This library provides a convenient interface to the integer set library. It allows reasoning sets and relations of integers bounded by affine constraints.
We mostly ported value, union-set, union-map, constraints, and everything related to the ast generation
We made some changes compared to cl-isl, to try to simplify usage.
For instance context
is a special variable and not required to be supplied as an argument.
For instance union-set-from-str
takes only a string, not a context and a string.
You also don’t have to collect memory. When your lisp object is collected, it calls the C free function associated.
Moreover, every value is copied by default. So you don’t have to copy arguments when you use functions.
For instance union-set-intersect
does not destroy arguments when used.
So (let ((a (union-set-from-str " { [1] } "))) (union-set-intersect a a))
is perfectly valid!
You can find everything we ported in the code/packages.lisp
file
We made a file for each object. You should probably do the same.
The simplest example is the file basic-map.lisp. It first define the object with define-isl-object
, then how to print it, and then some function. A file which defines a lot of functions for its object is union-map.lisp
Sometimes you want to do more than just transposing functions. Then you want to look at value.lisp.
Be careful, sometimes functions exist on the isl library, but aren’t listed on the pdf documentation. And sometimes the swig import is buggy.
(let ((a (isl:union-set-from-str " { [1] } "))
(b (isl:union-set-from-str " { [i] } ")))
(isl:union-set-intersect a b) ; #<CL-ISL:UNION-SET { [1] }>
(isl:union-set-union a b)) ; #<CL-ISL:UNION-SET { [i] }>
(let ((a (isl:union-set-from-str " { [1] } "))
(b (isl:union-set-from-str " { [i]: i >= 10 } ")))
(isl:union-set-intersect a b) ; #<CL-ISL:UNION-SET { }>
(isl:union-set-union a b)) ; #<CL-ISL:UNION-SET { [i0] : i0 >= 10; [1] }>
(let ((a (isl:union-set-from-str " { [i]: i >= -5 } "))
(b (isl:union-set-from-str " { [i]: i <= 5 } ")))
(isl:union-set-intersect a b); #<CL-ISL:UNION-SET { [i] : -5 <= i <= 5 }>
(isl:union-set-union a b)) ; #<CL-ISL:UNION-SET { [i] : i >= -5 or i <= 5 }>
(let ((a (isl:union-map-from-str " { [i] -> [i + 1] } ")))
(isl:union-map-apply-range a a)) ; #<CL-ISL:UNION-MAP { [i] -> [2 + i] }>
(let ((a (isl:union-map-from-str "{ A [2 ,8 ,1] -> B [5]; A [2 ,8 ,1] -> B [6]; B [5] -> B [5] } "))
(b (isl:union-set-from-str " { A [2 ,8 ,1]; B [5] } ")))
(isl:union-set-apply b a)) ; #<CL-ISL:UNION-SET { B[6]; B[5] }>
(let ((a (isl:union-map-from-str " { [i] -> [i+1] } "))
(b (isl:union-set-from-str " { [i]: i <= 9 } ")))
(isl:union-set-apply b a)) ; #<CL-ISL:UNION-SET { [i0] : i0 <= 10 }>
We created this library to use it in Loopus, a loop optimization framework for common lisp. One feature is the factorization of expression outside of loops (when possible).
For example, these two loops are equivalent
(loopus:for (i 0 5)
(print (aref array 0)))
(let ((element (aref array 0)))
(loopus:for (i 0 5)
(print element)))
This is because (aref array 0)
is pure (this is done by ir-specialize.lisp), and because array is not modified in the loop!
The analysis of if array is modified or not can be quite complex, for instance on the following loop we should still be able to factorize this expression.
(loopus:for (i 1 5)
(setf (aref array i) i)
(print (aref array 0)))
To do so, we detect where the pure statement may read in memory. Here the set of value it can read is { array[0] }
.
We then extract where the loop may write in memory. Here it’s { array[i]: 1 <= i < 5 }
.
If the intersection is empty, it’s ok, otherwise it’s not! Thankfully cl-isl has the function union-set-intersect
which takes two sets and return the intersection. We just have to check if the result is empty with union-set-emptyp
(todo add this function).
This approach also works to move a pure statement to the most outer loop possible. For instance consider this code
(loopus:for (i 1 10)
(setf (aref array i) (random 10))
(loopus:for (j 1 10)
(setf (aref array (+ i j) (random 10)))
(print (aref array 1)))))
The value that is read is array[1]
, the inner loop only writes to array[k]
with k >= 2
, so it’s fine to factorize this statement a loop above. But the very outloop can write to array[1]
, so we only factorize one level.
So to compute that with cl-isl, you can do:
- The domain of iteration of iteration the outerloop is
{ [i, j] : 1 <= i < 10 and 1 <= j < 10 }
- The domain of iteration of iteration the innerloop is
[i] -> { [i, j] : 1 <= j < 10 }
- What is read in the instruction in the innerloop is
{ [i, j] -> array[1] }
- What is written in the instruction in the inner loop is
{ [i, j] -> array[i+j] }
- The total of the thing written in the innerloop is
[i] -> { array[i0] : i < i0 <= 9 + i }
- Let’s compute the intersection for the innerloop, it’s
[i] -> { array[1] : -8 <= i <= 0 }
- Because the result depends on i, let’s add the information that i is between 1 and 10. The result is the empty set! So we can move this statement outside the innerloop.
- Just to double check, if the outerloop is from 0 to 10, the result would be
[i] -> { array[1] : i = 0 }
- Also, the total of the thing written in the outerloop is
{ array[i0] : 0 < i0 <= 18 and (i0 >= 2 or i0 <= 9) }
- So the intersection with
{ array[1] }
is non empty! So we can’t move outside the outerloop
The full repl session:
CL-USER> (in-package :cl-isl)
#<PACKAGE "CL-ISL">
;; The domain of iteration of iteration the outer loop is { [i, j] : 1 <= i < 10 and 1 <= j < 10 }
ISL> (setf domain-outer (union-set-from-str " { [i, j] : 1 <= i < 10 and 1 <= j < 10 } "))
#<UNION-SET { [i, j] : 0 < i <= 9 and 0 < j <= 9 }>
;; The domain of iteration of iteration the inner loop is [i] -> { [i, j] : 1 <= j < 10 }
ISL> (setf domain-inner (union-set-from-str " [i] -> { [i, j] : 1 <= j < 10 } "))
#<UNION-SET [i] -> { [i, j] : 0 < j <= 9 }>
;; What is read in the instruction in the inner loop is { [i, j] -> array[1] }
ISL> (setf vread (union-map-from-str " { [i, j] -> array[1] } "))
#<UNION-MAP { [i, j] -> array[1] }>
;; What is written in the instruction in the inner loop is { [i, j] -> array[i+j] }
ISL> (setf vwrite (union-map-from-str " { [i, j] -> array[i+j] } "))
#<UNION-MAP { [i, j] -> array[i + j] }>
;; The total of the thing written in the innerloop is [i] -> { array[i0] : i < i0 <= 9 + i }
ISL> (setf all-written-inner (union-set-apply domain-inner vwrite))
#<UNION-SET [i] -> { array[i0] : i < i0 <= 9 + i }>
;; Let's compute the intersection for the innerloop, it's [i] -> { array[1] : -8 <= i <= 0 }
ISL> (setf intersection-inner (union-set-intersect (union-set-apply domain-inner vread) all-written-inner))
#<UNION-SET [i] -> { array[1] : -8 <= i <= 0 }>
;; Because the result depends on i, let's add the information that i is between 1 and 10. The result is the empty set! So we can move this statement outside the innerloop.
ISL> (union-set-intersect-params intersection-inner (set-from-str "[i] -> {: 1 <= i < 10}"))
#<UNION-SET [i] -> { }>
;; Just to double check, if the outerloop is from 0 to 10, the result would be [i] -> { array[1] : i = 0 }
ISL> (union-set-intersect-params intersection-inner (set-from-str "[i] -> {: 0 <= i < 10}"))
#<UNION-SET [i] -> { array[1] : i = 0 }>
;; Also, the total of the thing written in the outerloop is { array[i0] : 0 < i0 <= 18 and (i0 >= 2 or i0 <= 9) }
ISL> (setf all-written-outer (union-set-union (union-set-apply domain-outer vwrite) (union-set-from-str " {array[i]: 1 <= i < 10}")))
#<UNION-SET { array[i0] : 0 < i0 <= 18 and (i0 >= 2 or i0 <= 9) }>
;; So the intersection with { array[1] } is non empty! So we can't move outside the outerloop
ISL> (union-set-intersect (union-set-from-str " { array[1] }") all-written-outer)
#<UNION-SET { array[1] }>
In practice in Loopus we do not do this computation yet, there is probably something less naive to do. But you can find how we used this library in the library Loopus in the files input-loopus-ir and output-loopus-ir.
- Documentation of isl the C library https://libisl.sourceforge.io/manual.pdf
- Tutorial of isl https://libisl.sourceforge.io/tutorial.pdf
- A wrapper around isl with which you can play on your browser https://compsys-tools.ens-lyon.fr/iscc