Thread: [ clisp-Bugs-3147908 ] too much consing when compiling IGNORE-ERRORS

Brought to you by: haible, hoehle, sds

clisp-devel

[ clisp-Bugs-3147908 ] too much consing when compiling IGNORE-ERRORS

From: SourceForge.net <no...@so...> - 2010-12-30 01:14:42

Bugs item #3147908, was opened at 2010-12-29 20:14
Message generated for change (Tracker Item Submitted) made by sds
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: clisp
Group: None
Status: Open
Resolution: None
Priority: 5
Private: No
Submitted By: Sam Steingold (sds)
Assigned to: Bruno Haible (haible)
Summary: too much consing when compiling IGNORE-ERRORS

Initial Comment:
in GNU CLISP 2.33.1 (2004-05-22) (and earlier)
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = (#(ERROR 15) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
16 byte-code instructions:
0     (BLOCK-OPEN 0 L24)
3     (HANDLER-OPEN 1 L15)                ; (#(ERROR 15) 2 . 1)
5     (LOAD&PUSH 9)
6     (LOAD&PUSH 9)
7     (CALLSR 1 45)                       ; =
10    (SKIP 4)
12    (BLOCK-CLOSE)
13    (SKIP&RET 3)
15    L15
15    (HANDLER-BEGIN&PUSH)
16    (NIL&PUSH)
17    (LOAD&PUSH 1)
18    (STACK-TO-MV 2)
20    (RETURN-FROM-I 0 0 2)
24    L24
24    (SKIP&RET 3)

in GNU CLISP 2.34 (2005-07-20) (and later):
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = #<COMPILED-FUNCTION :LAMBDA-1>
(CONST 2) = #<COMPILED-FUNCTION :LAMBDA-2>
(CONST 3) = (#(ERROR 32) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
25 byte-code instructions:
0     (NIL)
1     (MAKE-VECTOR1&PUSH 2)
3     (LOAD&STOREC 3 0 0)
7     (LOAD&STOREC 2 0 1)
11    (BLOCK-OPEN 0 L47)
14    (LOAD&PUSH 2)
15    (COPY-CLOSURE&PUSH 1 1)             ; #<COMPILED-FUNCTION :LAMBDA-1>
18    (LOAD&PUSH 4)
19    (COPY-CLOSURE&PUSH 2 1)             ; #<COMPILED-FUNCTION :LAMBDA-2>
22    (HANDLER-OPEN 3 L32)                ; (#(ERROR 32) 2 . 1)
24    (LOAD&PUSH 4)
25    (FUNCALL 0)
27    (SKIP 6)
29    (BLOCK-CLOSE)
30    (SKIP&RET 4)
32    L32
32    (HANDLER-BEGIN&PUSH)
33    (LOADI&PUSH 0 0 1)
37    (FUNCALL&PUSH 0)
39    (LOAD&PUSH 1)
40    (FUNCALL 1)
42    (SKIPSP 3 1)
45    (SKIP&RET 2)
47    L47
47    (SKIP&RET 4)

which is longer and conses more.

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355

[ clisp-Bugs-3147908 ] too much consing when compiling IGNORE-ERRORS

From: SourceForge.net <no...@so...> - 2010-12-30 14:12:58

Bugs item #3147908, was opened at 2010-12-29 20:14
Message generated for change (Comment added) made by sds
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: clisp
Group: None
Status: Open
Resolution: None
Priority: 5
Private: No
Submitted By: Sam Steingold (sds)
>Assigned to: Sam Steingold (sds)
Summary: too much consing when compiling IGNORE-ERRORS

Initial Comment:
in GNU CLISP 2.33.1 (2004-05-22) (and earlier)
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = (#(ERROR 15) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
16 byte-code instructions:
0     (BLOCK-OPEN 0 L24)
3     (HANDLER-OPEN 1 L15)                ; (#(ERROR 15) 2 . 1)
5     (LOAD&PUSH 9)
6     (LOAD&PUSH 9)
7     (CALLSR 1 45)                       ; =
10    (SKIP 4)
12    (BLOCK-CLOSE)
13    (SKIP&RET 3)
15    L15
15    (HANDLER-BEGIN&PUSH)
16    (NIL&PUSH)
17    (LOAD&PUSH 1)
18    (STACK-TO-MV 2)
20    (RETURN-FROM-I 0 0 2)
24    L24
24    (SKIP&RET 3)

in GNU CLISP 2.34 (2005-07-20) (and later):
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = #<COMPILED-FUNCTION :LAMBDA-1>
(CONST 2) = #<COMPILED-FUNCTION :LAMBDA-2>
(CONST 3) = (#(ERROR 32) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
25 byte-code instructions:
0     (NIL)
1     (MAKE-VECTOR1&PUSH 2)
3     (LOAD&STOREC 3 0 0)
7     (LOAD&STOREC 2 0 1)
11    (BLOCK-OPEN 0 L47)
14    (LOAD&PUSH 2)
15    (COPY-CLOSURE&PUSH 1 1)             ; #<COMPILED-FUNCTION :LAMBDA-1>
18    (LOAD&PUSH 4)
19    (COPY-CLOSURE&PUSH 2 1)             ; #<COMPILED-FUNCTION :LAMBDA-2>
22    (HANDLER-OPEN 3 L32)                ; (#(ERROR 32) 2 . 1)
24    (LOAD&PUSH 4)
25    (FUNCALL 0)
27    (SKIP 6)
29    (BLOCK-CLOSE)
30    (SKIP&RET 4)
32    L32
32    (HANDLER-BEGIN&PUSH)
33    (LOADI&PUSH 0 0 1)
37    (FUNCALL&PUSH 0)
39    (LOAD&PUSH 1)
40    (FUNCALL 1)
42    (SKIPSP 3 1)
45    (SKIP&RET 2)
47    L47
47    (SKIP&RET 4)

which is longer and conses more.

----------------------------------------------------------------------

>Comment By: Sam Steingold (sds)
Date: 2010-12-30 09:12

Message:
somehow I cannot compile 2.34 on my laptop (intparam.h:7:2: error: #error
"Integers of type int have no binary representation!!" &c) so my
investigation is limited to the lisp side.
looks like older clisp compiles the call to = and values inline which 2.34
creates closures for that (which is kind of crazy).
apparently caused by my 2005-01-21 patch

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355

[ clisp-Bugs-3147908 ] too much consing when compiling IGNORE-ERRORS

From: SourceForge.net <no...@so...> - 2010-12-31 02:25:11

Bugs item #3147908, was opened at 2010-12-30 02:14
Message generated for change (Comment added) made by haible
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: clisp
Group: None
Status: Open
Resolution: None
Priority: 5
Private: No
Submitted By: Sam Steingold (sds)
Assigned to: Sam Steingold (sds)
Summary: too much consing when compiling IGNORE-ERRORS

Initial Comment:
in GNU CLISP 2.33.1 (2004-05-22) (and earlier)
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = (#(ERROR 15) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
16 byte-code instructions:
0     (BLOCK-OPEN 0 L24)
3     (HANDLER-OPEN 1 L15)                ; (#(ERROR 15) 2 . 1)
5     (LOAD&PUSH 9)
6     (LOAD&PUSH 9)
7     (CALLSR 1 45)                       ; =
10    (SKIP 4)
12    (BLOCK-CLOSE)
13    (SKIP&RET 3)
15    L15
15    (HANDLER-BEGIN&PUSH)
16    (NIL&PUSH)
17    (LOAD&PUSH 1)
18    (STACK-TO-MV 2)
20    (RETURN-FROM-I 0 0 2)
24    L24
24    (SKIP&RET 3)

in GNU CLISP 2.34 (2005-07-20) (and later):
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = #<COMPILED-FUNCTION :LAMBDA-1>
(CONST 2) = #<COMPILED-FUNCTION :LAMBDA-2>
(CONST 3) = (#(ERROR 32) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
25 byte-code instructions:
0     (NIL)
1     (MAKE-VECTOR1&PUSH 2)
3     (LOAD&STOREC 3 0 0)
7     (LOAD&STOREC 2 0 1)
11    (BLOCK-OPEN 0 L47)
14    (LOAD&PUSH 2)
15    (COPY-CLOSURE&PUSH 1 1)             ; #<COMPILED-FUNCTION :LAMBDA-1>
18    (LOAD&PUSH 4)
19    (COPY-CLOSURE&PUSH 2 1)             ; #<COMPILED-FUNCTION :LAMBDA-2>
22    (HANDLER-OPEN 3 L32)                ; (#(ERROR 32) 2 . 1)
24    (LOAD&PUSH 4)
25    (FUNCALL 0)
27    (SKIP 6)
29    (BLOCK-CLOSE)
30    (SKIP&RET 4)
32    L32
32    (HANDLER-BEGIN&PUSH)
33    (LOADI&PUSH 0 0 1)
37    (FUNCALL&PUSH 0)
39    (LOAD&PUSH 1)
40    (FUNCALL 1)
42    (SKIPSP 3 1)
45    (SKIP&RET 2)
47    L47
47    (SKIP&RET 4)

which is longer and conses more.

----------------------------------------------------------------------

>Comment By: Bruno Haible (haible)
Date: 2010-12-31 03:25

Message:
Yes I agree it's most likely caused by this patch:

2005-01-21  Sam Steingold  <sd...@gn...>

        * condition.lisp (handler-bind): use the function syntax for
        %HANDLER-BIND
        * compiler.lisp (quote-p): new function
        (l-constantp, c-constantp): use it
        (c-form-table): do not handle HANDLER-BIND specially
        (c-HANDLER-BIND): handle HANDLER-BIND in function syntax
        * init.lisp (%expand-form): do not handle %HANDLER-BIND specially
        * clos-class3.lisp (reinitialize-instance-<defined-class>):
        use use the function syntax for %HANDLER-BIND
        * clos-genfun4.lisp (generic-function-undeterminedp): ditto
        * clos-methcomb2.lisp (any-method-combination-check-options):
ditto

The NEWS entry for this patch was: "Third party code walkers can now
handle HANDLER-BIND et al."
So the objective was to remove the need for code walkers to handle
sys::%handler-bind.
The patch itself was fine. The problem is that this macroexpansion:

> (macroexpand-1 (third (macroexpand-1 '(IGNORE-ERRORS (= X Y)))))
(LET
  ((#:G2727
      #'(LAMBDA NIL
          (PROGN
            #'(LAMBDA (CONDITION) (RETURN-FROM #:G2726 (VALUES NIL
CONDITION)))
   )    ) )
    (#:G2728 #'(LAMBDA NIL (PROGN (= X Y))))
  )
  (LOCALLY (DECLARE (COMPILE))
    (SYSTEM::%HANDLER-BIND #:G2728 'ERROR
      #'(LAMBDA (CONDITION) (FUNCALL (FUNCALL #:G2727) CONDITION))
) ) )

is not compiled to good code. This needs work in the compiler.

A smaller example is:
(disassemble (LAMBDA (X Y) (HANDLER-BIND () (= X Y))))

Macro expansion:

(LET ((#:G2784 #'(LAMBDA NIL (PROGN (= X Y)))))
  (LOCALLY (DECLARE (COMPILE)) (SYSTEM::%HANDLER-BIND #:G2784))
)

clisp's function inliner is not powerful enough to resolve this. I don't
remember whether it is because
it does not do data flow analysis or because the LOCALLY form introduces a
new declaration scope.
But there is a way to make this special case work better without beefing
up the function inliner nor the
data flow analysis...


----------------------------------------------------------------------

Comment By: Sam Steingold (sds)
Date: 2010-12-30 15:12

Message:
somehow I cannot compile 2.34 on my laptop (intparam.h:7:2: error: #error
"Integers of type int have no binary representation!!" &c) so my
investigation is limited to the lisp side.
looks like older clisp compiles the call to = and values inline which 2.34
creates closures for that (which is kind of crazy).
apparently caused by my 2005-01-21 patch

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355

[ clisp-Bugs-3147908 ] too much consing when compiling IGNORE-ERRORS

From: SourceForge.net <no...@so...> - 2011-03-28 22:05:49

Bugs item #3147908, was opened at 2010-12-29 20:14
Message generated for change (Comment added) made by sds
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: clisp
Group: None
Status: Open
Resolution: None
Priority: 5
Private: No
Submitted By: Sam Steingold (sds)
Assigned to: Sam Steingold (sds)
Summary: too much consing when compiling IGNORE-ERRORS

Initial Comment:
in GNU CLISP 2.33.1 (2004-05-22) (and earlier)
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = (#(ERROR 15) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
16 byte-code instructions:
0     (BLOCK-OPEN 0 L24)
3     (HANDLER-OPEN 1 L15)                ; (#(ERROR 15) 2 . 1)
5     (LOAD&PUSH 9)
6     (LOAD&PUSH 9)
7     (CALLSR 1 45)                       ; =
10    (SKIP 4)
12    (BLOCK-CLOSE)
13    (SKIP&RET 3)
15    L15
15    (HANDLER-BEGIN&PUSH)
16    (NIL&PUSH)
17    (LOAD&PUSH 1)
18    (STACK-TO-MV 2)
20    (RETURN-FROM-I 0 0 2)
24    L24
24    (SKIP&RET 3)

in GNU CLISP 2.34 (2005-07-20) (and later):
(disassemble (LAMBDA (X Y) (IGNORE-ERRORS (= X Y))))
Disassembly of function :LAMBDA
(CONST 0) = NIL
(CONST 1) = #<COMPILED-FUNCTION :LAMBDA-1>
(CONST 2) = #<COMPILED-FUNCTION :LAMBDA-2>
(CONST 3) = (#(ERROR 32) 2 . 1)
2 required arguments
0 optional arguments
No rest parameter
No keyword parameters
25 byte-code instructions:
0     (NIL)
1     (MAKE-VECTOR1&PUSH 2)
3     (LOAD&STOREC 3 0 0)
7     (LOAD&STOREC 2 0 1)
11    (BLOCK-OPEN 0 L47)
14    (LOAD&PUSH 2)
15    (COPY-CLOSURE&PUSH 1 1)             ; #<COMPILED-FUNCTION :LAMBDA-1>
18    (LOAD&PUSH 4)
19    (COPY-CLOSURE&PUSH 2 1)             ; #<COMPILED-FUNCTION :LAMBDA-2>
22    (HANDLER-OPEN 3 L32)                ; (#(ERROR 32) 2 . 1)
24    (LOAD&PUSH 4)
25    (FUNCALL 0)
27    (SKIP 6)
29    (BLOCK-CLOSE)
30    (SKIP&RET 4)
32    L32
32    (HANDLER-BEGIN&PUSH)
33    (LOADI&PUSH 0 0 1)
37    (FUNCALL&PUSH 0)
39    (LOAD&PUSH 1)
40    (FUNCALL 1)
42    (SKIPSP 3 1)
45    (SKIP&RET 2)
47    L47
47    (SKIP&RET 4)

which is longer and conses more.

----------------------------------------------------------------------

>Comment By: Sam Steingold (sds)
Date: 2011-03-28 18:05

Message:
the simple version
(defmacro handler-bind (clauses &body body)
  `(LOCALLY (DECLARE (COMPILE))
     (SYS::%HANDLER-BIND
      #'(LAMBDA () ,@body)
      ,@(mapcan #'(lambda (clause)
                    (destructuring-bind (typespec handler) clause
                      `(',typespec ,handler)))
                clauses))))
does not work.

before the patch c-handler-bind received the non-macro-expanded version
which made all it possible to inline everything.

----------------------------------------------------------------------

Comment By: Bruno Haible (haible)
Date: 2010-12-30 21:25

Message:
Yes I agree it's most likely caused by this patch:

2005-01-21  Sam Steingold  <sd...@gn...>

        * condition.lisp (handler-bind): use the function syntax for
        %HANDLER-BIND
        * compiler.lisp (quote-p): new function
        (l-constantp, c-constantp): use it
        (c-form-table): do not handle HANDLER-BIND specially
        (c-HANDLER-BIND): handle HANDLER-BIND in function syntax
        * init.lisp (%expand-form): do not handle %HANDLER-BIND specially
        * clos-class3.lisp (reinitialize-instance-<defined-class>):
        use use the function syntax for %HANDLER-BIND
        * clos-genfun4.lisp (generic-function-undeterminedp): ditto
        * clos-methcomb2.lisp (any-method-combination-check-options):
ditto

The NEWS entry for this patch was: "Third party code walkers can now
handle HANDLER-BIND et al."
So the objective was to remove the need for code walkers to handle
sys::%handler-bind.
The patch itself was fine. The problem is that this macroexpansion:

> (macroexpand-1 (third (macroexpand-1 '(IGNORE-ERRORS (= X Y)))))
(LET
  ((#:G2727
      #'(LAMBDA NIL
          (PROGN
            #'(LAMBDA (CONDITION) (RETURN-FROM #:G2726 (VALUES NIL
CONDITION)))
   )    ) )
    (#:G2728 #'(LAMBDA NIL (PROGN (= X Y))))
  )
  (LOCALLY (DECLARE (COMPILE))
    (SYSTEM::%HANDLER-BIND #:G2728 'ERROR
      #'(LAMBDA (CONDITION) (FUNCALL (FUNCALL #:G2727) CONDITION))
) ) )

is not compiled to good code. This needs work in the compiler.

A smaller example is:
(disassemble (LAMBDA (X Y) (HANDLER-BIND () (= X Y))))

Macro expansion:

(LET ((#:G2784 #'(LAMBDA NIL (PROGN (= X Y)))))
  (LOCALLY (DECLARE (COMPILE)) (SYSTEM::%HANDLER-BIND #:G2784))
)

clisp's function inliner is not powerful enough to resolve this. I don't
remember whether it is because
it does not do data flow analysis or because the LOCALLY form introduces a
new declaration scope.
But there is a way to make this special case work better without beefing
up the function inliner nor the
data flow analysis...


----------------------------------------------------------------------

Comment By: Sam Steingold (sds)
Date: 2010-12-30 09:12

Message:
somehow I cannot compile 2.34 on my laptop (intparam.h:7:2: error: #error
"Integers of type int have no binary representation!!" &c) so my
investigation is limited to the lisp side.
looks like older clisp compiles the call to = and values inline which 2.34
creates closures for that (which is kind of crazy).
apparently caused by my 2005-01-21 patch

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=101355&aid=3147908&group_id=1355