I've had an idea for a while of a different way we could evaluate
TH-like splices which would be more lightweight and easier to work with. The idea is to create a third quotation/splicing mechanism which has no introspection (like MetaML) but then to evaluate these quotes and splices in the optimiser rather than using the bytecode interpreter. I am motivated by the many examples of recursive functions in libraries, which when given a statically known argument should be able to be unrolled to produce much better code. It is understandable that the compiler doesn't try to do this itself but there should be an easy way for the user to direct the compiler to do so. (See also https://www.reddit.com/r/haskell/comments/7yvb43/ghc_compiletime_evaluation/) An example to be concrete: Take the power function such that power k n computes k^n. power :: Int -> Int -> Int power k n = if n == 0 then 1 else k * power k (n - 1) If we statically know n then we can create a staged version. We use R to indicate that an argument is dynamically known. power :: R Int -> Int -> R Int power k n = if n == 0 then .< 1 >. else .< ~k * ~(power k (n-1)) >. One way to implement this in Haskell is to use typed template haskell quoting and splicing. The key thing to notice about why this works is that in order to splice `power k (n-1)` we need to evaluate `power k (n-1)` so that we have something of type `R Int` which we can then splice back into the quote. The way this is currently implemented is that the bytecode interpreter runs these splices in order to find out what they evaluate to. The idea is that instead, we add another mode which runs splices in the core-to-core optimiser. The optimiser performs evaluation by beta reduction, inlining and constant folding. For simple definitions on algebraic data types it does a very good job of eliminating overhead. As long as the call is not recursive. If we can just get the optimiser to inline recursive calls in a controlled manner as well, it should do a good job on the unrolled definition. The benefits of this are that it works across all platforms and fits nicely already into the compilation pipeline. It also fits in nicely with the intuition that a "quote" means to stop evaluating and a "splice" means to evaluate. A disadvantage is that the optimiser is only a *partial* evaluator rather than an evaluator. It gets stuck evaluating things containing primitives and so on. I don't forsee this as a major issue but a limitation that library authors should be aware of when writing their staged programs. To go back to the power example, the recursive condition would have to be an inductively defined natural (data N = Z | S N) rather than an Int as the comparison operator for integers can't be evaluated by the optimiser. It is already rather easy to write staged programs which loop if you don't carefully consider the staging so this seems now worse than the status quo. Does this sound at all sensible to anyone else? Matt _______________________________________________ ghc-devs mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs |
Hi,
something like this would be great. I don’t have a sense yet of what “something” should be like. Am Dienstag, den 27.02.2018, 09:59 +0000 schrieb Matthew Pickering: > To go back to the power example, the recursive > condition would have to be an inductively defined natural (data N = Z > | S N) rather than an Int as the comparison operator for integers > can't be evaluated by the optimiser. Sure they can: $ cat ConstantFolding.hs {-# LANGUAGE TemplateHaskell #-} {-# OPTIONS_GHC -fplugin=Test.Inspection.Plugin #-} module ConstantFolding where import Test.Inspection ltInt :: Bool ltInt = (3::Int) > 2 ltInteger :: Bool ltInteger = (3::Integer) > 2 true :: Bool true = True inspect $ 'ltInt === 'true inspect $ 'ltInteger === 'true $ ghc -O ConstantFolding.hs [1 of 1] Compiling ConstantFolding ( ConstantFolding.hs, ConstantFolding.o ) ConstantFolding.hs:17:1: ltInt === true passed. ConstantFolding.hs:18:1: ltInteger === true passed. inspection testing successful expected successes: 2 As an alternative with a maybe simpler user interface (and probably less power), I wonder if we can create a magic function > compileTimeWHNF :: a -> a or > compileTimeNF :: a -> a and a GHC core plugin (or eventually built-in thing) that finds these magic functions and evaluates their arguments, using the simplifier. Cheers, Joachim -- Joachim Breitner [hidden email] http://www.joachim-breitner.de/ _______________________________________________ ghc-devs mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs signature.asc (849 bytes) Download Attachment |
Hey Matt, cool idea! Also it looks like such a tool could 'solve' stream fusion: https://ghc.haskell.org/trac/2018-02-27 18:01 GMT+01:00 Joachim Breitner <[hidden email]>: Hi, _______________________________________________ ghc-devs mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs |
In reply to this post by Joachim Breitner-2
You have identified a small inaccuracy in my original email.
It is true that ultimately 3 == 0 will be evaluated but these constant folding opportunities are only applied later after other inlining takes place. It becomes quite crucial under this scheme when different optimisations happen because when we inline a recursive function, we have to make sure to apply the right evaluation steps before trying to simplify the function body. In my hacked together implementation of this, constant folding happens too late so only the algebraic version teminated. Cheers, Matt On Tue, Feb 27, 2018 at 5:01 PM, Joachim Breitner <[hidden email]> wrote: > Hi, > > something like this would be great. I don’t have a sense yet of what > “something” should be like. > > > Am Dienstag, den 27.02.2018, 09:59 +0000 schrieb Matthew Pickering: >> To go back to the power example, the recursive >> condition would have to be an inductively defined natural (data N = Z >> | S N) rather than an Int as the comparison operator for integers >> can't be evaluated by the optimiser. > > Sure they can: > > $ cat ConstantFolding.hs > {-# LANGUAGE TemplateHaskell #-} > {-# OPTIONS_GHC -fplugin=Test.Inspection.Plugin #-} > module ConstantFolding where > > import Test.Inspection > > ltInt :: Bool > ltInt = (3::Int) > 2 > > ltInteger :: Bool > ltInteger = (3::Integer) > 2 > > true :: Bool > true = True > > > inspect $ 'ltInt === 'true > inspect $ 'ltInteger === 'true > > $ ghc -O ConstantFolding.hs > [1 of 1] Compiling ConstantFolding ( ConstantFolding.hs, > ConstantFolding.o ) > ConstantFolding.hs:17:1: ltInt === true passed. > ConstantFolding.hs:18:1: ltInteger === true passed. > inspection testing successful > expected successes: 2 > > > > As an alternative with a maybe simpler user interface (and probably > less power), I wonder if we can create a magic function >> compileTimeWHNF :: a -> a > or >> compileTimeNF :: a -> a > and a GHC core plugin (or eventually built-in thing) that finds these > magic functions and evaluates their arguments, using the simplifier. > > > Cheers, > Joachim > > -- > Joachim Breitner > [hidden email] > http://www.joachim-breitner.de/ ghc-devs mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs |
> In my hacked together implementation of this, constant folding happens
too late so only the algebraic version teminated. ah, very good point! Joachim Am 28. Februar 2018 06:26:35 EST schrieb Matthew Pickering <[hidden email]>: >You have identified a small inaccuracy in my original email. > >It is true that ultimately 3 == 0 will be evaluated but these constant >folding opportunities are only applied later after other inlining >takes place. > >It becomes quite crucial under this scheme when different >optimisations happen because when we inline a recursive function, we >have to make sure to apply the right evaluation steps before trying to >simplify the function body. >In my hacked together implementation of this, constant folding happens >too late so only the algebraic version teminated. > >Cheers, > >Matt > >On Tue, Feb 27, 2018 at 5:01 PM, Joachim Breitner ><[hidden email]> wrote: >> Hi, >> >> something like this would be great. I don’t have a sense yet of what >> “something” should be like. >> >> >> Am Dienstag, den 27.02.2018, 09:59 +0000 schrieb Matthew Pickering: >>> To go back to the power example, the recursive >>> condition would have to be an inductively defined natural (data N = >Z >>> | S N) rather than an Int as the comparison operator for integers >>> can't be evaluated by the optimiser. >> >> Sure they can: >> >> $ cat ConstantFolding.hs >> {-# LANGUAGE TemplateHaskell #-} >> {-# OPTIONS_GHC -fplugin=Test.Inspection.Plugin #-} >> module ConstantFolding where >> >> import Test.Inspection >> >> ltInt :: Bool >> ltInt = (3::Int) > 2 >> >> ltInteger :: Bool >> ltInteger = (3::Integer) > 2 >> >> true :: Bool >> true = True >> >> >> inspect $ 'ltInt === 'true >> inspect $ 'ltInteger === 'true >> >> $ ghc -O ConstantFolding.hs >> [1 of 1] Compiling ConstantFolding ( ConstantFolding.hs, >> ConstantFolding.o ) >> ConstantFolding.hs:17:1: ltInt === true passed. >> ConstantFolding.hs:18:1: ltInteger === true passed. >> inspection testing successful >> expected successes: 2 >> >> >> >> As an alternative with a maybe simpler user interface (and probably >> less power), I wonder if we can create a magic function >>> compileTimeWHNF :: a -> a >> or >>> compileTimeNF :: a -> a >> and a GHC core plugin (or eventually built-in thing) that finds these >> magic functions and evaluates their arguments, using the simplifier. >> >> >> Cheers, >> Joachim >> >> -- >> Joachim Breitner >> [hidden email] >> http://www.joachim-breitner.de/ ghc-devs mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs |
In reply to this post by Matthew Pickering
(Clearing up my inbox.)
Sorry to be slow on this Matthew. A difficulty I see is that the optimiser must exhaustively evaluate until it gets to a quote. And that is not easy if you had $$$(if (reverse [1,2,3] == [3,2,1]) then .. else .. ) you'd probably get stuck. Or at least you'd have to be able to inline reverse, and (==) and all the typeclass machinery for dealing with (==). A merit of the current setup is that evaluation in a splice can call arbitrary libraries, including compiled ones (not just bytecode!). Also the current setup lets you inspect the environment a bit with the Q monad. We already have typed and untyped TH; as yet I'm not convinced that the advantages of a third route would pay for the costs. But don't let me discourage you. I love being convinced! (Bandwidth to think about this is limited though, hence delay.) Simon | -----Original Message----- | From: Matthew Pickering <[hidden email]> | Sent: 27 February 2018 11:18 | To: Simon Peyton Jones <[hidden email]> | Subject: Re: An idea for a different style of metaprogramming | evaluation using the optimiser | | Sorry I wasn't clear. I'll try to flesh out the power example in much | more detail. The general idea is that instead of using the bytecode | interpreter in order to evaluate splices, the optimiser is often | sufficient. | | The benefits of this are: | | 1. Cross platform. | 2. Simpler to use as you are not manually creating ASTs | 3. More predictable as there are never any IO actions. | | The motivation is to precisely tell the inliner what to do rather than | trying to convince it to do what you like as it is quite | unpredictable. | | | Take the staged power example. | | ``` | power :: Int -> R Int -> R Int | power n k = if n == 0 | then .< 1 >. | else .< ~k * ~(power (n-1) k) >. | ``` | | I propose introducing two new pieces of syntax $$$(..) and [||| ... | |||] which are pronounced splice and quote. We also introduce a new | type constructor R, for representation. We can informally give the | types | of quote and splice as a -> R a and R a -> a respectively. | | What are the limitations of $$$() and [||| |||]? The only way to make | `R` fragments is by quoting. It is also not allowed to inspect the | quoted code. The only way to interact with it is by splicing. | | Which would then mean we staged power like so. | | | ``` | power :: Int -> R Int -> R Int | power n k = if n == 0 | then [||| 1 |||]. | else [||| $$$(k) * $$$(power k (n-1)) |||] | ``` | | Then to use power, we might call it like.. | | ``` | $$$(power 3 [||| x |||]) | ``` | | which we hope evaluates to x * x * x * 1. | | The difference I am proposing is that instead of the bytecode | interpreter evaluating the splice, it is the optimiser which does the | evaluation. | | Concretely, we interpret a splice meaning "evaluate as much as | possible" and a quote as meaning "don't evaluate me any more". | By evaluate as much as possible, this means to be very keen to inline | functions including recursive ones. This sounds bad but it is the user | who is in control by | where they put the annotations. | | So in this example, we would evaluate as follows.. E marks an | "evaluation context" where we are very keen to evaluate. | | ``` | $$$(power 1 [||| x |||]) | => | E: power 1 [||| x |||] | => | E: if 1 == 0 then [||| 1 |||] else ... | => (Eval condition) | E: [||| $$$([||| x |||]) * $$$(power (1-0)) k |||] | => (Quote removes evaluation context) | $$$([||| x |||]) * $$$(power (1-0) k) | => (Eval $$$(k)) | x * $$$(power (1-0) k) | => (Eval $$$(power (1-0) k) | k * E: power (1 - 0) k | => (Unroll as we are in E) | k * E: if 0 == 0 then [||| 1 |||] ... | => | k* E: [||| 1 |||] | => | k * 1 | ``` | | So we can completely evaluate the splices properly in the evaluator if | the definitions are simple enough. | In this example, that isn't actually going to be the case as the | optimiser doesn't evaluate `3 == 0` for integers because it is | implemented using a primitive. | | If we rewrite the example using an inductive data type for the | recursive argument then we can see that it would work correctly. | | ``` | data Nat = Z | S Nat | | power :: Nat -> R Int -> R Int | power n k = case n of | Z -> [||| 1 |||] | (S n) -> [||| $$$(k) * $$$(power n k) |||] | ``` | | Then it is perhaps clearer that the optimiser is all we need in order | to evaluate $$$(power (S Z) [||| x |||]). The current implementation | of $$ and [||] would invoke the bytecode interpreter to do this | evaluation but it is unnecessary for this simple example which the | optimiser could do just as well. | | ``` | $$$(power (S Z) [|| x ||]) | => | E: power (S Z) [|| x ||] | => | E: case (S Z) of | Z -> [||| 1 |||] | (S n) -> [||| $$$([||| x |||]) * $$$(power n [||| x |||]) |||] | => (Case of known constructor) | E: [||| $$$([|| x ||]) * $$$(power Z [||| x |||) |||] | => Quote | $$$([||| x |||]) * $$$(power Z [||| x |||]) | => | k * $$$(power Z [||| x |||]) | => Splice | k * E: power Z [||| x |||] | => Inline power as we in E | k * E (case Z of Z -> [||| 1 |||]; S n ... ->) | => Case of known constructor | k * E: [||| 1 |||] | => | k * 1 | ``` | | | Hope that clears some things up. | | | Matt | | On Tue, Feb 27, 2018 at 10:37 AM, Simon Peyton Jones | <[hidden email]> wrote: | > Matthew, I'm afraid I don't understand the proposal at all. Can you | give a few examples? | > | > S | > | > | -----Original Message----- | > | From: ghc-devs [mailto:[hidden email]] On Behalf Of | > | Matthew Pickering | > | Sent: 27 February 2018 09:59 | > | To: GHC developers <[hidden email]> | > | Subject: An idea for a different style of metaprogramming | evaluation | > | using the optimiser | > | | > | I've had an idea for a while of a different way we could evaluate | TH- | > | like splices which would be more lightweight and easier to work | with. | > | | > | The idea is to create a third quotation/splicing mechanism which | has | > | no introspection (like MetaML) but then to evaluate these quotes | and | > | splices in the optimiser rather than using the bytecode | interpreter. | > | | > | I am motivated by the many examples of recursive functions in | > | libraries, which when given a statically known argument should be | able | > | to be unrolled to produce much better code. It is understandable | that | > | the compiler doesn't try to do this itself but there should be an | easy | > | way for the user to direct the compiler to do so. (See also | > | | https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.r | > | | eddit.com%2Fr%2Fhaskell%2Fcomments%2F7yvb43%2Fghc_compiletime_evaluati | > | | on%2F&data=04%7C01%7Csimonpj%40microsoft.com%7C9979850e7fb142a7e11c08d | > | | 57dc8dcb8%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636553224038023 | > | | 725%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTi | > | I6Ik1haWwifQ%3D%3D%7C- | > | | 1&sdata=%2FOQOcFIXduGY9wtWWEkbqd3qQ1715oCdaim0pCeSgrI%3D&reserved=0) | > | | > | An example to be concrete: | > | | > | Take the power function such that power k n computes k^n. | > | | > | power :: Int -> Int -> Int | > | power k n = if n == 0 | > | then 1 | > | else k * power k (n - 1) | > | | > | If we statically know n then we can create a staged version. We | use R | > | to indicate that an argument is dynamically known. | > | | > | power :: R Int -> Int -> R Int | > | power k n = if n == 0 | > | then .< 1 >. | > | else .< ~k * ~(power k (n-1)) >. | > | | > | One way to implement this in Haskell is to use typed template | haskell | > | quoting and splicing. | > | The key thing to notice about why this works is that in order to | > | splice `power k (n-1)` we need to evaluate `power k (n-1)` so | that we | > | have something of type `R Int` which we can then splice back into | the | > | quote. | > | | > | The way this is currently implemented is that the bytecode | interpreter | > | runs these splices in order to find out what they evaluate to. | > | | > | The idea is that instead, we add another mode which runs splices | in | > | the core-to-core optimiser. The optimiser performs evaluation by | beta | > | reduction, inlining and constant folding. For simple definitions | on | > | algebraic data types it does a very good job of eliminating | overhead. | > | As long as the call is not recursive. If we can just get the | optimiser | > | to inline recursive calls in a controlled manner as well, it | should do | > | a good job on the unrolled definition. | > | | > | The benefits of this are that it works across all platforms and | fits | > | nicely already into the compilation pipeline. It also fits in | nicely | > | with the intuition that a "quote" means to stop evaluating and a | > | "splice" means to evaluate. | > | | > | A disadvantage is that the optimiser is only a *partial* | evaluator | > | rather than an evaluator. It gets stuck evaluating things | containing | > | primitives and so on. I don't forsee this as a major issue but a | > | limitation that library authors should be aware of when writing | their | > | staged programs. To go back to the power example, the recursive | > | condition would have to be an inductively defined natural (data N | = Z | > | | S N) rather than an Int as the comparison operator for integers | > | can't be evaluated by the optimiser. It is already rather easy to | > | write staged programs which loop if you don't carefully consider | the | > | staging so this seems now worse than the status quo. | > | | > | Does this sound at all sensible to anyone else? | > | | > | Matt | > | _______________________________________________ | > | ghc-devs mailing list | > | [hidden email] | > | | https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fmail.h | > | askell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fghc- | > | | devs&data=04%7C01%7Csimonpj%40microsoft.com%7C9979850e7fb142a7e11c08d5 | > | | 7dc8dcb8%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C6365532240380237 | > | | 25%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI | > | 6Ik1haWwifQ%3D%3D%7C- | > | | 1&sdata=CJgS3uH4hHw5Zups1fAiUtOnTzn%2ByCh1bg1auiMBcYM%3D&reserved=0 _______________________________________________ ghc-devs mailing list [hidden email] http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs |
Free forum by Nabble | Edit this page |