When either input has a small number of digits, and the exact product
is requested, the speed of numeric multiplication can be increased
significantly by using a faster direct multiplication algorithm. This
works by fully computing each result digit in turn, starting with the
least significant, and propagating the carry up. This save cycles by
not requiring a temporary buffer to store digit products, not making
multiple passes over the digits of the longer input, and not requiring
separate carry-propagation passes.
For now, this is used when the shorter input has 1-4 NBASE digits (up
to 13-16 decimal digits), and the longer input is of any size, which
covers a lot of common real-world cases. Also, the relative benefit
increases as the size of the longer input increases.
Possible future work would be to try extending the technique to larger
numbers of digits in the shorter input.
Joel Jacobson and Dean Rasheed.
Discussion: https://postgr.es/m/44d2ffca-d560-4919-b85a-4d07060946aa@app.fastmail.com
The numeric round() and trunc() functions clamp the scale argument to
the range between +/- NUMERIC_MAX_RESULT_SCALE (2000), which is much
smaller than the actual allowed range of type numeric. As a result,
they return incorrect results when asked to round/truncate more than
2000 digits before or after the decimal point.
Fix by using the correct upper and lower scale limits based on the
actual allowed (and documented) range of type numeric.
While at it, use the new NUMERIC_WEIGHT_MAX constant instead of
SHRT_MAX in all other overflow checks, and fix a comment thinko in
power_var() introduced by e54a758d24 -- the minimum value of
ln_dweight is -NUMERIC_DSCALE_MAX (-16383), not -SHRT_MAX, though this
doesn't affect the point being made in the comment, that the resulting
local_rscale value may exceed NUMERIC_MAX_DISPLAY_SCALE (1000).
Back-patch to all supported branches.
Dean Rasheed, reviewed by Joel Jacobson.
Discussion: https://postgr.es/m/CAEZATCXB%2BrDTuMjhK5ZxcouufigSc-X4tGJCBTMpZ3n%3DxxQuhg%40mail.gmail.com
This adds 3 new variants of the random() function:
random(min integer, max integer) returns integer
random(min bigint, max bigint) returns bigint
random(min numeric, max numeric) returns numeric
Each returns a random number x in the range min <= x <= max.
For the numeric function, the number of digits after the decimal point
is equal to the number of digits that "min" or "max" has after the
decimal point, whichever has more.
The main entry points for these functions are in a new C source file.
The existing random(), random_normal(), and setseed() functions are
moved there too, so that they can all share the same PRNG state, which
is kept private to that file.
Dean Rasheed, reviewed by Jian He, David Zhang, Aleksander Alekseev,
and Tomas Vondra.
Discussion: https://postgr.es/m/CAEZATCV89Vxuq93xQdmc0t-0Y2zeeNQTdsjbmV7dyFBPykbV4Q@mail.gmail.com
as determined by include-what-you-use (IWYU)
While IWYU also suggests to *add* a bunch of #include's (which is its
main purpose), this patch does not do that. In some cases, a more
specific #include replaces another less specific one.
Some manual adjustments of the automatic result:
- IWYU currently doesn't know about includes that provide global
variable declarations (like -Wmissing-variable-declarations), so
those includes are being kept manually.
- All includes for port(ability) headers are being kept for now, to
play it safe.
- No changes of catalog/pg_foo.h to catalog/pg_foo_d.h, to keep the
patch from exploding in size.
Note that this patch touches just *.c files, so nothing declared in
header files changes in hidden ways.
As a small example, in src/backend/access/transam/rmgr.c, some IWYU
pragma annotations are added to handle a special case there.
Discussion: https://www.postgresql.org/message-id/flat/af837490-6b2f-46df-ba05-37ea6a6653fc%40eisentraut.org
f0efa5aec added initReadOnlyStringInfo to allow a StringInfo to be
initialized from an existing buffer and also relaxed the requirement
that a StringInfo's buffer must be NUL terminated at data[len]. Now
that we have that, there's no need for these aggregate deserial
functions to use appendBinaryStringInfo() as that rather wastefully
palloc'd a new buffer and memcpy'd in the bytea's buffer. Instead, we can
just use the bytea's buffer and point the StringInfo directly to that
using the new initializer function.
In Amdahl's law, this speeds up the serial portion of parallel
aggregates and makes sum(numeric), avg(numeric), var_pop(numeric),
var_samp(numeric), variance(numeric), stddev_pop(numeric),
stddev_samp(numeric), stddev(numeric), array_agg(anyarray),
string_agg(text) and string_agg(bytea) scale better in parallel queries.
Author: David Rowley
Discussion: https://postgr.es/m/CAApHDvr%3De-YOigriSHHm324a40HPqcUhSp6pWWgjz5WwegR%3DcQ%40mail.gmail.com
The serialized representation of an internal aggregate state is a bytea
value. In each deserial function, in order to "receive" the bytea value
we appended it onto a short-lived StringInfoData using
appendBinaryStringInfo. This was a little wasteful as it meant having to
palloc memory, copy a (possibly long) series of bytes then later pfree
that memory. Instead of going to this extra trouble, we can just fake up
a StringInfoData and point the data directly at the bytea's payload. This
should help increase the performance of internal aggregate
deserialization.
Reviewed-by: Michael Paquier
Discussion: https://postgr.es/m/CAApHDvr=e-YOigriSHHm324a40HPqcUhSp6pWWgjz5WwegR=cQ@mail.gmail.com
I realized that the third overflow case I posited in commit b0e9e4d76
actually should be handled in a different way: rather than tolerating
the idea that the quotient could round to 1, we should clamp so that
the output cannot be more than "count" when we know that the operand is
less than bound2. That being the case, we don't need an overflow-aware
increment in that code path, which leads me to revert the movement of
the pg_add_s32_overflow() call. (The diff in width_bucket_float8
might be easier to read by comparing against b0e9e4d76^.)
What's more, width_bucket_numeric also has this problem of the quotient
potentially rounding to 1, so add a clamp there too.
As before, I'm not quite convinced that a back-patch is warranted.
Discussion: https://postgr.es/m/391415.1680268470@sss.pgh.pa.us
This allows underscores to be used in integer and numeric literals,
and their corresponding type input functions, for visual grouping.
For example:
1_500_000_000
3.14159_26535_89793
0xffff_ffff
0b_1001_0001
A single underscore is allowed between any 2 digits, or immediately
after the base prefix indicator of non-decimal integers, per SQL:202x
draft.
Peter Eisentraut and Dean Rasheed
Discussion: https://postgr.es/m/84aae844-dc55-a4be-86d9-4f0fa405cc97%40enterprisedb.com
The prior coding of int64_div_fast_to_numeric() had a number of bugs
that would cause it to fail under different circumstances, such as
with log10val2 <= 0, or log10val2 a multiple of 4, or in the "slow"
numeric path with log10val2 >= 10.
None of those could be triggered by any of our current code, which
only uses log10val2 = 3 or 6. However, they made it a hazard for any
future code that might use it. Also, since this is exported by
numeric.c, users writing their own C code might choose to use it.
Therefore fix, and back-patch to v14, where it was introduced.
Dean Rasheed, reviewed by Tom Lane.
Discussion: https://postgr.es/m/CAEZATCW8gXgW0tgPxPgHDPhVX71%2BSWFRkhnXy%2BTfGDsKLepu2g%40mail.gmail.com
Improve the comment explaining the choice of rscale in numeric_sqrt(),
and ensure that the code works consistently when other values of
NBASE/DEC_DIGITS are used.
Note that, in practice, we always expect DEC_DIGITS == 4, and this
does not change the computation in that case.
Joel Jacobson and Dean Rasheed
Discussion: https://postgr.es/m/06712c29-98e9-43b3-98da-f234d81c6e49%40app.fastmail.com
This enhances the numeric type input function, adding support for
hexadecimal, octal, and binary integers of any size, up to the limits
of the numeric type.
Since 6fcda9aba8, such non-decimal integers have been accepted by the
parser as integer literals and passed through to numeric_in(). This
commit gives numeric_in() the ability to handle them.
While at it, simplify the handling of NaN and infinities, reducing the
number of calls to pg_strncasecmp(), and arrange for pg_strncasecmp()
to not be called at all for regular numbers. This gives a significant
performance improvement for decimal inputs, more than offsetting the
small performance hit of checking for non-decimal input.
Discussion: https://postgr.es/m/CAEZATCV8XShnmT9HZy25C%2Bo78CVOFmUN5EM9FRAZ5xvYTggPMg%40mail.gmail.com
Because we added StaticAssertStmt() first before StaticAssertDecl(),
some uses as well as the instructions in c.h are now a bit backwards
from the "native" way static assertions are meant to be used in C.
This updates the guidance and moves some static assertions to better
places.
Specifically, since the addition of StaticAssertDecl(), we can put
static assertions at the file level. This moves a number of static
assertions out of function bodies, where they might have been stuck
out of necessity, to perhaps better places at the file level or in
header files.
Also, when the static assertion appears in a position where a
declaration is allowed, then using StaticAssertDecl() is more native
than StaticAssertStmt().
Reviewed-by: John Naylor <john.naylor@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/941a04e7-dd6f-c0e4-8cdf-a33b3338cbda%40enterprisedb.com
This patch converts the input functions for bool, int2, int4, int8,
float4, float8, numeric, and contrib/cube to the new soft-error style.
array_in and record_in are also converted. There's lots more to do,
but this is enough to provide proof-of-concept that the soft-error
API is usable, as well as reference examples for how to convert
input functions.
This patch is mostly by me, but it owes very substantial debt to
earlier work by Nikita Glukhov, Andrew Dunstan, and Amul Sul.
Thanks to Andres Freund for review.
Discussion: https://postgr.es/m/3bbbb0df-7382-bf87-9737-340ba096e034@postgrespro.ru
This makes the choice of result scale of numeric power() for integer
exponents consistent with the choice for non-integer exponents, and
with the result scale of other numeric functions. Specifically, the
result scale will be at least as large as the scale of either input,
and sufficient to ensure that the result has at least 16 significant
digits.
Formerly, the result scale was based only on the scale of the first
input, without taking into account the weight of the result. For
results with negative weight, that could lead to results with very few
or even no non-zero significant digits (e.g., 10.0 ^ (-18) produced
0.0000000000000000).
Fix this by moving responsibility for the choice of result scale into
power_var_int(), which already has code to estimate the result weight.
Per report by Adrian Klaver and suggested fix by Tom Lane.
No back-patch -- arguably this is a bug fix, but one which is easy to
work around, so it doesn't seem worth the risk of changing query
results in stable branches.
Discussion: https://postgr.es/m/12a40226-70ac-3a3b-3d3a-fdaf9e32d312%40aklaver.com
Make sure that function declarations use names that exactly match the
corresponding names from function definitions in optimizer, parser,
utility, libpq, and "commands" code, as well as in remaining library
code. Do the same for all code related to frontend programs (with the
exception of pg_dump/pg_dumpall related code).
Like other recent commits that cleaned up function parameter names, this
commit was written with help from clang-tidy. Later commits will handle
ecpg and pg_dump/pg_dumpall.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CAH2-WznJt9CMM9KJTMjJh_zbL5hD9oX44qdJ4aqZtjFi-zA3Tg@mail.gmail.com
Most of these are cases where we could call memcpy() or other libc
functions with a NULL pointer and a zero count, which is forbidden
by POSIX even though every production version of libc allows it.
We've fixed such things before in a piecemeal way, but apparently
never made an effort to try to get them all. I don't claim that
this patch does so either, but it gets every failure I observe in
check-world, using clang 12.0.1 on current RHEL8.
numeric.c has a different issue that the sanitizer doesn't like:
"ln(-1.0)" will compute log10(0) and then try to assign the
resulting -Inf to an integer variable. We don't actually use the
result in such a case, so there's no live bug.
Back-patch to all supported branches, with the idea that we might
start running a buildfarm member that tests this case. This includes
back-patching c1132aae3 (Check the size in COPY_POINTER_FIELD),
which previously silenced some of these issues in copyfuncs.c.
Discussion: https://postgr.es/m/CALNJ-vT9r0DSsAOw9OXVJFxLENoVS_68kJ5x0p44atoYH+H4dg@mail.gmail.com
Formerly div_var() had "fast path" short division code that was
significantly faster when the divisor was just one base-NBASE digit,
but otherwise used long division.
This commit adds a new function div_var_int() that divides by an
arbitrary 32-bit integer, using the fast short division algorithm, and
updates both div_var() and div_var_fast() to use it for one and two
digit divisors. In the case of div_var(), this is slightly faster in
the one-digit case, because it avoids some digit array copying, and is
much faster in the two-digit case where it replaces long division. For
div_var_fast(), it is much faster in both cases because the main
div_var_fast() algorithm is optimised for larger inputs.
Additionally, optimise exp() and ln() by using div_var_int(), allowing
a NumericVar to be replaced by an int in a couple of places, most
notably in the Taylor series code. This produces a significant speedup
of exp(), ln() and the numeric_big regression test.
Dean Rasheed, reviewed by Tom Lane.
Discussion: https://postgr.es/m/CAEZATCVwsBi-ND-t82Cuuh1=8ee6jdOpzsmGN+CUZB6yjLg9jw@mail.gmail.com
In the standard numeric division algorithm, the inner loop multiplies
the divisor by the next quotient digit and subtracts that from the
working dividend. As suggested by the original code comment, the
separate "carry" and "borrow" variables (from the multiplication and
subtraction steps respectively) can be folded together into a single
variable. Doing so significantly improves performance, as well as
simplifying the code.
Dean Rasheed, reviewed by Tom Lane.
Discussion: https://postgr.es/m/CAEZATCVwsBi-ND-t82Cuuh1=8ee6jdOpzsmGN+CUZB6yjLg9jw@mail.gmail.com
This loop is basically the same as the inner loop of mul_var(), which
was auto-vectorized in commit 8870917623, but the compiler will only
consider auto-vectorizing the div_var_fast() loop if the assignment
target div[qi + i] is replaced by div_qi[i], where div_qi = &div[qi].
Additionally, since the compiler doesn't know that qdigit is
guaranteed to fit in a 16-bit NumericDigit, cast it to NumericDigit
before multiplying to make the resulting auto-vectorized code more
efficient (avoiding unnecessary multiplication of the high 16 bits).
While at it, per suggestion from Tom Lane, change var1digit in
mul_var() to be a NumericDigit rather than an int for the same
reason. This actually makes no difference with modern gcc, but it
might help other compilers generate more efficient assembly.
Dean Rasheed, reviewed by Tom Lane.
Discussion: https://postgr.es/m/CAEZATCVwsBi-ND-t82Cuuh1=8ee6jdOpzsmGN+CUZB6yjLg9jw@mail.gmail.com
This fixes a loss of precision that occurs when the first input is
very close to 1, so that its logarithm is very small.
Formerly, during the initial low-precision calculation to estimate the
result weight, the logarithm was computed to a local rscale that was
capped to NUMERIC_MAX_DISPLAY_SCALE (1000). However, the base may be
as close as 1e-16383 to 1, hence its logarithm may be as small as
1e-16383, and so the local rscale needs to be allowed to exceed 16383,
otherwise all precision is lost, leading to a poor choice of rscale
for the full-precision calculation.
Fix this by removing the cap on the local rscale during the initial
low-precision calculation, as we already do in the full-precision
calculation. This doesn't change the fact that the initial calculation
is a low-precision approximation, computing the logarithm to around 8
significant digits, which is very fast, especially when the base is
very close to 1.
Patch by me, reviewed by Alvaro Herrera.
Discussion: https://postgr.es/m/CAEZATCV-Ceu%2BHpRMf416yUe4KKFv%3DtdgXQAe5-7S9tD%3D5E-T1g%40mail.gmail.com
Formerly, the numeric code tested whether an integer value of a larger
type would fit in a smaller type by casting it to the smaller type and
then testing if the reverse conversion produced the original value.
That's perfectly fine, except that it caused a test failure on
buildfarm animal castoroides, most likely due to a compiler bug.
Instead, do these tests by comparing against PG_INT16/32_MIN/MAX. That
matches existing code in other places, such as int84(), which is more
widely tested, and so is less likely to go wrong.
While at it, add regression tests covering the numeric-to-int8/4/2
conversions, and adjust the recently added tests to the style of
434ddfb79a (on the v11 branch) to make failures easier to diagnose.
Per buildfarm via Tom Lane, reviewed by Tom Lane.
Discussion: https://postgr.es/m/2394813.1628179479%40sss.pgh.pa.us
This fixes a long-standing bug when using to_char() to format a
numeric value in scientific notation -- if the value's exponent is
less than -NUMERIC_MAX_DISPLAY_SCALE-1 (-1001), it produced a
division-by-zero error.
The reason for this error was that get_str_from_var_sci() divides its
input by 10^exp, which it produced using power_var_int(). However, the
underflow test in power_var_int() causes it to return zero if the
result scale is too small. That's not a problem for power_var_int()'s
only other caller, power_var(), since that limits the rscale to 1000,
but in get_str_from_var_sci() the exponent can be much smaller,
requiring a much larger rscale. Fix by introducing a new function to
compute 10^exp directly, with no rscale limit. This also allows 10^exp
to be computed more efficiently, without any numeric multiplication,
division or rounding.
Discussion: https://postgr.es/m/CAEZATCWhojfH4whaqgUKBe8D5jNHB8ytzemL-PnRx+KCTyMXmg@mail.gmail.com
This fixes a couple of related problems that arise when raising
numbers to very large powers.
Firstly, when raising a negative number to a very large integer power,
the result should be well-defined, but the previous code would only
cope if the exponent was small enough to go through power_var_int().
Otherwise it would throw an internal error, attempting to take the
logarithm of a negative number. Fix this by adding suitable handling
to the general case in power_var() to cope with negative bases,
checking for integer powers there.
Next, when raising a (positive or negative) number whose absolute
value is slightly less than 1 to a very large power, the result should
approach zero as the power is increased. However, in some cases, for
sufficiently large powers, this would lose all precision and return 1
instead of 0. This was due to the way that the local_rscale was being
calculated for the final full-precision calculation:
local_rscale = rscale + (int) val - ln_dweight + 8
The first two terms on the right hand side are meant to give the
number of significant digits required in the result ("val" being the
estimated result weight). However, this failed to account for the fact
that rscale is clipped to a maximum of NUMERIC_MAX_DISPLAY_SCALE
(1000), and the result weight might be less then -1000, causing their
sum to be negative, leading to a loss of precision. Fix this by
forcing the number of significant digits calculated to be nonnegative.
It's OK for it to be zero (when the result weight is less than -1000),
since the local_rscale value then includes a few extra digits to
ensure an accurate result.
Finally, add additional underflow checks to exp_var() and power_var(),
so that they consistently return zero for cases like this where the
result is indistinguishable from zero. Some paths through this code
already returned zero in such cases, but others were throwing overflow
errors.
Dean Rasheed, reviewed by Yugo Nagata.
Discussion: http://postgr.es/m/CAEZATCW6Dvq7+3wN3tt5jLj-FyOcUgT5xNoOqce5=6Su0bCR0w@mail.gmail.com
Formerly, when specifying NUMERIC(precision, scale), the scale had to
be in the range [0, precision], which was per SQL spec. This commit
extends the range of allowed scales to [-1000, 1000], independent of
the precision (whose valid range remains [1, 1000]).
A negative scale implies rounding before the decimal point. For
example, a column might be declared with a scale of -3 to round values
to the nearest thousand. Note that the display scale remains
non-negative, so in this case the display scale will be zero, and all
digits before the decimal point will be displayed.
A scale greater than the precision supports fractional values with
zeros immediately after the decimal point.
Take the opportunity to tidy up the code that packs, unpacks and
validates the contents of a typmod integer, encapsulating it in a
small set of new inline functions.
Bump the catversion because the allowed contents of atttypmod have
changed for numeric columns. This isn't a change that requires a
re-initdb, but negative scale values in the typmod would confuse old
backends.
Dean Rasheed, with additional improvements by Tom Lane. Reviewed by
Tom Lane.
Discussion: https://postgr.es/m/CAEZATCWdNLgpKihmURF8nfofP0RFtAKJ7ktY6GcZOPnMfUoRqA@mail.gmail.com
This fixes an overflow error when using the numeric * operator if the
result has more than 16383 digits after the decimal point by rounding
the result. Overflow errors should only occur if the result has too
many digits *before* the decimal point.
Discussion: https://postgr.es/m/CAEZATCUmeFWCrq2dNzZpRj5+6LfN85jYiDoqm+ucSXhb9U2TbA@mail.gmail.com
Formerly various numeric aggregate functions supported parallel
aggregation by having each worker convert partial aggregate values to
Numeric and use numeric_send() as part of serializing their state.
That's problematic, since the range of Numeric is smaller than that of
NumericVar, so it's possible for it to overflow (on either side of the
decimal point) in cases that would succeed in non-parallel mode.
Fix by serializing NumericVars instead, to avoid the overflow risk and
ensure that parallel and non-parallel modes work the same.
A side benefit is that this improves the efficiency of the
serialization/deserialization code, which can make a noticeable
difference to performance with large numbers of parallel workers.
No back-patch due to risk from changing the binary format of the
aggregate serialization states, as well as lack of prior field
complaints and low probability of such overflows in practice.
Patch by me. Thanks to David Rowley for review and performance
testing, and Ranier Vilela for an additional suggestion.
Discussion: https://postgr.es/m/CAEZATCUmeFWCrq2dNzZpRj5+6LfN85jYiDoqm+ucSXhb9U2TbA@mail.gmail.com
The previous implementation of EXTRACT mapped internally to
date_part(), which returned type double precision (since it was
implemented long before the numeric type existed). This can lead to
imprecise output in some cases, so returning numeric would be
preferrable. Changing the return type of an existing function is a
bit risky, so instead we do the following: We implement a new set of
functions, which are now called "extract", in parallel to the existing
date_part functions. They work the same way internally but use
numeric instead of float8. The EXTRACT construct is now mapped by the
parser to these new extract functions. That way, dumps of views
etc. from old versions (which would use date_part) continue to work
unchanged, but new uses will map to the new extract functions.
Additionally, the reverse compilation of EXTRACT now reproduces the
original syntax, using the new mechanism introduced in
40c24bfef92530bd846e111c1742c2a54441c62c.
The following minor changes of behavior result from the new
implementation:
- The column name from an isolated EXTRACT call is now "extract"
instead of "date_part".
- Extract from date now rejects inappropriate field names such as
HOUR. It was previously mapped internally to extract from
timestamp, so it would silently accept everything appropriate for
timestamp.
- Return values when extracting fields with possibly fractional
values, such as second and epoch, now have the full scale that the
value has internally (so, for example, '1.000000' instead of just
'1').
Reported-by: Petr Fedorov <petr.fedorov@phystech.edu>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/flat/42b73d2d-da12-ba9f-570a-420e0cce19d9@phystech.edu
In power_var_int(), the computation of the number of significant
digits to use in the computation used log(Abs(exp)), which isn't safe
because Abs(exp) returns INT_MIN when exp is INT_MIN. Use fabs()
instead of Abs(), so that the exponent is cast to a double before the
absolute value is taken.
Back-patch to 9.6, where this was introduced (by 7d9a4737c2).
Discussion: https://postgr.es/m/CAEZATCVd6pMkz=BrZEgBKyqqJrt2xghr=fNc8+Z=5xC6cgWrWA@mail.gmail.com
Multiply before dividing, not the reverse, so that cases that should
produce exact results do produce exact results. (width_bucket_float8
got this right already.) Even when the result is inexact, this avoids
making it more inexact, since only the division step introduces any
imprecision.
While at it, fix compute_bucket() to not uselessly repeat the sign
check already done by its caller, and avoid duplicating the
multiply/divide steps by adjusting variable usage.
Per complaint from Martin Visser. Although this seems like a bug fix,
I'm hesitant to risk changing width_bucket()'s results in stable
branches, so no back-patch.
Discussion: https://postgr.es/m/6FA5117D-6AED-4656-8FEF-B74AC18FAD85@brytlyt.com
While the calculation is not well-defined if the bounds arguments are
infinite, there is a perfectly sane outcome if the test operand is
infinite: it's just like any other value that's before the first bucket
or after the last one. width_bucket_float8() got this right, but
I was too hasty about the case when adding infinities to numerics
(commit a57d312a7), so that width_bucket_numeric() just rejected it.
Fix that, and sync the relevant error message strings.
No back-patch needed, since infinities-in-numeric haven't shipped yet.
Discussion: https://postgr.es/m/2465409.1602170063@sss.pgh.pa.us
This essentially reverts a micro-optimization I made years ago,
as part of the much larger commit d72f6c750. It's doubtful
that there was any hard evidence for it being helpful even then,
and the case is even more dubious now that modern compilers
are so much smarter about inlining memset().
The proximate reason for undoing it is to get rid of the type punning
inherent in MemSet, for fear that that may cause problems now that
we're applying additional optimization switches to numeric.c.
At the very least this'll silence some warnings from a few old
buildfarm animals.
(It's probably past time for another look at whether MemSet is still
worth anything at all, but I do not propose to tackle that question
right now.)
Discussion: https://postgr.es/m/CAJ3gD9evtA_vBo+WMYMyT-u=keHX7-r8p2w7OSRfXf42LTwCZQ@mail.gmail.com
Experimentation shows that clang will auto-vectorize the critical
multiplication loop if the termination condition is written "i2 < limit"
rather than "i2 <= limit". This seems unbelievably stupid, but I've
reproduced it on both clang 9.0.1 (RHEL8) and 11.0.3 (macOS Catalina).
gcc doesn't care, so tweak the code to do it that way.
Discussion: https://postgr.es/m/CAJ3gD9evtA_vBo+WMYMyT-u=keHX7-r8p2w7OSRfXf42LTwCZQ@mail.gmail.com
Compile numeric.c with -ftree-vectorize where available, and adjust
the innermost loop of mul_var() so that it is amenable to being
auto-vectorized. (Mainly, that involves making it process the arrays
left-to-right not right-to-left.)
Applying -ftree-vectorize actually makes numeric.o smaller, at least
with my compiler (gcc 8.3.1 on x86_64), and it's a little faster too.
Independently of that, fixing the inner loop to be vectorizable also
makes things a bit faster. But doing both is a huge win for
multiplications with lots of digits. For me, the numeric regression
test is the same speed to within measurement noise, but numeric_big
is a full 45% faster.
We also looked into applying -funroll-loops, but that makes numeric.o
bloat quite a bit, and the additional speed improvement is very
marginal.
Amit Khandekar, reviewed and edited a little by me
Discussion: https://postgr.es/m/CAJ3gD9evtA_vBo+WMYMyT-u=keHX7-r8p2w7OSRfXf42LTwCZQ@mail.gmail.com
Add infinities that behave the same as they do in the floating-point
data types. Aside from any intrinsic usefulness these may have,
this closes an important gap in our ability to convert floating
values to numeric and/or replace float-based APIs with numeric.
The new values are represented by bit patterns that were formerly
not used (although old code probably would take them for NaNs).
So there shouldn't be any pg_upgrade hazard.
Patch by me, reviewed by Dean Rasheed and Andrew Gierth
Discussion: https://postgr.es/m/606717.1591924582@sss.pgh.pa.us
