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

issue ALLOW-LOCAL-INLINE, version 4

This version incorporates some minor changes suggested by Pitman.

Forum:		Compiler
References:	CLtL p. 156, 159
Edit History:   21 Sept. 88 V1 by David Gray
                27 Oct.  88 V2 by David Gray - new proposal
		 9 Nov.  88 V3 by David Gray - expanded problem
			description and discussion sections.
		30 Dec.  88 V4 by Sandra Loosemore -- suggestions from Pitman
Problem Description:

  The proposal PROCLAIM-INLINE-WHERE:BEFORE (which was accepted by X3J13
  on 10/12/88) clarifies the use of INLINE proclamations, but there
  remains a similar problem with the use of a local (DECLARE (INLINE
  ...)):  how can the compiler expand the function inline if it didn't
  know that the necessary information should have been saved when the
  function was compiled?

  Note that an INLINE proclamation does two things:

    (1) It tells the compiler to do extra things when it sees the
        function -definition-, to make it possible to code the function

    (2) It tells the compiler to code -calls- to the function inline.

  In order for local INLINE declarations to be useful, we need part 1
  without part 2.

  Clarify that to define a function FOO which is not INLOINE by default
  but for which (DECLARE (INLINE FOO)) will make FOO be locally inlined,
  the proper definition sequence is:
   (DEFUN foo ...)

  The INLINE proclamation preceding the DEFUN ensures that compiler will
  save the information necessary for inline expansion, and the NOTINLINE
  proclamation following the DEFUN prevents it from being expanded
  inline everywhere.  

  Note that while implementations are never required to perform inline
  expansion of function calls, many implementations that do support
  inline expansion will not be able to respond to local INLINE requests
  if this technique is not followed.


  Local INLINE declarations are of little use without some way of
  alerting the compiler to the possibility of inline expansion before
  the function is compiled.  This seems the simplest solution since it
  just clarifies existing practice instead of adding a new feature to
  the language.

  A compiler could use some heuristic to save the definitions of functions
  that are short enough to look like good candidates for inline expansion,
  but then the user is never sure what to expect.  It is possible that a
  compiler could simply save all definitions (assuming availability
  of adequate storage space) but we shouldn't require that.

 Test Cases/Examples: 

  Given the following input to COMPILE-FILE, does F1 get expanded inline
  in F2, and does F3 get expanded inline in F4?

    (defun f1 (a) (+ a 100))
    (defun f2 (b) (declare (inline f1)) (f1 b))
    (proclaim '(inline f3))
    (defun f3 (a) (+ a 100))
    (proclaim '(notinline f3))
    (defun f4 (b) (f3 b))			;F3 is not inline.
    (defun f5 (c) (declare (inline f3)) (f3 c)) ;F3 is locally inline.
    (defun f6 (c) (f3 c)) 			;The local effect is not 
						;  persistent.
 Current Practice:
  In the example above, using Symbolics, Lucid, or Explorer, F1 is not
  expanded in F2, but F3 is expanded in F5.

 Cost to implementors:
  None, since this is a clarification in accordance with current

 Cost to users:


  Users will be able to use (DECLARE (INLINE ...)) with greater assurance
  that it will really do something.

 Costs of Non-Adoption: 

  Users will not know how to reliably request inline expansion on a
  local basis.  This technique is not obvious, and even the need
  for it likely to be apparent only to people who understand something
  about how the compiler does inline expansion.


  Version 1 of this issue included proposal
  This was met with a lack of enthusiasm since it was pointed out that
  the same effect could be obtained by using a combination of INLINE and

  This is may not be completely true, however, since people's thinking
  about NOTINLINE has evolved in the direction of a declaration that
  tells the compiler "assume nothing about this function".  Thus, a
  NOTINLINE proclamation might suppress some optimizations that would
  have occurred if there had never been an INLINE and NOTINLINE.

  Ideally, it would be nice to have multiple levels of control instead
  of just INLINE or NOTINLINE -- something like:
    * always inline
    * maybe inline (let the compiler decide)
    * just save the definition for possible local inline
    * default
    * never inline

  Pitman has said that he generally approves of the direction of this
  proposal, but he has also expressed concerns about how the persistance
  of INLINE proclamations may cause confusion when functions are redefined
  in an incremental development environment.