Re: eliminatable UNWIND-PROTECT

[Sam S. <sd...@gn...>]

> * Bruno Haible <oe...@py...t> [2003-10-15 13:01:08 +0200]:
>
> Sam wrote:
>> Now, could you please explain this?
>>
>> [6]> (defun 1st (x) (with-open-file (s x) (write-line (read-line s))))
>> 1ST
>> [7]> (disassemble '1st)
>>
>>
>> Disassembly of function 1ST
>> (CONST 0) = #<SYSTEM-FUNCTION VALUES>
>> (CONST 1) = CLOSE
>> (CONST 2) = :ABORT
>> 1 required argument
>> 0 optional arguments
>> No rest parameter
>> No keyword parameters
>> 33 byte-code instructions:
>> 0     (LOAD&PUSH 1)
>> 1     (PUSH-UNBOUND 6)
>> 3     (CALLS1&PUSH 249)                   ; OPEN
>> 5     (UNWIND-PROTECT-OPEN L38)
>> 7     (CONST 0)                           ; #<SYSTEM-FUNCTION VALUES>
>> 8     (MVCALLP)
>> 9     (LOAD&PUSH 3)
>> 10    (PUSH-UNBOUND 3)
>> 12    (CALLS1&PUSH 119)                   ; READ-LINE
>> 14    (PUSH-UNBOUND 3)
>> 16    (CALLS1 138)                        ; WRITE-LINE
>> 18    (MV-TO-STACK)
>> 19    (LOADI 0 0 2)
>> 23    (JMPIFNOT L28)
>> 25    (PUSH)
>> 26    (CALL1 1)                           ; CLOSE
>> 28    L28
>> 28    (MVCALL)
>> 29    (UNWIND-PROTECT-NORMAL-EXIT)
>> 30    (LOADI 0 0 0)
>> 34    (JMPIFNOT L50)
>> 36    (JMP L44)
>> 38    L38
>> 38    (LOADI 0 0 0)
>> 42    (JMPIFNOT L50)
>> 44    L44
>> 44    (PUSH)
>> 45    (CONST&PUSH 2)                      ; :ABORT
>> 46    (T&PUSH)
>> 47    (CALL 3 1)                          ; CLOSE
>> 50    L50
>> 50    (UNWIND-PROTECT-CLOSE)
>> 51    (SKIP&RET 3)
>> NIL
>>
>>
>> 2. instructions 30/34/36 appear to be equivalent to 38/42/44; i.e.,
>>    30/34/36/38 can be eliminated altogether (and 5 point to L44)
>
> You are right. This is a consequence of a heuristic that Michael built in:
> the jump optimization only merges
>
>       (I1) ... (Ik) ... (In) LNNN
> and   (Ik+1) ... (In) (JMP LNNN)
>
> to
>
>       (I1) ... (Ik) LKKK ... (In) LNNN
> and   (JMP LKKK)
>
> if and only if n-k >= 2, i.e. if there are two instructions to be
> saved.  And similarly it can happen that a single instruction is
> copied from one code-list to another if that makes a jump go
> away. While this makes the code slightly larger, it ensures that the
> relative frequency of jumps (i.e. of no-op instructions) is kept to <=
> 33%. Which is important for speed.

wait a second - if the 33% ratio is maintained by keeping the
denominator larger (instead of numerator small), nothing is gained.

> You are welcome to refine this heuristic so that it recognizes a
> complex case like this one where nothing is won by copying the (LOADI
> 0 0 0) instruction.

most of this message is Greek to me.
sorry.
maybe you could point me to the part of compiler.lisp which you are
talking about?
(or, better yet, DIY and I will learn from your patch, like I learned
about *current-value* and *current-vars* in traverse-anode from your
2003-05-14 patch)

-- 
Sam Steingold (http://www.podval.org/~sds) running w2k
<http://www.camera.org> <http://www.iris.org.il> <http://www.memri.org/>
<http://www.mideasttruth.com/> <http://www.honestreporting.com>
MS Windows: error: the operation completed successfully.