Обсуждение: Special-case executor expression steps for common combinations
The attached patch adds special-case expression steps for common sets of steps in the executor to shave a few cycles off during execution, and make the JIT generated code simpler. * Adds EEOP_FUNCEXPR_STRICT_1 and EEOP_FUNCEXPR_STRICT_2 for function calls of strict functions with 1 or 2 arguments (EEOP_FUNCEXPR_STRICT remains used for > 2 arguments). * Adds EEOP_AGG_STRICT_INPUT_CHECK_ARGS_1 which is a special case for the common case of one arg aggs. * Replace EEOP_DONE with EEOP_DONE_RETURN and EEOP_DONE_NO_RETURN to be able to skip extra setup for steps which are only interested in the side effects. Stressing the EEOP_FUNCEXPR_STRICT_* steps specifically shows a 1.5% improvement and pgbench over the branch shows a ~1% improvement in TPS (both measured over 6 runs with outliers removed). EEOP_FUNCEXPR_STRICT_* (10M iterations): master : (7503.317, 7553.691, 7634.524) patched : (7422.756, 7455.120, 7492.393) pgbench: master : (3653.83, 3792.97, 3863.70) patched : (3743.04, 3830.02, 3869.80) This patch was extracted from a larger body of work from Andres [0] aiming at providing the necessary executor infrastructure for making JIT expression caching possible. This patch, and more which are to be submitted, is however separate in the sense that it is not part of the infrastructure, it's an improvements on its own. Thoughts? -- Daniel Gustafsson [0]: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
Вложения
On 12/10/2023 12:48, Daniel Gustafsson wrote: > The attached patch adds special-case expression steps for common sets of steps > in the executor to shave a few cycles off during execution, and make the JIT > generated code simpler. > > * Adds EEOP_FUNCEXPR_STRICT_1 and EEOP_FUNCEXPR_STRICT_2 for function calls of > strict functions with 1 or 2 arguments (EEOP_FUNCEXPR_STRICT remains used for > > 2 arguments). > * Adds EEOP_AGG_STRICT_INPUT_CHECK_ARGS_1 which is a special case for the > common case of one arg aggs. Are these relevant when JITting? I'm a little sad if the JIT compiler cannot unroll these on its own. Is there something we could do to hint it, so that it could treat the number of arguments as a constant? I understand that this can give a small boost in interpreter mode, so maybe we should do it in any case. But I'd like to know if we're missing a trick with the JITter, before we mask it with this. > * Replace EEOP_DONE with EEOP_DONE_RETURN and EEOP_DONE_NO_RETURN to be able to > skip extra setup for steps which are only interested in the side effects. I'm a little surprised if this makes a measurable performance difference, but sure, why not. It seems nice to be more explicit when you don't expect a return value. -- Heikki Linnakangas Neon (https://neon.tech)
On Thu, 12 Oct 2023 at 22:54, Daniel Gustafsson <daniel@yesql.se> wrote: > EEOP_FUNCEXPR_STRICT_* (10M iterations): > master : (7503.317, 7553.691, 7634.524) > patched : (7422.756, 7455.120, 7492.393) > > pgbench: > master : (3653.83, 3792.97, 3863.70) > patched : (3743.04, 3830.02, 3869.80) > > Thoughts? Did any of these tests compile the expression with JIT? If not, how does the performance compare for a query that JITs the expression? David
Hi, On 2023-10-12 13:24:27 +0300, Heikki Linnakangas wrote: > On 12/10/2023 12:48, Daniel Gustafsson wrote: > > The attached patch adds special-case expression steps for common sets of steps > > in the executor to shave a few cycles off during execution, and make the JIT > > generated code simpler. > > > > * Adds EEOP_FUNCEXPR_STRICT_1 and EEOP_FUNCEXPR_STRICT_2 for function calls of > > strict functions with 1 or 2 arguments (EEOP_FUNCEXPR_STRICT remains used for > > > 2 arguments). > > * Adds EEOP_AGG_STRICT_INPUT_CHECK_ARGS_1 which is a special case for the > > common case of one arg aggs. > > Are these relevant when JITting? I'm a little sad if the JIT compiler cannot > unroll these on its own. Is there something we could do to hint it, so that > it could treat the number of arguments as a constant? I think it's mainly important for interpreted execution. > > skip extra setup for steps which are only interested in the side effects. > > I'm a little surprised if this makes a measurable performance difference, > but sure, why not. It seems nice to be more explicit when you don't expect a > return value. IIRC this is more interesting for JIT than the above, because it allows LLVM to know that the return value isn't needed and thus doesn't need to be computed. Greetings, Andres Freund
> On 12 Oct 2023, at 19:52, Andres Freund <andres@anarazel.de> wrote: > On 2023-10-12 13:24:27 +0300, Heikki Linnakangas wrote: >> On 12/10/2023 12:48, Daniel Gustafsson wrote: >>> The attached patch adds special-case expression steps for common sets of steps >>> in the executor to shave a few cycles off during execution, and make the JIT >>> generated code simpler. >>> >>> * Adds EEOP_FUNCEXPR_STRICT_1 and EEOP_FUNCEXPR_STRICT_2 for function calls of >>> strict functions with 1 or 2 arguments (EEOP_FUNCEXPR_STRICT remains used for >>>> 2 arguments). >>> * Adds EEOP_AGG_STRICT_INPUT_CHECK_ARGS_1 which is a special case for the >>> common case of one arg aggs. >> >> Are these relevant when JITting? I'm a little sad if the JIT compiler cannot >> unroll these on its own. Is there something we could do to hint it, so that >> it could treat the number of arguments as a constant? > > I think it's mainly important for interpreted execution. Agreed. >>> skip extra setup for steps which are only interested in the side effects. >> >> I'm a little surprised if this makes a measurable performance difference, >> but sure, why not. It seems nice to be more explicit when you don't expect a >> return value. Right, performance benefits aside it does improve readability IMHO. > IIRC this is more interesting for JIT than the above, because it allows LLVM > to know that the return value isn't needed and thus doesn't need to be > computed. Correct, this is important to the JIT code which no longer has to perform two Loads and one Store in order to get nothing, but can instead fastpath to building a zero returnvalue. -- Daniel Gustafsson
On 10/12/23 11:48 AM, Daniel Gustafsson wrote: > Thoughts? I have looked at the patch and it still applies, builds and passes the test cases and I personally think these optimizations are pretty much no-brainers that we should do and it is a pity nobody has had the time to review this patch. 1) The no-return case should help with the JIT, making jitted code faster. 2) The specialized strict steps helps with many common queries in the interpreted mode. The code itself looks really good (great work!) but I have two comments on it. 1) I think the patch should be split into two. The two different optimizations are not related at all other than that they create specialized versions of expressions steps. Having them as separate makes the commit history easier to read for future developers. 2) We could generate functions which return void rather than NULL and therefore not have to do a return at all but I am not sure that small optimization and extra clarity would be worth the hassle. The current approach with adding Assert() is ok with me. Daniel, what do you think? Andreas
On 6/20/24 5:22 PM, Andreas Karlsson wrote: > On 10/12/23 11:48 AM, Daniel Gustafsson wrote: >> Thoughts? > > I have looked at the patch and it still applies, builds and passes the > test cases and I personally think these optimizations are pretty much > no-brainers that we should do and it is a pity nobody has had the time > to review this patch. Forgot to write that I am planning to also try to do so benchmarks to see if I can reproduce the speedups. :) Andreas
> On 20 Jun 2024, at 17:22, Andreas Karlsson <andreas@proxel.se> wrote: > > On 10/12/23 11:48 AM, Daniel Gustafsson wrote: >> Thoughts? > > I have looked at the patch and it still applies, builds and passes the test cases and I personally think these optimizationsare pretty much no-brainers that we should do and it is a pity nobody has had the time to review this patch. > > 1) The no-return case should help with the JIT, making jitted code faster. > > 2) The specialized strict steps helps with many common queries in the interpreted mode. > > The code itself looks really good (great work!) but I have two comments on it. Thanks for review! > 1) I think the patch should be split into two. The two different optimizations are not related at all other than that theycreate specialized versions of expressions steps. Having them as separate makes the commit history easier to read forfuture developers. That's a good point, the attached v2 splits it into two separate commits. > 2) We could generate functions which return void rather than NULL and therefore not have to do a return at all but I amnot sure that small optimization and extra clarity would be worth the hassle. The current approach with adding Assert()is ok with me. Daniel, what do you think? I'm not sure that would move the needle enough to warrant the extra complexity. It could be worth pursuing, but it can be done separately from this. -- Daniel Gustafsson
Вложения
On 7/4/24 6:26 PM, Daniel Gustafsson wrote: >> 2) We could generate functions which return void rather than NULL and therefore not have to do a return at all but I amnot sure that small optimization and extra clarity would be worth the hassle. The current approach with adding Assert()is ok with me. Daniel, what do you think? > > I'm not sure that would move the needle enough to warrant the extra complexity. > It could be worth pursuing, but it can be done separately from this. Agreed. I looked some more at the patch and have a suggestion for code style. Attaching the diff. Do with them what you wish, for me they make to code easier to read. Andreas
Вложения
I have bench marked the two patches now and failed to measure any speedup or slowdown from the first patch (removing return) but I think it is a good idea anyway. For the second patch (optimize strict) I managed to measure a ~1% speed up for the following query "SELECT sum(x + y + 1) FROM t;" over one million rows. I would say both patches are ready for committer modulo my proposed style fixes. Andreas