[gforth] / gforth / glocals.fs  

gforth: gforth/glocals.fs


1 : anton 1.1 \ Local variables are quite important for writing readable programs, but
2 :     \ IMO (anton) they are the worst part of the standard. There they are very
3 :     \ restricted and have an ugly interface.
4 :    
5 :     \ So, we implement the locals wordset, but do not recommend using
6 :     \ locals-ext (which is a really bad user interface for locals).
7 :    
8 :     \ We also have a nice and powerful user-interface for locals: locals are
9 :     \ defined with
10 :    
11 :     \ { local1 local2 ... }
12 :     \ or
13 :     \ { local1 local2 ... -- ... }
14 :     \ (anything after the -- is just a comment)
15 :    
16 :     \ Every local in this list consists of an optional type specification
17 :     \ and a name. If there is only the name, it stands for a cell-sized
18 :     \ value (i.e., you get the value of the local variable, not it's
19 :     \ address). The following type specifiers stand before the name:
20 :    
21 :     \ Specifier Type Access
22 :     \ W: Cell value
23 :     \ W^ Cell address
24 :     \ D: Double value
25 :     \ D^ Double address
26 :     \ F: Float value
27 :     \ F^ Float address
28 :     \ C: Char value
29 :     \ C^ Char address
30 :    
31 :     \ The local variables are initialized with values from the appropriate
32 :     \ stack. In contrast to the examples in the standard document our locals
33 :     \ take the arguments in the expected way: The last local gets the top of
34 :     \ stack, the second last gets the second stack item etc. An example:
35 :    
36 :     \ : CX* { F: Ar F: Ai F: Br F: Bi -- Cr Ci }
37 :     \ \ complex multiplication
38 :     \ Ar Br f* Ai Bi f* f-
39 :     \ Ar Bi f* Ai Br f* f+ ;
40 :    
41 :     \ There will also be a way to add user types, but it is not yet decided,
42 :     \ how. Ideas are welcome.
43 :    
44 :     \ Locals defined in this manner live until (!! see below).
45 :     \ Their names can be used during this time to get
46 :     \ their value or address; The addresses produced in this way become
47 :     \ invalid at the end of the lifetime.
48 :    
49 :     \ Values can be changed with TO, but this is not recomended (TO is a
50 :     \ kludge and words lose the single-assignment property, which makes them
51 :     \ harder to analyse).
52 :    
53 :     \ As for the internals, we use a special locals stack. This eliminates
54 :     \ the problems and restrictions of reusing the return stack and allows
55 :     \ to store floats as locals: the return stack is not guaranteed to be
56 :     \ aligned correctly, but our locals stack must be float-aligned between
57 :     \ words.
58 :    
59 :     \ Other things about the internals are pretty unclear now.
60 :    
61 :     \ Currently locals may only be
62 :     \ defined at the outer level and TO is not supported.
63 :    
64 :     include float.fs
65 :     include search-order.fs
66 :    
67 : anton 1.3 : compile-@local ( n -- )
68 :     case
69 :     0 of postpone @local0 endof
70 :     4 of postpone @local4 endof
71 :     8 of postpone @local8 endof
72 :     12 of postpone @local12 endof
73 :     ( otherwise ) dup postpone @local# ,
74 :     endcase ;
75 :    
76 :     : compile-f@local ( n -- )
77 :     case
78 :     0 of postpone f@local0 endof
79 :     8 of postpone f@local8 endof
80 :     ( otherwise ) dup postpone f@local# ,
81 :     endcase ;
82 :    
83 : anton 1.1 \ the locals stack grows downwards (see primitives)
84 :     \ of the local variables of a group (in braces) the leftmost is on top,
85 :     \ i.e. by going onto the locals stack the order is reversed.
86 :     \ there are alignment gaps if necessary.
87 :     \ lp must have the strictest alignment (usually float) across calls;
88 :     \ for simplicity we align it strictly for every group.
89 :    
90 : anton 1.5 slowvoc @
91 :     slowvoc on \ we want a linked list for the vocabulary locals
92 : anton 1.1 vocabulary locals \ this contains the local variables
93 : anton 1.3 ' locals >body ' locals-list >body !
94 : anton 1.5 slowvoc !
95 : anton 1.1
96 :     create locals-buffer 1000 allot \ !! limited and unsafe
97 :     \ here the names of the local variables are stored
98 :     \ we would have problems storing them at the normal dp
99 :    
100 :     variable locals-dp \ so here's the special dp for locals.
101 :    
102 :     : alignlp-w ( n1 -- n2 )
103 :     \ cell-align size and generate the corresponding code for aligning lp
104 : anton 1.3 aligned dup adjust-locals-size ;
105 : anton 1.1
106 :     : alignlp-f ( n1 -- n2 )
107 : anton 1.3 faligned dup adjust-locals-size ;
108 : anton 1.1
109 :     \ a local declaration group (the braces stuff) is compiled by calling
110 :     \ the appropriate compile-pushlocal for the locals, starting with the
111 :     \ righmost local; the names are already created earlier, the
112 :     \ compile-pushlocal just inserts the offsets from the frame base.
113 :    
114 :     : compile-pushlocal-w ( a-addr -- ) ( run-time: w -- )
115 :     \ compiles a push of a local variable, and adjusts locals-size
116 :     \ stores the offset of the local variable to a-addr
117 :     locals-size @ alignlp-w cell+ dup locals-size !
118 :     swap !
119 :     postpone >l ;
120 :    
121 :     : compile-pushlocal-f ( a-addr -- ) ( run-time: f -- )
122 :     locals-size @ alignlp-f float+ dup locals-size !
123 :     swap !
124 :     postpone f>l ;
125 :    
126 :     : compile-pushlocal-d ( a-addr -- ) ( run-time: w1 w2 -- )
127 :     locals-size @ alignlp-w cell+ cell+ dup locals-size !
128 :     swap !
129 :     postpone swap postpone >l postpone >l ;
130 :    
131 :     : compile-pushlocal-c ( a-addr -- ) ( run-time: w -- )
132 : anton 1.3 -1 chars compile-lp+!
133 : anton 1.1 locals-size @ swap !
134 :     postpone lp@ postpone c! ;
135 :    
136 :     : create-local ( " name" -- a-addr )
137 :     \ defines the local "name"; the offset of the local shall be stored in a-addr
138 :     create
139 :     immediate
140 :     here 0 , ( place for the offset ) ;
141 :    
142 : anton 1.3 : lp-offset ( n1 -- n2 )
143 :     \ converts the offset from the frame start to an offset from lp and
144 :     \ i.e., the address of the local is lp+locals_size-offset
145 :     locals-size @ swap - ;
146 :    
147 : anton 1.1 : lp-offset, ( n -- )
148 :     \ converts the offset from the frame start to an offset from lp and
149 :     \ adds it as inline argument to a preceding locals primitive
150 : anton 1.3 lp-offset , ;
151 : anton 1.1
152 :     vocabulary locals-types \ this contains all the type specifyers, -- and }
153 :     locals-types definitions
154 :    
155 :     : W:
156 :     create-local ( "name" -- a-addr xt )
157 :     \ xt produces the appropriate locals pushing code when executed
158 :     ['] compile-pushlocal-w
159 :     does> ( Compilation: -- ) ( Run-time: -- w )
160 :     \ compiles a local variable access
161 : anton 1.3 @ lp-offset compile-@local ;
162 : anton 1.1
163 :     : W^
164 :     create-local ( "name" -- a-addr xt )
165 :     ['] compile-pushlocal-w
166 :     does> ( Compilation: -- ) ( Run-time: -- w )
167 :     postpone laddr# @ lp-offset, ;
168 :    
169 :     : F:
170 :     create-local ( "name" -- a-addr xt )
171 :     ['] compile-pushlocal-f
172 :     does> ( Compilation: -- ) ( Run-time: -- w )
173 : anton 1.3 @ lp-offset compile-f@local ;
174 : anton 1.1
175 :     : F^
176 :     create-local ( "name" -- a-addr xt )
177 :     ['] compile-pushlocal-f
178 :     does> ( Compilation: -- ) ( Run-time: -- w )
179 :     postpone laddr# @ lp-offset, ;
180 :    
181 :     : D:
182 :     create-local ( "name" -- a-addr xt )
183 :     ['] compile-pushlocal-d
184 :     does> ( Compilation: -- ) ( Run-time: -- w )
185 :     postpone laddr# @ lp-offset, postpone 2@ ;
186 :    
187 :     : D^
188 :     create-local ( "name" -- a-addr xt )
189 :     ['] compile-pushlocal-d
190 :     does> ( Compilation: -- ) ( Run-time: -- w )
191 :     postpone laddr# @ lp-offset, ;
192 :    
193 :     : C:
194 :     create-local ( "name" -- a-addr xt )
195 :     ['] compile-pushlocal-c
196 :     does> ( Compilation: -- ) ( Run-time: -- w )
197 :     postpone laddr# @ lp-offset, postpone c@ ;
198 :    
199 :     : C^
200 :     create-local ( "name" -- a-addr xt )
201 :     ['] compile-pushlocal-c
202 :     does> ( Compilation: -- ) ( Run-time: -- w )
203 :     postpone laddr# @ lp-offset, ;
204 :    
205 :     \ you may want to make comments in a locals definitions group:
206 :     ' \ alias \ immediate
207 :     ' ( alias ( immediate
208 :    
209 :     forth definitions
210 :    
211 :     \ the following gymnastics are for declaring locals without type specifier.
212 :     \ we exploit a feature of our dictionary: every wordlist
213 :     \ has it's own methods for finding words etc.
214 :     \ So we create a vocabulary new-locals, that creates a 'w:' local named x
215 :     \ when it is asked if it contains x.
216 :    
217 :     also locals-types
218 :    
219 :     : new-locals-find ( caddr u w -- nfa )
220 :     \ this is the find method of the new-locals vocabulary
221 :     \ make a new local with name caddr u; w is ignored
222 :     \ the returned nfa denotes a word that produces what W: produces
223 :     \ !! do the whole thing without nextname
224 : anton 1.3 drop nextname
225 :     ['] W: >name ;
226 : anton 1.1
227 :     previous
228 :    
229 :     : new-locals-reveal ( -- )
230 :     true abort" this should not happen: new-locals-reveal" ;
231 :    
232 :     create new-locals-map ' new-locals-find A, ' new-locals-reveal A,
233 :    
234 :     vocabulary new-locals
235 :     new-locals-map ' new-locals >body cell+ A! \ !! use special access words
236 :    
237 :     variable old-dpp
238 :    
239 :     \ and now, finally, the user interface words
240 :     : { ( -- addr wid 0 )
241 :     dp old-dpp !
242 :     locals-dp dpp !
243 :     also new-locals
244 :     also get-current locals definitions locals-types
245 :     0 TO locals-wordlist
246 :     0 postpone [ ; immediate
247 :    
248 :     locals-types definitions
249 :    
250 :     : } ( addr wid 0 a-addr1 xt1 ... -- )
251 :     \ ends locals definitions
252 :     ] old-dpp @ dpp !
253 :     begin
254 :     dup
255 :     while
256 :     execute
257 :     repeat
258 :     drop
259 :     locals-size @ alignlp-f locals-size ! \ the strictest alignment
260 :     set-current
261 :     previous previous
262 :     locals-list TO locals-wordlist ;
263 :    
264 :     : -- ( addr wid 0 ... -- )
265 :     }
266 :     [char] } word drop ;
267 :    
268 :     forth definitions
269 :    
270 :     \ A few thoughts on automatic scopes for locals and how they can be
271 :     \ implemented:
272 :    
273 :     \ We have to combine locals with the control structures. My basic idea
274 :     \ was to start the life of a local at the declaration point. The life
275 :     \ would end at any control flow join (THEN, BEGIN etc.) where the local
276 :     \ is lot live on both input flows (note that the local can still live in
277 :     \ other, later parts of the control flow). This would make a local live
278 :     \ as long as you expected and sometimes longer (e.g. a local declared in
279 :     \ a BEGIN..UNTIL loop would still live after the UNTIL).
280 :    
281 :     \ The following example illustrates the problems of this approach:
282 :    
283 :     \ { z }
284 :     \ if
285 :     \ { x }
286 :     \ begin
287 :     \ { y }
288 :     \ [ 1 cs-roll ] then
289 :     \ ...
290 :     \ until
291 :    
292 :     \ x lives only until the BEGIN, but the compiler does not know this
293 :     \ until it compiles the UNTIL (it can deduce it at the THEN, because at
294 :     \ that point x lives in no thread, but that does not help much). This is
295 :     \ solved by optimistically assuming at the BEGIN that x lives, but
296 :     \ warning at the UNTIL that it does not. The user is then responsible
297 :     \ for checking that x is only used where it lives.
298 :    
299 :     \ The produced code might look like this (leaving out alignment code):
300 :    
301 :     \ >l ( z )
302 :     \ ?branch <then>
303 :     \ >l ( x )
304 :     \ <begin>:
305 :     \ >l ( y )
306 :     \ lp+!# 8 ( RIP: x,y )
307 :     \ <then>:
308 :     \ ...
309 :     \ lp+!# -4 ( adjust lp to <begin> state )
310 :     \ ?branch <begin>
311 :     \ lp+!# 4 ( undo adjust )
312 :    
313 :     \ The BEGIN problem also has another incarnation:
314 :    
315 :     \ AHEAD
316 :     \ BEGIN
317 :     \ x
318 :     \ [ 1 CS-ROLL ] THEN
319 :     \ { x }
320 :     \ ...
321 :     \ UNTIL
322 :    
323 :     \ should be legal: The BEGIN is not a control flow join in this case,
324 :     \ since it cannot be entered from the top; therefore the definition of x
325 :     \ dominates the use. But the compiler processes the use first, and since
326 :     \ it does not look ahead to notice the definition, it will complain
327 :     \ about it. Here's another variation of this problem:
328 :    
329 :     \ IF
330 :     \ { x }
331 :     \ ELSE
332 :     \ ...
333 :     \ AHEAD
334 :     \ BEGIN
335 :     \ x
336 :     \ [ 2 CS-ROLL ] THEN
337 :     \ ...
338 :     \ UNTIL
339 :    
340 :     \ In this case x is defined before the use, and the definition dominates
341 :     \ the use, but the compiler does not know this until it processes the
342 :     \ UNTIL. So what should the compiler assume does live at the BEGIN, if
343 :     \ the BEGIN is not a control flow join? The safest assumption would be
344 :     \ the intersection of all locals lists on the control flow
345 :     \ stack. However, our compiler assumes that the same variables are live
346 :     \ as on the top of the control flow stack. This covers the following case:
347 :    
348 :     \ { x }
349 :     \ AHEAD
350 :     \ BEGIN
351 :     \ x
352 :     \ [ 1 CS-ROLL ] THEN
353 :     \ ...
354 :     \ UNTIL
355 :    
356 :     \ If this assumption is too optimistic, the compiler will warn the user.
357 :    
358 : anton 1.3 \ Implementation: migrated to kernal.fs
359 : anton 1.1
360 :     \ THEN (another control flow from before joins the current one):
361 :     \ The new locals-list is the intersection of the current locals-list and
362 :     \ the orig-local-list. The new locals-size is the (alignment-adjusted)
363 :     \ size of the new locals-list. The following code is generated:
364 :     \ lp+!# (current-locals-size - orig-locals-size)
365 :     \ <then>:
366 :     \ lp+!# (orig-locals-size - new-locals-size)
367 :    
368 :     \ Of course "lp+!# 0" is not generated. Still this is admittedly a bit
369 :     \ inefficient, e.g. if there is a locals declaration between IF and
370 :     \ ELSE. However, if ELSE generates an appropriate "lp+!#" before the
371 :     \ branch, there will be none after the target <then>.
372 :    
373 : anton 1.3 \ explicit scoping
374 : anton 1.1
375 : anton 1.3 : scope ( -- scope )
376 :     cs-push-part scopestart ; immediate
377 :    
378 :     : endscope ( scope -- )
379 :     scope?
380 : anton 1.1 drop
381 : anton 1.3 locals-list @ common-list
382 :     dup list-size adjust-locals-size
383 :     locals-list ! ; immediate
384 : anton 1.1
385 : anton 1.3 \ adapt the hooks
386 : anton 1.1
387 : anton 1.3 : locals-:-hook ( sys -- sys addr xt n )
388 :     \ addr is the nfa of the defined word, xt its xt
389 : anton 1.1 DEFERS :-hook
390 :     last @ lastcfa @
391 :     clear-leave-stack
392 :     0 locals-size !
393 :     locals-buffer locals-dp !
394 : anton 1.3 0 locals-list !
395 :     dead-code off
396 :     defstart ;
397 : anton 1.1
398 : anton 1.3 : locals-;-hook ( sys addr xt sys -- sys )
399 :     def?
400 : anton 1.1 0 TO locals-wordlist
401 : anton 1.3 0 adjust-locals-size ( not every def ends with an exit )
402 : anton 1.1 lastcfa ! last !
403 :     DEFERS ;-hook ;
404 :    
405 :     ' locals-:-hook IS :-hook
406 :     ' locals-;-hook IS ;-hook
407 :    
408 :     \ The words in the locals dictionary space are not deleted until the end
409 :     \ of the current word. This is a bit too conservative, but very simple.
410 :    
411 :     \ There are a few cases to consider: (see above)
412 :    
413 :     \ after AGAIN, AHEAD, EXIT (the current control flow is dead):
414 :     \ We have to special-case the above cases against that. In this case the
415 :     \ things above are not control flow joins. Everything should be taken
416 :     \ over from the live flow. No lp+!# is generated.
417 :    
418 :     \ !! The lp gymnastics for UNTIL are also a real problem: locals cannot be
419 :     \ used in signal handlers (or anything else that may be called while
420 :     \ locals live beyond the lp) without changing the locals stack.
421 :    
422 :     \ About warning against uses of dead locals. There are several options:
423 :    
424 :     \ 1) Do not complain (After all, this is Forth;-)
425 :    
426 :     \ 2) Additional restrictions can be imposed so that the situation cannot
427 :     \ arise; the programmer would have to introduce explicit scoping
428 :     \ declarations in cases like the above one. I.e., complain if there are
429 :     \ locals that are live before the BEGIN but not before the corresponding
430 :     \ AGAIN (replace DO etc. for BEGIN and UNTIL etc. for AGAIN).
431 :    
432 :     \ 3) The real thing: i.e. complain, iff a local lives at a BEGIN, is
433 :     \ used on a path starting at the BEGIN, and does not live at the
434 :     \ corresponding AGAIN. This is somewhat hard to implement. a) How does
435 :     \ the compiler know when it is working on a path starting at a BEGIN
436 :     \ (consider "{ x } if begin [ 1 cs-roll ] else x endif again")? b) How
437 :     \ is the usage info stored?
438 :    
439 :     \ For now I'll resort to alternative 2. When it produces warnings they
440 :     \ will often be spurious, but warnings should be rare. And better
441 :     \ spurious warnings now and then than days of bug-searching.
442 :    
443 :     \ Explicit scoping of locals is implemented by cs-pushing the current
444 :     \ locals-list and -size (and an unused cell, to make the size equal to
445 :     \ the other entries) at the start of the scope, and restoring them at
446 :     \ the end of the scope to the intersection, like THEN does.
447 :    
448 :    
449 :     \ And here's finally the ANS standard stuff
450 :    
451 :     : (local) ( addr u -- )
452 : anton 1.3 \ a little space-inefficient, but well deserved ;-)
453 :     \ In exchange, there are no restrictions whatsoever on using (local)
454 : anton 1.4 \ as long as you use it in a definition
455 : anton 1.3 dup
456 :     if
457 :     nextname POSTPONE { [ also locals-types ] W: } [ previous ]
458 :     else
459 :     2drop
460 :     endif ;
461 : anton 1.1
462 : anton 1.4 : >definer ( xt -- definer )
463 :     \ this gives a unique identifier for the way the xt was defined
464 :     \ words defined with different does>-codes have different definers
465 :     \ the definer can be used for comparison and in definer!
466 :     dup >code-address [ ' bits >code-address ] Literal =
467 :     \ !! this definition will not work on some implementations for `bits'
468 :     if \ if >code-address delivers the same value for all does>-def'd words
469 :     >does-code 1 or \ bit 0 marks special treatment for does codes
470 :     else
471 :     >code-address
472 :     then ;
473 :    
474 :     : definer! ( definer xt -- )
475 :     \ gives the word represented by xt the behaviour associated with definer
476 :     over 1 and if
477 :     does-code!
478 :     else
479 :     code-address!
480 :     then ;
481 :    
482 :     \ !! untested
483 :     : TO ( c|w|d|r "name" -- )
484 :     \ !! state smart
485 :     0 0 0. 0.0e0 { c: clocal w: wlocal d: dlocal f: flocal }
486 :     ' dup >definer
487 :     state @
488 :     if
489 :     case
490 :     [ ' locals-wordlist >definer ] literal \ value
491 :     OF >body POSTPONE Aliteral POSTPONE ! ENDOF
492 :     [ ' clocal >definer ] literal
493 :     OF POSTPONE laddr# >body @ lp-offset, POSTPONE c! ENDOF
494 :     [ ' wlocal >definer ] literal
495 :     OF POSTPONE laddr# >body @ lp-offset, POSTPONE ! ENDOF
496 :     [ ' dlocal >definer ] literal
497 :     OF POSTPONE laddr# >body @ lp-offset, POSTPONE d! ENDOF
498 :     [ ' flocal >definer ] literal
499 :     OF POSTPONE laddr# >body @ lp-offset, POSTPONE f! ENDOF
500 :     abort" can only store TO value or local value"
501 :     endcase
502 :     else
503 :     [ ' locals-wordlist >definer ] literal =
504 :     if
505 :     >body !
506 :     else
507 :     abort" can only store TO value"
508 :     endif
509 :     endif ; immediate
510 : anton 1.1
511 : pazsan 1.6 : locals|
512 :     BEGIN sname 2dup s" |" compare 0= WHILE
513 :     (local) REPEAT drop 0 (local) ; immediate restrict

CVS Admin

Powered by ViewCVS 1.0-dev
(Powered by ViewCVS)

ViewCVS and CVS Help