from Alexander Lakhin. This has been wrong
since we invented expanded objects, so back-patch to all
supported branches.
Discussion: https://postgr.es/m/18722-fd9e645448cc78b4@postgresql.org
In EEOP_BOOL_AND_STEP* and EEOP_BOOL_OR_STEP*, we emitted pointlesss
store instructions to store to resnull/resvalue values that were just
loaded from the same fields in the previous instructions. They will
surely get optimized away by LLVM if any optimizations are enabled,
but it's better to not emit them in the first place. In
EEOP_BOOL_NOT_STEP, similar story with resnull.
In EEOP_NULLIF, when it returns NULL, there was also a redundant store
to resvalue just after storing a 0 to it. The value of resvalue
doesn't matter when resnull is set, so in fact even storing the 0 is
unnecessary, but I kept that because we tend to do that for general
tidiness.
Author: Xing Guo <higuoxing@gmail.com>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://www.postgresql.org/message-id/CACpMh%2BC%3Dg13WdvzLRSponsVWGgxwDSMzQWM4Gz0heOyaA0-N6g@mail.gmail.com
Supply a new memory manager for RuntimeDyld, to avoid crashes in
generated code caused by memory placement that can overflow a 32 bit
data type. This is a drop-in replacement for the
llvm::SectionMemoryManager class in the LLVM library, with Michael
Smith's proposed fix from
https://www.github.com/llvm/llvm-project/pull/71968.
We hereby slurp it into our own source tree, after moving into a new
namespace llvm::backport and making some minor adjustments so that it
can be compiled with older LLVM versions as far back as 12. It's harder
to make it work on even older LLVM versions, but it doesn't seem likely
that people are really using them so that is not investigated for now.
The problem could also be addressed by switching to JITLink instead of
RuntimeDyld, and that is the LLVM project's recommended solution as
the latter is about to be deprecated. We'll have to do that soon enough
anyway, and then when the LLVM version support window advances far
enough in a few years we'll be able to delete this code. Unfortunately
that wouldn't be enough for PostgreSQL today: in most relevant versions
of LLVM, JITLink is missing or incomplete.
Several other projects have already back-ported this fix into their fork
of LLVM, which is a vote of confidence despite the lack of commit into
LLVM as of today. We don't have our own copy of LLVM so we can't do
exactly what they've done; instead we have a copy of the whole patched
class so we can pass an instance of it to RuntimeDyld.
The LLVM project hasn't chosen to commit the fix yet, and even if it
did, it wouldn't be back-ported into the releases of LLVM that most of
our users care about, so there is not much point in waiting any longer
for that. If they make further changes and commit it to LLVM 19 or 20,
we'll still need this for older versions, but we may want to
resynchronize our copy and update some comments.
The changes that we've had to make to our copy can be seen by diffing
our SectionMemoryManager.{h,cpp} files against the ones in the tree of
the pull request. Per the LLVM project's license requirements, a copy
is in SectionMemoryManager.LICENSE.
This should fix the spate of crash reports we've been receiving lately
from users on large memory ARM systems.
Back-patch to all supported releases.
Co-authored-by: Thomas Munro <thomas.munro@gmail.com>
Co-authored-by: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Reviewed-by: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se> (license aspects)
Reported-by: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Discussion: https://postgr.es/m/CAO6_Xqr63qj%3DSx7HY6ZiiQ6R_JbX%2B-p6sTPwDYwTWZjUmjsYBg%40mail.gmail.com
adf97c156 made it so ExprStates could support hashing and changed Hash
Join to use that instead of manually extracting Datums from tuples and
hashing them one column at a time.
When hashing multiple columns or expressions, the code added in that
commit stored the intermediate hash value in the ExprState's resvalue
field. That was a mistake as steps may be injected into the ExprState
between each hashing step that look at or overwrite the stored
intermediate hash value. EEOP_PARAM_SET is an example of such a step.
Here we fix this by adding a new dedicated field for storing
intermediate hash values and adjust the code so that all apart from the
final hashing step store their result in the intermediate field.
In passing, rename a variable so that it's more aligned to the
surrounding code and also so a few lines stay within the 80 char margin.
Reported-by: Andres Freund
Reviewed-by: Alena Rybakina <a.rybakina@postgrespro.ru>
Discussion: https://postgr.es/m/CAApHDvqo9eenEFXND5zZ9JxO_k4eTA4jKMGxSyjdTrsmYvnmZw@mail.gmail.com
LLVM's opaque pointer change began in LLVM 14, but remained optional
until LLVM 16. When commit 37d5babb added opaque pointer support, we
didn't turn it on for LLVM 14 and 15 yet because we didn't want to risk
weird bitcode incompatibility problems in released branches of
PostgreSQL. (That might have been overly cautious, I don't know.)
Now that PostgreSQL 18 has dropped support for LLVM versions < 14, and
since it hasn't been released yet and no extensions or bitcode have been
built against it in the wild yet, we can be more aggressive. We can rip
out the support code and build system clutter that made opaque pointer
use optional.
Author: Thomas Munro <thomas.munro@gmail.com>
Reviewed-by: Peter Eisentraut <peter@eisentraut.org>
Discussions: https://postgr.es/m/CA%2BhUKGLhNs5geZaVNj2EJ79Dx9W8fyWUU3HxcpZy55sMGcY%3DiA%40mail.gmail.com
Here we add ExprState support for obtaining a 32-bit hash value from a
list of expressions. This allows both faster hashing and also JIT
compilation of these expressions. This is especially useful when hash
joins have multiple join keys as the previous code called ExecEvalExpr on
each hash join key individually and that was inefficient as tuple
deformation would have only taken into account one key at a time, which
could lead to walking the tuple once for each join key. With the new
code, we'll determine the maximum attribute required and deform the tuple
to that point only once.
Some performance tests done with this change have shown up to a 20%
performance increase of a query containing a Hash Join without JIT
compilation and up to a 26% performance increase when JIT is enabled and
optimization and inlining were performed by the JIT compiler. The
performance increase with 1 join column was less with a 14% increase
with and without JIT. This test was done using a fairly small hash
table and a large number of hash probes. The increase will likely be
less with large tables, especially ones larger than L3 cache as memory
pressure is more likely to be the limiting factor there.
This commit only addresses Hash Joins, but lays expression evaluation
and JIT compilation infrastructure for other hashing needs such as Hash
Aggregate.
Author: David Rowley
Reviewed-by: Alexey Dvoichenkov <alexey@hyperplane.net>
Reviewed-by: Tels <nospam-pg-abuse@bloodgate.com>
Discussion: https://postgr.es/m/CAApHDvoexAxgQFNQD_GRkr2O_eJUD1-wUGm%3Dm0L%2BGc%3DT%3DkEa4g%40mail.gmail.com
Otherwise, this would break if using C and C++ compilers from
different families and they understand different options. It already
used the right flags for compiling, this is only for linking. Also,
the meson setup already did this correctly.
Reported-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/228700.1722717983@sss.pgh.pa.us
Until now we generated an ExprState for each parameter to a SubPlan and
evaluated them one-by-one ExecScanSubPlan. That's sub-optimal as creating lots
of small ExprStates
a) makes JIT compilation more expensive
b) wastes memory
c) is a bit slower to execute
This commit arranges to evaluate parameters to a SubPlan as part of the
ExprState referencing a SubPlan, using the new EEOP_PARAM_SET expression
step. We emit one EEOP_PARAM_SET for each argument to a subplan, just before
the EEOP_SUBPLAN step.
It likely is worth using EEOP_PARAM_SET in other places as well, e.g. for
SubPlan outputs, nestloop parameters and - more ambitiously - to get rid of
ExprContext->domainValue/caseValue/ecxt_agg*. But that's for later.
Author: Andres Freund <andres@anarazel.de>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Alena Rybakina <lena.ribackina@yandex.ru>
Discussion: https://postgr.es/m/20230225214401.346ancgjqc3zmvek@awork3.anarazel.de
The problem fixed by commit 53c8d6c9 would have been noticed if we'd
been running LLVM's verify pass on generated IR. Doing so also reveals
a complaint about incorrect name mangling, fixed here. Only enabled for
LLVM 17+ because it uses the new pass manager API.
Suggested-by: Dmitry Dolgov <9erthalion6@gmail.com>
Discussion: https://postgr.es/m/CAFj8pRACpVFr7LMdVYENUkScG5FCYMZDDdSGNU-tch%2Bw98OxYg%40mail.gmail.com
Instead of looking up casts at parse time for converting the result
of JsonPath* query functions to the specified or the default
RETURNING type, always perform the conversion at runtime using either
the target type's input function or the function
json_populate_type().
There are two motivations for this change:
1. json_populate_type() coerces to types with typmod such that any
string values that exceed length limit cause an error instead of
silent truncation, which is necessary to be standard-conforming.
2. It was possible to end up with a cast expression that doesn't
support soft handling of errors causing bugs in the of handling
ON ERROR clause.
JsonExpr.coercion_expr which would store the cast expression is no
longer necessary, so remove.
Bump catversion because stored rules change because of the above
removal.
Reported-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Jian He <jian.universality@gmail.com>
Discussion: Discussion: https://postgr.es/m/202405271326.5a5rprki64aw%40alvherre.pgsql
The configure check for HAVE_DECL_LLVMORCREGISTERPERF was removed by
e9a9843e13, but some code guarded by it was left. (That commit
removed the "register" calls but left the "unregister" calls.) That
code cannot be reached anymore, so remove it.
Reported-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://www.postgresql.org/message-id/5539b16c-cff7-46d5-9621-c3fb6b549e9e@iki.fi
Commit 72559438 started copying more attributes from AttributeTemplate
to the functions we generate on the fly. In the case of deform
functions, which return void, this meant that "noundef", from
AttributeTemplate's return value (a Datum) was copied to a void type.
Older LLVM releases were OK with that, but LLVM 18 crashes.
Update our llvm_copy_attributes() function to skip copying the attribute
for the return value, if the target function returns void.
Thanks to Dmitry Dolgov for help chasing this down.
Back-patch to all supported releases, like 72559438.
Reported-by: Pavel Stehule <pavel.stehule@gmail.com>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Discussion: https://postgr.es/m/CAFj8pRACpVFr7LMdVYENUkScG5FCYMZDDdSGNU-tch%2Bw98OxYg%40mail.gmail.com
This introduces the following SQL/JSON functions for querying JSON
data using jsonpath expressions:
JSON_EXISTS(), which can be used to apply a jsonpath expression to a
JSON value to check if it yields any values.
JSON_QUERY(), which can be used to to apply a jsonpath expression to
a JSON value to get a JSON object, an array, or a string. There are
various options to control whether multi-value result uses array
wrappers and whether the singleton scalar strings are quoted or not.
JSON_VALUE(), which can be used to apply a jsonpath expression to a
JSON value to return a single scalar value, producing an error if it
multiple values are matched.
Both JSON_VALUE() and JSON_QUERY() functions have options for
handling EMPTY and ERROR conditions, which can be used to specify
the behavior when no values are matched and when an error occurs
during jsonpath evaluation, respectively.
Author: Nikita Glukhov <n.gluhov@postgrespro.ru>
Author: Teodor Sigaev <teodor@sigaev.ru>
Author: Oleg Bartunov <obartunov@gmail.com>
Author: Alexander Korotkov <aekorotkov@gmail.com>
Author: Andrew Dunstan <andrew@dunslane.net>
Author: Amit Langote <amitlangote09@gmail.com>
Author: Peter Eisentraut <peter@eisentraut.org>
Author: Jian He <jian.universality@gmail.com>
Reviewers have included (in no particular order):
Andres Freund, Alexander Korotkov, Pavel Stehule, Andrew Alsup,
Erik Rijkers, Zihong Yu, Himanshu Upadhyaya, Daniel Gustafsson,
Justin Pryzby, Álvaro Herrera, Jian He, Anton A. Melnikov,
Nikita Malakhov, Peter Eisentraut, Tomas Vondra
Discussion: https://postgr.es/m/cd0bb935-0158-78a7-08b5-904886deac4b@postgrespro.ru
Discussion: https://postgr.es/m/20220616233130.rparivafipt6doj3@alap3.anarazel.de
Discussion: https://postgr.es/m/abd9b83b-aa66-f230-3d6d-734817f0995d%40postgresql.org
Discussion: https://postgr.es/m/CA+HiwqHROpf9e644D8BRqYvaAPmgBZVup-xKMDPk-nd4EpgzHw@mail.gmail.com
Discussion: https://postgr.es/m/CA+HiwqE4XTdfb1nW=Ojoy_tQSRhYt-q_kb6i5d4xcKyrLC1Nbg@mail.gmail.com
