[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: shiftf

At 11:31 1/27/93 +0017, Kim Barrett wrote:
>> MCL's shiftf macro produces unneccesary complicated form in the simple
>> cases.  The most frequent use of shift, is probably for simple pointer
>> switching in list structures: Replacing a value while saving the old
>> value.  (The two-argument (place, newvalue) special case of shiftf used
>> to have its own name, swapf)
>> Here's what MCL's shiftf produces:
>> ? (macroexpand '(shiftf (cddr liste) nil))
>> (LET* ((#:G91 LISTE) 
>>        (#:G89 (MULTIPLE-VALUE-LIST (CDDR #:G91)))) 
>>     (PROGN (CCL::SET-CDDR #:G91 #:G90))) 
>>   (VALUES-LIST #:G89))
>> Here's a simpler expansion suggestion (Allegro 4.1 produces something
>> similar to this):
>> (LET* ((#:G91 LISTE) 
>>        (#:G89 (CDDR #:G91))
>>        (#:G90 NIL))
>>   (CCL::SET-CDDR #:G91 #:G90)
>>   #:G89)
>You didn't mention which version of MCL you are using.  SETF and related macros 
>had a lot of work done on them not too long ago.  In MCL2.0p2 (and probably in 
>MCL2.0, but p2 is what I've got in front of me right now) the expansion is
>(LET* ((#:G30466 LISTE))
>    (MULTIPLE-VALUE-BIND (#:G30465)
>        NIL
>      (PROGN (SET-CDDR #:G30466 #:G30465)))))
>The MULTIPLE-VALUE-BIND with a single variable is trivially transformed into a 
>LET.  If the compiler handling for MULTIPLE-VALUE-PROG! recognizes that CDDR 
>always returns one value and makes the obvious transformation, then compiling 
>this expansion should produce the same code as your simpler expansion, without 
>requiring that the SHIFTF macro (and presumably lots of other similar macros) 
>itself try to optimize the expansion.  The reason for all the 
>multiple-value-mumbles in the expansion is to support multiple-valued place 
>forms (see X3J13 issue SETF-MULTIPLE-STORE-VARIABLES).  Rather than try to make 
>the macros optimize this stuff, its much better to make the compiler smart 
>about some of these special cases where multiple value forms can be transformed 
>into single value forms.  Unfortunately, the compiler doesn't currently seem to 
>know about CDDR being single valued, so things could be a bit better.  But the 
>"bug" is there, not in SHIFTF.

Unpatched MCL 2.0 produces the expansion to which Espen referred. Patch 2
modifies SHIFTF & ROTATEF to produce Kim's better code. Mark announced how to
get patch 2 about a month ago. Ask and I'll resend. I have prepared a patch that
makes the compiler a little bit smarter about turning MULTIPLE-VALUE-PROG1 into
PROG1. In particular, it makes the patch 2 expansion for (shift (cddr liste) nil)
compile into the same code as the Allegro 4.1 expansion. Ask for