1 \chapter[chap:normalization]{Normalization}
2 % A helper to print a single example in the half the page width. The example
3 % text should be in a buffer whose name is given in an argument.
5 % The align=right option really does left-alignment, but without the program
6 % will end up on a single line. The strut=no option prevents a bunch of empty
7 % space at the start of the frame.
9 \framed[offset=1mm,align=right,strut=no,background=box,frame=off]{
10 \setuptyping[option=LAM,style=sans,before=,after=,strip=auto]
12 \setuptyping[option=none,style=\tttf,strip=auto]
16 \define[4]\transexample{
17 \placeexample[here][ex:trans:#1]{#2}
18 \startcombination[2*1]
19 {\example{#3}}{Original program}
20 {\example{#4}}{Transformed program}
24 The first step in the core to \small{VHDL} translation process, is normalization. We
25 aim to bring the core description into a simpler form, which we can
26 subsequently translate into \small{VHDL} easily. This normal form is needed because
27 the full core language is more expressive than \small{VHDL} in some areas and because
28 core can describe expressions that do not have a direct hardware
31 \todo{Describe core properties not supported in \VHDL, and describe how the
32 \VHDL we want to generate should look like.}
35 \todo{Refresh or refer to distinct hardware per application principle}
36 The transformations described here have a well-defined goal: To bring the
37 program in a well-defined form that is directly translatable to hardware,
38 while fully preserving the semantics of the program. We refer to this form as
39 the \emph{normal form} of the program. The formal definition of this normal
42 \placedefinition{}{A program is in \emph{normal form} if none of the
43 transformations from this chapter apply.}
45 Of course, this is an \quote{easy} definition of the normal form, since our
46 program will end up in normal form automatically. The more interesting part is
47 to see if this normal form actually has the properties we would like it to
50 But, before getting into more definitions and details about this normal form,
51 let's try to get a feeling for it first. The easiest way to do this is by
52 describing the things we want to not have in a normal form.
55 \item Any \emph{polymorphism} must be removed. When laying down hardware, we
56 can't generate any signals that can have multiple types. All types must be
57 completely known to generate hardware.
59 \item Any \emph{higher order} constructions must be removed. We can't
60 generate a hardware signal that contains a function, so all values,
61 arguments and returns values used must be first order.
63 \item Any complex \emph{nested scopes} must be removed. In the \small{VHDL}
64 description, every signal is in a single scope. Also, full expressions are
65 not supported everywhere (in particular port maps can only map signal
66 names and constants, not complete expressions). To make the \small{VHDL}
67 generation easy, a separate binder must be bound to ever application or
71 \todo{Intermezzo: functions vs plain values}
73 A very simple example of a program in normal form is given in
74 \in{example}[ex:MulSum]. As you can see, all arguments to the function (which
75 will become input ports in the final hardware) are at the outer level.
76 This means that the body of the inner lambda abstraction is never a
77 function, but always a plain value.
79 As the body of the inner lambda abstraction, we see a single (recursive)
80 let expression, that binds two variables (\lam{mul} and \lam{sum}). These
81 variables will be signals in the final hardware, bound to the output port
82 of the \lam{*} and \lam{+} components.
84 The final line (the \quote{return value} of the function) selects the
85 \lam{sum} signal to be the output port of the function. This \quote{return
86 value} can always only be a variable reference, never a more complex
89 \todo{Add generated VHDL}
92 alu :: Bit -> Word -> Word -> Word
101 \startuseMPgraphic{MulSum}
102 save a, b, c, mul, add, sum;
105 newCircle.a(btex $a$ etex) "framed(false)";
106 newCircle.b(btex $b$ etex) "framed(false)";
107 newCircle.c(btex $c$ etex) "framed(false)";
108 newCircle.sum(btex $res$ etex) "framed(false)";
111 newCircle.mul(btex * etex);
112 newCircle.add(btex + etex);
114 a.c - b.c = (0cm, 2cm);
115 b.c - c.c = (0cm, 2cm);
116 add.c = c.c + (2cm, 0cm);
117 mul.c = midpoint(a.c, b.c) + (2cm, 0cm);
118 sum.c = add.c + (2cm, 0cm);
121 % Draw objects and lines
122 drawObj(a, b, c, mul, add, sum);
124 ncarc(a)(mul) "arcangle(15)";
125 ncarc(b)(mul) "arcangle(-15)";
131 \placeexample[here][ex:MulSum]{Simple architecture consisting of a
132 multiplier and a subtractor.}
133 \startcombination[2*1]
134 {\typebufferlam{MulSum}}{Core description in normal form.}
135 {\boxedgraphic{MulSum}}{The architecture described by the normal form.}
138 The previous example described composing an architecture by calling other
139 functions (operators), resulting in a simple architecture with components and
140 connections. There is of course also some mechanism for choice in the normal
141 form. In a normal Core program, the \emph{case} expression can be used in a
142 few different ways to describe choice. In normal form, this is limited to a
145 \in{Example}[ex:AddSubAlu] shows an example describing a
146 simple \small{ALU}, which chooses between two operations based on an opcode
147 bit. The main structure is similar to \in{example}[ex:MulSum], but this
148 time the \lam{res} variable is bound to a case expression. This case
149 expression scrutinizes the variable \lam{opcode} (and scrutinizing more
150 complex expressions is not supported). The case expression can select a
151 different variable based on the constructor of \lam{opcode}.
153 \startbuffer[AddSubAlu]
154 alu :: Bit -> Word -> Word -> Word
166 \startuseMPgraphic{AddSubAlu}
167 save opcode, a, b, add, sub, mux, res;
170 newCircle.opcode(btex $opcode$ etex) "framed(false)";
171 newCircle.a(btex $a$ etex) "framed(false)";
172 newCircle.b(btex $b$ etex) "framed(false)";
173 newCircle.res(btex $res$ etex) "framed(false)";
175 newCircle.add(btex + etex);
176 newCircle.sub(btex - etex);
179 opcode.c - a.c = (0cm, 2cm);
180 add.c - a.c = (4cm, 0cm);
181 sub.c - b.c = (4cm, 0cm);
182 a.c - b.c = (0cm, 3cm);
183 mux.c = midpoint(add.c, sub.c) + (1.5cm, 0cm);
184 res.c - mux.c = (1.5cm, 0cm);
187 % Draw objects and lines
188 drawObj(opcode, a, b, res, add, sub, mux);
190 ncline(a)(add) "posA(e)";
191 ncline(b)(sub) "posA(e)";
192 nccurve(a)(sub) "posA(e)", "angleA(0)";
193 nccurve(b)(add) "posA(e)", "angleA(0)";
194 nccurve(add)(mux) "posB(inpa)", "angleB(0)";
195 nccurve(sub)(mux) "posB(inpb)", "angleB(0)";
196 nccurve(opcode)(mux) "posB(n)", "angleA(0)", "angleB(-90)";
197 ncline(mux)(res) "posA(out)";
200 \placeexample[here][ex:AddSubAlu]{Simple \small{ALU} supporting two operations.}
201 \startcombination[2*1]
202 {\typebufferlam{AddSubAlu}}{Core description in normal form.}
203 {\boxedgraphic{AddSubAlu}}{The architecture described by the normal form.}
206 As a more complete example, consider \in{example}[ex:NormalComplete]. This
207 example contains everything that is supported in normal form, with the
208 exception of builtin higher order functions. The graphical version of the
209 architecture contains a slightly simplified version, since the state tuple
210 packing and unpacking have been left out. Instead, two seperate registers are
211 drawn. Also note that most synthesis tools will further optimize this
212 architecture by removing the multiplexers at the register input and
213 instead put some gates in front of the register's clock input, but we want
214 to show the architecture as close to the description as possible.
216 As you can see from the previous examples, the generation of the final
217 architecture from the normal form is straightforward. In each of the
218 examples, there is a direct match between the normal form structure,
219 the generated VHDL and the architecture shown in the images.
221 \startbuffer[NormalComplete]
224 -> State (Word, Word)
225 -> (State (Word, Word), Word)
227 -- All arguments are an inital lambda (address, data, packed state)
229 -- There are nested let expressions at top level
231 -- Unpack the state by coercion (\eg, cast from
232 -- State (Word, Word) to (Word, Word))
233 s = sp ▶ (Word, Word)
234 -- Extract both registers from the state
235 r1 = case s of (a, b) -> a
236 r2 = case s of (a, b) -> b
237 -- Calling some other user-defined function.
239 -- Conditional connections
251 -- pack the state by coercion (\eg, cast from
252 -- (Word, Word) to State (Word, Word))
253 sp' = s' ▶ State (Word, Word)
254 -- Pack our return value
261 \startuseMPgraphic{NormalComplete}
262 save a, d, r, foo, muxr, muxout, out;
265 newCircle.a(btex \lam{a} etex) "framed(false)";
266 newCircle.d(btex \lam{d} etex) "framed(false)";
267 newCircle.out(btex \lam{out} etex) "framed(false)";
269 %newCircle.add(btex + etex);
270 newBox.foo(btex \lam{foo} etex);
271 newReg.r1(btex $\lam{r1}$ etex) "dx(4mm)", "dy(6mm)";
272 newReg.r2(btex $\lam{r2}$ etex) "dx(4mm)", "dy(6mm)", "reflect(true)";
274 % Reflect over the vertical axis
275 reflectObj(muxr1)((0,0), (0,1));
278 rotateObj(muxout)(-90);
280 d.c = foo.c + (0cm, 1.5cm);
281 a.c = (xpart r2.c + 2cm, ypart d.c - 0.5cm);
282 foo.c = midpoint(muxr1.c, muxr2.c) + (0cm, 2cm);
283 muxr1.c = r1.c + (0cm, 2cm);
284 muxr2.c = r2.c + (0cm, 2cm);
285 r2.c = r1.c + (4cm, 0cm);
287 muxout.c = midpoint(r1.c, r2.c) - (0cm, 2cm);
288 out.c = muxout.c - (0cm, 1.5cm);
290 % % Draw objects and lines
291 drawObj(a, d, foo, r1, r2, muxr1, muxr2, muxout, out);
294 nccurve(foo)(muxr1) "angleA(-90)", "posB(inpa)", "angleB(180)";
295 nccurve(foo)(muxr2) "angleA(-90)", "posB(inpb)", "angleB(0)";
296 nccurve(muxr1)(r1) "posA(out)", "angleA(180)", "posB(d)", "angleB(0)";
297 nccurve(r1)(muxr1) "posA(out)", "angleA(0)", "posB(inpb)", "angleB(180)";
298 nccurve(muxr2)(r2) "posA(out)", "angleA(0)", "posB(d)", "angleB(180)";
299 nccurve(r2)(muxr2) "posA(out)", "angleA(180)", "posB(inpa)", "angleB(0)";
300 nccurve(r1)(muxout) "posA(out)", "angleA(0)", "posB(inpb)", "angleB(-90)";
301 nccurve(r2)(muxout) "posA(out)", "angleA(180)", "posB(inpa)", "angleB(-90)";
303 nccurve(a)(muxout) "angleA(-90)", "angleB(180)", "posB(sel)";
304 nccurve(a)(muxr1) "angleA(180)", "angleB(-90)", "posB(sel)";
305 nccurve(a)(muxr2) "angleA(180)", "angleB(-90)", "posB(sel)";
306 ncline(muxout)(out) "posA(out)";
309 \todo{Don't split registers in this image?}
310 \placeexample[here][ex:NormalComplete]{Simple architecture consisting of an adder and a
312 \startcombination[2*1]
313 {\typebufferlam{NormalComplete}}{Core description in normal form.}
314 {\boxedgraphic{NormalComplete}}{The architecture described by the normal form.}
319 \subsection[sec:normalization:intendednormalform]{Intended normal form definition}
320 Now we have some intuition for the normal form, we can describe how we want
321 the normal form to look like in a slightly more formal manner. The following
322 EBNF-like description completely captures the intended structure (and
323 generates a subset of GHC's core format).
325 Some clauses have an expression listed in parentheses. These are conditions
326 that need to apply to the clause.
328 \defref{intended normal form definition}
329 \todo{Fix indentation}
331 \italic{normal} := \italic{lambda}
332 \italic{lambda} := λvar.\italic{lambda} (representable(var))
334 \italic{toplet} := letrec [\italic{binding}...] in var (representable(var))
335 \italic{binding} := var = \italic{rhs} (representable(rhs))
336 -- State packing and unpacking by coercion
337 | var0 = var1 ▶ State ty (lvar(var1))
338 | var0 = var1 ▶ ty (var1 :: State ty ∧ lvar(var1))
339 \italic{rhs} := userapp
342 | case var of C a0 ... an -> ai (lvar(var))
344 | case var of (lvar(var))
345 [ DEFAULT -> var ] (lvar(var))
346 C0 w0,0 ... w0,n -> var0
348 Cm wm,0 ... wm,n -> varm (\forall{}i \forall{}j, wi,j \neq vari, lvar(vari))
349 \italic{userapp} := \italic{userfunc}
350 | \italic{userapp} {userarg}
351 \italic{userfunc} := var (gvar(var))
352 \italic{userarg} := var (lvar(var))
353 \italic{builtinapp} := \italic{builtinfunc}
354 | \italic{builtinapp} \italic{builtinarg}
355 \italic{builtinfunc} := var (bvar(var))
356 \italic{builtinarg} := var (representable(var) ∧ lvar(var))
357 | \italic{partapp} (partapp :: a -> b)
358 | \italic{coreexpr} (¬representable(coreexpr) ∧ ¬(coreexpr :: a -> b))
359 \italic{partapp} := \italic{userapp} | \italic{builtinapp}
362 \todo{There can still be other casts around (which the code can handle,
363 e.g., ignore), which still need to be documented here}
365 When looking at such a program from a hardware perspective, the top level
366 lambda's define the input ports. The variable reference in the body of
367 the recursive let expression is the output port. Most function
368 applications bound by the let expression define a component
369 instantiation, where the input and output ports are mapped to local
370 signals or arguments. Some of the others use a builtin construction (\eg
371 the \lam{case} expression) or call a builtin function (\eg \lam{+} or
372 \lam{map}). For these, a hardcoded \small{VHDL} translation is
375 \section[sec:normalization:transformation]{Transformation notation}
376 To be able to concisely present transformations, we use a specific format
377 for them. It is a simple format, similar to one used in logic reasoning.
379 Such a transformation description looks like the following.
384 <original expression>
385 -------------------------- <expression conditions>
386 <transformed expresssion>
391 This format desribes a transformation that applies to \lam{<original
392 expresssion>} and transforms it into \lam{<transformed expression>}, assuming
393 that all conditions apply. In this format, there are a number of placeholders
394 in pointy brackets, most of which should be rather obvious in their meaning.
395 Nevertheless, we will more precisely specify their meaning below:
397 \startdesc{<original expression>} The expression pattern that will be matched
398 against (subexpressions of) the expression to be transformed. We call this a
399 pattern, because it can contain \emph{placeholders} (variables), which match
400 any expression or binder. Any such placeholder is said to be \emph{bound} to
401 the expression it matches. It is convention to use an uppercase letter (\eg
402 \lam{M} or \lam{E}) to refer to any expression (including a simple variable
403 reference) and lowercase letters (\eg \lam{v} or \lam{b}) to refer to
404 (references to) binders.
406 For example, the pattern \lam{a + B} will match the expression
407 \lam{v + (2 * w)} (binding \lam{a} to \lam{v} and \lam{B} to
408 \lam{(2 * w)}), but not \lam{(2 * w) + v}.
411 \startdesc{<expression conditions>}
412 These are extra conditions on the expression that is matched. These
413 conditions can be used to further limit the cases in which the
414 transformation applies, commonly to prevent a transformation from
415 causing a loop with itself or another transformation.
417 Only if these conditions are \emph{all} true, the transformation
421 \startdesc{<context conditions>}
422 These are a number of extra conditions on the context of the function. In
423 particular, these conditions can require some (other) top level function to be
424 present, whose value matches the pattern given here. The format of each of
425 these conditions is: \lam{binder = <pattern>}.
427 Typically, the binder is some placeholder bound in the \lam{<original
428 expression>}, while the pattern contains some placeholders that are used in
429 the \lam{transformed expression}.
431 Only if a top level binder exists that matches each binder and pattern,
432 the transformation applies.
435 \startdesc{<transformed expression>}
436 This is the expression template that is the result of the transformation. If, looking
437 at the above three items, the transformation applies, the \lam{<original
438 expression>} is completely replaced with the \lam{<transformed expression>}.
439 We call this a template, because it can contain placeholders, referring to
440 any placeholder bound by the \lam{<original expression>} or the
441 \lam{<context conditions>}. The resulting expression will have those
442 placeholders replaced by the values bound to them.
444 Any binder (lowercase) placeholder that has no value bound to it yet will be
445 bound to (and replaced with) a fresh binder.
448 \startdesc{<context additions>}
449 These are templates for new functions to add to the context. This is a way
450 to have a transformation create new top level functions.
452 Each addition has the form \lam{binder = template}. As above, any
453 placeholder in the addition is replaced with the value bound to it, and any
454 binder placeholder that has no value bound to it yet will be bound to (and
455 replaced with) a fresh binder.
458 As an example, we'll look at η-abstraction:
462 -------------- \lam{E} does not occur on a function position in an application
463 λx.E x \lam{E} is not a lambda abstraction.
466 η-abstraction is a well known transformation from lambda calculus. What
467 this transformation does, is take any expression that has a function type
468 and turn it into a lambda expression (giving an explicit name to the
469 argument). There are some extra conditions that ensure that this
470 transformation does not apply infinitely (which are not necessarily part
471 of the conventional definition of η-abstraction).
473 Consider the following function, which is a fairly obvious way to specify a
474 simple ALU (Note that \in{example}[ex:AddSubAlu] shows the normal form of this
475 function). The parentheses around the \lam{+} and \lam{-} operators are
476 commonly used in Haskell to show that the operators are used as normal
477 functions, instead of \emph{infix} operators (\eg, the operators appear
478 before their arguments, instead of in between).
481 alu :: Bit -> Word -> Word -> Word
482 alu = λopcode. case opcode of
487 There are a few subexpressions in this function to which we could possibly
488 apply the transformation. Since the pattern of the transformation is only
489 the placeholder \lam{E}, any expression will match that. Whether the
490 transformation applies to an expression is thus solely decided by the
491 conditions to the right of the transformation.
493 We will look at each expression in the function in a top down manner. The
494 first expression is the entire expression the function is bound to.
497 λopcode. case opcode of
502 As said, the expression pattern matches this. The type of this expression is
503 \lam{Bit -> Word -> Word -> Word}, which matches \lam{a -> b} (Note that in
504 this case \lam{a = Bit} and \lam{b = Word -> Word -> Word}).
506 Since this expression is at top level, it does not occur at a function
507 position of an application. However, The expression is a lambda abstraction,
508 so this transformation does not apply.
510 The next expression we could apply this transformation to, is the body of
511 the lambda abstraction:
519 The type of this expression is \lam{Word -> Word -> Word}, which again
520 matches \lam{a -> b}. The expression is the body of a lambda expression, so
521 it does not occur at a function position of an application. Finally, the
522 expression is not a lambda abstraction but a case expression, so all the
523 conditions match. There are no context conditions to match, so the
524 transformation applies.
526 By now, the placeholder \lam{E} is bound to the entire expression. The
527 placeholder \lam{x}, which occurs in the replacement template, is not bound
528 yet, so we need to generate a fresh binder for that. Let's use the binder
529 \lam{a}. This results in the following replacement expression:
537 Continuing with this expression, we see that the transformation does not
538 apply again (it is a lambda expression). Next we look at the body of this
547 Here, the transformation does apply, binding \lam{E} to the entire
548 expression and \lam{x} to the fresh binder \lam{b}, resulting in the
557 Again, the transformation does not apply to this lambda abstraction, so we
558 look at its body. For brevity, we'll put the case statement on one line from
562 (case opcode of Low -> (+); High -> (-)) a b
565 The type of this expression is \lam{Word}, so it does not match \lam{a -> b}
566 and the transformation does not apply. Next, we have two options for the
567 next expression to look at: The function position and argument position of
568 the application. The expression in the argument position is \lam{b}, which
569 has type \lam{Word}, so the transformation does not apply. The expression in
570 the function position is:
573 (case opcode of Low -> (+); High -> (-)) a
576 Obviously, the transformation does not apply here, since it occurs in
577 function position (which makes the second condition false). In the same
578 way the transformation does not apply to both components of this
579 expression (\lam{case opcode of Low -> (+); High -> (-)} and \lam{a}), so
580 we'll skip to the components of the case expression: The scrutinee and
581 both alternatives. Since the opcode is not a function, it does not apply
584 The first alternative is \lam{(+)}. This expression has a function type
585 (the operator still needs two arguments). It does not occur in function
586 position of an application and it is not a lambda expression, so the
587 transformation applies.
589 We look at the \lam{<original expression>} pattern, which is \lam{E}.
590 This means we bind \lam{E} to \lam{(+)}. We then replace the expression
591 with the \lam{<transformed expression>}, replacing all occurences of
592 \lam{E} with \lam{(+)}. In the \lam{<transformed expression>}, the This gives us the replacement expression:
593 \lam{λx.(+) x} (A lambda expression binding \lam{x}, with a body that
594 applies the addition operator to \lam{x}).
596 The complete function then becomes:
598 (case opcode of Low -> λa1.(+) a1; High -> (-)) a
601 Now the transformation no longer applies to the complete first alternative
602 (since it is a lambda expression). It does not apply to the addition
603 operator again, since it is now in function position in an application. It
604 does, however, apply to the application of the addition operator, since
605 that is neither a lambda expression nor does it occur in function
606 position. This means after one more application of the transformation, the
610 (case opcode of Low -> λa1.λb1.(+) a1 b1; High -> (-)) a
613 The other alternative is left as an exercise to the reader. The final
614 function, after applying η-abstraction until it does no longer apply is:
617 alu :: Bit -> Word -> Word -> Word
618 alu = λopcode.λa.b. (case opcode of
619 Low -> λa1.λb1 (+) a1 b1
620 High -> λa2.λb2 (-) a2 b2) a b
623 \subsection{Transformation application}
624 In this chapter we define a number of transformations, but how will we apply
625 these? As stated before, our normal form is reached as soon as no
626 transformation applies anymore. This means our application strategy is to
627 simply apply any transformation that applies, and continuing to do that with
628 the result of each transformation.
630 In particular, we define no particular order of transformations. Since
631 transformation order should not influence the resulting normal form,
632 \todo{This is not really true, but would like it to be...} this leaves
633 the implementation free to choose any application order that results in
634 an efficient implementation.
636 When applying a single transformation, we try to apply it to every (sub)expression
637 in a function, not just the top level function body. This allows us to
638 keep the transformation descriptions concise and powerful.
640 \subsection{Definitions}
641 In the following sections, we will be using a number of functions and
642 notations, which we will define here.
644 \todo{Define substitution (notation)}
646 \subsubsection{Concepts}
647 A \emph{global variable} is any variable (binder) that is bound at the
648 top level of a program, or an external module. A \emph{local variable} is any
649 other variable (\eg, variables local to a function, which can be bound by
650 lambda abstractions, let expressions and pattern matches of case
651 alternatives). Note that this is a slightly different notion of global versus
652 local than what \small{GHC} uses internally.
653 \defref{global variable} \defref{local variable}
655 A \emph{hardware representable} (or just \emph{representable}) type or value
656 is (a value of) a type that we can generate a signal for in hardware. For
657 example, a bit, a vector of bits, a 32 bit unsigned word, etc. Values that are
658 not runtime representable notably include (but are not limited to): Types,
659 dictionaries, functions.
660 \defref{representable}
662 A \emph{builtin function} is a function supplied by the Cλash framework, whose
663 implementation is not valid Cλash. The implementation is of course valid
664 Haskell, for simulation, but it is not expressable in Cλash.
665 \defref{builtin function} \defref{user-defined function}
667 For these functions, Cλash has a \emph{builtin hardware translation}, so calls
668 to these functions can still be translated. These are functions like
669 \lam{map}, \lam{hwor} and \lam{length}.
671 A \emph{user-defined} function is a function for which we do have a Cλash
672 implementation available.
674 \subsubsection{Predicates}
675 Here, we define a number of predicates that can be used below to concisely
676 specify conditions.\refdef{global variable}
678 \emph{gvar(expr)} is true when \emph{expr} is a variable that references a
679 global variable. It is false when it references a local variable.
681 \refdef{local variable}\emph{lvar(expr)} is the complement of \emph{gvar}; it is true when \emph{expr}
682 references a local variable, false when it references a global variable.
684 \refdef{representable}\emph{representable(expr)} or \emph{representable(var)} is true when
685 \emph{expr} or \emph{var} is \emph{representable}.
687 \subsection[sec:normalization:uniq]{Binder uniqueness}
688 A common problem in transformation systems, is binder uniqueness. When not
689 considering this problem, it is easy to create transformations that mix up
690 bindings and cause name collisions. Take for example, the following core
694 (λa.λb.λc. a * b * c) x c
697 By applying β-reduction (see \in{section}[sec:normalization:beta]) once,
698 we can simplify this expression to:
704 Now, we have replaced the \lam{a} binder with a reference to the \lam{x}
705 binder. No harm done here. But note that we see multiple occurences of the
706 \lam{c} binder. The first is a binding occurence, to which the second refers.
707 The last, however refers to \emph{another} instance of \lam{c}, which is
708 bound somewhere outside of this expression. Now, if we would apply beta
709 reduction without taking heed of binder uniqueness, we would get:
715 This is obviously not what was supposed to happen! The root of this problem is
716 the reuse of binders: Identical binders can be bound in different scopes, such
717 that only the inner one is \quote{visible} in the inner expression. In the example
718 above, the \lam{c} binder was bound outside of the expression and in the inner
719 lambda expression. Inside that lambda expression, only the inner \lam{c} is
722 There are a number of ways to solve this. \small{GHC} has isolated this
723 problem to their binder substitution code, which performs \emph{deshadowing}
724 during its expression traversal. This means that any binding that shadows
725 another binding on a higher level is replaced by a new binder that does not
726 shadow any other binding. This non-shadowing invariant is enough to prevent
727 binder uniqueness problems in \small{GHC}.
729 In our transformation system, maintaining this non-shadowing invariant is
730 a bit harder to do (mostly due to implementation issues, the prototype doesn't
731 use \small{GHC}'s subsitution code). Also, the following points can be
735 \item Deshadowing does not guarantee overall uniqueness. For example, the
736 following (slightly contrived) expression shows the identifier \lam{x} bound in
737 two seperate places (and to different values), even though no shadowing
741 (let x = 1 in x) + (let x = 2 in x)
744 \item In our normal form (and the resulting \small{VHDL}), all binders
745 (signals) within the same function (entity) will end up in the same
746 scope. To allow this, all binders within the same function should be
749 \item When we know that all binders in an expression are unique, moving around
750 or removing a subexpression will never cause any binder conflicts. If we have
751 some way to generate fresh binders, introducing new subexpressions will not
752 cause any problems either. The only way to cause conflicts is thus to
753 duplicate an existing subexpression.
756 Given the above, our prototype maintains a unique binder invariant. This
757 means that in any given moment during normalization, all binders \emph{within
758 a single function} must be unique. To achieve this, we apply the following
761 \todo{Define fresh binders and unique supplies}
764 \item Before starting normalization, all binders in the function are made
765 unique. This is done by generating a fresh binder for every binder used. This
766 also replaces binders that did not cause any conflict, but it does ensure that
767 all binders within the function are generated by the same unique supply.
768 \refdef{fresh binder}
769 \item Whenever a new binder must be generated, we generate a fresh binder that
770 is guaranteed to be different from \emph{all binders generated so far}. This
771 can thus never introduce duplication and will maintain the invariant.
772 \item Whenever (a part of) an expression is duplicated (for example when
773 inlining), all binders in the expression are replaced with fresh binders
774 (using the same method as at the start of normalization). These fresh binders
775 can never introduce duplication, so this will maintain the invariant.
776 \item Whenever we move part of an expression around within the function, there
777 is no need to do anything special. There is obviously no way to introduce
778 duplication by moving expressions around. Since we know that each of the
779 binders is already unique, there is no way to introduce (incorrect) shadowing
783 \section{Transform passes}
784 In this section we describe the actual transforms.
786 Each transformation will be described informally first, explaining
787 the need for and goal of the transformation. Then, we will formally define
788 the transformation using the syntax introduced in
789 \in{section}[sec:normalization:transformation].
791 \subsection{General cleanup}
792 These transformations are general cleanup transformations, that aim to
793 make expressions simpler. These transformations usually clean up the
794 mess left behind by other transformations or clean up expressions to
795 expose new transformation opportunities for other transformations.
797 Most of these transformations are standard optimizations in other
798 compilers as well. However, in our compiler, most of these are not just
799 optimizations, but they are required to get our program into intended
802 \subsubsection[sec:normalization:beta]{β-reduction}
803 \defref{beta-reduction}
804 β-reduction is a well known transformation from lambda calculus, where it is
805 the main reduction step. It reduces applications of lambda abstractions,
806 removing both the lambda abstraction and the application.
808 In our transformation system, this step helps to remove unwanted lambda
809 abstractions (basically all but the ones at the top level). Other
810 transformations (application propagation, non-representable inlining) make
811 sure that most lambda abstractions will eventually be reducable by
814 Note that β-reduction also works on type lambda abstractions and type
815 applications as well. This means the substitution below also works on
816 type variables, in the case that the binder is a type variable and teh
817 expression applied to is a type.
834 \transexample{beta}{β-reduction}{from}{to}
844 \transexample{beta-type}{β-reduction for type abstractions}{from}{to}
846 \subsubsection{Empty let removal}
847 This transformation is simple: It removes recursive lets that have no bindings
848 (which usually occurs when unused let binding removal removes the last
851 Note that there is no need to define this transformation for
852 non-recursive lets, since they always contain exactly one binding.
862 \subsubsection[sec:normalization:simplelet]{Simple let binding removal}
863 This transformation inlines simple let bindings, that bind some
864 binder to some other binder instead of a more complex expression (\ie
867 This transformation is not needed to get an expression into intended
868 normal form (since these bindings are part of the intended normal
869 form), but makes the resulting \small{VHDL} a lot shorter.
880 ----------------------------- \lam{b} is a variable reference
881 letrec \lam{ai} ≠ \lam{b}
894 \subsubsection{Unused let binding removal}
895 This transformation removes let bindings that are never used.
896 Occasionally, \GHC's desugarer introduces some unused let bindings.
898 This normalization pass should really be unneeded to get into intended normal form
899 (since unused bindings are not forbidden by the normal form), but in practice
900 the desugarer or simplifier emits some unused bindings that cannot be
901 normalized (e.g., calls to a \type{PatError}\todo{Check this name}). Also,
902 this transformation makes the resulting \small{VHDL} a lot shorter.
904 \todo{Don't use old-style numerals in transformations}
913 M \lam{ai} does not occur free in \lam{M}
914 ---------------------------- \forall j, 0 ≤ j ≤ n, j ≠ i (\lam{ai} does not occur free in \lam{Ej})
928 \subsubsection{Cast propagation / simplification}
929 This transform pushes casts down into the expression as far as possible.
930 Since its exact role and need is not clear yet, this transformation is
933 \todo{Cast propagation}
935 \subsubsection{Top level binding inlining}
936 This transform takes simple top level bindings generated by the
937 \small{GHC} compiler. \small{GHC} sometimes generates very simple
938 \quote{wrapper} bindings, which are bound to just a variable
939 reference, or a partial application to constants or other variable
942 Note that this transformation is completely optional. It is not
943 required to get any function into intended normal form, but it does help making
944 the resulting VHDL output easier to read (since it removes a bunch of
945 components that are really boring).
947 This transform takes any top level binding generated by the compiler,
948 whose normalized form contains only a single let binding.
951 x = λa0 ... λan.let y = E in y
954 -------------------------------------- \lam{x} is generated by the compiler
955 λa0 ... λan.let y = E in y
959 (+) :: Word -> Word -> Word
960 (+) = GHC.Num.(+) @Word \$dNum
965 GHC.Num.(+) @ Alu.Word \$dNum a b
968 \transexample{toplevelinline}{Top level binding inlining}{from}{to}
970 \in{Example}[ex:trans:toplevelinline] shows a typical application of
971 the addition operator generated by \GHC. The type and dictionary
972 arguments used here are described in
973 \in{Section}[section:prototype:polymorphism].
975 Without this transformation, there would be a \lam{(+)} entity
976 in the \VHDL which would just add its inputs. This generates a
977 lot of overhead in the \VHDL, which is particularly annoying
978 when browsing the generated RTL schematic (especially since most
979 non-alphanumerics, like all characters in \lam{(+)}, are not
980 allowed in \VHDL architecture names\footnote{Technically, it is
981 allowed to use non-alphanumerics when using extended
982 identifiers, but it seems that none of the tooling likes
983 extended identifiers in filenames, so it effectively doesn't
984 work.}, so the entity would be called \quote{w7aA7f} or
985 something similarly unreadable and autogenerated).
987 \subsection{Program structure}
988 These transformations are aimed at normalizing the overall structure
989 into the intended form. This means ensuring there is a lambda abstraction
990 at the top for every argument (input port or current state), putting all
991 of the other value definitions in let bindings and making the final
992 return value a simple variable reference.
994 \subsubsection[sec:normalization:eta]{η-abstraction}
995 This transformation makes sure that all arguments of a function-typed
996 expression are named, by introducing lambda expressions. When combined with
997 β-reduction and non-representable binding inlining, all function-typed
998 expressions should be lambda abstractions or global identifiers.
1002 -------------- \lam{E} is not the first argument of an application.
1003 λx.E x \lam{E} is not a lambda abstraction.
1004 \lam{x} is a variable that does not occur free in \lam{E}.
1014 foo = λa.λx.(case a of
1019 \transexample{eta}{η-abstraction}{from}{to}
1021 \subsubsection[sec:normalization:appprop]{Application propagation}
1022 This transformation is meant to propagate application expressions downwards
1023 into expressions as far as possible. This allows partial applications inside
1024 expressions to become fully applied and exposes new transformation
1025 opportunities for other transformations (like β-reduction and
1028 Since all binders in our expression are unique (see
1029 \in{section}[sec:normalization:uniq]), there is no risk that we will
1030 introduce unintended shadowing by moving an expression into a lower
1031 scope. Also, since only move expression into smaller scopes (down into
1032 our expression), there is no risk of moving a variable reference out
1033 of the scope in which it is defined.
1036 (letrec binds in E) M
1037 ------------------------
1057 \transexample{appproplet}{Application propagation for a let expression}{from}{to}
1085 \transexample{apppropcase}{Application propagation for a case expression}{from}{to}
1087 \subsubsection[sec:normalization:letrecurse]{Let recursification}
1088 This transformation makes all non-recursive lets recursive. In the
1089 end, we want a single recursive let in our normalized program, so all
1090 non-recursive lets can be converted. This also makes other
1091 transformations simpler: They can simply assume all lets are
1099 ------------------------------------------
1106 \subsubsection{Let flattening}
1107 This transformation puts nested lets in the same scope, by lifting the
1108 binding(s) of the inner let into the outer let. Eventually, this will
1109 cause all let bindings to appear in the same scope.
1111 This transformation only applies to recursive lets, since all
1112 non-recursive lets will be made recursive (see
1113 \in{section}[sec:normalization:letrecurse]).
1115 Since we are joining two scopes together, there is no risk of moving a
1116 variable reference out of the scope where it is defined.
1122 ai = (letrec bindings in M)
1127 ------------------------------------------
1162 \transexample{letflat}{Let flattening}{from}{to}
1164 \subsubsection{Return value simplification}
1165 This transformation ensures that the return value of a function is always a
1166 simple local variable reference.
1168 Currently implemented using lambda simplification, let simplification, and
1169 top simplification. Should change into something like the following, which
1170 works only on the result of a function instead of any subexpression. This is
1171 achieved by the contexts, like \lam{x = E}, though this is strictly not
1172 correct (you could read this as "if there is any function \lam{x} that binds
1173 \lam{E}, any \lam{E} can be transformed, while we only mean the \lam{E} that
1174 is bound by \lam{x}. This might need some extra notes or something).
1176 Note that the return value is not simplified if its not representable.
1177 Otherwise, this would cause a direct loop with the inlining of
1178 unrepresentable bindings. If the return value is not
1179 representable because it has a function type, η-abstraction should
1180 make sure that this transformation will eventually apply. If the value
1181 is not representable for other reasons, the function result itself is
1182 not representable, meaning this function is not translatable anyway.
1185 x = E \lam{E} is representable
1186 ~ \lam{E} is not a lambda abstraction
1187 E \lam{E} is not a let expression
1188 --------------------------- \lam{E} is not a local variable reference
1194 ~ \lam{E} is representable
1195 E \lam{E} is not a let expression
1196 --------------------------- \lam{E} is not a local variable reference
1201 x = λv0 ... λvn.let ... in E
1202 ~ \lam{E} is representable
1203 E \lam{E} is not a local variable reference
1204 -----------------------------
1213 x = letrec x = add 1 2 in x
1216 \transexample{retvalsimpl}{Return value simplification}{from}{to}
1218 \todo{More examples}
1220 \subsection[sec:normalization:argsimpl]{Representable arguments simplification}
1221 This section contains just a single transformation that deals with
1222 representable arguments in applications. Non-representable arguments are
1223 handled by the transformations in
1224 \in{section}[sec:normalization:nonrep].
1226 This transformation ensures that all representable arguments will become
1227 references to local variables. This ensures they will become references
1228 to local signals in the resulting \small{VHDL}, which is required due to
1229 limitations in the component instantiation code in \VHDL (one can only
1230 assign a signal or constant to an input port). By ensuring that all
1231 arguments are always simple variable references, we always have a signal
1232 available to map to the input ports.
1234 To reduce a complex expression to a simple variable reference, we create
1235 a new let expression around the application, which binds the complex
1236 expression to a new variable. The original function is then applied to
1239 \refdef{global variable}
1240 Note that references to \emph{global variables} (like a top level
1241 function without arguments, but also an argumentless dataconstructors
1242 like \lam{True}) are also simplified. Only local variables generate
1243 signals in the resulting architecture. Even though argumentless
1244 dataconstructors generate constants in generated \VHDL code and could be
1245 mapped to an input port directly, they are still simplified to make the
1246 normal form more regular.
1248 \refdef{representable}
1251 -------------------- \lam{N} is representable
1252 letrec x = N in M x \lam{N} is not a local variable reference
1254 \refdef{local variable}
1261 letrec x = add a 1 in add x 1
1264 \transexample{argsimpl}{Argument simplification}{from}{to}
1266 \subsection[sec:normalization:builtins]{Builtin functions}
1267 This section deals with (arguments to) builtin functions. In the
1268 intended normal form definition\refdef{intended normal form definition}
1269 we can see that there are three sorts of arguments a builtin function
1273 \item A representable local variable reference. This is the most
1274 common argument to any function. The argument simplification
1275 transformation described in \in{section}[sec:normalization:argsimpl]
1276 makes sure that \emph{any} representable argument to \emph{any}
1277 function (including builtin functions) is turned into a local variable
1279 \item (A partial application of) a top level function (either builtin on
1280 user-defined). The function extraction transformation described in
1281 this section takes care of turning every functiontyped argument into
1282 (a partial application of) a top level function.
1283 \item Any expression that is not representable and does not have a
1284 function type. Since these can be any expression, there is no
1285 transformation needed. Note that this category is exactly all
1286 expressions that are not transformed by the transformations for the
1287 previous two categories. This means that \emph{any} core expression
1288 that is used as an argument to a builtin function will be either
1289 transformed into one of the above categories, or end up in this
1290 categorie. In any case, the result is in normal form.
1293 As noted, the argument simplification will handle any representable
1294 arguments to a builtin function. The following transformation is needed
1295 to handle non-representable arguments with a function type, all other
1296 non-representable arguments don't need any special handling.
1298 \subsubsection[sec:normalization:funextract]{Function extraction}
1299 This transform deals with function-typed arguments to builtin
1301 Since builtin functions cannot be specialized (see
1302 \in{section}[sec:normalization:specialize]) to remove the arguments,
1303 these arguments are extracted into a new global function instead. In
1304 other words, we create a new top level function that has exactly the
1305 extracted argument as its body. This greatly simplifies the
1306 translation rules needed for builtin functions, since they only need
1307 to handle (partial applications of) top level functions.
1309 Any free variables occuring in the extracted arguments will become
1310 parameters to the new global function. The original argument is replaced
1311 with a reference to the new function, applied to any free variables from
1312 the original argument.
1314 This transformation is useful when applying higher order builtin functions
1315 like \hs{map} to a lambda abstraction, for example. In this case, the code
1316 that generates \small{VHDL} for \hs{map} only needs to handle top level functions and
1317 partial applications, not any other expression (such as lambda abstractions or
1318 even more complicated expressions).
1321 M N \lam{M} is (a partial aplication of) a builtin function.
1322 --------------------- \lam{f0 ... fn} are all free local variables of \lam{N}
1323 M (x f0 ... fn) \lam{N :: a -> b}
1324 ~ \lam{N} is not a (partial application of) a top level function
1329 addList = λb.λxs.map (λa . add a b) xs
1333 addList = λb.λxs.map (f b) xs
1338 \transexample{funextract}{Function extraction}{from}{to}
1340 Note that the function \lam{f} will still need normalization after
1343 \subsection{Case normalisation}
1344 \subsubsection{Scrutinee simplification}
1345 This transform ensures that the scrutinee of a case expression is always
1346 a simple variable reference.
1351 ----------------- \lam{E} is not a local variable reference
1370 \transexample{letflat}{Case normalisation}{from}{to}
1373 \subsubsection{Case simplification}
1374 This transformation ensures that all case expressions become normal form. This
1375 means they will become one of:
1377 \item An extractor case with a single alternative that picks a single field
1378 from a datatype, \eg \lam{case x of (a, b) -> a}.
1379 \item A selector case with multiple alternatives and only wild binders, that
1380 makes a choice between expressions based on the constructor of another
1381 expression, \eg \lam{case x of Low -> a; High -> b}.
1384 \defref{wild binder}
1387 C0 v0,0 ... v0,m -> E0
1389 Cn vn,0 ... vn,m -> En
1390 --------------------------------------------------- \forall i \forall j, 0 ≤ i ≤ n, 0 ≤ i < m (\lam{wi,j} is a wild (unused) binder)
1392 v0,0 = case E of C0 v0,0 .. v0,m -> v0,0
1394 v0,m = case E of C0 v0,0 .. v0,m -> v0,m
1396 vn,m = case E of Cn vn,0 .. vn,m -> vn,m
1402 C0 w0,0 ... w0,m -> x0
1404 Cn wn,0 ... wn,m -> xn
1406 \todo{Check the subscripts of this transformation}
1408 Note that this transformation applies to case statements with any
1409 scrutinee. If the scrutinee is a complex expression, this might result
1410 in duplicate hardware. An extra condition to only apply this
1411 transformation when the scrutinee is already simple (effectively
1412 causing this transformation to be only applied after the scrutinee
1413 simplification transformation) might be in order.
1415 \fxnote{This transformation specified like this is complicated and misses
1416 conditions to prevent looping with itself. Perhaps it should be split here for
1435 \transexample{selcasesimpl}{Selector case simplification}{from}{to}
1443 b = case a of (,) b c -> b
1444 c = case a of (,) b c -> c
1451 \transexample{excasesimpl}{Extractor case simplification}{from}{to}
1453 \refdef{selector case}
1454 In \in{example}[ex:trans:excasesimpl] the case expression is expanded
1455 into multiple case expressions, including a pretty useless expression
1456 (that is neither a selector or extractor case). This case can be
1457 removed by the Case removal transformation in
1458 \in{section}[sec:transformation:caseremoval].
1460 \subsubsection[sec:transformation:caseremoval]{Case removal}
1461 This transform removes any case statements with a single alternative and
1464 These "useless" case statements are usually leftovers from case simplification
1465 on extractor case (see the previous example).
1470 ---------------------- \lam{\forall i, 0 ≤ i ≤ m} (\lam{vi} does not occur free in E)
1483 \transexample{caserem}{Case removal}{from}{to}
1485 \subsection[sec:normalization:nonrep]{Removing unrepresentable values}
1486 The transformations in this section are aimed at making all the
1487 values used in our expression representable. There are two main
1488 transformations that are applied to \emph{all} unrepresentable let
1489 bindings and function arguments. These are meant to address three
1490 different kinds of unrepresentable values: Polymorphic values, higher
1491 order values and literals. The transformation are described generically:
1492 They apply to all non-representable values. However, non-representable
1493 values that don't fall into one of these three categories will be moved
1494 around by these transformations but are unlikely to completely
1495 disappear. They usually mean the program was not valid in the first
1496 place, because unsupported types were used (for example, a program using
1499 Each of these three categories will be detailed below, followed by the
1500 actual transformations.
1502 \subsubsection{Removing Polymorphism}
1503 As noted in \in{section}[sec:prototype:polymporphism],
1504 polymorphism is made explicit in Core through type and
1505 dictionary arguments. To remove the polymorphism from a
1506 function, we can simply specialize the polymorphic function for
1507 the particular type applied to it. The same goes for dictionary
1508 arguments. To remove polymorphism from let bound values, we
1509 simply inline the let bindings that have a polymorphic type,
1510 which should (eventually) make sure that the polymorphic
1511 expression is applied to a type and/or dictionary, which can
1512 then be removed by β-reduction (\in{section}[sec:normalization:beta]).
1514 Since both type and dictionary arguments are not representable,
1515 \refdef{representable}
1516 the non-representable argument specialization and
1517 non-representable let binding inlining transformations below
1518 take care of exactly this.
1520 There is one case where polymorphism cannot be completely
1521 removed: Builtin functions are still allowed to be polymorphic
1522 (Since we have no function body that we could properly
1523 specialize). However, the code that generates \VHDL for builtin
1524 functions knows how to handle this, so this is not a problem.
1526 \subsubsection{Defunctionalization}
1527 These transformations remove higher order expressions from our
1528 program, making all values first-order.
1530 Higher order values are always introduced by lambda abstractions, none
1531 of the other Core expression elements can introduce a function type.
1532 However, other expressions can \emph{have} a function type, when they
1533 have a lambda expression in their body.
1535 For example, the following expression is a higher order expression
1536 that is not a lambda expression itself:
1538 \refdef{id function}
1545 The reference to the \lam{id} function shows that we can introduce a
1546 higher order expression in our program without using a lambda
1547 expression directly. However, inside the definition of the \lam{id}
1548 function, we can be sure that a lambda expression is present.
1550 Looking closely at the definition of our normal form in
1551 \in{section}[sec:normalization:intendednormalform], we can see that
1552 there are three possibilities for higher order values to appear in our
1553 intended normal form:
1556 \item[item:toplambda] Lambda abstractions can appear at the highest level of a
1557 top level function. These lambda abstractions introduce the
1558 arguments (input ports / current state) of the function.
1559 \item[item:builtinarg] (Partial applications of) top level functions can appear as an
1560 argument to a builtin function.
1561 \item[item:completeapp] (Partial applications of) top level functions can appear in
1562 function position of an application. Since a partial application
1563 cannot appear anywhere else (except as builtin function arguments),
1564 all partial applications are applied, meaning that all applications
1565 will become complete applications. However, since application of
1566 arguments happens one by one, in the expression:
1570 the subexpression \lam{f 1} has a function type. But this is
1571 allowed, since it is inside a complete application.
1574 We will take a typical function with some higher order values as an
1575 example. The following function takes two arguments: a \lam{Bit} and a
1576 list of numbers. Depending on the first argument, each number in the
1577 list is doubled, or the list is returned unmodified. For the sake of
1578 the example, no polymorphism is shown. In reality, at least map would
1582 λy.let double = λx. x + x in
1588 This example shows a number of higher order values that we cannot
1589 translate to \VHDL directly. The \lam{double} binder bound in the let
1590 expression has a function type, as well as both of the alternatives of
1591 the case expression. The first alternative is a partial application of
1592 the \lam{map} builtin function, whereas the second alternative is a
1595 To reduce all higher order values to one of the above items, a number
1596 of transformations we've already seen are used. The η-abstraction
1597 transformation from \in{section}[sec:normalization:eta] ensures all
1598 function arguments are introduced by lambda abstraction on the highest
1599 level of a function. These lambda arguments are allowed because of
1600 \in{item}[item:toplambda] above. After η-abstraction, our example
1601 becomes a bit bigger:
1604 λy.λq.(let double = λx. x + x in
1611 η-abstraction also introduces extra applications (the application of
1612 the let expression to \lam{q} in the above example). These
1613 applications can then propagated down by the application propagation
1614 transformation (\in{section}[sec:normalization:appprop]). In our
1615 example, the \lam{q} and \lam{r} variable will be propagated into the
1616 let expression and then into the case expression:
1619 λy.λq.let double = λx. x + x in
1625 This propagation makes higher order values become applied (in
1626 particular both of the alternatives of the case now have a
1627 representable type. Completely applied top level functions (like the
1628 first alternative) are now no longer invalid (they fall under
1629 \in{item}[item:completeapp] above). (Completely) applied lambda
1630 abstractions can be removed by β-abstraction. For our example,
1631 applying β-abstraction results in the following:
1634 λy.λq.let double = λx. x + x in
1640 As you can see in our example, all of this moves applications towards
1641 the higher order values, but misses higher order functions bound by
1642 let expressions. The applications cannot be moved towards these values
1643 (since they can be used in multiple places), so the values will have
1644 to be moved towards the applications. This is achieved by inlining all
1645 higher order values bound by let applications, by the
1646 non-representable binding inlining transformation below. When applying
1647 it to our example, we get the following:
1651 Low -> map (λx. x + x) q
1655 We've nearly eliminated all unsupported higher order values from this
1656 expressions. The one that's remaining is the first argument to the
1657 \lam{map} function. Having higher order arguments to a builtin
1658 function like \lam{map} is allowed in the intended normal form, but
1659 only if the argument is a (partial application) of a top level
1660 function. This is easily done by introducing a new top level function
1661 and put the lambda abstraction inside. This is done by the function
1662 extraction transformation from
1663 \in{section}[sec:normalization:funextract].
1671 This also introduces a new function, that we have called \lam{func}:
1677 Note that this does not actually remove the lambda, but now it is a
1678 lambda at the highest level of a function, which is allowed in the
1679 intended normal form.
1681 There is one case that has not been discussed yet. What if the
1682 \lam{map} function in the example above was not a builtin function
1683 but a user-defined function? Then extracting the lambda expression
1684 into a new function would not be enough, since user-defined functions
1685 can never have higher order arguments. For example, the following
1686 expression shows an example:
1689 twice :: (Word -> Word) -> Word -> Word
1690 twice = λf.λa.f (f a)
1692 main = λa.app (λx. x + x) a
1695 This example shows a function \lam{twice} that takes a function as a
1696 first argument and applies that function twice to the second argument.
1697 Again, we've made the function monomorphic for clarity, even though
1698 this function would be a lot more useful if it was polymorphic. The
1699 function \lam{main} uses \lam{twice} to apply a lambda epression twice.
1701 When faced with a user defined function, a body is available for that
1702 function. This means we could create a specialized version of the
1703 function that only works for this particular higher order argument
1704 (\ie, we can just remove the argument and call the specialized
1705 function without the argument). This transformation is detailed below.
1706 Applying this transformation to the example gives:
1709 twice' :: Word -> Word
1710 twice' = λb.(λf.λa.f (f a)) (λx. x + x) b
1715 The \lam{main} function is now in normal form, since the only higher
1716 order value there is the top level lambda expression. The new
1717 \lam{twice'} function is a bit complex, but the entire original body of
1718 the original \lam{twice} function is wrapped in a lambda abstraction
1719 and applied to the argument we've specialized for (\lam{λx. x + x})
1720 and the other arguments. This complex expression can fortunately be
1721 effectively reduced by repeatedly applying β-reduction:
1724 twice' :: Word -> Word
1725 twice' = λb.(b + b) + (b + b)
1728 This example also shows that the resulting normal form might not be as
1729 efficient as we might hope it to be (it is calculating \lam{(b + b)}
1730 twice). This is discussed in more detail in
1731 \in{section}[sec:normalization:duplicatework].
1733 \subsubsection{Literals}
1734 There are a limited number of literals available in Haskell and Core.
1735 \refdef{enumerated types} When using (enumerating) algebraic
1736 datatypes, a literal is just a reference to the corresponding data
1737 constructor, which has a representable type (the algebraic datatype)
1738 and can be translated directly. This also holds for literals of the
1739 \hs{Bool} Haskell type, which is just an enumerated type.
1741 There is, however, a second type of literal that does not have a
1742 representable type: Integer literals. Cλash supports using integer
1743 literals for all three integer types supported (\hs{SizedWord},
1744 \hs{SizedInt} and \hs{RangedWord}). This is implemented using
1745 Haskell's \hs{Num} typeclass, which offers a \hs{fromInteger} method
1746 that converts any \hs{Integer} to the Cλash datatypes.
1748 When \GHC sees integer literals, it will automatically insert calls to
1749 the \hs{fromInteger} method in the resulting Core expression. For
1750 example, the following expression in Haskell creates a 32 bit unsigned
1751 word with the value 1. The explicit type signature is needed, since
1752 there is no context for \GHC to determine the type from otherwise.
1758 This Haskell code results in the following Core expression:
1761 fromInteger @(SizedWord D32) \$dNum (smallInteger 10)
1764 The literal 10 will have the type \lam{GHC.Prim.Int\#}, which is
1765 converted into an \lam{Integer} by \lam{smallInteger}. Finally, the
1766 \lam{fromInteger} function will finally convert this into a
1767 \lam{SizedWord D32}.
1769 Both the \lam{GHC.Prim.Int\#} and \lam{Integer} types are not
1770 representable, and cannot be translated directly. Fortunately, there
1771 is no need to translate them, since \lam{fromInteger} is a builtin
1772 function that knows how to handle these values. However, this does
1773 require that the \lam{fromInteger} function is directly applied to
1774 these non-representable literal values, otherwise errors will occur.
1775 For example, the following expression is not in the intended normal
1776 form, since one of the let bindings has an unrepresentable type
1780 let l = smallInteger 10 in fromInteger @(SizedWord D32) \$dNum l
1783 By inlining these let-bindings, we can ensure that unrepresentable
1784 literals bound by a let binding end up in an application of the
1785 appropriate builtin function, where they are allowed. Since it is
1786 possible that the application of that function is in a different
1787 function than the definition of the literal value, we will always need
1788 to specialize away any unrepresentable literals that are used as
1789 function arguments. The following two transformations do exactly this.
1791 \subsubsection{Non-representable binding inlining}
1792 This transform inlines let bindings that are bound to a
1793 non-representable value. Since we can never generate a signal
1794 assignment for these bindings (we cannot declare a signal assignment
1795 with a non-representable type, for obvious reasons), we have no choice
1796 but to inline the binding to remove it.
1798 As we have seen in the previous sections, inlining these bindings
1799 solves (part of) the polymorphism, higher order values and
1800 unrepresentable literals in an expression.
1811 -------------------------- \lam{Ei} has a non-representable type.
1813 a0 = E0 [ai=>Ei] \vdots
1814 ai-1 = Ei-1 [ai=>Ei]
1815 ai+1 = Ei+1 [ai=>Ei]
1834 x = fromInteger (smallInteger 10)
1836 (λb -> add b 1) (add 1 x)
1839 \transexample{nonrepinline}{Nonrepresentable binding inlining}{from}{to}
1841 \subsubsection[sec:normalization:specialize]{Function specialization}
1842 This transform removes arguments to user-defined functions that are
1843 not representable at runtime. This is done by creating a
1844 \emph{specialized} version of the function that only works for one
1845 particular value of that argument (in other words, the argument can be
1848 Specialization means to create a specialized version of the called
1849 function, with one argument already filled in. As a simple example, in
1850 the following program (this is not actual Core, since it directly uses
1851 a literal with the unrepresentable type \lam{GHC.Prim.Int\#}).
1858 We could specialize the function \lam{f} against the literal argument
1859 1, with the following result:
1866 In some way, this transformation is similar to β-reduction, but it
1867 operates across function boundaries. It is also similar to
1868 non-representable let binding inlining above, since it sort of
1869 \quote{inlines} an expression into a called function.
1871 Special care must be taken when the argument has any free variables.
1872 If this is the case, the original argument should not be removed
1873 completely, but replaced by all the free variables of the expression.
1874 In this way, the original expression can still be evaluated inside the
1877 To prevent us from propagating the same argument over and over, a
1878 simple local variable reference is not propagated (since is has
1879 exactly one free variable, itself, we would only replace that argument
1882 This shows that any free local variables that are not runtime
1883 representable cannot be brought into normal form by this transform. We
1884 rely on an inlining or β-reduction transformation to replace such a
1885 variable with an expression we can propagate again.
1890 x Y0 ... Yi ... Yn \lam{Yi} is not representable
1891 --------------------------------------------- \lam{Yi} is not a local variable reference
1892 x' y0 ... yi-1 f0 ... fm Yi+1 ... Yn \lam{f0 ... fm} are all free local vars of \lam{Yi}
1893 ~ \lam{T0 ... Tn} are the types of \lam{Y0 ... Yn}
1894 x' = λ(y0 :: T0) ... λ(yi-1 :: Ty-1). λf0 ... λfm. λ(yi+1 :: Ty+1) ... λ(yn :: Tn).
1895 E y0 ... yi-1 Yi yi+1 ... yn
1898 This is a bit of a complex transformation. It transforms an
1899 application of the function \lam{x}, where one of the arguments
1900 (\lam{Y_i}) is not representable. A new
1901 function \lam{x'} is created that wraps the body of the old function.
1902 The body of the new function becomes a number of nested lambda
1903 abstractions, one for each of the original arguments that are left
1906 The ith argument is replaced with the free variables of
1907 \lam{Y_i}. Note that we reuse the same binders as those used in
1908 \lam{Y_i}, since we can then just use \lam{Y_i} inside the new
1909 function body and have all of the variables it uses be in scope.
1911 The argument that we are specializing for, \lam{Y_i}, is put inside
1912 the new function body. The old function body is applied to it. Since
1913 we use this new function only in place of an application with that
1914 particular argument \lam{Y_i}, behaviour should not change.
1916 Note that the types of the arguments of our new function are taken
1917 from the types of the \emph{actual} arguments (\lam{T0 ... Tn}). This
1918 means that any polymorphism in the arguments is removed, even when the
1919 corresponding explicit type lambda is not removed
1920 yet.\refdef{type lambda}
1922 \todo{Examples. Perhaps reference the previous sections}
1925 \section{Unsolved problems}
1926 The above system of transformations has been implemented in the prototype
1927 and seems to work well to compile simple and more complex examples of
1928 hardware descriptions. \todo{Ref christiaan?} However, this normalization
1929 system has not seen enough review and work to be complete and work for
1930 every Core expression that is supplied to it. A number of problems
1931 have already been identified and are discussed in this section.
1933 \subsection[sec:normalization:duplicatework]{Work duplication}
1934 A possible problem of β-reduction is that it could duplicate work.
1935 When the expression applied is not a simple variable reference, but
1936 requires calculation and the binder the lambda abstraction binds to
1937 is used more than once, more hardware might be generated than strictly
1940 As an example, consider the expression:
1946 When applying β-reduction to this expression, we get:
1952 which of course calculates \lam{(a * b)} twice.
1954 A possible solution to this would be to use the following alternative
1955 transformation, which is of course no longer normal β-reduction. The
1956 followin transformation has not been tested in the prototype, but is
1957 given here for future reference:
1965 This doesn't seem like much of an improvement, but it does get rid of
1966 the lambda expression (and the associated higher order value), while
1967 at the same time introducing a new let binding. Since the result of
1968 every application or case expression must be bound by a let expression
1969 in the intended normal form anyway, this is probably not a problem. If
1970 the argument happens to be a variable reference, then simple let
1971 binding removal (\in{section}[sec:normalization:simplelet]) will
1972 remove it, making the result identical to that of the original
1973 β-reduction transformation.
1975 When also applying argument simplification to the above example, we
1976 get the following expression:
1984 Looking at this, we could imagine an alternative approach: Create a
1985 transformation that removes let bindings that bind identical values.
1986 In the above expression, the \lam{y} and \lam{z} variables could be
1987 merged together, resulting in the more efficient expression:
1990 let y = (a * b) in y + y
1993 \subsection{Non-determinism}
1994 As an example, again consider the following expression:
2000 We can apply both β-reduction (\in{section}[sec:normalization:beta])
2001 as well as argument simplification
2002 (\in{section}[sec:normalization:argsimpl]) to this expression.
2004 When applying argument simplification first and then β-reduction, we
2005 get the following expression:
2008 let y = (a * b) in y + y
2011 When applying β-reduction first and then argument simplification, we
2012 get the following expression:
2020 As you can see, this is a different expression. This means that the
2021 order of expressions, does in fact change the resulting normal form,
2022 which is something that we would like to avoid. In this particular
2023 case one of the alternatives is even clearly more efficient, so we
2024 would of course like the more efficient form to be the normal form.
2026 For this particular problem, the solutions for duplication of work
2027 seem from the previous section seem to fix the determinism of our
2028 transformation system as well. However, it is likely that there are
2029 other occurences of this problem.
2032 We do not fully understand the use of cast expressions in Core, so
2033 there are probably expressions involving cast expressions that cannot
2034 be brought into intended normal form by this transformation system.
2036 The uses of casts in the core system should be investigated more and
2037 transformations will probably need updating to handle them in all
2040 \section[sec:normalization:properties]{Provable properties}
2041 When looking at the system of transformations outlined above, there are a
2042 number of questions that we can ask ourselves. The main question is of course:
2043 \quote{Does our system work as intended?}. We can split this question into a
2044 number of subquestions:
2047 \item[q:termination] Does our system \emph{terminate}? Since our system will
2048 keep running as long as transformations apply, there is an obvious risk that
2049 it will keep running indefinitely. This typically happens when one
2050 transformation produces a result that is transformed back to the original
2051 by another transformation, or when one or more transformations keep
2052 expanding some expression.
2053 \item[q:soundness] Is our system \emph{sound}? Since our transformations
2054 continuously modify the expression, there is an obvious risk that the final
2055 normal form will not be equivalent to the original program: Its meaning could
2057 \item[q:completeness] Is our system \emph{complete}? Since we have a complex
2058 system of transformations, there is an obvious risk that some expressions will
2059 not end up in our intended normal form, because we forgot some transformation.
2060 In other words: Does our transformation system result in our intended normal
2061 form for all possible inputs?
2062 \item[q:determinism] Is our system \emph{deterministic}? Since we have defined
2063 no particular order in which the transformation should be applied, there is an
2064 obvious risk that different transformation orderings will result in
2065 \emph{different} normal forms. They might still both be intended normal forms
2066 (if our system is \emph{complete}) and describe correct hardware (if our
2067 system is \emph{sound}), so this property is less important than the previous
2068 three: The translator would still function properly without it.
2071 Unfortunately, the final transformation system has only been
2072 developed in the final part of the research, leaving no more time
2073 for verifying these properties. In fact, it is likely that the
2074 current transformation system still violates some of these
2075 properties in some cases and should be improved (or extra conditions
2076 on the input hardware descriptions should be formulated).
2078 This is most likely the case with the completeness and determinism
2079 properties, perhaps als the termination property. The soundness
2080 property probably holds, since it is easier to manually verify (each
2081 transformation can be reviewed separately).
2083 Even though no complete proofs have been made, some ideas for
2084 possible proof strategies are shown below.
2086 \subsection{Graph representation}
2087 Before looking into how to prove these properties, we'll look at our
2088 transformation system from a graph perspective. The nodes of the graph are
2089 all possible Core expressions. The (directed) edges of the graph are
2090 transformations. When a transformation α applies to an expression \lam{A} to
2091 produce an expression \lam{B}, we add an edge from the node for \lam{A} to the
2092 node for \lam{B}, labeled α.
2094 \startuseMPgraphic{TransformGraph}
2098 newCircle.a(btex \lam{(λx.λy. (+) x y) 1} etex);
2099 newCircle.b(btex \lam{λy. (+) 1 y} etex);
2100 newCircle.c(btex \lam{(λx.(+) x) 1} etex);
2101 newCircle.d(btex \lam{(+) 1} etex);
2104 c.c = b.c + (4cm, 0cm);
2105 a.c = midpoint(b.c, c.c) + (0cm, 4cm);
2106 d.c = midpoint(b.c, c.c) - (0cm, 3cm);
2108 % β-conversion between a and b
2109 ncarc.a(a)(b) "name(bred)";
2110 ObjLabel.a(btex $\xrightarrow[normal]{}{β}$ etex) "labpathname(bred)", "labdir(rt)";
2111 ncarc.b(b)(a) "name(bexp)", "linestyle(dashed withdots)";
2112 ObjLabel.b(btex $\xleftarrow[normal]{}{β}$ etex) "labpathname(bexp)", "labdir(lft)";
2114 % η-conversion between a and c
2115 ncarc.a(a)(c) "name(ered)";
2116 ObjLabel.a(btex $\xrightarrow[normal]{}{η}$ etex) "labpathname(ered)", "labdir(rt)";
2117 ncarc.c(c)(a) "name(eexp)", "linestyle(dashed withdots)";
2118 ObjLabel.c(btex $\xleftarrow[normal]{}{η}$ etex) "labpathname(eexp)", "labdir(lft)";
2120 % η-conversion between b and d
2121 ncarc.b(b)(d) "name(ered)";
2122 ObjLabel.b(btex $\xrightarrow[normal]{}{η}$ etex) "labpathname(ered)", "labdir(rt)";
2123 ncarc.d(d)(b) "name(eexp)", "linestyle(dashed withdots)";
2124 ObjLabel.d(btex $\xleftarrow[normal]{}{η}$ etex) "labpathname(eexp)", "labdir(lft)";
2126 % β-conversion between c and d
2127 ncarc.c(c)(d) "name(bred)";
2128 ObjLabel.c(btex $\xrightarrow[normal]{}{β}$ etex) "labpathname(bred)", "labdir(rt)";
2129 ncarc.d(d)(c) "name(bexp)", "linestyle(dashed withdots)";
2130 ObjLabel.d(btex $\xleftarrow[normal]{}{β}$ etex) "labpathname(bexp)", "labdir(lft)";
2132 % Draw objects and lines
2133 drawObj(a, b, c, d);
2136 \placeexample[right][ex:TransformGraph]{Partial graph of a lambda calculus
2137 system with β and η reduction (solid lines) and expansion (dotted lines).}
2138 \boxedgraphic{TransformGraph}
2140 Of course our graph is unbounded, since we can construct an infinite amount of
2141 Core expressions. Also, there might potentially be multiple edges between two
2142 given nodes (with different labels), though seems unlikely to actually happen
2145 See \in{example}[ex:TransformGraph] for the graph representation of a very
2146 simple lambda calculus that contains just the expressions \lam{(λx.λy. (+) x
2147 y) 1}, \lam{λy. (+) 1 y}, \lam{(λx.(+) x) 1} and \lam{(+) 1}. The
2148 transformation system consists of β-reduction and η-reduction (solid edges) or
2149 β-expansion and η-expansion (dotted edges).
2151 \todo{Define β-reduction and η-reduction?}
2153 Note that the normal form of such a system consists of the set of nodes
2154 (expressions) without outgoing edges, since those are the expression to which
2155 no transformation applies anymore. We call this set of nodes the \emph{normal
2156 set}. The set of nodes containing expressions in intended normal
2157 form \refdef{intended normal form} is called the \emph{intended
2160 From such a graph, we can derive some properties easily:
2162 \item A system will \emph{terminate} if there is no path of infinite length
2163 in the graph (this includes cycles, but can also happen without cycles).
2164 \item Soundness is not easily represented in the graph.
2165 \item A system is \emph{complete} if all of the nodes in the normal set have
2166 the intended normal form. The inverse (that all of the nodes outside of
2167 the normal set are \emph{not} in the intended normal form) is not
2168 strictly required. In other words, our normal set must be a
2169 subset of the intended normal form, but they do not need to be
2172 \item A system is deterministic if all paths starting at a particular
2173 node, which end in a node in the normal set, end at the same node.
2176 When looking at the \in{example}[ex:TransformGraph], we see that the system
2177 terminates for both the reduction and expansion systems (but note that, for
2178 expansion, this is only true because we've limited the possible
2179 expressions. In comlete lambda calculus, there would be a path from
2180 \lam{(λx.λy. (+) x y) 1} to \lam{(λx.λy.(λz.(+) z) x y) 1} to
2181 \lam{(λx.λy.(λz.(λq.(+) q) z) x y) 1} etc.)
2183 If we would consider the system with both expansion and reduction, there
2184 would no longer be termination either, since there would be cycles all
2187 The reduction and expansion systems have a normal set of containing just
2188 \lam{(+) 1} or \lam{(λx.λy. (+) x y) 1} respectively. Since all paths in
2189 either system end up in these normal forms, both systems are \emph{complete}.
2190 Also, since there is only one node in the normal set, it must obviously be
2191 \emph{deterministic} as well.
2193 \todo{Add content to these sections}
2194 \subsection{Termination}
2195 In general, proving termination of an arbitrary program is a very
2196 hard problem. \todo{Ref about arbitrary termination} Fortunately,
2197 we only have to prove termination for our specific transformation
2200 A common approach for these kinds of proofs is to associate a
2201 measure with each possible expression in our system. If we can
2202 show that each transformation strictly decreases this measure
2203 (\ie, the expression transformed to has a lower measure than the
2204 expression transformed from). \todo{ref about measure-based
2205 termination proofs / analysis}
2207 A good measure for a system consisting of just β-reduction would
2208 be the number of lambda expressions in the expression. Since every
2209 application of β-reduction removes a lambda abstraction (and there
2210 is always a bounded number of lambda abstractions in every
2211 expression) we can easily see that a transformation system with
2212 just β-reduction will always terminate.
2214 For our complete system, this measure would be fairly complex
2215 (probably the sum of a lot of things). Since the (conditions on)
2216 our transformations are pretty complex, we would need to include
2217 both simple things like the number of let expressions as well as
2218 more complex things like the number of case expressions that are
2219 not yet in normal form.
2221 No real attempt has been made at finding a suitable measure for
2224 \subsection{Soundness}
2225 Soundness is a property that can be proven for each transformation
2226 separately. Since our system only runs separate transformations
2227 sequentially, if each of our transformations leaves the
2228 \emph{meaning} of the expression unchanged, then the entire system
2229 will of course leave the meaning unchanged and is thus
2232 The current prototype has only been verified in an ad-hoc fashion
2233 by inspecting (the code for) each transformation. A more formal
2234 verification would be more appropriate.
2236 To be able to formally show that each transformation properly
2237 preserves the meaning of every expression, we require an exact
2238 definition of the \emph{meaning} of every expression, so we can
2239 compare them. Currently there seems to be no formal definition of
2240 the meaning or semantics of \GHC's core language, only informal
2241 descriptions are available.
2243 It should be possible to have a single formal definition of
2244 meaning for Core for both normal Core compilation by \GHC and for
2245 our compilation to \VHDL. The main difference seems to be that in
2246 hardware every expression is always evaluated, while in software
2247 it is only evaluated if needed, but it should be possible to
2248 assign a meaning to core expressions that assumes neither.
2250 Since each of the transformations can be applied to any
2251 subexpression as well, there is a constraint on our meaning
2252 definition: The meaning of an expression should depend only on the
2253 meaning of subexpressions, not on the expressions themselves. For
2254 example, the meaning of the application in \lam{f (let x = 4 in
2255 x)} should be the same as the meaning of the application in \lam{f
2256 4}, since the argument subexpression has the same meaning (though
2257 the actual expression is different).
2259 \subsection{Completeness}
2260 Proving completeness is probably not hard, but it could be a lot
2261 of work. We have seen above that to prove completeness, we must
2262 show that the normal set of our graph representation is a subset
2263 of the intended normal set.
2265 However, it is hard to systematically generate or reason about the
2266 normal set, since it is defined as any nodes to which no
2267 transformation applies. To determine this set, each transformation
2268 must be considered and when a transformation is added, the entire
2269 set should be re-evaluated. This means it is hard to show that
2270 each node in the normal set is also in the intended normal set.
2271 Reasoning about our intended normal set is easier, since we know
2272 how to generate it from its definition. \refdef{intended normal
2275 Fortunately, we can also prove the complement (which is
2276 equivalent, since $A \subseteq B \Leftrightarrow \overline{B}
2277 \subseteq \overline{A}$): Show that the set of nodes not in
2278 intended normal form is a subset of the set of nodes not in normal
2279 form. In other words, show that for every expression that is not
2280 in intended normal form, that there is at least one transformation
2281 that applies to it (since that means it is not in normal form
2282 either and since $A \subseteq C \Leftrightarrow \forall x (x \in A
2283 \rightarrow x \in C)$).
2285 By systematically reviewing the entire Core language definition
2286 along with the intended normal form definition (both of which have
2287 a similar structure), it should be possible to identify all
2288 possible (sets of) core expressions that are not in intended
2289 normal form and identify a transformation that applies to it.
2291 This approach is especially useful for proving completeness of our
2292 system, since if expressions exist to which none of the
2293 transformations apply (\ie if the system is not yet complete), it
2294 is immediately clear which expressions these are and adding
2295 (or modifying) transformations to fix this should be relatively
2298 As observed above, applying this approach is a lot of work, since
2299 we need to check every (set of) transformation(s) separately.
2301 \todo{Perhaps do a few steps of the proofs as proof-of-concept}
2303 % vim: set sw=2 sts=2 expandtab: