The WHERE clause introduced by 31f38174 was not described.  While on it,
split the grammar of \copy FROM and TO into two distinct parts for
clarity as they support different set of options.

Author: Vignesh C
Discussion: https://postgr.es/m/CALDaNm3zWr=OmxeNqOqfT=uZTSdam_j-gkX94CL8eTNfgUtf6A@mail.gmail.com
Backpatch-through: 12

The term "hash_mem" will take on new significance when pending work to
add a new hash_mem_multiplier GUC is committed.  Rename a local variable
that happens to have been called hash_mem now to avoid confusion.

Teach Valgrind memcheck to maintain the "defined-ness" of each shared
buffer based on whether the backend holds at least one pin at the point
it is accessed by access method code.  Bugs like the one fixed by commit
b0229f26 can be detected using this new instrumentation.

Note that backends running with Valgrind naturally have their own
independent ideas about whether any given byte in shared memory is safe
or unsafe to access.  There is no risk that concurrent access by
multiple backends to the same shared memory will confuse Valgrind's
instrumentation, because everything already works at the process level
(or at the memory mapping level, if you prefer).

Author: Álvaro Herrera, Peter Geoghegan
Reviewed-By: Anastasia Lubennikova
Discussion: https://postgr.es/m/20150723195349.GW5596@postgresql.org
Discussion: https://postgr.es/m/CAH2-WzkLgyN3zBvRZ1pkNJThC=xi_0gpWRUb_45eexLH1+k2_Q@mail.gmail.com

pg_dump produces custom-format archive files that lack data offsets
when it is unable to seek its output.  Up to now that's been a hazard
for pg_restore.  But if pg_restore is able to seek in the archive
file, there is no reason to throw up our hands when asked to restore
data blocks out of order.  Instead, whenever we are searching for a
data block, record the locations of the blocks we passed over (that
is, fill in the missing data-offset fields in our in-memory copy of
the TOC data).  Then, when we hit a case that requires going
backwards, we can just seek back.

Also track the furthest point that we've searched to, and seek back
to there when beginning a search for a new data block.  This avoids
possible O(N^2) time consumption, by ensuring that each data block
is examined at most twice.  (On Unix systems, that's at most twice
per parallel-restore job; but since Windows uses threads here, the
threads can share block location knowledge, reducing the amount of
duplicated work.)

We can also improve the code a bit by using fseeko() to skip over
data blocks during the search.

This is all of some use even in simple restores, but it's really
significant for parallel pg_restore.  In that case, we require
seekability of the input already, and we will very probably need
to do out-of-order restores.

Back-patch to v12, as this fixes a regression introduced by commit
548e5097.  Before that, parallel restore avoided requesting
out-of-order restores, so it would work on a data-offset-less
archive.  Now it will again.

Ideally this patch would include some test coverage, but there are
other open bugs that need to be fixed before we can extend our
coverage of parallel restore very much.  Plan to revisit that later.

David Gilman and Tom Lane; reviewed by Justin Pryzby

Discussion: https://postgr.es/m/CALBH9DDuJ+scZc4MEvw5uO-=vRyR2=QF9+Yh=3hPEnKHWfS81A@mail.gmail.com

We do not really need to track the file position by hand.  We were
already relying on ftello() whenever the archive file is seekable,
while if it's not seekable we don't need the file position info
anyway because we're not going to be able to re-write the TOC.

Moreover, that tracking was buggy since it failed to account for
the effects of fseeko().  Somewhat remarkably, that seems not to
have made for any live bugs up to now.  We could fix the oversights,
but it seems better to just get rid of the whole error-prone mess.

In itself this is merely code cleanup.  However, it's necessary
infrastructure for an upcoming bug-fix patch (because that code
*does* need valid file position after fseeko).  The bug fix
needs to go back as far as v12; hence, back-patch that far.

Discussion: https://postgr.es/m/CALBH9DDuJ+scZc4MEvw5uO-=vRyR2=QF9+Yh=3hPEnKHWfS81A@mail.gmail.com

There is no advantage to attempting deduplication for a unique index
during CREATE INDEX, since there cannot possibly be any duplicates.
Doing so wastes cycles due to unnecessary copying.  Make sure that we
avoid it consistently.

We already avoided unique index deduplication in the case where there
were some spool2 tuples to merge.  That didn't account for the fact that
spool2 is removed early/unset in the common case where it has no tuples
that need to be merged (i.e. it failed to account for the "spool2 turns
out to be unnecessary" optimization in _bt_spools_heapscan()).

Oversight in commit 0d861bbb, which added nbtree deduplication

Backpatch: 13-, where nbtree deduplication was introduced.

We had two occurrences of "Mitteleuropäische Zeit" in Europe.txt,
though the corresponding entries in Default were spelled
"Mitteleuropaeische Zeit".  Standardize on the latter spelling to
avoid questions of which encoding to use.

While here, correct a couple of other trivial inconsistencies between
the Default file and the supposedly-matching entries in the *.txt
files, as exposed by some checking with comm(1).  Also, add BDST to
the Europe.txt file; it previously was only listed in Default.
None of this has any direct functional effect.

Per complaint from Christoph Berg.  As usual for timezone data patches,
apply to all branches.

Discussion: https://postgr.es/m/20200716100743.GE3534683@msg.df7cb.de

Commit 1e53fe0e has unified the usage of the config-file reload flag by
using the same signal handler function for the SIGHUP signal at many places
in the code.  By mistake, it used the wrong SIGNAL in apply launcher
process for the SIGHUP signal handler function.

Author: Bharath Rupireddy
Reviewed-by: Dilip Kumar
Backpatch-through: 13, where it was introduced
Discussion: https://postgr.es/m/CALj2ACVzHCRnS20bOiEHaLtP5PVBENZQn4khdsSJQgOv_GM-LA@mail.gmail.com

This representation saves 8 bytes per tuple compared to HeapTuple, and
avoids the need to allocate, copy and free on the receiving side.

Gather can emit the returned MinimalTuple directly, but GatherMerge now
needs to make an explicit copy because it buffers multiple tuples at a
time.  That should be no worse than before.
Reviewed-by: Soumyadeep Chakraborty <soumyadeep2007@gmail.com>
Discussion: https://postgr.es/m/CA%2BhUKG%2B8T_ggoUTAE-U%3DA%2BOcPc4%3DB0nPPHcSfffuQhvXXjML6w%40mail.gmail.com

This allows the huge page size to be set explicitly.  The default is 0,
meaning it will use the system default, as before.

Author: Odin Ugedal <odin@ugedal.com>
Discussion: https://postgr.es/m/20200608154639.20254-1-odin%40ugedal.com

Windows has junction points which function as symbolic links for
directories. This patch introduces a new function TestLib::dir_symlink()
which creates a junction point on Windows and a standard Unix type
symbolic link elsewhere.

The function TestLib::perl2host is also modified, first to use cygpath
where it's available (e.g. msys2) and second to allow it to succeed if
the gandparent directory exists but the parent does not.

Given these changes the only symlink tests that need to be skipped on
Windows are those related to permissions or to use of readlink. The
relevant tests for pg_basebackup and pg_rewind are therefore adjusted
accordingly.

Andrew Dunstan, reviewed by Peter Eisentraut and Michael Paquier.

Discussion: https://postgr.es/m/c50a646c-d9bb-7c62-a4bf-8256ff6ff338@2ndquadrant.com

pg_test_fsync has always opened files using the text mode on Windows, as
this is the default mode used if not enforced by _setmode().

This fixes a failure when running pg_test_fsync down to 12 because
O_DSYNC and the text mode are not able to work together nicely.  We
fixed the handling of O_DSYNC in 12~ for the tool by switching to the
concurrent-safe version of fopen() in src/port/ with 0ba06e0b.  And
40cfe860, by enforcing the text mode for compatibility reasons if O_TEXT
or O_BINARY are not specified by the caller, broke pg_test_fsync.  For
all versions, this avoids any translation overhead, and pg_test_fsync
should test binary writes, so it is a gain in all cases.

Note that O_DSYNC is still not handled correctly in ~11, leading to
pg_test_fsync to show insanely high numbers for open_datasync() (using
this property it is easy to notice that the binary mode is much
faster).  This would require a backpatch of 0ba06e0b and 40cfe860, which
could potentially break existing applications, so this is left out.

There are no TAP tests for this tool yet, so I have checked all builds
manually using MSVC.  We could invent a new option to run a single
transaction instead of using a duration of 1s to make the tests a
maximum short, but this is left as future work.

Thanks to Bruce Momjian for the discussion.

Reported-by: Jeff Janes
Author: Michael Paquier
Discussion: https://postgr.es/m/16526-279ded30a230d275@postgresql.org
Backpatch-through: 9.5

Similar to daa9fe8a, instead of
repeating the almost same large query in each version branch, use one
query and add a few columns to the SELECT list depending on the
version.  This saves a lot of duplication.
Reviewed-by: Fabien COELHO <coelho@cri.ensmp.fr>
Discussion: https://www.postgresql.org/message-id/flat/6594334b-40fd-14f1-6bc5-877afa3feed5@2ndquadrant.com

When working with an online source cluster, pg_rewind gets a list of all
the files in the source data directory using a WITH RECURSIVE query,
returning a NULL result for a file's metadata if it gets removed between
the moment it is listed in a directory and the moment its metadata is
obtained with pg_stat_file() (say a recycled WAL segment).  The query
result was processed in such a way that for each tuple we checked only
that the first file's metadata was NULL.  This could have two
consequences, both resulting in a failure of the rewind:
- If the first tuple referred to a removed file, all files from the
source would be ignored.
- Any file actually missing would not be considered as such.

While on it, rework slightly the code so as no values are saved if we
know that a file is going to be skipped.

Issue introduced by b36805f3, so backpatch down to 9.5.

Author: Justin Pryzby, Michael Paquier
Reviewed-by: Daniel Gustafsson, Masahiko Sawada
Discussion: https://postgr.es/m/20200713061010.GC23581@telsasoft.com
Backpatch-through: 9.5

One change for getObjectIdentity() has been missed in 2a10fdc4, causing
the module to not compile properly.  This was actually the only problem,
and it happens that it is easy enough to check the compilation of the
module on Debian after installing libselinux1-dev.

Per buildfarm member rhinoceros.

When using the following functions, users could see various types of
errors of the type "cache lookup failed for OID XXX" with elog(), that
can only be used for internal errors:
* pg_describe_object()
* pg_identify_object()
* pg_identify_object_as_address()

The set of APIs managing object addresses for all object types are made
smarter by gaining a new argument "missing_ok" that allows any caller to
control if an error is raised or not on an undefined object.  The SQL
functions listed above are changed to handle the case where an object is
missing.

Regression tests are added for all object types for the cases where
these are undefined.  Before this commit, these cases failed with cache
lookup errors, and now they basically return NULL (minus the name of the
object type requested).

Author: Michael Paquier
Reviewed-by: Aleksander Alekseev, Dmitry Dolgov, Daniel Gustafsson,
Álvaro Herrera, Kyotaro Horiguchi
Discussion: https://postgr.es/m/CAB7nPqSZxrSmdHK-rny7z8mi=EAFXJ5J-0RbzDw6aus=wB5azQ@mail.gmail.com

reparameterize_path_by_child() failed to reparameterize BitmapAnd
and BitmapOr paths.  This matters only if such a path is chosen as
the inside of a nestloop partition-wise join, where we have to pass
in parameters from the outside of the nestloop.  If that did happen,
we generated a bad plan that would likely lead to crashes at execution.

This is not entirely reparameterize_path_by_child()'s fault though;
it's the victim of an ancient decision (my ancient decision, I think)
to not bother filling in param_info in BitmapAnd/Or path nodes.  That
caused the function to believe that such nodes and their children
contain no parameter references and so need not be processed.

In hindsight that decision looks pretty penny-wise and pound-foolish:
while it saves a few cycles during path node setup, we do commonly
need the information later.  In particular, by reversing the decision
and requiring valid param_info data in all nodes of a bitmap path
tree, we can get rid of indxpath.c's get_bitmap_tree_required_outer()
function, which computed the data on-demand.  It's not unlikely that
that nets out as a savings of cycles in many scenarios.  A couple
of other things in indxpath.c can be simplified as well.

While here, get rid of some cases in reparameterize_path_by_child()
that are visibly dead or useless, given that we only care about
reparameterizing paths that can be on the inside of a parameterized
nestloop.  This case reminds one of the maxim that untested code
probably does not work, so I'm unwilling to leave unreachable code
in this function.  (I did leave the T_Gather case in place even
though it's not reached in the regression tests.  It's not very
clear to me when the planner might prefer to put Gather below
rather than above a nestloop, but at least in principle the case
might be interesting.)

Per bug #16536, originally from Arne Roland but with a test case
by Andrew Gierth.  Back-patch to v11 where this code came in.

Discussion: https://postgr.es/m/16536-2213ee0b3aad41fd@postgresql.org

Three groups of issues needed to be addressed:

load_external_function() and related functions returned PGFunction,
even though not necessarily all callers are looking for a function of
type PGFunction.  Since these functions are really just wrappers
around dlsym(), change to return void * just like dlsym().

In dynahash.c, we are using strlcpy() where a function with a
signature like memcpy() is expected.  This should be safe, as the new
comment there explains, but the cast needs to be augmented to avoid
the warning.

In PL/Python, methods all need to be cast to PyCFunction, per Python
API, but this now runs afoul of these warnings.  (This issue also
exists in core CPython.)

To fix the second and third case, we add a new type pg_funcptr_t that
is defined specifically so that gcc accepts it as a special function
pointer that can be cast to any other function pointer without the
warning.

Also add -Wcast-function-type to the standard warning flags, subject
to configure check.
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/flat/1e97628e-6447-b4fd-e230-d109cec2d584%402ndquadrant.com

Removing the unused 'miinfo' parameter has been raised a couple of times
now.  It was decided in the 2nd discussion below that we're going to leave
it alone.  It seems like it might be useful to add a comment to mention
this fact so that nobody wastes any time in the future proposing its
removal again.

Discussion: https://postgr.es/m/CAApHDvpCf-qR5HC1rXskUM4ToV+3YDb4-n1meY=vpAHsRS_1PA@mail.gmail.com
Discussion: https://postgr.es/m/CAE9k0P%3DFvcDswnSVtRpSyZMpcAWC%3DGp%3DifZ0HdfPaRQ%3D__LBtw%40mail.gmail.com

An ALTER TABLE to validate a foreign key in which another subcommand
already caused a pending table rewrite could fail due to ALTER TABLE
attempting to validate the foreign key before the actual table rewrite
takes place.  This situation could result in an error such as:

ERROR:  could not read block 0 in file "base/nnnnn/nnnnn": read only 0 of 8192 bytes

The failure here was due to the SPI call which validates the foreign key
trying to access an index which is yet to be rebuilt.

Similarly, we also incorrectly tried to validate CHECK constraints before
the heap had been rewritten.

The fix for both is to delay constraint validation until phase 3, after
the table has been rewritten.  For CHECK constraints this means a slight
behavioral change.  Previously ALTER TABLE VALIDATE CONSTRAINT on
inheritance tables would be validated from the bottom up.  This was
different from the order of evaluation when a new CHECK constraint was
added.  The changes made here aligns the VALIDATE CONSTRAINT evaluation
order for inheritance tables to be the same as ADD CONSTRAINT, which is
generally top-down.

Reported-by: Nazli Ugur Koyluoglu, using SQLancer
Discussion: https://postgr.es/m/CAApHDvp%3DZXv8wiRyk_0rWr00skhGkt8vXDrHJYXRMft3TjkxCA%40mail.gmail.com
Backpatch-through: 9.5 (all supported versions)

The following header files missed the shot:
- jsonfuncs.h, as of ce0425b1.
- jsonapi.h, as of beb46990.
- llvmjit_emit.h as of 7ec0d80c.
- partdesc.h, as of 1bb5e782.

Author: Jesse Zhang
Discussion: https://postgr.es/m/CAGf+fX4-8xULEOz09DE2dZGjT+q8VJ--rqfTpvcFwc+A4fc-3Q@mail.gmail.com

Incorrect function names were referenced.  As this fixes some portions
of tableam.h, that is mentioned in the docs as something to look at when
implementing a table AM, backpatch down to 12 where this has been
introduced.

Author: Hironobu Suzuki
Discussion: https://postgr.es/m/8fe6d672-28dd-3f1d-7aed-ac2f6d599d3f@interdb.jp
Backpatch-through: 12

The qual pushdown logic assumed that all Vars in a restriction clause
must be Vars referencing subquery outputs; but since we introduced
LATERAL, it's possible for such a Var to be a lateral reference instead.
This led to an assertion failure in debug builds.  In a non-debug
build, there might be no ill effects (if qual_is_pushdown_safe decided
the qual was unsafe anyway), or we could get failures later due to
construction of an invalid plan.  I've not gone to much length to
characterize the possible failures, but at least segfaults in the
executor have been observed.

Given that this has been busted since 9.3 and it took this long for
anybody to notice, I judge that the case isn't worth going to great
lengths to optimize.  Hence, fix by just teaching qual_is_pushdown_safe
that such quals are unsafe to push down, matching the previous behavior
when it accidentally didn't fail.

Per report from Tom Ellis.  Back-patch to all supported branches.

Discussion: https://postgr.es/m/20200713175124.GQ8220@cloudinit-builder

Remove previous hack in KeepLogSeg that added a case to deal with a
(badly represented) invalid segment number.  This was added for the sake
of GetWALAvailability.  But it's not needed if in that function we
initialize the segment number to be retreated to the currently being
written segment, so do that instead.

Per valgrind-running buildfarm member skink, and some sparc64 animals.

Discussion: https://postgr.es/m/1724648.1594230917@sss.pgh.pa.us

GSS-related resources should be cleaned up in pqDropConnection,
not freePGconn, else the wrong things happen when resetting
a connection or trying to switch to a different server.
It's also critical to reset conn->gssenc there.

During connection setup, initialize conn->try_gss at the correct
place, else switching to a different server won't work right.

Remove now-redundant cleanup of GSS resources around one (and, for
some reason, only one) pqDropConnection call in connectDBStart.

Per report from Kyotaro Horiguchi that psql would freeze up,
rather than successfully resetting a GSS-encrypted connection
after a server restart.

This is YA oversight in commit b0b39f72, so back-patch to v12.

Discussion: https://postgr.es/m/20200710.173803.435804731896516388.horikyota.ntt@gmail.com

 * Strategy number and purpose are essential information for opfamily operator.
   So, show those columns in non-verbose output.
 * "Left/right arg type" \dAp column names are confusing, because those type
   don't necessary match to function arguments.  Rename them to "Registered
   left/right type".
 * Replace manual assembling of operator/procedure names with casts to
   regoperator/regprocedure.
 * Add schema-qualification for pg_catalog functions and tables.

Reported-by: Peter Eisentraut, Tom Lane
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/2edc7b27-031f-b2b6-0db2-864241c91cb9%402ndquadrant.com
Backpatch-through: 13

This is a replacement for 4cad2534. Instead of projecting all tuples
going into a HashAgg, only remove unnecessary attributes when actually
spilling. This avoids the regression for the in-memory case.

Discussion: https://postgr.es/m/a2fb7dfeb4f50aa0a123e42151ee3013933cb802.camel%40j-davis.com
Backpatch-through: 13

This reverts commit 4cad2534 due to a
performance regression. It will be replaced by a new approach in an
upcoming commit.

Reported-by: Andres Freund
Discussion: https://postgr.es/m/20200614181418.mx4bvljmfkkhoqzl@alap3.anarazel.de
Backpatch-through: 13

The stats with this commit was available only for WALSenders, however,
users might want to see for backends doing logical decoding via SQL API.
Then, users might want to reset and access these stats across server
restart which was not possible with the current patch.

List of commits reverted:

caa3c424   Don't call elog() while holding spinlock.
e641b2a9   Doc: Update the documentation for spilled transaction
statistics.
5883f5fe   Fix unportable printf format introduced in commit 9290ad19.
9290ad19   Track statistics for spilling of changes from ReorderBuffer.

Additionaly, remove the release notes entry for this feature.

Backpatch-through: 13, where it was introduced
Discussion: https://postgr.es/m/CA+fd4k5_pPAYRTDrO2PbtTOe0eHQpBvuqmCr8ic39uTNmR49Eg@mail.gmail.com

Switching the regression tests to use tstzrange() has proved to not be a
good idea for environments where the timestamp precision is low, as
internal range checks exclude the upper bound.  So, if the commit
timestamp of a transaction matched with now() from the next query,
the test would fail.  This changes to use two bound checks instead of
the range function, where the upper bound is inclusive.

Per buildfarm member jacana.

Discussion: https://postgr.es/m/20200712122507.GD21680@paquier.xyz

Replication origins created by regression tests should have names
starting with "regress_", and the test introduced in b1e48bbe for commit
timestamps did not do that.

Per buildfarm member longfin.

Discussion: https://postgr.es/m/20200712122507.GD21680@paquier.xyz

This includes two changes:
- Addition of a new function pg_xact_commit_timestamp_origin() able, for
a given transaction ID, to return the commit timestamp and replication
origin of this transaction.  An equivalent function existed in
pglogical.
- Addition of the replication origin to pg_last_committed_xact().

The commit timestamp manager includes already APIs able to return the
replication origin of a transaction on top of its commit timestamp, but
the code paths for replication origins were never stressed as those
functions have never looked for a replication origin, and the SQL
functions available have never included this information since their
introduction in 73c986ad.

While on it, refactor a test of modules/commit_ts/ to use tstzrange() to
check that a transaction timestamp is within the wanted range, making
the test a bit easier to read.

Bump catalog version.

Author: Movead Li
Reviewed-by: Madan Kumar, Michael Paquier
Discussion: https://postgr.es/m/2020051116430836450630@highgo.ca

The raw_buf and line_buf buffers aren't used when reading binary format,
so skip allocating them.  raw_buf is 64K so that seems like a worthwhile
savings.  An unused line_buf only wastes 1K, but as long as we're checking
it's free to avoid allocating that too.

Bharath Rupireddy, tweaked a bit by me

Discussion: https://postgr.es/m/CALj2ACXcCKaGPY0whowqrJ4OPJvDnTssgpGCzvuFQu5z0CXb-g@mail.gmail.com

Parallel pg_restore has always supposed that ACL items for different
objects are independent and can be restored in parallel without
conflicts.  However, there is one case where this fails: because
REVOKE on a table is defined to also revoke the privilege(s) at
column level, we can't restore per-column ACLs till after we restore
any table-level privileges on their table.  Failure to honor this
restriction can lead to "tuple concurrently updated" errors during
parallel restore, or even to the per-column ACLs silently disappearing
because the table-level REVOKE is executed afterwards.

To fix, add a dependency from each column-level ACL item to its table's
ACL item, if there is one.  Note that this doesn't fix the hazard
for pre-existing archive files, only for ones made with a corrected
pg_dump.  Given that the bug's been there quite awhile without
field reports, I think this is acceptable.

This requires changing the API of pg_dump's dumpACL() function.
To keep its argument list from getting even longer, I removed the
"CatalogId objCatId" argument, which has been unused for ages.

Per report from Justin Pryzby.  Back-patch to all supported branches.

Discussion: https://postgr.es/m/20200706050129.GW4107@telsasoft.com

Reported-by: Lee Dong Wook <sh95119@gmail.com>

"relkind" normally refers to the char field from pg_class.  However, in
the parse nodes AlterTableStmt and CreateTableAsStmt, "relkind" was used
for a field of type enum ObjectType, that could refer to other object
types than those possible for a relkind.  Such fields being usually
named "objtype", switch the name in both structures to make things more
consistent.  Note that this led to some confusion in functions that
also operate on a RangeTableEntry object, which also has a field named
"relkind".

This naming goes back to commit 09d4e96d, where only OBJECT_TABLE and
OBJECT_INDEX were used.  This got extended later to use as well
OBJECT_TYPE with e440e12c, not really a relation kind.

Author: Mark Dilger
Reviewed-by: Daniel Gustafsson, Álvaro Herrera, Michael Paquier
Discussion: https://postgr.es/m/609181AE-E399-47C7-9221-856E0F96BF93@enterprisedb.com

SQL standard doesn't define numeric Inf or NaN values.  It appears even more
ridiculous to support then in jsonpath assuming JSON doesn't support these
values as well.  This commit forbids returning NaN from .double(), which was
previously allowed.  NaN can't be result of inner-jsonpath computation over
non-NaNs.  So, we can not expect NaN in the jsonpath output.

Reported-by: Tom Lane
Discussion: https://postgr.es/m/203949.1591879542%40sss.pgh.pa.us
Author: Alexander Korotkov
Reviewed-by: Tom Lane
Backpatch-through: 12

When jsonpath .double() method detects that numeric or string can't be
converted to double precision, it throws an error.  This commit makes these
errors explicitly express the reason of failure.

Discussion: https://postgr.es/m/CAPpHfdtqJtiSXkP7tOXez18NxhLUH_-75bL8%3DOce4Ki%2Bbv7V6Q%40mail.gmail.com
Author: Alexander Korotkov
Reviewed-by: Tom Lane
Backpatch-through: 12