Polishing the README and iterator protocol.

* README.md: added some small documentation and loop expansion.
 * goof.scm: Changed the iterator protocol to not use unnecessary :acc and :fors.
 * iterators.scm: see above.
 * ref-let.scm: a new macro to allow multiple values and pattern matching to co-exist for body-bindings.
This commit is contained in:
Linus 2020-12-16 19:54:55 +01:00
parent aab9fcabb0
commit 2d35c9d6cf
4 changed files with 220 additions and 19 deletions

152
README.md
View file

@ -7,7 +7,7 @@ goof-loops aims to be an amalgamation of the racket for loops and Alex Shinn's (
```
(loop ((:for a (in 0 b))
(:for b (in 1 (+ a b)))
(count (up-from 0 (to 1000)))
(:for count (up-from 0 (to 1000)))
(:acc acc (listing b)))
=> acc
(display b) (newline))
@ -44,13 +44,13 @@ Accumulators can be in any of the loop's stages:
### syntactical
for-clauses are split into :for and :let clauses. This is because the addition of subloops means we have to treat accumulators differently.
for-clauses are split into :for and :acc clauses. This is because the addition of subloops means we have to treat accumulators differently.
while and until are removed in favour of :break.
:when and :unless are added to better control when the loop body is executed (and accumulators accumulated)
with-clauses are removed in favour of (:forvar (in init [step [stop]])) or (:acc var (folding init [step])) in case of accumulators.
with-clauses are removed in favour of (:for var (in init [step [stop]])) or (:acc var (folding init [step])) in case of accumulators.
### Regressions compared to foof-loop
@ -96,12 +96,158 @@ Named updates also work.
;; => (values (1 3 5) (2 4))
```
### Simple forms
I also provide simplified forms for many common operations. Omitting :for is allowed, and :acc clauses are not allowed.
```
(loop/list ((a (up-from 0 3)))
a)
;; => (0 1 2)
(loop/sum ((:for a (up-from 1 4))) a)
;; => 6
(loop/product ((a (in-list '(2 3 4))))
a)
;; => 24
(loop/first ((a (in-list '(a b c 3 4 d))) :when (integer? a))
(display a)
a)
;; => displays 3 and returns 3.
(loop/last ((a (in-list '(a b c d e f))) :break (eq? a 'e))
a)
;; => 'd
(loop/and ((a (in-list '(1 2 3 'error))))
(< a 3))
;; => #f
(loop/or ((a (in-list '(1 2 3 4))))
(symbol? a))
;; => #f
(loop/list/parallel ((a (in-list '(42 41 43))))
(expensive-function a))
;; => same result as loop/list, but faster if the problem parallelizes well
```
### Loop expansion
A goof loop expands into something looking like this:
```
(let* (<outer-let>)
(letrec ((final-function (lambda (<final-binding>) <final-expr>))
(goof-loop (lambda (<accumulator> ... <loop-var> ...)
(if (or <check> ...)
(begin
<for-clause-finalizer> ...
(final-function (<accumulator-finalizer> <accumulator>) ...))
(let ((<body-binding> ... <body-binding-expr>) ...)
(let ((<user-binding> ... <user-binding-expr>) ...)
(match-let ((<parenthesised-pattern> <match-expr>))
(if (and <when-expr> ...)
(cond
((or <user-break> ...)
<for-clause-finalizer> ...
(final-function (<accumulator-finalizer> <accumulator>) ...))
(else
<loop-body>
(goof-loop <accumulate> ... <loop-var-next> ...))
(goof-loop <accumulator> ... <loop-var-next> ...))))))))
(goof-loop <accumulator-init> ... <loop-var-init> ...)))
```
<outer-let>: are provided by accumulators or for clauses for bindings that are not passed as an argument to the loop, for example a vector. The vector is bound here, and the index into the vector is the thing iterated over.
<final-binding> and <final-expr>: When the iteration ends, this function is called with the results of the :acc clauses. In the case of (:acc lst-acc (listing ...)), the name of the accumulator is never lst-acc in the loop body, but only in the <final-expr>. In case of (listing ...) the accumulated results are reversed before the final function.
<accumulator> and <loop-variable>: <accumulator> holds the current state of an accumulator clause. This is not necessarily the same binding as the user provided as the name, as described above. <loop-var> is the current state of a :for clause.
<check>: Checks for :for-clauses. In the case of (in-list ...) this would check for (not (pair? ...)).
<for-clause-finalizer>: some :for clauses need to be finalized. In the case of (in-file ...) the open file handle is closed at any point where the iteration stops.
<accumulator-finalizer>: <accumulator-finalizer> is any preprocessing done to <accumulator> before passing it on to the final-function. In the case of (listing ...) that would be (reverse ...).
<body-binding> and <body-binding-expr>:<body-binding> are the names the user provided for the body bindings. In the case of (:for a (in-list '(1 2 3))) the body binding would be (a (car name-of-loop-variable)). The body binding may be an (ice-9 match) pattern. More on that below.
<parenthesised-pattern> and <match-expr>: If a <user-binding> is not an identifier, it is presumed to be a match-let pattern. The result is bound to a variable and matched against this match-let.
<when-expr>: the user supplied :when or :unless guard expression.
<user-break>: user-supplied :break guard.
<loop-body>, <accumulate>, and <loop-var-next>: The user supplied body of the loop. If the loop is not named (i.e: in loops where the user controls the iteration) an expression for the next loop iteration is added to the body. <accumulate> is the expression the accumulator clause provided to accumulate a new value. For (:acc acc (listing elem)) that is (cons elem acc). <loop-var-next> is the expression evaluated to get the next iteration's loop variable. In the case of (in-list lst) that is (cdr lst). If a loop name is provided there is no implicit next loop.
<accumulator-init> and <loop-var-init>: <accumulator-init> are ALL accumulator init values, including the ones in subloops. For (listing ...) that is the empty list. <loop-var-init> is the initial loop vars.
In case of subloops, those are placed instead of <loop-body>. They use the same final-function, and instead of quitting when any <check> triggers they go out to the outer loop.
### Speed
Speed is good. Despite the rather involved expansion above, due to dead-code elimination, the actual expansion shows some good code:
```
> ,opt (loop ((:for a (in-list '(1 2 3 4)))
:when (even? a)
(:acc acc (listing a)))
=> acc)
$1 = (let loopy-loop ((cursor-1 '()) (cursor '(1 2 3 4)))
(if (pair? cursor)
(let ((a (car cursor)) (succ (cdr cursor)))
(if (even? a)
(loopy-loop (cons a cursor-1) succ)
(loopy-loop cursor-1 succ)))
(reverse cursor-1)))
;; loop/list, being less general, produces faster code that can be more easily unroled and optimized.
> ,opt (loop/list ((a (in-list '(1 2 3 4)))
:when (even? a))
a)
$2 = (list 2 4)
;; Removing the opportunity to completely remove the loop
> ,opt (loop/list ((a (in-list (read)))
:when (even? a))
a)
$5 = (let loopy-loop ((cursor (read)))
(if (pair? cursor)
(let ((a (car cursor)) (succ (cdr cursor)))
(if (even? a)
(cons a (loopy-loop succ))
(loopy-loop
;; The code expansion of the partition procedure above produces
(define (partition list predicate)
(let loopy-loop ((satisfied '()) (unsatisfied '()) (cursor list))
(if (pair? cursor)
(let ((element (car cursor)) (succ (cdr cursor)))
(if (predicate element)
(loopy-loop (cons element satisfied)
unsatisfied
succ)
(loopy-loop satisfied
(cons element unsatisfied)
succ)))
(values (reverse satisfied) (reverse unsatisfied)))))
```
## Todo
Tests and documentation.
Fix the inlining behavious of some of the :for iterators.
add generator support for all provided iterators
## foof, what a guy
I have previously expressed some admiration for Alex and I will do it again. The source of chibi loop is extremely elegant, and all but the hairiest part is written in syntax-rules. Not only has he written my two favourite SRFIs, his input in all the other discussions I have seen is always on-point, pragmatic and generally fantastic. He neither knows of this project, nor embraces it in any way. Y'all should go look at the source of (chibi loop) though.