clang only uses the 'i1' type for scalar booleans, not for pointers to
booleans (as the pointer might be pointing into a larger memory
allocation). Therefore a pointer-to-bool needs to the "storage" boolean.
There's no known case of wrong code generation due to this, but it seems quite
possible that it could cause problems (see e.g. 72559438f9).
Author: Andres Freund
Discussion: https://postgr.es/m/20201207212142.wz5tnbk2jsaqzogb@alap3.anarazel.de
Backpatch: 11-, where jit support was added
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 72559438f9 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 72559438f9
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
When commit bd3daddaf introduced AlternativeSubPlans, I had some
ambitions towards allowing the choice of subplan to change during
execution. That has not happened, or even been thought about, in the
ensuing twelve years; so it seems like a failed experiment. So let's
rip that out and resolve the choice of subplan at the end of planning
(in setrefs.c) rather than during executor startup. This has a number
of positive benefits:
* Removal of a few hundred lines of executor code, since
AlternativeSubPlans need no longer be supported there.
* Removal of executor-startup overhead (particularly, initialization
of subplans that won't be used).
* Removal of incidental costs of having a larger plan tree, such as
tree-scanning and copying costs in the plancache; not to mention
setrefs.c's own costs of processing the discarded subplans.
* EXPLAIN no longer has to print a weird (and undocumented)
representation of an AlternativeSubPlan choice; it sees only the
subplan actually used. This should mean less confusion for users.
* Since setrefs.c knows which subexpression of a plan node it's
working on at any instant, it's possible to adjust the estimated
number of executions of the subplan based on that. For example,
we should usually estimate more executions of a qual expression
than a targetlist expression. The implementation used here is
pretty simplistic, because we don't want to expend a lot of cycles
on the issue; but it's better than ignoring the point entirely,
as the executor had to.
That last point might possibly result in shifting the choice
between hashed and non-hashed EXISTS subplans in a few cases,
but in general this patch isn't meant to change planner choices.
Since we're doing the resolution so late, it's really impossible
to change any plan choices outside the AlternativeSubPlan itself.
Patch by me; thanks to David Rowley for review.
Discussion: https://postgr.es/m/1992952.1592785225@sss.pgh.pa.us
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
303640199d and back-patched as commit
bcbc27251d and commit
f7013683d9, so for now, no back-patch.
Bharath Rupireddy
Discussion: http://postgr.es/m/CALj2ACWk7j4F2v2fxxYfrroOF=AdFNPr1WsV+AGtHAFQOqm_pw@mail.gmail.com
Includes some manual cleanup of places that pgindent messed up,
most of which weren't per project style anyway.
Notably, it seems some people didn't absorb the style rules of
commit c9d297751, because there were a bunch of new occurrences
of function calls with a newline just after the left paren, all
with faulty expectations about how the rest of the call would get
indented.
Optionally push a step to check for a NULL pointer to the pergroup
state.
This will be important for disk-based hash aggregation in combination
with grouping sets. When memory limits are reached, a given tuple may
find its per-group state for some grouping sets but not others. For
the former, it advances the per-group state as normal; for the latter,
it skips evaluation and the calling code will have to spill the tuple
and reprocess it in a later batch.
Add the NULL check as a separate expression step because in some
common cases it's not needed.
Discussion: https://postgr.es/m/20200221202212.ssb2qpmdgrnx52sj%40alap3.anarazel.de
Our usual practice for "poor man's enum" catalog columns is to define
macros for the possible values and use those, not literal constants,
in C code. But for some reason lost in the mists of time, this was
never done for typalign/attalign or typstorage/attstorage. It's never
too late to make it better though, so let's do that.
The reason I got interested in this right now is the need to duplicate
some uses of the TYPSTORAGE constants in an upcoming ALTER TYPE patch.
But in general, this sort of change aids greppability and readability,
so it's a good idea even without any specific motivation.
I may have missed a few places that could be converted, and it's even
more likely that pending patches will re-introduce some hard-coded
references. But that's not fatal --- there's no expectation that
we'd actually change any of these values. We can clean up stragglers
over time.
Discussion: https://postgr.es/m/16457.1583189537@sss.pgh.pa.us
Do so by combining the various steps that are part of aggregate
transition function invocation into one larger step. As some of the
current steps are only necessary for some aggregates, have one variant
of the aggregate transition step for each possible combination.
To avoid further manual copies of code in the different transition
step implementations, move most of the code into helper functions
marked as "always inline".
The benefit of this change is an increase in performance when
aggregating lots of rows. This comes in part due to the reduced number
of indirect jumps due to the reduced number of steps, and in part by
reducing redundant setup code across steps. This mainly benefits
interpreted execution, but the code generated by JIT is also improved
a bit.
As a nice side-effect it also ends up making the code a bit simpler.
A small additional optimization is removing the need to set
aggstate->curaggcontext before calling ExecAggInitGroup, choosing to
instead passign curaggcontext as an argument. It was, in contrast to
other aggregate related functions, only needed to fetch a memory
context to copy the transition value into.
Author: Andres Freund
Discussion:
https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.dehttps://postgr.es/m/5c371df7cee903e8cd4c685f90c6c72086d3a2dc.camel@j-davis.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
This merges the code emission for a number of opcodes by handling the
behavioural difference more locally. This reduces code, and also
improves the generated code a bit in some cases, by removing redundant
constants.
Author: Andres Freund
Discussion: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
It's already tracked via ExprState->parent, so we don't need to also
include it in ExprEvalStep. When that code originally was written
ExprState->parent didn't exist, but it since has been introduced in
6719b238e8.
Author: Andres Freund
Discussion: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
This mostly consists of using C99 style for loops, moving variables
into narrower scopes, and a smattering of other minor improvements.
Done separately to make it easier to review patches with actual
functional changes.
Author: Andres Freund
Discussion: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
Tuple conversion support in tupconvert.c is able to convert rowtypes
between two relations, inner and outer, which are logically equivalent
but have a different ordering or even dropped columns (used mainly for
inheritance tree and partitions). This makes use of attribute mappings,
which are simple arrays made of AttrNumber elements with a length
matching the number of attributes of the outer relation. The length of
the attribute mapping has been treated as completely independent of the
mapping itself until now, making it easy to pass down an incorrect
mapping length.
This commit refactors the code related to attribute mappings and moves
it into an independent facility called attmap.c, extracted from
tupconvert.c. This merges the attribute mapping with its length,
avoiding to try to guess what is the length of a mapping to use as this
is computed once, when the map is built.
This will avoid mistakes like what has been fixed in dc816e58, which has
used an incorrect mapping length by matching it with the number of
attributes of an inner relation (a child partition) instead of an outer
relation (a partitioned table).
Author: Michael Paquier
Reviewed-by: Amit Langote
Discussion: https://postgr.es/m/20191121042556.GD153437@paquier.xyz
When maintaining or merging patches, one of the most common sources
for conflicts are the list of objects in makefiles. Especially when
the split across lines has been changed on both sides, which is
somewhat common due to attempting to stay below 80 columns, those
conflicts are unnecessarily laborious to resolve.
By splitting, and alphabetically sorting, OBJS style lines into one
object per line, conflicts should be less frequent, and easier to
resolve when they still occur.
Author: Andres Freund
Discussion: https://postgr.es/m/20191029200901.vww4idgcxv74cwes@alap3.anarazel.de
That avoids unnecessary work during both interpreted execution, and
JIT compiled expression evaluation. Both benefit from fewer expression
steps needing be processed, and for interpreted execution there now is
a fastpath dedicated to just fetching a value from a virtual
slot. That's e.g. beneficial for hashjoins over nodes that perform
projections, as the hashed columns are currently fetched individually.
Author: Soumyadeep Chakraborty, Andres Freund
Discussion: https://postgr.es/m/CAE-ML+9OKSN71+mHtfMD-L24oDp8dGTfaVjDU6U+j+FNAW5kRQ@mail.gmail.com
Bleeding-edge LLVM has stopped supplying replacements for various
C++14 library features, for people on older C++ versions. Since we're
not ready to require C++14 yet, just use plain old new instead of
make_unique. As revealed by buildfarm animal seawasp.
Back-patch to 11.
Reviewed-by: Andres Freund
Discussion: https://postgr.es/m/CA%2BhUKGJWG7unNqmkxg7nC5o3o-0p2XP6co4r%3D9epqYMm8UY4Mw%40mail.gmail.com
Early previews of LLVM 9 reveal that our Min() macro causes compiler
errors in LLVM headers reached by the #include directives in
llvmjit_inline.cpp. Let's just undefine it. Per buildfarm animal
seawasp. Back-patch to 11.
Reviewed-by: Fabien Coelho, Tom Lane
Discussion: https://postgr.es/m/20190606173216.GA6306%40alvherre.pgsql
While discussing comment improvements (see next commit) by Justin
Pryzby, Tom complained about a few details of the logic to infer the
length of the NULL bitmap when building the JITed tuple deforming
function. That bitmap allows to avoid checking the tuple header's
natts, a check which often causes a pipeline stall
Improvements:
a) As long as missing columns aren't taken into account, we can
continue to infer the length of the NULL bitmap from NOT NULL
columns following it. Previously we stopped at the first missing
column. It's unlikely to matter much in practice, but the
alternative would have been to document why we stop.
b) For robustness reasons it seems better to also check against
attisdropped - RemoveAttributeById() sets attnotnull to false, but
an additional check is trivial.
c) Improve related comments
Discussion: https://postgr.es/m/20637.1555957068@sss.pgh.pa.us
Backpatch: -
Create a new header optimizer/optimizer.h, which exposes just the
planner functions that can be used "at arm's length", without need
to access Paths or the other planner-internal data structures defined
in nodes/relation.h. This is intended to provide the whole planner
API seen by most of the rest of the system; although FDWs still need
to use additional stuff, and more thought is also needed about just
what selfuncs.c should rely on.
The main point of doing this now is to limit the amount of new
#include baggage that will be needed by "planner support functions",
which I expect to introduce later, and which will be in relevant
datatype modules rather than anywhere near the planner.
This commit just moves relevant declarations into optimizer.h from
other header files (a couple of which go away because everything
got moved), and adjusts #include lists to match. There's further
cleanup that could be done if we want to decide that some stuff
being exposed by optimizer.h doesn't belong in the planner at all,
but I'll leave that for another day.
Discussion: https://postgr.es/m/11460.1548706639@sss.pgh.pa.us
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
The function generated to perform JIT compiled tuple deforming failed
when HeapTupleHeader's t_hoff was bigger than a signed int8. I'd
failed to realize that LLVM's getelementptr would treat an int8 index
argument as signed, rather than unsigned. That means that a hoff
larger than 127 would result in a negative offset being applied. Fix
that by widening the index to 32bit.
Add a testcase with a wide table. Don't drop it, as it seems useful to
verify other tools deal properly with wide tables.
Thanks to Justin Pryzby for both reporting a bug and then reducing it
to a reproducible testcase!
Reported-By: Justin Pryzby
Author: Andres Freund
Discussion: https://postgr.es/m/20181115223959.GB10913@telsasoft.com
Backpatch: 11, just as jit compilation was
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
Virtual tuple table slots never need tuple deforming. Therefore, if we
know at expression compilation time, that a certain slot will always
be virtual, there's no need to create a tuple deforming routine for
it.
Author: Andres Freund
Discussion: https://postgr.es/m/20181105210039.hh4vvi4vwoq5ba2q@alap3.anarazel.de
Previously this information was computed when JIT compiling an
expression. But the information is useful for assertions in the
non-JIT case too (for assertions), therefore it makes sense to move
it.
This will, in a followup commit, allow to treat different slot types
differently. E.g. for virtual slots there's no need to generate a JIT
function to deform the slot.
Author: Andres Freund
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
This only became a problem with 4c640f4f38, which didn't synchronize
the value agg_strict_input_check.nargs is set to, with the guard
condition for emitting the operation.
Besides such instructions being unnecessary overhead, currently the
LLVM JIT provider doesn't support them. It seems more sensible to
avoid generating such instruction than supporting them. Add assertions
to make it easier to debug a potential further occurance.
Discussion: https://postgr.es/m/2a505161-2727-2473-7c46-591ed108ac52@email.cz
Backpatch: 11-, like 4c640f4f38.